block-migration.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-common.h"
  17 #include "block/block_int.h"
  18 #include "hw/hw.h"
  19 #include "qemu/queue.h"
  20 #include "qemu/timer.h"
  21 #include "migration/block.h"
  22 #include "migration/migration.h"
  23 #include "sysemu/blockdev.h"
  24 #include <assert.h>
  25
  26 #define BLOCK_SIZE                       (1 << 20)
  27 #define BDRV_SECTORS_PER_DIRTY_CHUNK     (BLOCK_SIZE >> BDRV_SECTOR_BITS)
  28
  29 #define BLK_MIG_FLAG_DEVICE_BLOCK       0x01
  30 #define BLK_MIG_FLAG_EOS                0x02
  31 #define BLK_MIG_FLAG_PROGRESS           0x04
  32 #define BLK_MIG_FLAG_ZERO_BLOCK         0x08
  33
  34 #define MAX_IS_ALLOCATED_SEARCH 65536
  35
  36 //#define DEBUG_BLK_MIGRATION
  37
  38 #ifdef DEBUG_BLK_MIGRATION
  39 #define DPRINTF(fmt, ...) \
  40     do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
  41 #else
  42 #define DPRINTF(fmt, ...) \
  43     do { } while (0)
  44 #endif
  45
  46 typedef struct BlkMigDevState {
  47     /* Written during setup phase.  Can be read without a lock.  */
  48     BlockDriverState *bs;
  49     int shared_base;
  50     int64_t total_sectors;
  51     QSIMPLEQ_ENTRY(BlkMigDevState) entry;
  52
  53     /* Only used by migration thread.  Does not need a lock.  */
  54     int bulk_completed;
  55     int64_t cur_sector;
  56     int64_t cur_dirty;
  57
  58     /* Protected by block migration lock.  */
  59     unsigned long *aio_bitmap;
  60     int64_t completed_sectors;
  61     BdrvDirtyBitmap *dirty_bitmap;
  62 } BlkMigDevState;
  63
  64 typedef struct BlkMigBlock {
  65     /* Only used by migration thread.  */
  66     uint8_t *buf;
  67     BlkMigDevState *bmds;
  68     int64_t sector;
  69     int nr_sectors;
  70     struct iovec iov;
  71     QEMUIOVector qiov;
  72     BlockDriverAIOCB *aiocb;
  73
  74     /* Protected by block migration lock.  */
  75     int ret;
  76     QSIMPLEQ_ENTRY(BlkMigBlock) entry;
  77 } BlkMigBlock;
  78
  79 typedef struct BlkMigState {
  80     /* Written during setup phase.  Can be read without a lock.  */
  81     int blk_enable;
  82     int shared_base;
  83     QSIMPLEQ_HEAD(bmds_list, BlkMigDevState) bmds_list;
  84     int64_t total_sector_sum;
  85     bool zero_blocks;
  86
  87     /* Protected by lock.  */
  88     QSIMPLEQ_HEAD(blk_list, BlkMigBlock) blk_list;
  89     int submitted;
  90     int read_done;
  91
  92     /* Only used by migration thread.  Does not need a lock.  */
  93     int transferred;
  94     int prev_progress;
  95     int bulk_completed;
  96
  97     /* Lock must be taken _inside_ the iothread lock.  */
  98     QemuMutex lock;
  99 } BlkMigState;
 100
 101 static BlkMigState block_mig_state;
 102
 103 static void blk_mig_lock(void)
 104 {
 105     qemu_mutex_lock(&block_mig_state.lock);
 106 }
 107
 108 static void blk_mig_unlock(void)
 109 {
 110     qemu_mutex_unlock(&block_mig_state.lock);
 111 }
 112
 113 /* Must run outside of the iothread lock during the bulk phase,
 114  * or the VM will stall.
