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/cutils.h"
  20 #include "qemu/queue.h"
  21 #include "block.h"
  22 #include "migration/misc.h"
  23 #include "migration.h"
  24 #include "migration/register.h"
  25 #include "qemu-file.h"
  26 #include "migration/vmstate.h"
  27 #include "sysemu/block-backend.h"
  28
  29 #define BLOCK_SIZE                       (1 << 20)
  30 #define BDRV_SECTORS_PER_DIRTY_CHUNK     (BLOCK_SIZE >> BDRV_SECTOR_BITS)
  31
  32 #define BLK_MIG_FLAG_DEVICE_BLOCK       0x01
  33 #define BLK_MIG_FLAG_EOS                0x02
  34 #define BLK_MIG_FLAG_PROGRESS           0x04
  35 #define BLK_MIG_FLAG_ZERO_BLOCK         0x08
  36
  37 #define MAX_IS_ALLOCATED_SEARCH (65536 * BDRV_SECTOR_SIZE)
  38
  39 #define MAX_IO_BUFFERS 512
  40 #define MAX_PARALLEL_IO 16
  41
  42 //#define DEBUG_BLK_MIGRATION
  43
  44 #ifdef DEBUG_BLK_MIGRATION
  45 #define DPRINTF(fmt, ...) \
  46     do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
  47 #else
  48 #define DPRINTF(fmt, ...) \
  49     do { } while (0)
  50 #endif
  51
  52 typedef struct BlkMigDevState {
  53     /* Written during setup phase.  Can be read without a lock.  */
  54     BlockBackend *blk;
  55     char *blk_name;
  56     int shared_base;
  57     int64_t total_sectors;
  58     QSIMPLEQ_ENTRY(BlkMigDevState) entry;
  59     Error *blocker;
  60
  61     /* Only used by migration thread.  Does not need a lock.  */
  62     int bulk_completed;
  63     int64_t cur_sector;
  64     int64_t cur_dirty;
  65
  66     /* Data in the aio_bitmap is protected by block migration lock.
  67      * Allocation and free happen during setup and cleanup respectively.
  68      */
  69     unsigned long *aio_bitmap;
  70
  71     /* Protected by block migration lock.  */
  72     int64_t completed_sectors;
  73
  74     /* During migration this is protected by iothread lock / AioContext.
  75      * Allocation and free happen during setup and cleanup respectively.
  76      */
  77     BdrvDirtyBitmap *dirty_bitmap;
  78 } BlkMigDevState;
  79
  80 typedef struct BlkMigBlock {
  81     /* Only used by migration thread.  */
  82     uint8_t *buf;
  83     BlkMigDevState *bmds;
  84     int64_t sector;
  85     int nr_sectors;
  86     QEMUIOVector qiov;
  87     BlockAIOCB *aiocb;
  88
  89     /* Protected by block migration lock.  */
  90     int ret;
  91     QSIMPLEQ_ENTRY(BlkMigBlock) entry;
  92 } BlkMigBlock;
  93
  94 typedef struct BlkMigState {
  95     QSIMPLEQ_HEAD(, BlkMigDevState) bmds_list;
  96     int64_t total_sector_sum;
  97     bool zero_blocks;
  98
  99     /* Protected by lock.  */
 100     QSIMPLEQ_HEAD(, BlkMigBlock) blk_list;
 101     int submitted;
 102     int read_done;
 103
 104     /* Only used by migration thread.  Does not need a lock.  */
 105     int transferred;
 106     int prev_progress;
 107     int bulk_completed;
 108
 109     /* Lock must be taken _inside_ the iothread lock and any AioContexts.  */
 110     QemuMutex lock;
 111 } BlkMigState;
 112
 113 static BlkMigState block_mig_state;
 114
 115 static void blk_mig_lock(void)
 116 {
 117     qemu_mutex_lock(&block_mig_state.lock);
 118 }
 119
 120 static void blk_mig_unlock(void)
 121 {
 122     qemu_mutex_unlock(&block_mig_state.lock);
 123 }
 124
 125 /* Must run outside of the iothread lock during the bulk phase,
 126  * or the VM will stall.
