qemu/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 "sysemu/block-backend.h"
  30
  31 #define BLOCK_SIZE                       (1 << 20)
  32 #define BDRV_SECTORS_PER_DIRTY_CHUNK     (BLOCK_SIZE >> BDRV_SECTOR_BITS)
  33
  34 #define BLK_MIG_FLAG_DEVICE_BLOCK       0x01
  35 #define BLK_MIG_FLAG_EOS                0x02
  36 #define BLK_MIG_FLAG_PROGRESS           0x04
  37 #define BLK_MIG_FLAG_ZERO_BLOCK         0x08
  38
  39 #define MAX_IS_ALLOCATED_SEARCH 65536
  40
  41 #define MAX_INFLIGHT_IO 512
  42
  43 //#define DEBUG_BLK_MIGRATION
  44
  45 #ifdef DEBUG_BLK_MIGRATION
  46 #define DPRINTF(fmt, ...) \
  47     do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
  48 #else
  49 #define DPRINTF(fmt, ...) \
  50     do { } while (0)
  51 #endif
  52
  53 typedef struct BlkMigDevState {
  54     /* Written during setup phase.  Can be read without a lock.  */
  55     BlockDriverState *bs;
  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     struct iovec iov;
  87     QEMUIOVector qiov;
  88     BlockAIOCB *aiocb;
  89
  90     /* Protected by block migration lock.  */
  91     int ret;
  92     QSIMPLEQ_ENTRY(BlkMigBlock) entry;
  93 } BlkMigBlock;
  94
  95 typedef struct BlkMigState {
  96     /* Written during setup phase.  Can be read without a lock.  */
  97     int blk_enable;
  98     int shared_base;
  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(bdrv_get_device_name(blk->bmds->bs));
 149     qemu_put_byte(f, len);
 150     qemu_put_buffer(f, (uint8_t *)bdrv_get_device_name(blk->bmds->bs), 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 < bdrv_nb_sectors(bmds->bs)) {
 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     BlockDriverState *bs = bmds->bs;
 239     int64_t bitmap_size;
 240
 241     bitmap_size = bdrv_nb_sectors(bs) + 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     BlockDriverState *bs = bmds->bs;
 272     BlkMigBlock *blk;
 273     int nr_sectors;
 274
 275     if (bmds->shared_base) {
 276         qemu_mutex_lock_iothread();
 277         aio_context_acquire(bdrv_get_aio_context(bs));
 278         while (cur_sector < total_sectors &&
 279                !bdrv_is_allocated(bs, cur_sector, MAX_IS_ALLOCATED_SEARCH,
 280                                   &nr_sectors)) {
 281             cur_sector += nr_sectors;
 282         }
 283         aio_context_release(bdrv_get_aio_context(bs));
 284         qemu_mutex_unlock_iothread();
 285     }
 286
 287     if (cur_sector >= total_sectors) {
 288         bmds->cur_sector = bmds->completed_sectors = total_sectors;
 289         return 1;
 290     }
 291
 292     bmds->completed_sectors = cur_sector;
 293
 294     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
 295
 296     /* we are going to transfer a full block even if it is not allocated */
 297     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 298
 299     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 300         nr_sectors = total_sectors - cur_sector;
 301     }
 302
 303     blk = g_new(BlkMigBlock, 1);
 304     blk->buf = g_malloc(BLOCK_SIZE);
 305     blk->bmds = bmds;
 306     blk->sector = cur_sector;
 307     blk->nr_sectors = nr_sectors;
 308
 309     blk->iov.iov_base = blk->buf;
 310     blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 311     qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 312
 313     blk_mig_lock();
 314     block_mig_state.submitted++;
 315     blk_mig_unlock();
 316
 317     /* We do not know if bs is under the main thread (and thus does
 318      * not acquire the AioContext when doing AIO) or rather under
 319      * dataplane.  Thus acquire both the iothread mutex and the
 320      * AioContext.
 321      *
 322      * This is ugly and will disappear when we make bdrv_* thread-safe,
 323      * without the need to acquire the AioContext.