 115  */
 116
 117 static void blk_send(QEMUFile *f, BlkMigBlock * blk)
 118 {
 119     int len;
 120     uint64_t flags = BLK_MIG_FLAG_DEVICE_BLOCK;
 121
 122     if (block_mig_state.zero_blocks &&
 123         buffer_is_zero(blk->buf, BLOCK_SIZE)) {
 124         flags |= BLK_MIG_FLAG_ZERO_BLOCK;
 125     }
 126
 127     /* sector number and flags */
 128     qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
 129                      | flags);
 130
 131     /* device name */
 132     len = strlen(blk->bmds->bs->device_name);
 133     qemu_put_byte(f, len);
 134     qemu_put_buffer(f, (uint8_t *)blk->bmds->bs->device_name, len);
 135
 136     /* if a block is zero we need to flush here since the network
 137      * bandwidth is now a lot higher than the storage device bandwidth.
 138      * thus if we queue zero blocks we slow down the migration */
 139     if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 140         qemu_fflush(f);
 141         return;
 142     }
 143
 144     qemu_put_buffer(f, blk->buf, BLOCK_SIZE);
 145 }
 146
 147 int blk_mig_active(void)
 148 {
 149     return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
 150 }
 151
 152 uint64_t blk_mig_bytes_transferred(void)
 153 {
 154     BlkMigDevState *bmds;
 155     uint64_t sum = 0;
 156
 157     blk_mig_lock();
 158     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 159         sum += bmds->completed_sectors;
 160     }
 161     blk_mig_unlock();
 162     return sum << BDRV_SECTOR_BITS;
 163 }
 164
 165 uint64_t blk_mig_bytes_remaining(void)
 166 {
 167     return blk_mig_bytes_total() - blk_mig_bytes_transferred();
 168 }
 169
 170 uint64_t blk_mig_bytes_total(void)
 171 {
 172     BlkMigDevState *bmds;
 173     uint64_t sum = 0;
 174
 175     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 176         sum += bmds->total_sectors;
 177     }
 178     return sum << BDRV_SECTOR_BITS;
 179 }
 180
 181
 182 /* Called with migration lock held.  */
 183
 184 static int bmds_aio_inflight(BlkMigDevState *bmds, int64_t sector)
 185 {
 186     int64_t chunk = sector / (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK;
 187
 188     if ((sector << BDRV_SECTOR_BITS) < bdrv_getlength(bmds->bs)) {
 189         return !!(bmds->aio_bitmap[chunk / (sizeof(unsigned long) * 8)] &
 190             (1UL << (chunk % (sizeof(unsigned long) * 8))));
 191     } else {
 192         return 0;
 193     }
 194 }
 195
 196 /* Called with migration lock held.  */
 197
 198 static void bmds_set_aio_inflight(BlkMigDevState *bmds, int64_t sector_num,
 199                              int nb_sectors, int set)
 200 {
 201     int64_t start, end;
 202     unsigned long val, idx, bit;
 203
 204     start = sector_num / BDRV_SECTORS_PER_DIRTY_CHUNK;
 205     end = (sector_num + nb_sectors - 1) / BDRV_SECTORS_PER_DIRTY_CHUNK;
 206
 207     for (; start <= end; start++) {
 208         idx = start / (sizeof(unsigned long) * 8);
 209         bit = start % (sizeof(unsigned long) * 8);
 210         val = bmds->aio_bitmap[idx];
 211         if (set) {
 212             val |= 1UL << bit;
 213         } else {
 214             val &= ~(1UL << bit);
 215         }
 216         bmds->aio_bitmap[idx] = val;
 217     }
 218 }
 219
 220 static void alloc_aio_bitmap(BlkMigDevState *bmds)
 221 {
 222     BlockDriverState *bs = bmds->bs;
 223     int64_t bitmap_size;
 224
 225     bitmap_size = (bdrv_getlength(bs) >> BDRV_SECTOR_BITS) +
 226             BDRV_SECTORS_PER_DIRTY_CHUNK * 8 - 1;
 227     bitmap_size /= BDRV_SECTORS_PER_DIRTY_CHUNK * 8;
 228
 229     bmds->aio_bitmap = g_malloc0(bitmap_size);
 230 }
 231
 232 /* Never hold migration lock when yielding to the main loop!  */
 233
 234 static void blk_mig_read_cb(void *opaque, int ret)
 235 {
 236     BlkMigBlock *blk = opaque;
 237
 238     blk_mig_lock();
 239     blk->ret = ret;
 240
 241     QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry);
 242     bmds_set_aio_inflight(blk->bmds, blk->sector, blk->nr_sectors, 0);
 243
 244     block_mig_state.submitted--;
 245     block_mig_state.read_done++;
 246     assert(block_mig_state.submitted >= 0);
 247     blk_mig_unlock();
 248 }
 249
