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-common.h"
  19 #include "block/block.h"
  20 #include "qemu/error-report.h"
  21 #include "qemu/main-loop.h"
  22 #include "hw/hw.h"
  23 #include "qemu/cutils.h"
  24 #include "qemu/queue.h"
  25 #include "qemu/timer.h"
  26 #include "migration/block.h"
  27 #include "migration/migration.h"
  28 #include "sysemu/blockdev.h"
  29 #include "migration/qemu-file.h"
  30 #include "migration/vmstate.h"
  31 #include "sysemu/block-backend.h"
  32
  33 #define BLOCK_SIZE                       (1 << 20)
  34 #define BDRV_SECTORS_PER_DIRTY_CHUNK     (BLOCK_SIZE >> BDRV_SECTOR_BITS)
  35
  36 #define BLK_MIG_FLAG_DEVICE_BLOCK       0x01
  37 #define BLK_MIG_FLAG_EOS                0x02
  38 #define BLK_MIG_FLAG_PROGRESS           0x04
  39 #define BLK_MIG_FLAG_ZERO_BLOCK         0x08
  40
  41 #define MAX_IS_ALLOCATED_SEARCH 65536
  42
  43 #define MAX_INFLIGHT_IO 512
  44
  45 //#define DEBUG_BLK_MIGRATION
  46
  47 #ifdef DEBUG_BLK_MIGRATION
  48 #define DPRINTF(fmt, ...) \
  49     do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
  50 #else
  51 #define DPRINTF(fmt, ...) \
  52     do { } while (0)
  53 #endif
  54
  55 typedef struct BlkMigDevState {
  56     /* Written during setup phase.  Can be read without a lock.  */
  57     BlockBackend *blk;
  58     char *blk_name;
  59     int shared_base;
  60     int64_t total_sectors;
  61     QSIMPLEQ_ENTRY(BlkMigDevState) entry;
  62     Error *blocker;
  63
  64     /* Only used by migration thread.  Does not need a lock.  */
  65     int bulk_completed;
  66     int64_t cur_sector;
  67     int64_t cur_dirty;
  68
  69     /* Data in the aio_bitmap is protected by block migration lock.
  70      * Allocation and free happen during setup and cleanup respectively.
  71      */
  72     unsigned long *aio_bitmap;
  73
  74     /* Protected by block migration lock.  */
  75     int64_t completed_sectors;
  76
  77     /* During migration this is protected by iothread lock / AioContext.
  78      * Allocation and free happen during setup and cleanup respectively.
  79      */
  80     BdrvDirtyBitmap *dirty_bitmap;
  81 } BlkMigDevState;
  82
  83 typedef struct BlkMigBlock {
  84     /* Only used by migration thread.  */
  85     uint8_t *buf;
  86     BlkMigDevState *bmds;
  87     int64_t sector;
  88     int nr_sectors;
  89     struct iovec iov;
  90     QEMUIOVector qiov;
  91     BlockAIOCB *aiocb;
  92
  93     /* Protected by block migration lock.  */
  94     int ret;
  95     QSIMPLEQ_ENTRY(BlkMigBlock) entry;
  96 } BlkMigBlock;
  97
  98 typedef struct BlkMigState {
  99     QSIMPLEQ_HEAD(bmds_list, BlkMigDevState) bmds_list;
 100     int64_t total_sector_sum;
 101     bool zero_blocks;
 102
 103     /* Protected by lock.  */
 104     QSIMPLEQ_HEAD(blk_list, BlkMigBlock) blk_list;
 105     int submitted;
 106     int read_done;
 107
 108     /* Only used by migration thread.  Does not need a lock.  */
 109     int transferred;
 110     int prev_progress;
 111     int bulk_completed;
 112
 113     /* Lock must be taken _inside_ the iothread lock and any AioContexts.  */
 114     QemuMutex lock;
 115 } BlkMigState;
 116
 117 static BlkMigState block_mig_state;
 118
 119 static void blk_mig_lock(void)
 120 {
 121     qemu_mutex_lock(&block_mig_state.lock);
 122 }
 123
 124 static void blk_mig_unlock(void)
 125 {
 126     qemu_mutex_unlock(&block_mig_state.lock);
 127 }
 128
 129 /* Must run outside of the iothread lock during the bulk phase,
 130  * or the VM will stall.