 127  */
 128
 129 static void blk_send(QEMUFile *f, BlkMigBlock * blk)
 130 {
 131     int len;
 132     uint64_t flags = BLK_MIG_FLAG_DEVICE_BLOCK;
 133
 134     if (block_mig_state.zero_blocks &&
 135         buffer_is_zero(blk->buf, BLOCK_SIZE)) {
 136         flags |= BLK_MIG_FLAG_ZERO_BLOCK;
 137     }
 138
 139     /* sector number and flags */
 140     qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
 141                      | flags);
 142
 143     /* device name */
 144     len = strlen(blk->bmds->blk_name);
 145     qemu_put_byte(f, len);
 146     qemu_put_buffer(f, (uint8_t *) blk->bmds->blk_name, len);
 147
 148     /* if a block is zero we need to flush here since the network
 149      * bandwidth is now a lot higher than the storage device bandwidth.
 150      * thus if we queue zero blocks we slow down the migration */
 151     if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 152         qemu_fflush(f);
 153         return;
 154     }
 155
 156     qemu_put_buffer(f, blk->buf, BLOCK_SIZE);
 157 }
 158
 159 int blk_mig_active(void)
 160 {
 161     return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
 162 }
 163
 164 int blk_mig_bulk_active(void)
 165 {
 166     return blk_mig_active() && !block_mig_state.bulk_completed;
 167 }
 168
 169 uint64_t blk_mig_bytes_transferred(void)
 170 {
 171     BlkMigDevState *bmds;
 172     uint64_t sum = 0;
 173
 174     blk_mig_lock();
 175     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 176         sum += bmds->completed_sectors;
 177     }
 178     blk_mig_unlock();
 179     return sum << BDRV_SECTOR_BITS;
 180 }
 181
 182 uint64_t blk_mig_bytes_remaining(void)
 183 {
 184     return blk_mig_bytes_total() - blk_mig_bytes_transferred();
 185 }
 186
 187 uint64_t blk_mig_bytes_total(void)
 188 {
 189     BlkMigDevState *bmds;
 190     uint64_t sum = 0;
 191
 192     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 193         sum += bmds->total_sectors;
 194     }
 195     return sum << BDRV_SECTOR_BITS;
 196 }
 197
 198
 199 /* Called with migration lock held.  */
 200
 201 static int bmds_aio_inflight(BlkMigDevState *bmds, int64_t sector)
 202 {
 203     int64_t chunk = sector / (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK;
 204
 205     if (sector < blk_nb_sectors(bmds->blk)) {
 206         return !!(bmds->aio_bitmap[chunk / (sizeof(unsigned long) * 8)] &
 207             (1UL << (chunk % (sizeof(unsigned long) * 8))));
 208     } else {
 209         return 0;
 210     }
 211 }
 212
 213 /* Called with migration lock held.  */
 214
 215 static void bmds_set_aio_inflight(BlkMigDevState *bmds, int64_t sector_num,
 216                              int nb_sectors, int set)
 217 {
 218     int64_t start, end;
 219     unsigned long val, idx, bit;
 220
 221     start = sector_num / BDRV_SECTORS_PER_DIRTY_CHUNK;
 222     end = (sector_num + nb_sectors - 1) / BDRV_SECTORS_PER_DIRTY_CHUNK;
 223
 224     for (; start <= end; start++) {
 225         idx = start / (sizeof(unsigned long) * 8);
 226         bit = start % (sizeof(unsigned long) * 8);
 227         val = bmds->aio_bitmap[idx];
 228         if (set) {
 229             val |= 1UL << bit;
 230         } else {
 231             val &= ~(1UL << bit);
 232         }
 233         bmds->aio_bitmap[idx] = val;
 234     }
 235 }
 236
 237 static void alloc_aio_bitmap(BlkMigDevState *bmds)
 238 {
 239     BlockBackend *bb = bmds->blk;
 240     int64_t bitmap_size;
 241
 242     bitmap_size = blk_nb_sectors(bb) + BDRV_SECTORS_PER_DIRTY_CHUNK * 8 - 1;
 243     bitmap_size /= BDRV_SECTORS_PER_DIRTY_CHUNK * 8;
 244
 245     bmds->aio_bitmap = g_malloc0(bitmap_size);
 246 }
 247
 248 /* Never hold migration lock when yielding to the main loop!  */
 249
 250 static void blk_mig_read_cb(void *opaque, int ret)
 251 {
 252     BlkMigBlock *blk = opaque;
 253
 254     blk_mig_lock();
 255     blk->ret = ret;
 256
 257     QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry);
 258     bmds_set_aio_inflight(blk->bmds, blk->sector, blk->nr_sectors, 0);