 250 /* Called with no lock taken.  */
 251
 252 static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds)
 253 {
 254     int64_t total_sectors = bmds->total_sectors;
 255     int64_t cur_sector = bmds->cur_sector;
 256     BlockDriverState *bs = bmds->bs;
 257     BlkMigBlock *blk;
 258     int nr_sectors;
 259
 260     if (bmds->shared_base) {
 261         qemu_mutex_lock_iothread();
 262         while (cur_sector < total_sectors &&
 263                !bdrv_is_allocated(bs, cur_sector, MAX_IS_ALLOCATED_SEARCH,
 264                                   &nr_sectors)) {
 265             cur_sector += nr_sectors;
 266         }
 267         qemu_mutex_unlock_iothread();
 268     }
 269
 270     if (cur_sector >= total_sectors) {
 271         bmds->cur_sector = bmds->completed_sectors = total_sectors;
 272         return 1;
 273     }
 274
 275     bmds->completed_sectors = cur_sector;
 276
 277     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
 278
 279     /* we are going to transfer a full block even if it is not allocated */
 280     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 281
 282     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 283         nr_sectors = total_sectors - cur_sector;
 284     }
 285
 286     blk = g_malloc(sizeof(BlkMigBlock));
 287     blk->buf = g_malloc(BLOCK_SIZE);
 288     blk->bmds = bmds;
 289     blk->sector = cur_sector;
 290     blk->nr_sectors = nr_sectors;
 291
 292     blk->iov.iov_base = blk->buf;
 293     blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 294     qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 295
 296     blk_mig_lock();
 297     block_mig_state.submitted++;
 298     blk_mig_unlock();
 299
 300     qemu_mutex_lock_iothread();
 301     blk->aiocb = bdrv_aio_readv(bs, cur_sector, &blk->qiov,
 302                                 nr_sectors, blk_mig_read_cb, blk);
 303
 304     bdrv_reset_dirty(bs, cur_sector, nr_sectors);
 305     qemu_mutex_unlock_iothread();
 306
 307     bmds->cur_sector = cur_sector + nr_sectors;
 308     return (bmds->cur_sector >= total_sectors);
 309 }
 310
 311 /* Called with iothread lock taken.  */
 312
 313 static void set_dirty_tracking(void)
 314 {
 315     BlkMigDevState *bmds;
 316
 317     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 318         bmds->dirty_bitmap = bdrv_create_dirty_bitmap(bmds->bs, BLOCK_SIZE);
 319     }
 320 }
 321
 322 static void unset_dirty_tracking(void)
 323 {
 324     BlkMigDevState *bmds;
 325
 326     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 327         bdrv_release_dirty_bitmap(bmds->bs, bmds->dirty_bitmap);
 328     }
 329 }
 330
 331 static void init_blk_migration_it(void *opaque, BlockDriverState *bs)
 332 {
 333     BlkMigDevState *bmds;
 334     int64_t sectors;
 335
 336     if (!bdrv_is_read_only(bs)) {
 337         sectors = bdrv_getlength(bs) >> BDRV_SECTOR_BITS;
 338         if (sectors <= 0) {
 339             return;
 340         }
 341
 342         bmds = g_malloc0(sizeof(BlkMigDevState));
 343         bmds->bs = bs;
 344         bmds->bulk_completed = 0;
 345         bmds->total_sectors = sectors;
 346         bmds->completed_sectors = 0;
 347         bmds->shared_base = block_mig_state.shared_base;
 348         alloc_aio_bitmap(bmds);
 349         bdrv_set_in_use(bs, 1);
 350         bdrv_ref(bs);
 351
 352         block_mig_state.total_sector_sum += sectors;
 353
 354         if (bmds->shared_base) {
 355             DPRINTF("Start migration for %s with shared base image\n",
 356                     bs->device_name);
 357         } else {
 358             DPRINTF("Start full migration for %s\n", bs->device_name);
 359         }
 360
 361         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 362     }
 363 }
 364
 365 static void init_blk_migration(QEMUFile *f)
 366 {
 367     block_mig_state.submitted = 0;
 368     block_mig_state.read_done = 0;
 369     block_mig_state.transferred = 0;
 370     block_mig_state.total_sector_sum = 0;
 371     block_mig_state.prev_progress = -1;
 372     block_mig_state.bulk_completed = 0;
 373     block_mig_state.zero_blocks = migrate_zero_blocks();
 374