 131  */
 132
 133 static void blk_send(QEMUFile *f, BlkMigBlock * blk)
 134 {
 135     int len;
 136     uint64_t flags = BLK_MIG_FLAG_DEVICE_BLOCK;
 137
 138     if (block_mig_state.zero_blocks &&
 139         buffer_is_zero(blk->buf, BLOCK_SIZE)) {
 140         flags |= BLK_MIG_FLAG_ZERO_BLOCK;
 141     }
 142
 143     /* sector number and flags */
 144     qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
 145                      | flags);
 146
 147     /* device name */
 148     len = strlen(blk->bmds->blk_name);
 149     qemu_put_byte(f, len);
 150     qemu_put_buffer(f, (uint8_t *) blk->bmds->blk_name, len);
 151
 152     /* if a block is zero we need to flush here since the network
 153      * bandwidth is now a lot higher than the storage device bandwidth.
 154      * thus if we queue zero blocks we slow down the migration */
 155     if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 156         qemu_fflush(f);
 157         return;
 158     }
 159
 160     qemu_put_buffer(f, blk->buf, BLOCK_SIZE);
 161 }
 162
 163 int blk_mig_active(void)
 164 {
 165     return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
 166 }
 167
 168 uint64_t blk_mig_bytes_transferred(void)
 169 {
 170     BlkMigDevState *bmds;
 171     uint64_t sum = 0;
 172
 173     blk_mig_lock();
 174     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 175         sum += bmds->completed_sectors;
 176     }
 177     blk_mig_unlock();
 178     return sum << BDRV_SECTOR_BITS;
 179 }
 180
 181 uint64_t blk_mig_bytes_remaining(void)
 182 {
 183     return blk_mig_bytes_total() - blk_mig_bytes_transferred();
 184 }
 185
 186 uint64_t blk_mig_bytes_total(void)
 187 {
 188     BlkMigDevState *bmds;
 189     uint64_t sum = 0;
 190
 191     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 192         sum += bmds->total_sectors;
 193     }
 194     return sum << BDRV_SECTOR_BITS;
 195 }
 196
 197
 198 /* Called with migration lock held.  */
 199
 200 static int bmds_aio_inflight(BlkMigDevState *bmds, int64_t sector)
 201 {
 202     int64_t chunk = sector / (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK;
 203
 204     if (sector < blk_nb_sectors(bmds->blk)) {
 205         return !!(bmds->aio_bitmap[chunk / (sizeof(unsigned long) * 8)] &
 206             (1UL << (chunk % (sizeof(unsigned long) * 8))));
 207     } else {
 208         return 0;
 209     }
 210 }
 211
 212 /* Called with migration lock held.  */
 213
 214 static void bmds_set_aio_inflight(BlkMigDevState *bmds, int64_t sector_num,
 215                              int nb_sectors, int set)
 216 {
 217     int64_t start, end;
 218     unsigned long val, idx, bit;
 219
 220     start = sector_num / BDRV_SECTORS_PER_DIRTY_CHUNK;
 221     end = (sector_num + nb_sectors - 1) / BDRV_SECTORS_PER_DIRTY_CHUNK;
 222
 223     for (; start <= end; start++) {
 224         idx = start / (sizeof(unsigned long) * 8);
 225         bit = start % (sizeof(unsigned long) * 8);
 226         val = bmds->aio_bitmap[idx];
 227         if (set) {
 228             val |= 1UL << bit;
 229         } else {
 230             val &= ~(1UL << bit);
 231         }
 232         bmds->aio_bitmap[idx] = val;
 233     }
 234 }
 235
 236 static void alloc_aio_bitmap(BlkMigDevState *bmds)
 237 {
 238     BlockBackend *bb = bmds->blk;
 239     int64_t bitmap_size;
 240
 241     bitmap_size = blk_nb_sectors(bb) + BDRV_SECTORS_PER_DIRTY_CHUNK * 8 - 1;
 242     bitmap_size /= BDRV_SECTORS_PER_DIRTY_CHUNK * 8;
 243
 244     bmds->aio_bitmap = g_malloc0(bitmap_size);
 245 }
 246
 247 /* Never hold migration lock when yielding to the main loop!  */
 248
 249 static void blk_mig_read_cb(void *opaque, int ret)
 250 {
 251     BlkMigBlock *blk = opaque;
 252
 253     blk_mig_lock();
 254     blk->ret = ret;
 255
 256     QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry);
 257     bmds_set_aio_inflight(blk->bmds, blk->sector, blk->nr_sectors, 0);
 258
 259     block_mig_state.submitted--;