 259
 260     block_mig_state.submitted--;
 261     block_mig_state.read_done++;
 262     assert(block_mig_state.submitted >= 0);
 263     blk_mig_unlock();
 264 }
 265
 266 /* Called with no lock taken.  */
 267
 268 static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds)
 269 {
 270     int64_t total_sectors = bmds->total_sectors;
 271     int64_t cur_sector = bmds->cur_sector;
 272     BlockBackend *bb = bmds->blk;
 273     BlkMigBlock *blk;
 274     int nr_sectors;
 275     int64_t count;
 276
 277     if (bmds->shared_base) {
 278         qemu_mutex_lock_iothread();
 279         aio_context_acquire(blk_get_aio_context(bb));
 280         /* Skip unallocated sectors; intentionally treats failure or
 281          * partial sector as an allocated sector */
 282         while (cur_sector < total_sectors &&
 283                !bdrv_is_allocated(blk_bs(bb), cur_sector * BDRV_SECTOR_SIZE,
 284                                   MAX_IS_ALLOCATED_SEARCH, &count)) {
 285             if (count < BDRV_SECTOR_SIZE) {
 286                 break;
 287             }
 288             cur_sector += count >> BDRV_SECTOR_BITS;
 289         }
 290         aio_context_release(blk_get_aio_context(bb));
 291         qemu_mutex_unlock_iothread();
 292     }
 293
 294     if (cur_sector >= total_sectors) {
 295         bmds->cur_sector = bmds->completed_sectors = total_sectors;
 296         return 1;
 297     }
 298
 299     bmds->completed_sectors = cur_sector;
 300
 301     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
 302
 303     /* we are going to transfer a full block even if it is not allocated */
 304     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 305
 306     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 307         nr_sectors = total_sectors - cur_sector;
 308     }
 309
 310     blk = g_new(BlkMigBlock, 1);
 311     blk->buf = g_malloc(BLOCK_SIZE);
 312     blk->bmds = bmds;
 313     blk->sector = cur_sector;
 314     blk->nr_sectors = nr_sectors;
 315
 316     qemu_iovec_init_buf(&blk->qiov, blk->buf, nr_sectors * BDRV_SECTOR_SIZE);
 317
 318     blk_mig_lock();
 319     block_mig_state.submitted++;
 320     blk_mig_unlock();
 321
 322     /* We do not know if bs is under the main thread (and thus does
 323      * not acquire the AioContext when doing AIO) or rather under
 324      * dataplane.  Thus acquire both the iothread mutex and the
 325      * AioContext.
 326      *
 327      * This is ugly and will disappear when we make bdrv_* thread-safe,
 328      * without the need to acquire the AioContext.
 329      */
 330     qemu_mutex_lock_iothread();
 331     aio_context_acquire(blk_get_aio_context(bmds->blk));
 332     bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, cur_sector * BDRV_SECTOR_SIZE,
 333                             nr_sectors * BDRV_SECTOR_SIZE);
 334     blk->aiocb = blk_aio_preadv(bb, cur_sector * BDRV_SECTOR_SIZE, &blk->qiov,
 335                                 0, blk_mig_read_cb, blk);
 336     aio_context_release(blk_get_aio_context(bmds->blk));
 337     qemu_mutex_unlock_iothread();
 338
 339     bmds->cur_sector = cur_sector + nr_sectors;
 340     return (bmds->cur_sector >= total_sectors);
 341 }
 342
 343 /* Called with iothread lock taken.  */
 344
 345 static int set_dirty_tracking(void)
 346 {
 347     BlkMigDevState *bmds;
 348     int ret;
 349
 350     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 351         bmds->dirty_bitmap = bdrv_create_dirty_bitmap(blk_bs(bmds->blk),
 352                                                       BLOCK_SIZE, NULL, NULL);
 353         if (!bmds->dirty_bitmap) {
 354             ret = -errno;
 355             goto fail;
 356         }
 357     }
 358     return 0;
 359
 360 fail:
 361     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 362         if (bmds->dirty_bitmap) {
 363             bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
 364         }
 365     }
 366     return ret;
 367 }
 368
 369 /* Called with iothread lock taken.  */