 324      */
 325     qemu_mutex_lock_iothread();
 326     aio_context_acquire(bdrv_get_aio_context(bmds->bs));
 327     blk->aiocb = bdrv_aio_readv(bs, cur_sector, &blk->qiov,
 328                                 nr_sectors, blk_mig_read_cb, blk);
 329
 330     bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, cur_sector, nr_sectors);
 331     aio_context_release(bdrv_get_aio_context(bmds->bs));
 332     qemu_mutex_unlock_iothread();
 333
 334     bmds->cur_sector = cur_sector + nr_sectors;
 335     return (bmds->cur_sector >= total_sectors);
 336 }
 337
 338 /* Called with iothread lock taken.  */
 339
 340 static int set_dirty_tracking(void)
 341 {
 342     BlkMigDevState *bmds;
 343     int ret;
 344
 345     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 346         aio_context_acquire(bdrv_get_aio_context(bmds->bs));
 347         bmds->dirty_bitmap = bdrv_create_dirty_bitmap(bmds->bs, BLOCK_SIZE,
 348                                                       NULL, NULL);
 349         aio_context_release(bdrv_get_aio_context(bmds->bs));
 350         if (!bmds->dirty_bitmap) {
 351             ret = -errno;
 352             goto fail;
 353         }
 354     }
 355     return 0;
 356
 357 fail:
 358     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 359         if (bmds->dirty_bitmap) {
 360             aio_context_acquire(bdrv_get_aio_context(bmds->bs));
 361             bdrv_release_dirty_bitmap(bmds->bs, bmds->dirty_bitmap);
 362             aio_context_release(bdrv_get_aio_context(bmds->bs));
 363         }
 364     }
 365     return ret;
 366 }
 367
 368 /* Called with iothread lock taken.  */
 369
 370 static void unset_dirty_tracking(void)
 371 {
 372     BlkMigDevState *bmds;
 373
 374     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 375         aio_context_acquire(bdrv_get_aio_context(bmds->bs));
 376         bdrv_release_dirty_bitmap(bmds->bs, bmds->dirty_bitmap);
 377         aio_context_release(bdrv_get_aio_context(bmds->bs));
 378     }
 379 }
 380
 381 static void init_blk_migration(QEMUFile *f)
 382 {
 383     BlockDriverState *bs;
 384     BlkMigDevState *bmds;
 385     int64_t sectors;
 386
 387     block_mig_state.submitted = 0;
 388     block_mig_state.read_done = 0;
 389     block_mig_state.transferred = 0;
 390     block_mig_state.total_sector_sum = 0;
 391     block_mig_state.prev_progress = -1;
 392     block_mig_state.bulk_completed = 0;
 393     block_mig_state.zero_blocks = migrate_zero_blocks();
 394
 395     for (bs = bdrv_next(NULL); bs; bs = bdrv_next(bs)) {
 396         if (bdrv_is_read_only(bs)) {
 397             continue;
 398         }
 399
 400         sectors = bdrv_nb_sectors(bs);
 401         if (sectors <= 0) {
 402             return;
 403         }
 404
 405         bmds = g_new0(BlkMigDevState, 1);
 406         bmds->bs = bs;
 407         bmds->bulk_completed = 0;
 408         bmds->total_sectors = sectors;
 409         bmds->completed_sectors = 0;
 410         bmds->shared_base = block_mig_state.shared_base;
 411         alloc_aio_bitmap(bmds);
 412         error_setg(&bmds->blocker, "block device is in use by migration");
 413         bdrv_op_block_all(bs, bmds->blocker);
 414         bdrv_ref(bs);
 415
 416         block_mig_state.total_sector_sum += sectors;
 417
 418         if (bmds->shared_base) {
 419             DPRINTF("Start migration for %s with shared base image\n",
 420                     bdrv_get_device_name(bs));
 421         } else {
 422             DPRINTF("Start full migration for %s\n", bdrv_get_device_name(bs));
 423         }
 424
 425         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 426     }
 427 }
 428
 429 /* Called with no lock taken.  */
 430
 431 static int blk_mig_save_bulked_block(QEMUFile *f)
 432 {
 433     int64_t completed_sector_sum = 0;
 434     BlkMigDevState *bmds;
 435     int progress;
 436     int ret = 0;
 437