 260     block_mig_state.read_done++;
 261     assert(block_mig_state.submitted >= 0);
 262     blk_mig_unlock();
 263 }
 264
 265 /* Called with no lock taken.  */
 266
 267 static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds)
 268 {
 269     int64_t total_sectors = bmds->total_sectors;
 270     int64_t cur_sector = bmds->cur_sector;
 271     BlockBackend *bb = bmds->blk;
 272     BlkMigBlock *blk;
 273     int nr_sectors;
 274
 275     if (bmds->shared_base) {
 276         qemu_mutex_lock_iothread();
 277         aio_context_acquire(blk_get_aio_context(bb));
 278         /* Skip unallocated sectors; intentionally treats failure as
 279          * an allocated sector */
 280         while (cur_sector < total_sectors &&
 281                !bdrv_is_allocated(blk_bs(bb), cur_sector,
 282                                   MAX_IS_ALLOCATED_SEARCH, &nr_sectors)) {
 283             cur_sector += nr_sectors;
 284         }
 285         aio_context_release(blk_get_aio_context(bb));
 286         qemu_mutex_unlock_iothread();
 287     }
 288
 289     if (cur_sector >= total_sectors) {
 290         bmds->cur_sector = bmds->completed_sectors = total_sectors;
 291         return 1;
 292     }
 293
 294     bmds->completed_sectors = cur_sector;
 295
 296     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
 297
 298     /* we are going to transfer a full block even if it is not allocated */
 299     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 300
 301     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 302         nr_sectors = total_sectors - cur_sector;
 303     }
 304
 305     blk = g_new(BlkMigBlock, 1);
 306     blk->buf = g_malloc(BLOCK_SIZE);
 307     blk->bmds = bmds;
 308     blk->sector = cur_sector;
 309     blk->nr_sectors = nr_sectors;
 310
 311     blk->iov.iov_base = blk->buf;
 312     blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 313     qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 314
 315     blk_mig_lock();
 316     block_mig_state.submitted++;
 317     blk_mig_unlock();
 318
 319     /* We do not know if bs is under the main thread (and thus does
 320      * not acquire the AioContext when doing AIO) or rather under
 321      * dataplane.  Thus acquire both the iothread mutex and the
 322      * AioContext.
 323      *
 324      * This is ugly and will disappear when we make bdrv_* thread-safe,
 325      * without the need to acquire the AioContext.
 326      */
 327     qemu_mutex_lock_iothread();
 328     aio_context_acquire(blk_get_aio_context(bmds->blk));
 329     blk->aiocb = blk_aio_preadv(bb, cur_sector * BDRV_SECTOR_SIZE, &blk->qiov,
 330                                 0, blk_mig_read_cb, blk);
 331
 332     bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, cur_sector, nr_sectors);
 333     aio_context_release(blk_get_aio_context(bmds->blk));
 334     qemu_mutex_unlock_iothread();
 335
 336     bmds->cur_sector = cur_sector + nr_sectors;
 337     return (bmds->cur_sector >= total_sectors);
 338 }
 339
 340 /* Called with iothread lock taken.  */
 341
 342 static int set_dirty_tracking(void)
 343 {
 344     BlkMigDevState *bmds;
 345     int ret;
 346
 347     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 348         aio_context_acquire(blk_get_aio_context(bmds->blk));
 349         bmds->dirty_bitmap = bdrv_create_dirty_bitmap(blk_bs(bmds->blk),
 350                                                       BLOCK_SIZE, NULL, NULL);
 351         aio_context_release(blk_get_aio_context(bmds->blk));
 352         if (!bmds->dirty_bitmap) {
 353             ret = -errno;
 354             goto fail;
 355         }
 356     }
 357     return 0;
 358
 359 fail:
 360     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 361         if (bmds->dirty_bitmap) {
 362             aio_context_acquire(blk_get_aio_context(bmds->blk));
 363             bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
 364             aio_context_release(blk_get_aio_context(bmds->blk));