 370
 371 static void unset_dirty_tracking(void)
 372 {
 373     BlkMigDevState *bmds;
 374
 375     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 376         bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
 377     }
 378 }
 379
 380 static int init_blk_migration(QEMUFile *f)
 381 {
 382     BlockDriverState *bs;
 383     BlkMigDevState *bmds;
 384     int64_t sectors;
 385     BdrvNextIterator it;
 386     int i, num_bs = 0;
 387     struct {
 388         BlkMigDevState *bmds;
 389         BlockDriverState *bs;
 390     } *bmds_bs;
 391     Error *local_err = NULL;
 392     int ret;
 393
 394     block_mig_state.submitted = 0;
 395     block_mig_state.read_done = 0;
 396     block_mig_state.transferred = 0;
 397     block_mig_state.total_sector_sum = 0;
 398     block_mig_state.prev_progress = -1;
 399     block_mig_state.bulk_completed = 0;
 400     block_mig_state.zero_blocks = migrate_zero_blocks();
 401
 402     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
 403         num_bs++;
 404     }
 405     bmds_bs = g_malloc0(num_bs * sizeof(*bmds_bs));
 406
 407     for (i = 0, bs = bdrv_first(&it); bs; bs = bdrv_next(&it), i++) {
 408         if (bdrv_is_read_only(bs)) {
 409             continue;
 410         }
 411
 412         sectors = bdrv_nb_sectors(bs);
 413         if (sectors <= 0) {
 414             ret = sectors;
 415             bdrv_next_cleanup(&it);
 416             goto out;
 417         }
 418
 419         bmds = g_new0(BlkMigDevState, 1);
 420         bmds->blk = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 421         bmds->blk_name = g_strdup(bdrv_get_device_name(bs));
 422         bmds->bulk_completed = 0;
 423         bmds->total_sectors = sectors;
 424         bmds->completed_sectors = 0;
 425         bmds->shared_base = migrate_use_block_incremental();
 426
 427         assert(i < num_bs);
 428         bmds_bs[i].bmds = bmds;
 429         bmds_bs[i].bs = bs;
 430
 431         block_mig_state.total_sector_sum += sectors;
 432
 433         if (bmds->shared_base) {
 434             DPRINTF("Start migration for %s with shared base image\n",
 435                     bdrv_get_device_name(bs));
 436         } else {
 437             DPRINTF("Start full migration for %s\n", bdrv_get_device_name(bs));
 438         }
 439
 440         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 441     }
 442
 443     /* Can only insert new BDSes now because doing so while iterating block
 444      * devices may end up in a deadlock (iterating the new BDSes, too). */
 445     for (i = 0; i < num_bs; i++) {
 446         BlkMigDevState *bmds = bmds_bs[i].bmds;
 447         BlockDriverState *bs = bmds_bs[i].bs;
 448
 449         if (bmds) {
 450             ret = blk_insert_bs(bmds->blk, bs, &local_err);
 451             if (ret < 0) {
 452                 error_report_err(local_err);
 453                 goto out;
 454             }
 455
 456             alloc_aio_bitmap(bmds);
 457             error_setg(&bmds->blocker, "block device is in use by migration");
 458             bdrv_op_block_all(bs, bmds->blocker);
 459         }
 460     }
 461
 462     ret = 0;
 463 out:
 464     g_free(bmds_bs);
 465     return ret;
 466 }
 467
 468 /* Called with no lock taken.  */
 469
 470 static int blk_mig_save_bulked_block(QEMUFile *f)
 471 {
 472     int64_t completed_sector_sum = 0;
 473     BlkMigDevState *bmds;
 474     int progress;
 475     int ret = 0;
 476
 477     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 478         if (bmds->bulk_completed == 0) {
 479             if (mig_save_device_bulk(f, bmds) == 1) {
 480                 /* completed bulk section for this device */
 481                 bmds->bulk_completed = 1;
 482             }
 483             completed_sector_sum += bmds->completed_sectors;
 484             ret = 1;
 485             break;
 486         } else {
 487             completed_sector_sum += bmds->completed_sectors;
 488         }
 489     }
 490