 438     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 439         if (bmds->bulk_completed == 0) {
 440             if (mig_save_device_bulk(f, bmds) == 1) {
 441                 /* completed bulk section for this device */
 442                 bmds->bulk_completed = 1;
 443             }
 444             completed_sector_sum += bmds->completed_sectors;
 445             ret = 1;
 446             break;
 447         } else {
 448             completed_sector_sum += bmds->completed_sectors;
 449         }
 450     }
 451
 452     if (block_mig_state.total_sector_sum != 0) {
 453         progress = completed_sector_sum * 100 /
 454                    block_mig_state.total_sector_sum;
 455     } else {
 456         progress = 100;
 457     }
 458     if (progress != block_mig_state.prev_progress) {
 459         block_mig_state.prev_progress = progress;
 460         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 461                          | BLK_MIG_FLAG_PROGRESS);
 462         DPRINTF("Completed %d %%\r", progress);
 463     }
 464
 465     return ret;
 466 }
 467
 468 static void blk_mig_reset_dirty_cursor(void)
 469 {
 470     BlkMigDevState *bmds;
 471
 472     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 473         bmds->cur_dirty = 0;
 474     }
 475 }
 476
 477 /* Called with iothread lock and AioContext taken.  */
 478
 479 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
 480                                  int is_async)
 481 {
 482     BlkMigBlock *blk;
 483     int64_t total_sectors = bmds->total_sectors;
 484     int64_t sector;
 485     int nr_sectors;
 486     int ret = -EIO;
 487
 488     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
 489         blk_mig_lock();
 490         if (bmds_aio_inflight(bmds, sector)) {
 491             blk_mig_unlock();
 492             bdrv_drain(bmds->bs);
 493         } else {
 494             blk_mig_unlock();
 495         }
 496         if (bdrv_get_dirty(bmds->bs, bmds->dirty_bitmap, sector)) {
 497
 498             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 499                 nr_sectors = total_sectors - sector;
 500             } else {
 501                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 502             }
 503             blk = g_new(BlkMigBlock, 1);
 504             blk->buf = g_malloc(BLOCK_SIZE);
 505             blk->bmds = bmds;
 506             blk->sector = sector;
 507             blk->nr_sectors = nr_sectors;
 508
 509             if (is_async) {
 510                 blk->iov.iov_base = blk->buf;
 511                 blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 512                 qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 513
 514                 blk->aiocb = bdrv_aio_readv(bmds->bs, sector, &blk->qiov,
 515                                             nr_sectors, blk_mig_read_cb, blk);
 516
 517                 blk_mig_lock();
 518                 block_mig_state.submitted++;
 519                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 520                 blk_mig_unlock();
 521             } else {
 522                 ret = bdrv_read(bmds->bs, sector, blk->buf, nr_sectors);
 523                 if (ret < 0) {
 524                     goto error;
 525                 }
 526                 blk_send(f, blk);
 527
 528                 g_free(blk->buf);
 529                 g_free(blk);
 530             }
 531
 532             bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, sector, nr_sectors);
 533             break;
 534         }
 535         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 536         bmds->cur_dirty = sector;
 537     }
 538
 539     return (bmds->cur_dirty >= bmds->total_sectors);
 540
 541 error:
 542     DPRINTF("Error reading sector %" PRId64 "\n", sector);
 543     g_free(blk->buf);
 544     g_free(blk);
 545     return ret;
 546 }
 547
 548 /* Called with iothread lock taken.