 365         }
 366     }
 367     return ret;
 368 }
 369
 370 /* Called with iothread lock taken.  */
 371
 372 static void unset_dirty_tracking(void)
 373 {
 374     BlkMigDevState *bmds;
 375
 376     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 377         aio_context_acquire(blk_get_aio_context(bmds->blk));
 378         bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
 379         aio_context_release(blk_get_aio_context(bmds->blk));
 380     }
 381 }
 382
 383 static int init_blk_migration(QEMUFile *f)
 384 {
 385     BlockDriverState *bs;
 386     BlkMigDevState *bmds;
 387     int64_t sectors;
 388     BdrvNextIterator it;
 389     int i, num_bs = 0;
 390     struct {
 391         BlkMigDevState *bmds;
 392         BlockDriverState *bs;
 393     } *bmds_bs;
 394     Error *local_err = NULL;
 395     int ret;
 396
 397     block_mig_state.submitted = 0;
 398     block_mig_state.read_done = 0;
 399     block_mig_state.transferred = 0;
 400     block_mig_state.total_sector_sum = 0;
 401     block_mig_state.prev_progress = -1;
 402     block_mig_state.bulk_completed = 0;
 403     block_mig_state.zero_blocks = migrate_zero_blocks();
 404
 405     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
 406         num_bs++;
 407     }
 408     bmds_bs = g_malloc0(num_bs * sizeof(*bmds_bs));
 409
 410     for (i = 0, bs = bdrv_first(&it); bs; bs = bdrv_next(&it), i++) {
 411         if (bdrv_is_read_only(bs)) {
 412             continue;
 413         }
 414
 415         sectors = bdrv_nb_sectors(bs);
 416         if (sectors <= 0) {
 417             ret = sectors;
 418             goto out;
 419         }
 420
 421         bmds = g_new0(BlkMigDevState, 1);
 422         bmds->blk = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 423         bmds->blk_name = g_strdup(bdrv_get_device_name(bs));
 424         bmds->bulk_completed = 0;
 425         bmds->total_sectors = sectors;
 426         bmds->completed_sectors = 0;
 427         bmds->shared_base = migrate_use_block_incremental();
 428
 429         assert(i < num_bs);
 430         bmds_bs[i].bmds = bmds;
 431         bmds_bs[i].bs = bs;
 432
 433         block_mig_state.total_sector_sum += sectors;
 434
 435         if (bmds->shared_base) {
 436             DPRINTF("Start migration for %s with shared base image\n",
 437                     bdrv_get_device_name(bs));
 438         } else {
 439             DPRINTF("Start full migration for %s\n", bdrv_get_device_name(bs));
 440         }
 441
 442         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 443     }
 444
 445     /* Can only insert new BDSes now because doing so while iterating block
 446      * devices may end up in a deadlock (iterating the new BDSes, too). */
 447     for (i = 0; i < num_bs; i++) {
 448         BlkMigDevState *bmds = bmds_bs[i].bmds;
 449         BlockDriverState *bs = bmds_bs[i].bs;
 450
 451         if (bmds) {
 452             ret = blk_insert_bs(bmds->blk, bs, &local_err);
 453             if (ret < 0) {
 454                 error_report_err(local_err);
 455                 goto out;
 456             }
 457
 458             alloc_aio_bitmap(bmds);
 459             error_setg(&bmds->blocker, "block device is in use by migration");
 460             bdrv_op_block_all(bs, bmds->blocker);
 461         }
 462     }
 463
 464     ret = 0;
 465 out:
 466     g_free(bmds_bs);
 467     return ret;
 468 }
 469
 470 /* Called with no lock taken.  */
 471
 472 static int blk_mig_save_bulked_block(QEMUFile *f)
 473 {
 474     int64_t completed_sector_sum = 0;
 475     BlkMigDevState *bmds;
 476     int progress;
 477     int ret = 0;
 478
 479     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 480         if (bmds->bulk_completed == 0) {
 481             if (mig_save_device_bulk(f, bmds) == 1) {
 482                 /* completed bulk section for this device */
 483                 bmds->bulk_completed = 1;
 484             }