 491     if (block_mig_state.total_sector_sum != 0) {
 492         progress = completed_sector_sum * 100 /
 493                    block_mig_state.total_sector_sum;
 494     } else {
 495         progress = 100;
 496     }
 497     if (progress != block_mig_state.prev_progress) {
 498         block_mig_state.prev_progress = progress;
 499         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 500                          | BLK_MIG_FLAG_PROGRESS);
 501         DPRINTF("Completed %d %%\r", progress);
 502     }
 503
 504     return ret;
 505 }
 506
 507 static void blk_mig_reset_dirty_cursor(void)
 508 {
 509     BlkMigDevState *bmds;
 510
 511     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 512         bmds->cur_dirty = 0;
 513     }
 514 }
 515
 516 /* Called with iothread lock and AioContext taken.  */
 517
 518 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
 519                                  int is_async)
 520 {
 521     BlkMigBlock *blk;
 522     BlockDriverState *bs = blk_bs(bmds->blk);
 523     int64_t total_sectors = bmds->total_sectors;
 524     int64_t sector;
 525     int nr_sectors;
 526     int ret = -EIO;
 527
 528     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
 529         blk_mig_lock();
 530         if (bmds_aio_inflight(bmds, sector)) {
 531             blk_mig_unlock();
 532             blk_drain(bmds->blk);
 533         } else {
 534             blk_mig_unlock();
 535         }
 536         bdrv_dirty_bitmap_lock(bmds->dirty_bitmap);
 537         if (bdrv_get_dirty_locked(bs, bmds->dirty_bitmap,
 538                                   sector * BDRV_SECTOR_SIZE)) {
 539             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 540                 nr_sectors = total_sectors - sector;
 541             } else {
 542                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 543             }
 544             bdrv_reset_dirty_bitmap_locked(bmds->dirty_bitmap,
 545                                            sector * BDRV_SECTOR_SIZE,
 546                                            nr_sectors * BDRV_SECTOR_SIZE);
 547             bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 548
 549             blk = g_new(BlkMigBlock, 1);
 550             blk->buf = g_malloc(BLOCK_SIZE);
 551             blk->bmds = bmds;
 552             blk->sector = sector;
 553             blk->nr_sectors = nr_sectors;
 554
 555             if (is_async) {
 556                 qemu_iovec_init_buf(&blk->qiov, blk->buf,
 557                                     nr_sectors * BDRV_SECTOR_SIZE);
 558
 559                 blk->aiocb = blk_aio_preadv(bmds->blk,
 560                                             sector * BDRV_SECTOR_SIZE,
 561                                             &blk->qiov, 0, blk_mig_read_cb,
 562                                             blk);
 563
 564                 blk_mig_lock();
 565                 block_mig_state.submitted++;
 566                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 567                 blk_mig_unlock();
 568             } else {
 569                 ret = blk_pread(bmds->blk, sector * BDRV_SECTOR_SIZE, blk->buf,
 570                                 nr_sectors * BDRV_SECTOR_SIZE);
 571                 if (ret < 0) {
 572                     goto error;
 573                 }
 574                 blk_send(f, blk);
 575
 576                 g_free(blk->buf);
 577                 g_free(blk);
 578             }
 579
 580             sector += nr_sectors;
 581             bmds->cur_dirty = sector;
 582             break;
 583         }
 584
 585         bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 586         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 587         bmds->cur_dirty = sector;
 588     }
 589
 590     return (bmds->cur_dirty >= bmds->total_sectors);
 591
 592 error:
 593     DPRINTF("Error reading sector %" PRId64 "\n", sector);
 594     g_free(blk->buf);
 595     g_free(blk);
 596     return ret;
 597 }
 598
 599 /* Called with iothread lock taken.