 375     bdrv_iterate(init_blk_migration_it, NULL);
 376 }
 377
 378 /* Called with no lock taken.  */
 379
 380 static int blk_mig_save_bulked_block(QEMUFile *f)
 381 {
 382     int64_t completed_sector_sum = 0;
 383     BlkMigDevState *bmds;
 384     int progress;
 385     int ret = 0;
 386
 387     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 388         if (bmds->bulk_completed == 0) {
 389             if (mig_save_device_bulk(f, bmds) == 1) {
 390                 /* completed bulk section for this device */
 391                 bmds->bulk_completed = 1;
 392             }
 393             completed_sector_sum += bmds->completed_sectors;
 394             ret = 1;
 395             break;
 396         } else {
 397             completed_sector_sum += bmds->completed_sectors;
 398         }
 399     }
 400
 401     if (block_mig_state.total_sector_sum != 0) {
 402         progress = completed_sector_sum * 100 /
 403                    block_mig_state.total_sector_sum;
 404     } else {
 405         progress = 100;
 406     }
 407     if (progress != block_mig_state.prev_progress) {
 408         block_mig_state.prev_progress = progress;
 409         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 410                          | BLK_MIG_FLAG_PROGRESS);
 411         DPRINTF("Completed %d %%\r", progress);
 412     }
 413
 414     return ret;
 415 }
 416
 417 static void blk_mig_reset_dirty_cursor(void)
 418 {
 419     BlkMigDevState *bmds;
 420
 421     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 422         bmds->cur_dirty = 0;
 423     }
 424 }
 425
 426 /* Called with iothread lock taken.  */
 427
 428 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
 429                                  int is_async)
 430 {
 431     BlkMigBlock *blk;
 432     int64_t total_sectors = bmds->total_sectors;
 433     int64_t sector;
 434     int nr_sectors;
 435     int ret = -EIO;
 436
 437     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
 438         blk_mig_lock();
 439         if (bmds_aio_inflight(bmds, sector)) {
 440             blk_mig_unlock();
 441             bdrv_drain_all();
 442         } else {
 443             blk_mig_unlock();
 444         }
 445         if (bdrv_get_dirty(bmds->bs, bmds->dirty_bitmap, sector)) {
 446
 447             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 448                 nr_sectors = total_sectors - sector;
 449             } else {
 450                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 451             }
 452             blk = g_malloc(sizeof(BlkMigBlock));
 453             blk->buf = g_malloc(BLOCK_SIZE);
 454             blk->bmds = bmds;
 455             blk->sector = sector;
 456             blk->nr_sectors = nr_sectors;
 457
 458             if (is_async) {
 459                 blk->iov.iov_base = blk->buf;
 460                 blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 461                 qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 462
 463                 blk->aiocb = bdrv_aio_readv(bmds->bs, sector, &blk->qiov,
 464                                             nr_sectors, blk_mig_read_cb, blk);
 465
 466                 blk_mig_lock();
 467                 block_mig_state.submitted++;
 468                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 469                 blk_mig_unlock();
 470             } else {
 471                 ret = bdrv_read(bmds->bs, sector, blk->buf, nr_sectors);
 472                 if (ret < 0) {
 473                     goto error;
 474                 }
 475                 blk_send(f, blk);
 476
 477                 g_free(blk->buf);
 478                 g_free(blk);
 479             }
 480
 481             bdrv_reset_dirty(bmds->bs, sector, nr_sectors);
 482             break;
 483         }
 484         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 485         bmds->cur_dirty = sector;
 486     }
 487
 488     return (bmds->cur_dirty >= bmds->total_sectors);
 489
 490 error:
 491     DPRINTF("Error reading sector %" PRId64 "\n", sector);
 492     g_free(blk->buf);
 493     g_free(blk);
 494     return ret;
 495 }
 496
 497 /* Called with iothread lock taken.