 485             completed_sector_sum += bmds->completed_sectors;
 486             ret = 1;
 487             break;
 488         } else {
 489             completed_sector_sum += bmds->completed_sectors;
 490         }
 491     }
 492
 493     if (block_mig_state.total_sector_sum != 0) {
 494         progress = completed_sector_sum * 100 /
 495                    block_mig_state.total_sector_sum;
 496     } else {
 497         progress = 100;
 498     }
 499     if (progress != block_mig_state.prev_progress) {
 500         block_mig_state.prev_progress = progress;
 501         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 502                          | BLK_MIG_FLAG_PROGRESS);
 503         DPRINTF("Completed %d %%\r", progress);
 504     }
 505
 506     return ret;
 507 }
 508
 509 static void blk_mig_reset_dirty_cursor(void)
 510 {
 511     BlkMigDevState *bmds;
 512
 513     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 514         bmds->cur_dirty = 0;
 515     }
 516 }
 517
 518 /* Called with iothread lock and AioContext taken.  */
 519
 520 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
 521                                  int is_async)
 522 {
 523     BlkMigBlock *blk;
 524     BlockDriverState *bs = blk_bs(bmds->blk);
 525     int64_t total_sectors = bmds->total_sectors;
 526     int64_t sector;
 527     int nr_sectors;
 528     int ret = -EIO;
 529
 530     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
 531         blk_mig_lock();
 532         if (bmds_aio_inflight(bmds, sector)) {
 533             blk_mig_unlock();
 534             blk_drain(bmds->blk);
 535         } else {
 536             blk_mig_unlock();
 537         }
 538         if (bdrv_get_dirty(bs, bmds->dirty_bitmap, sector)) {
 539
 540             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 541                 nr_sectors = total_sectors - sector;
 542             } else {
 543                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 544             }
 545             blk = g_new(BlkMigBlock, 1);
 546             blk->buf = g_malloc(BLOCK_SIZE);
 547             blk->bmds = bmds;
 548             blk->sector = sector;
 549             blk->nr_sectors = nr_sectors;
 550
 551             if (is_async) {
 552                 blk->iov.iov_base = blk->buf;
 553                 blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 554                 qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 555
 556                 blk->aiocb = blk_aio_preadv(bmds->blk,
 557                                             sector * BDRV_SECTOR_SIZE,
 558                                             &blk->qiov, 0, blk_mig_read_cb,
 559                                             blk);
 560
 561                 blk_mig_lock();
 562                 block_mig_state.submitted++;
 563                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 564                 blk_mig_unlock();
 565             } else {
 566                 ret = blk_pread(bmds->blk, sector * BDRV_SECTOR_SIZE, blk->buf,
 567                                 nr_sectors * BDRV_SECTOR_SIZE);
 568                 if (ret < 0) {
 569                     goto error;
 570                 }
 571                 blk_send(f, blk);
 572
 573                 g_free(blk->buf);
 574                 g_free(blk);
 575             }
 576
 577             bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, sector, nr_sectors);
 578             sector += nr_sectors;
 579             bmds->cur_dirty = sector;
 580
 581             break;
 582         }
 583         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 584         bmds->cur_dirty = sector;
 585     }
 586
 587     return (bmds->cur_dirty >= bmds->total_sectors);
 588
 589 error:
 590     DPRINTF("Error reading sector %" PRId64 "\n", sector);
 591     g_free(blk->buf);
 592     g_free(blk);
 593     return ret;
 594 }
 595
 596 /* Called with iothread lock taken.