 600  *
 601  * return value:
 602  * 0: too much data for max_downtime
 603  * 1: few enough data for max_downtime
 604 */
 605 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 606 {
 607     BlkMigDevState *bmds;
 608     int ret = 1;
 609
 610     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 611         aio_context_acquire(blk_get_aio_context(bmds->blk));
 612         ret = mig_save_device_dirty(f, bmds, is_async);
 613         aio_context_release(blk_get_aio_context(bmds->blk));
 614         if (ret <= 0) {
 615             break;
 616         }
 617     }
 618
 619     return ret;
 620 }
 621
 622 /* Called with no locks taken.  */
 623
 624 static int flush_blks(QEMUFile *f)
 625 {
 626     BlkMigBlock *blk;
 627     int ret = 0;
 628
 629     DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
 630             __func__, block_mig_state.submitted, block_mig_state.read_done,
 631             block_mig_state.transferred);
 632
 633     blk_mig_lock();
 634     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 635         if (qemu_file_rate_limit(f)) {
 636             break;
 637         }
 638         if (blk->ret < 0) {
 639             ret = blk->ret;
 640             break;
 641         }
 642
 643         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 644         blk_mig_unlock();
 645         blk_send(f, blk);
 646         blk_mig_lock();
 647
 648         g_free(blk->buf);
 649         g_free(blk);
 650
 651         block_mig_state.read_done--;
 652         block_mig_state.transferred++;
 653         assert(block_mig_state.read_done >= 0);
 654     }
 655     blk_mig_unlock();
 656
 657     DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __func__,
 658             block_mig_state.submitted, block_mig_state.read_done,
 659             block_mig_state.transferred);
 660     return ret;
 661 }
 662
 663 /* Called with iothread lock taken.  */
 664
 665 static int64_t get_remaining_dirty(void)
 666 {
 667     BlkMigDevState *bmds;
 668     int64_t dirty = 0;
 669
 670     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 671         aio_context_acquire(blk_get_aio_context(bmds->blk));
 672         dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
 673         aio_context_release(blk_get_aio_context(bmds->blk));
 674     }
 675
 676     return dirty;
 677 }
 678
 679
 680
 681 /* Called with iothread lock taken.  */
 682 static void block_migration_cleanup_bmds(void)
 683 {
 684     BlkMigDevState *bmds;
 685     AioContext *ctx;
 686
 687     unset_dirty_tracking();
 688
 689     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 690         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 691         bdrv_op_unblock_all(blk_bs(bmds->blk), bmds->blocker);
 692         error_free(bmds->blocker);
 693
 694         /* Save ctx, because bmds->blk can disappear during blk_unref.  */
 695         ctx = blk_get_aio_context(bmds->blk);
 696         aio_context_acquire(ctx);
 697         blk_unref(bmds->blk);
 698         aio_context_release(ctx);
 699
 700         g_free(bmds->blk_name);
 701         g_free(bmds->aio_bitmap);
 702         g_free(bmds);
 703     }
 704 }
 705
 706 /* Called with iothread lock taken.  */
 707 static void block_migration_cleanup(void *opaque)
 708 {
 709     BlkMigBlock *blk;
 710
 711     bdrv_drain_all();
 712
 713     block_migration_cleanup_bmds();
 714
 715     blk_mig_lock();
 716     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 717         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 718         g_free(blk->buf);
 719         g_free(blk);
 720     }
 721     blk_mig_unlock();
 722 }
 723
 724 static int block_save_setup(QEMUFile *f, void *opaque)
 725 {
 726     int ret;
 727
 728     DPRINTF("Enter save live setup submitted %d transferred %d\n",
 729             block_mig_state.submitted, block_mig_state.transferred);
 730
 731     qemu_mutex_lock_iothread();
 732     ret = init_blk_migration(f);
 733     if (ret < 0) {
 734         qemu_mutex_unlock_iothread();
 735         return ret;
 736     }
 737
 738     /* start track dirty blocks */
 739     ret = set_dirty_tracking();
 740
 741     qemu_mutex_unlock_iothread();
 742
 743     if (ret) {
 744         return ret;
 745     }
 746
 747     ret = flush_blks(f);
 748     blk_mig_reset_dirty_cursor();
 749     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 750
 751     return ret;
 752 }
 753
 754 static int block_save_iterate(QEMUFile *f, void *opaque)
 755 {
 756     int ret;
 757     int64_t last_ftell = qemu_ftell(f);
 758     int64_t delta_ftell;
 759
 760     DPRINTF("Enter save live iterate submitted %d transferred %d\n",
 761             block_mig_state.submitted, block_mig_state.transferred);
 762
 763     ret = flush_blks(f);
 764     if (ret) {
 765         return ret;
 766     }
 767
 768     blk_mig_reset_dirty_cursor();
 769
 770     /* control the rate of transfer */
 771     blk_mig_lock();
 772     while (block_mig_state.read_done * BLOCK_SIZE <
 773            qemu_file_get_rate_limit(f) &&
 774            block_mig_state.submitted < MAX_PARALLEL_IO &&
 775            (block_mig_state.submitted + block_mig_state.read_done) <
 776            MAX_IO_BUFFERS) {
 777         blk_mig_unlock();
 778         if (block_mig_state.bulk_completed == 0) {
 779             /* first finish the bulk phase */
 780             if (blk_mig_save_bulked_block(f) == 0) {
 781                 /* finished saving bulk on all devices */
 782                 block_mig_state.bulk_completed = 1;
 783             }
 784             ret = 0;
 785         } else {
 786             /* Always called with iothread lock taken for
 787              * simplicity, block_save_complete also calls it.