 498  *
 499  * return value:
 500  * 0: too much data for max_downtime
 501  * 1: few enough data for max_downtime
 502 */
 503 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 504 {
 505     BlkMigDevState *bmds;
 506     int ret = 1;
 507
 508     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 509         ret = mig_save_device_dirty(f, bmds, is_async);
 510         if (ret <= 0) {
 511             break;
 512         }
 513     }
 514
 515     return ret;
 516 }
 517
 518 /* Called with no locks taken.  */
 519
 520 static int flush_blks(QEMUFile *f)
 521 {
 522     BlkMigBlock *blk;
 523     int ret = 0;
 524
 525     DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
 526             __FUNCTION__, block_mig_state.submitted, block_mig_state.read_done,
 527             block_mig_state.transferred);
 528
 529     blk_mig_lock();
 530     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 531         if (qemu_file_rate_limit(f)) {
 532             break;
 533         }
 534         if (blk->ret < 0) {
 535             ret = blk->ret;
 536             break;
 537         }
 538
 539         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 540         blk_mig_unlock();
 541         blk_send(f, blk);
 542         blk_mig_lock();
 543
 544         g_free(blk->buf);
 545         g_free(blk);
 546
 547         block_mig_state.read_done--;
 548         block_mig_state.transferred++;
 549         assert(block_mig_state.read_done >= 0);
 550     }
 551     blk_mig_unlock();
 552
 553     DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__,
 554             block_mig_state.submitted, block_mig_state.read_done,
 555             block_mig_state.transferred);
 556     return ret;
 557 }
 558
 559 /* Called with iothread lock taken.  */
 560
 561 static int64_t get_remaining_dirty(void)
 562 {
 563     BlkMigDevState *bmds;
 564     int64_t dirty = 0;
 565
 566     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 567         dirty += bdrv_get_dirty_count(bmds->bs, bmds->dirty_bitmap);
 568     }
 569
 570     return dirty << BDRV_SECTOR_BITS;
 571 }
 572
 573 /* Called with iothread lock taken.  */
 574
 575 static void blk_mig_cleanup(void)
 576 {
 577     BlkMigDevState *bmds;
 578     BlkMigBlock *blk;
 579
 580     bdrv_drain_all();
 581
 582     unset_dirty_tracking();
 583
 584     blk_mig_lock();
 585     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 586         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 587         bdrv_set_in_use(bmds->bs, 0);
 588         bdrv_unref(bmds->bs);
 589         g_free(bmds->aio_bitmap);
 590         g_free(bmds);
 591     }
 592
 593     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 594         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 595         g_free(blk->buf);
 596         g_free(blk);
 597     }
 598     blk_mig_unlock();
 599 }
 600
 601 static void block_migration_cancel(void *opaque)
 602 {
 603     blk_mig_cleanup();
 604 }
 605
 606 static int block_save_setup(QEMUFile *f, void *opaque)
 607 {
 608     int ret;
 609
 610     DPRINTF("Enter save live setup submitted %d transferred %d\n",
 611             block_mig_state.submitted, block_mig_state.transferred);
 612
 613     qemu_mutex_lock_iothread();
 614     init_blk_migration(f);
 615
 616     /* start track dirty blocks */
 617     set_dirty_tracking();
 618     qemu_mutex_unlock_iothread();
 619
 620     ret = flush_blks(f);
 621     blk_mig_reset_dirty_cursor();
 622     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 623
 624     return ret;
 625 }
 626
 627 static int block_save_iterate(QEMUFile *f, void *opaque)
 628 {
 629     int ret;
 630     int64_t last_ftell = qemu_ftell(f);
 631
 632     DPRINTF("Enter save live iterate submitted %d transferred %d\n",