 597  *
 598  * return value:
 599  * 0: too much data for max_downtime
 600  * 1: few enough data for max_downtime
 601 */
 602 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 603 {
 604     BlkMigDevState *bmds;
 605     int ret = 1;
 606
 607     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 608         aio_context_acquire(blk_get_aio_context(bmds->blk));
 609         ret = mig_save_device_dirty(f, bmds, is_async);
 610         aio_context_release(blk_get_aio_context(bmds->blk));
 611         if (ret <= 0) {
 612             break;
 613         }
 614     }
 615
 616     return ret;
 617 }
 618
 619 /* Called with no locks taken.  */
 620
 621 static int flush_blks(QEMUFile *f)
 622 {
 623     BlkMigBlock *blk;
 624     int ret = 0;
 625
 626     DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
 627             __FUNCTION__, block_mig_state.submitted, block_mig_state.read_done,
 628             block_mig_state.transferred);
 629
 630     blk_mig_lock();
 631     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 632         if (qemu_file_rate_limit(f)) {
 633             break;
 634         }
 635         if (blk->ret < 0) {
 636             ret = blk->ret;
 637             break;
 638         }
 639
 640         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 641         blk_mig_unlock();
 642         blk_send(f, blk);
 643         blk_mig_lock();
 644
 645         g_free(blk->buf);
 646         g_free(blk);
 647
 648         block_mig_state.read_done--;
 649         block_mig_state.transferred++;
 650         assert(block_mig_state.read_done >= 0);
 651     }
 652     blk_mig_unlock();
 653
 654     DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__,
 655             block_mig_state.submitted, block_mig_state.read_done,
 656             block_mig_state.transferred);
 657     return ret;
 658 }
 659
 660 /* Called with iothread lock taken.  */
 661
 662 static int64_t get_remaining_dirty(void)
 663 {
 664     BlkMigDevState *bmds;
 665     int64_t dirty = 0;
 666
 667     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 668         aio_context_acquire(blk_get_aio_context(bmds->blk));
 669         dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
 670         aio_context_release(blk_get_aio_context(bmds->blk));
 671     }
 672
 673     return dirty << BDRV_SECTOR_BITS;
 674 }
 675
 676 /* Called with iothread lock taken.  */
 677
 678 static void block_migration_cleanup(void *opaque)
 679 {
 680     BlkMigDevState *bmds;
 681     BlkMigBlock *blk;
 682     AioContext *ctx;
 683
 684     bdrv_drain_all();
 685
 686     unset_dirty_tracking();
 687
 688     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 689         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 690         bdrv_op_unblock_all(blk_bs(bmds->blk), bmds->blocker);
 691         error_free(bmds->blocker);
 692
 693         /* Save ctx, because bmds->blk can disappear during blk_unref.  */
 694         ctx = blk_get_aio_context(bmds->blk);
 695         aio_context_acquire(ctx);
 696         blk_unref(bmds->blk);
 697         aio_context_release(ctx);
 698
 699         g_free(bmds->blk_name);
 700         g_free(bmds->aio_bitmap);
 701         g_free(bmds);
 702     }
 703
 704     blk_mig_lock();
 705     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 706         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 707         g_free(blk->buf);
 708         g_free(blk);
 709     }
 710     blk_mig_unlock();
 711 }
 712
 713 static int block_save_setup(QEMUFile *f, void *opaque)
 714 {
 715     int ret;
 716
 717     DPRINTF("Enter save live setup submitted %d transferred %d\n",
 718             block_mig_state.submitted, block_mig_state.transferred);
 719
 720     qemu_mutex_lock_iothread();
 721     ret = init_blk_migration(f);
 722     if (ret < 0) {
 723         qemu_mutex_unlock_iothread();
 724         return ret;
 725     }
 726
 727     /* start track dirty blocks */
 728     ret = set_dirty_tracking();
 729
 730     qemu_mutex_unlock_iothread();
 731
 732     if (ret) {
 733         return ret;
 734     }
 735
 736     ret = flush_blks(f);
 737     blk_mig_reset_dirty_cursor();
 738     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 739
 740     return ret;
 741 }
 742
 743 static int block_save_iterate(QEMUFile *f, void *opaque)
 744 {
 745     int ret;
 746     int64_t last_ftell = qemu_ftell(f);
 747     int64_t delta_ftell;
 748
 749     DPRINTF("Enter save live iterate submitted %d transferred %d\n",
 750             block_mig_state.submitted, block_mig_state.transferred);
 751
 752     ret = flush_blks(f);
 753     if (ret) {
 754         return ret;
 755     }
 756
 757     blk_mig_reset_dirty_cursor();
 758
 759     /* control the rate of transfer */
 760     blk_mig_lock();
 761     while ((block_mig_state.submitted +
 762             block_mig_state.read_done) * BLOCK_SIZE <
 763            qemu_file_get_rate_limit(f) &&
 764            (block_mig_state.submitted +
 765             block_mig_state.read_done) <
 766            MAX_INFLIGHT_IO) {
 767         blk_mig_unlock();
 768         if (block_mig_state.bulk_completed == 0) {
 769             /* first finish the bulk phase */
 770             if (blk_mig_save_bulked_block(f) == 0) {
 771                 /* finished saving bulk on all devices */
 772                 block_mig_state.bulk_completed = 1;
 773             }
 774             ret = 0;
 775         } else {
 776             /* Always called with iothread lock taken for
 777              * simplicity, block_save_complete also calls it.