 788              */
 789             qemu_mutex_lock_iothread();
 790             ret = blk_mig_save_dirty_block(f, 1);
 791             qemu_mutex_unlock_iothread();
 792         }
 793         if (ret < 0) {
 794             return ret;
 795         }
 796         blk_mig_lock();
 797         if (ret != 0) {
 798             /* no more dirty blocks */
 799             break;
 800         }
 801     }
 802     blk_mig_unlock();
 803
 804     ret = flush_blks(f);
 805     if (ret) {
 806         return ret;
 807     }
 808
 809     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 810     delta_ftell = qemu_ftell(f) - last_ftell;
 811     if (delta_ftell > 0) {
 812         return 1;
 813     } else if (delta_ftell < 0) {
 814         return -1;
 815     } else {
 816         return 0;
 817     }
 818 }
 819
 820 /* Called with iothread lock taken.  */
 821
 822 static int block_save_complete(QEMUFile *f, void *opaque)
 823 {
 824     int ret;
 825
 826     DPRINTF("Enter save live complete submitted %d transferred %d\n",
 827             block_mig_state.submitted, block_mig_state.transferred);
 828
 829     ret = flush_blks(f);
 830     if (ret) {
 831         return ret;
 832     }
 833
 834     blk_mig_reset_dirty_cursor();
 835
 836     /* we know for sure that save bulk is completed and
 837        all async read completed */
 838     blk_mig_lock();
 839     assert(block_mig_state.submitted == 0);
 840     blk_mig_unlock();
 841
 842     do {
 843         ret = blk_mig_save_dirty_block(f, 0);
 844         if (ret < 0) {
 845             return ret;
 846         }
 847     } while (ret == 0);
 848
 849     /* report completion */
 850     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 851
 852     DPRINTF("Block migration completed\n");
 853
 854     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 855
 856     /* Make sure that our BlockBackends are gone, so that the block driver
 857      * nodes can be inactivated. */
 858     block_migration_cleanup_bmds();
 859
 860     return 0;
 861 }
 862
 863 static void block_save_pending(QEMUFile *f, void *opaque, uint64_t max_size,
 864                                uint64_t *res_precopy_only,
 865                                uint64_t *res_compatible,
 866                                uint64_t *res_postcopy_only)
 867 {
 868     /* Estimate pending number of bytes to send */
 869     uint64_t pending;
 870
 871     qemu_mutex_lock_iothread();
 872     pending = get_remaining_dirty();
 873     qemu_mutex_unlock_iothread();
 874
 875     blk_mig_lock();
 876     pending += block_mig_state.submitted * BLOCK_SIZE +
 877                block_mig_state.read_done * BLOCK_SIZE;
 878     blk_mig_unlock();
 879
 880     /* Report at least one block pending during bulk phase */
 881     if (pending <= max_size && !block_mig_state.bulk_completed) {
 882         pending = max_size + BLOCK_SIZE;
 883     }
 884
 885     DPRINTF("Enter save live pending  %" PRIu64 "\n", pending);
 886     /* We don't do postcopy */
 887     *res_precopy_only += pending;
 888 }
 889
 890 static int block_load(QEMUFile *f, void *opaque, int version_id)
 891 {
 892     static int banner_printed;
 893     int len, flags;
 894     char device_name[256];
 895     int64_t addr;
 896     BlockBackend *blk, *blk_prev = NULL;
 897     Error *local_err = NULL;
 898     uint8_t *buf;
 899     int64_t total_sectors = 0;
 900     int nr_sectors;
 901     int ret;
 902     BlockDriverInfo bdi;
 903     int cluster_size = BLOCK_SIZE;
 904
 905     do {
 906         addr = qemu_get_be64(f);
 907
 908         flags = addr & ~BDRV_SECTOR_MASK;
 909         addr >>= BDRV_SECTOR_BITS;
 910
 911         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 912             /* get device name */
 913             len = qemu_get_byte(f);
 914             qemu_get_buffer(f, (uint8_t *)device_name, len);
 915             device_name[len] = '\0';
 916
 917             blk = blk_by_name(device_name);
 918             if (!blk) {
 919                 fprintf(stderr, "Error unknown block device %s\n",
 920                         device_name);
 921                 return -EINVAL;
 922             }
 923
 924             if (blk != blk_prev) {
 925                 blk_prev = blk;
 926                 total_sectors = blk_nb_sectors(blk);