 633             block_mig_state.submitted, block_mig_state.transferred);
 634
 635     ret = flush_blks(f);
 636     if (ret) {
 637         return ret;
 638     }
 639
 640     blk_mig_reset_dirty_cursor();
 641
 642     /* control the rate of transfer */
 643     blk_mig_lock();
 644     while ((block_mig_state.submitted +
 645             block_mig_state.read_done) * BLOCK_SIZE <
 646            qemu_file_get_rate_limit(f)) {
 647         blk_mig_unlock();
 648         if (block_mig_state.bulk_completed == 0) {
 649             /* first finish the bulk phase */
 650             if (blk_mig_save_bulked_block(f) == 0) {
 651                 /* finished saving bulk on all devices */
 652                 block_mig_state.bulk_completed = 1;
 653             }
 654             ret = 0;
 655         } else {
 656             /* Always called with iothread lock taken for
 657              * simplicity, block_save_complete also calls it.
 658              */
 659             qemu_mutex_lock_iothread();
 660             ret = blk_mig_save_dirty_block(f, 1);
 661             qemu_mutex_unlock_iothread();
 662         }
 663         if (ret < 0) {
 664             return ret;
 665         }
 666         blk_mig_lock();
 667         if (ret != 0) {
 668             /* no more dirty blocks */
 669             break;
 670         }
 671     }
 672     blk_mig_unlock();
 673
 674     ret = flush_blks(f);
 675     if (ret) {
 676         return ret;
 677     }
 678
 679     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 680     return qemu_ftell(f) - last_ftell;
 681 }
 682
 683 /* Called with iothread lock taken.  */
 684
 685 static int block_save_complete(QEMUFile *f, void *opaque)
 686 {
 687     int ret;
 688
 689     DPRINTF("Enter save live complete submitted %d transferred %d\n",
 690             block_mig_state.submitted, block_mig_state.transferred);
 691
 692     ret = flush_blks(f);
 693     if (ret) {
 694         return ret;
 695     }
 696
 697     blk_mig_reset_dirty_cursor();
 698
 699     /* we know for sure that save bulk is completed and
 700        all async read completed */
 701     blk_mig_lock();
 702     assert(block_mig_state.submitted == 0);
 703     blk_mig_unlock();
 704
 705     do {
 706         ret = blk_mig_save_dirty_block(f, 0);
 707         if (ret < 0) {
 708             return ret;
 709         }
 710     } while (ret == 0);
 711
 712     /* report completion */
 713     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 714
 715     DPRINTF("Block migration completed\n");
 716
 717     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 718
 719     blk_mig_cleanup();
 720     return 0;
 721 }
 722
 723 static uint64_t block_save_pending(QEMUFile *f, void *opaque, uint64_t max_size)
 724 {
 725     /* Estimate pending number of bytes to send */
 726     uint64_t pending;
 727
 728     qemu_mutex_lock_iothread();
 729     blk_mig_lock();
 730     pending = get_remaining_dirty() +
 731                        block_mig_state.submitted * BLOCK_SIZE +
 732                        block_mig_state.read_done * BLOCK_SIZE;
 733
 734     /* Report at least one block pending during bulk phase */
 735     if (pending == 0 && !block_mig_state.bulk_completed) {
 736         pending = BLOCK_SIZE;
 737     }
 738     blk_mig_unlock();
 739     qemu_mutex_unlock_iothread();
 740
 741     DPRINTF("Enter save live pending  %" PRIu64 "\n", pending);
 742     return pending;
 743 }
 744
 745 static int block_load(QEMUFile *f, void *opaque, int version_id)
 746 {
 747     static int banner_printed;
 748     int len, flags;
 749     char device_name[256];
 750     int64_t addr;
 751     BlockDriverState *bs, *bs_prev = NULL;
 752     uint8_t *buf;
 753     int64_t total_sectors = 0;
 754     int nr_sectors;
 755     int ret;
 756
 757     do {