 778              */
 779             qemu_mutex_lock_iothread();
 780             ret = blk_mig_save_dirty_block(f, 1);
 781             qemu_mutex_unlock_iothread();
 782         }
 783         if (ret < 0) {
 784             return ret;
 785         }
 786         blk_mig_lock();
 787         if (ret != 0) {
 788             /* no more dirty blocks */
 789             break;
 790         }
 791     }
 792     blk_mig_unlock();
 793
 794     ret = flush_blks(f);
 795     if (ret) {
 796         return ret;
 797     }
 798
 799     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 800     delta_ftell = qemu_ftell(f) - last_ftell;
 801     if (delta_ftell > 0) {
 802         return 1;
 803     } else if (delta_ftell < 0) {
 804         return -1;
 805     } else {
 806         return 0;
 807     }
 808 }
 809
 810 /* Called with iothread lock taken.  */
 811
 812 static int block_save_complete(QEMUFile *f, void *opaque)
 813 {
 814     int ret;
 815
 816     DPRINTF("Enter save live complete submitted %d transferred %d\n",
 817             block_mig_state.submitted, block_mig_state.transferred);
 818
 819     ret = flush_blks(f);
 820     if (ret) {
 821         return ret;
 822     }
 823
 824     blk_mig_reset_dirty_cursor();
 825
 826     /* we know for sure that save bulk is completed and
 827        all async read completed */
 828     blk_mig_lock();
 829     assert(block_mig_state.submitted == 0);
 830     blk_mig_unlock();
 831
 832     do {
 833         ret = blk_mig_save_dirty_block(f, 0);
 834         if (ret < 0) {
 835             return ret;
 836         }
 837     } while (ret == 0);
 838
 839     /* report completion */
 840     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 841
 842     DPRINTF("Block migration completed\n");
 843
 844     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 845
 846     return 0;
 847 }
 848
 849 static void block_save_pending(QEMUFile *f, void *opaque, uint64_t max_size,
 850                                uint64_t *non_postcopiable_pending,
 851                                uint64_t *postcopiable_pending)
 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 * BLOCK_SIZE +
 862                block_mig_state.read_done * BLOCK_SIZE;
 863     blk_mig_unlock();
 864
 865     /* Report at least one block pending during bulk phase */
 866     if (pending <= max_size && !block_mig_state.bulk_completed) {
 867         pending = max_size + BLOCK_SIZE;
 868     }
 869
 870     DPRINTF("Enter save live pending  %" PRIu64 "\n", pending);
 871     /* We don't do postcopy */
 872     *non_postcopiable_pending += 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 = BLOCK_SIZE;
 889
 890     do {
 891         addr = qemu_get_be64(f);
 892
 893         flags = addr & ~BDRV_SECTOR_MASK;
 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_invalidate_cache(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 <= BLOCK_SIZE &&
 927                     BLOCK_SIZE % bdi.cluster_size == 0) {
 928                     cluster_size = bdi.cluster_size;
 929                 } else {
 930                     cluster_size = 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(BLOCK_SIZE);
 950                 qemu_get_buffer(f, buf, BLOCK_SIZE);
 951                 for (i = 0; i < 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 < 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, cur_buf,
 963                                          cluster_size, 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: %#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_use_block();
 999 }
1000
1001 static SaveVMHandlers savevm_block_handlers = {
1002     .save_live_setup = block_save_setup,
1003     .save_live_iterate = block_save_iterate,
1004     .save_live_complete_precopy = block_save_complete,
1005     .save_live_pending = block_save_pending,
1006     .load_state = block_load,
1007     .cleanup = block_migration_cleanup,
1008     .is_active = block_is_active,
1009 };
1010
1011 void blk_mig_init(void)
1012 {
1013     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
1014     QSIMPLEQ_INIT(&block_mig_state.blk_list);
1015     qemu_mutex_init(&block_mig_state.lock);
1016
1017     register_savevm_live(NULL, "block", 0, 1, &savevm_block_handlers,
1018                          &block_mig_state);
1019 }