 927                 if (total_sectors <= 0) {
 928                     error_report("Error getting length of block device %s",
 929                                  device_name);
 930                     return -EINVAL;
 931                 }
 932
 933                 blk_invalidate_cache(blk, &local_err);
 934                 if (local_err) {
 935                     error_report_err(local_err);
 936                     return -EINVAL;
 937                 }
 938
 939                 ret = bdrv_get_info(blk_bs(blk), &bdi);
 940                 if (ret == 0 && bdi.cluster_size > 0 &&
 941                     bdi.cluster_size <= BLOCK_SIZE &&
 942                     BLOCK_SIZE % bdi.cluster_size == 0) {
 943                     cluster_size = bdi.cluster_size;
 944                 } else {
 945                     cluster_size = BLOCK_SIZE;
 946                 }
 947             }
 948
 949             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 950                 nr_sectors = total_sectors - addr;
 951             } else {
 952                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 953             }
 954
 955             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 956                 ret = blk_pwrite_zeroes(blk, addr * BDRV_SECTOR_SIZE,
 957                                         nr_sectors * BDRV_SECTOR_SIZE,
 958                                         BDRV_REQ_MAY_UNMAP);
 959             } else {
 960                 int i;
 961                 int64_t cur_addr;
 962                 uint8_t *cur_buf;
 963
 964                 buf = g_malloc(BLOCK_SIZE);
 965                 qemu_get_buffer(f, buf, BLOCK_SIZE);
 966                 for (i = 0; i < BLOCK_SIZE / cluster_size; i++) {
 967                     cur_addr = addr * BDRV_SECTOR_SIZE + i * cluster_size;
 968                     cur_buf = buf + i * cluster_size;
 969
 970                     if ((!block_mig_state.zero_blocks ||
 971                         cluster_size < BLOCK_SIZE) &&
 972                         buffer_is_zero(cur_buf, cluster_size)) {
 973                         ret = blk_pwrite_zeroes(blk, cur_addr,
 974                                                 cluster_size,
 975                                                 BDRV_REQ_MAY_UNMAP);
 976                     } else {
 977                         ret = blk_pwrite(blk, cur_addr, cur_buf,
 978                                          cluster_size, 0);
 979                     }
 980                     if (ret < 0) {
 981                         break;
 982                     }
 983                 }
 984                 g_free(buf);
 985             }
 986
 987             if (ret < 0) {
 988                 return ret;
 989             }
 990         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 991             if (!banner_printed) {
 992                 printf("Receiving block device images\n");
 993                 banner_printed = 1;
 994             }
 995             printf("Completed %d %%%c", (int)addr,
 996                    (addr == 100) ? '\n' : '\r');
 997             fflush(stdout);
 998         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
 999             fprintf(stderr, "Unknown block migration flags: %#x\n", flags);
1000             return -EINVAL;
1001         }
1002         ret = qemu_file_get_error(f);
1003         if (ret != 0) {
1004             return ret;
1005         }
1006     } while (!(flags & BLK_MIG_FLAG_EOS));
1007
1008     return 0;
1009 }
1010
1011 static bool block_is_active(void *opaque)
1012 {
1013     return migrate_use_block();
1014 }
1015
1016 static SaveVMHandlers savevm_block_handlers = {
1017     .save_setup = block_save_setup,
1018     .save_live_iterate = block_save_iterate,
1019     .save_live_complete_precopy = block_save_complete,
1020     .save_live_pending = block_save_pending,
1021     .load_state = block_load,
1022     .save_cleanup = block_migration_cleanup,
1023     .is_active = block_is_active,
1024 };
1025
1026 void blk_mig_init(void)
1027 {
1028     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
1029     QSIMPLEQ_INIT(&block_mig_state.blk_list);
1030     qemu_mutex_init(&block_mig_state.lock);
1031
1032     register_savevm_live(NULL, "block", 0, 1, &savevm_block_handlers,
1033                          &block_mig_state);
1034 }