 758         addr = qemu_get_be64(f);
 759
 760         flags = addr & ~BDRV_SECTOR_MASK;
 761         addr >>= BDRV_SECTOR_BITS;
 762
 763         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 764             /* get device name */
 765             len = qemu_get_byte(f);
 766             qemu_get_buffer(f, (uint8_t *)device_name, len);
 767             device_name[len] = '\0';
 768
 769             bs = bdrv_find(device_name);
 770             if (!bs) {
 771                 fprintf(stderr, "Error unknown block device %s\n",
 772                         device_name);
 773                 return -EINVAL;
 774             }
 775
 776             if (bs != bs_prev) {
 777                 bs_prev = bs;
 778                 total_sectors = bdrv_getlength(bs) >> BDRV_SECTOR_BITS;
 779                 if (total_sectors <= 0) {
 780                     error_report("Error getting length of block device %s",
 781                                  device_name);
 782                     return -EINVAL;
 783                 }
 784             }
 785
 786             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 787                 nr_sectors = total_sectors - addr;
 788             } else {
 789                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 790             }
 791
 792             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 793                 ret = bdrv_write_zeroes(bs, addr, nr_sectors,
 794                                         BDRV_REQ_MAY_UNMAP);
 795             } else {
 796                 buf = g_malloc(BLOCK_SIZE);
 797                 qemu_get_buffer(f, buf, BLOCK_SIZE);
 798                 ret = bdrv_write(bs, addr, buf, nr_sectors);
 799                 g_free(buf);
 800             }
 801
 802             if (ret < 0) {
 803                 return ret;
 804             }
 805         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 806             if (!banner_printed) {
 807                 printf("Receiving block device images\n");
 808                 banner_printed = 1;
 809             }
 810             printf("Completed %d %%%c", (int)addr,
 811                    (addr == 100) ? '\n' : '\r');
 812             fflush(stdout);
 813         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
 814             fprintf(stderr, "Unknown block migration flags: %#x\n", flags);
 815             return -EINVAL;
 816         }
 817         ret = qemu_file_get_error(f);
 818         if (ret != 0) {
 819             return ret;
 820         }
 821     } while (!(flags & BLK_MIG_FLAG_EOS));
 822
 823     return 0;
 824 }
 825
 826 static void block_set_params(const MigrationParams *params, void *opaque)
 827 {
 828     block_mig_state.blk_enable = params->blk;
 829     block_mig_state.shared_base = params->shared;
 830
 831     /* shared base means that blk_enable = 1 */
 832     block_mig_state.blk_enable |= params->shared;
 833 }
 834
 835 static bool block_is_active(void *opaque)
 836 {
 837     return block_mig_state.blk_enable == 1;
 838 }
 839
 840 SaveVMHandlers savevm_block_handlers = {
 841     .set_params = block_set_params,
 842     .save_live_setup = block_save_setup,
 843     .save_live_iterate = block_save_iterate,
 844     .save_live_complete = block_save_complete,
 845     .save_live_pending = block_save_pending,
 846     .load_state = block_load,
 847     .cancel = block_migration_cancel,
 848     .is_active = block_is_active,
 849 };
 850
 851 void blk_mig_init(void)
 852 {
 853     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
 854     QSIMPLEQ_INIT(&block_mig_state.blk_list);
 855     qemu_mutex_init(&block_mig_state.lock);
 856
 857     register_savevm_live(NULL, "block", 0, 1, &savevm_block_handlers,
 858                          &block_mig_state);
 859 }