 549  *
 550  * return value:
 551  * 0: too much data for max_downtime
 552  * 1: few enough data for max_downtime
 553 */
 554 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 555 {
 556     BlkMigDevState *bmds;
 557     int ret = 1;
 558
 559     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 560         aio_context_acquire(bdrv_get_aio_context(bmds->bs));
 561         ret = mig_save_device_dirty(f, bmds, is_async);
 562         aio_context_release(bdrv_get_aio_context(bmds->bs));
 563         if (ret <= 0) {
 564             break;
 565         }
 566     }
 567
 568     return ret;
 569 }
 570
 571 /* Called with no locks taken.  */
 572
 573 static int flush_blks(QEMUFile *f)
 574 {
 575     BlkMigBlock *blk;
 576     int ret = 0;
 577
 578     DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
 579             __FUNCTION__, block_mig_state.submitted, block_mig_state.read_done,
 580             block_mig_state.transferred);
 581
 582     blk_mig_lock();
 583     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 584         if (qemu_file_rate_limit(f)) {
 585             break;
 586         }
 587         if (blk->ret < 0) {
 588             ret = blk->ret;
 589             break;
 590         }
 591
 592         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 593         blk_mig_unlock();
 594         blk_send(f, blk);
 595         blk_mig_lock();
 596
 597         g_free(blk->buf);
 598         g_free(blk);
 599
 600         block_mig_state.read_done--;
 601         block_mig_state.transferred++;
 602         assert(block_mig_state.read_done >= 0);
 603     }
 604     blk_mig_unlock();
 605
 606     DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__,
 607             block_mig_state.submitted, block_mig_state.read_done,
 608             block_mig_state.transferred);
 609     return ret;
 610 }
 611
 612 /* Called with iothread lock taken.  */
 613
 614 static int64_t get_remaining_dirty(void)
 615 {
 616     BlkMigDevState *bmds;
 617     int64_t dirty = 0;
 618
 619     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 620         aio_context_acquire(bdrv_get_aio_context(bmds->bs));
 621         dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
 622         aio_context_release(bdrv_get_aio_context(bmds->bs));
 623     }
 624
 625     return dirty << BDRV_SECTOR_BITS;
 626 }
 627
 628 /* Called with iothread lock taken.  */
 629
 630 static void block_migration_cleanup(void *opaque)
 631 {
 632     BlkMigDevState *bmds;
 633     BlkMigBlock *blk;
 634     AioContext *ctx;
 635
 636     bdrv_drain_all();
 637
 638     unset_dirty_tracking();
 639
 640     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 641         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 642         bdrv_op_unblock_all(bmds->bs, bmds->blocker);
 643         error_free(bmds->blocker);
 644
 645         /* Save ctx, because bmds->bs can disappear during bdrv_unref.  */
 646         ctx = bdrv_get_aio_context(bmds->bs);
 647         aio_context_acquire(ctx);
 648         bdrv_unref(bmds->bs);
 649         aio_context_release(ctx);
 650
 651         g_free(bmds->aio_bitmap);
 652         g_free(bmds);
 653     }
 654
 655     blk_mig_lock();
 656     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 657         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 658         g_free(blk->buf);
 659         g_free(blk);
 660     }
 661     blk_mig_unlock();
 662 }
 663
 664 static int block_save_setup(QEMUFile *f, void *opaque)
 665 {
 666     int ret;
 667
 668     DPRINTF("Enter save live setup submitted %d transferred %d\n",
 669             block_mig_state.submitted, block_mig_state.transferred);
 670
 671     qemu_mutex_lock_iothread();
 672     init_blk_migration(f);
 673
 674     /* start track dirty blocks */
 675     ret = set_dirty_tracking();
 676
 677     qemu_mutex_unlock_iothread();
 678
 679     if (ret) {
 680         return ret;
 681     }
 682
 683     ret = flush_blks(f);
 684     blk_mig_reset_dirty_cursor();
 685     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 686
 687     return ret;
 688 }
 689
 690 static int block_save_iterate(QEMUFile *f, void *opaque)
 691 {
 692     int ret;
 693     int64_t last_ftell = qemu_ftell(f);
 694     int64_t delta_ftell;
 695
 696     DPRINTF("Enter save live iterate submitted %d transferred %d\n",
 697             block_mig_state.submitted, block_mig_state.transferred);
 698
 699     ret = flush_blks(f);
 700     if (ret) {
 701         return ret;
 702     }
 703
 704     blk_mig_reset_dirty_cursor();
 705
 706     /* control the rate of transfer */
 707     blk_mig_lock();
 708     while ((block_mig_state.submitted +
 709             block_mig_state.read_done) * BLOCK_SIZE <
 710            qemu_file_get_rate_limit(f) &&
 711            (block_mig_state.submitted +
 712             block_mig_state.read_done) <
 713            MAX_INFLIGHT_IO) {
 714         blk_mig_unlock();
 715         if (block_mig_state.bulk_completed == 0) {
 716             /* first finish the bulk phase */
 717             if (blk_mig_save_bulked_block(f) == 0) {
 718                 /* finished saving bulk on all devices */
 719                 block_mig_state.bulk_completed = 1;
 720             }
 721             ret = 0;
 722         } else {
 723             /* Always called with iothread lock taken for
 724              * simplicity, block_save_complete also calls it.
 725              */
 726             qemu_mutex_lock_iothread();
 727             ret = blk_mig_save_dirty_block(f, 1);
 728             qemu_mutex_unlock_iothread();
 729         }
 730         if (ret < 0) {
 731             return ret;
 732         }
 733         blk_mig_lock();
 734         if (ret != 0) {
 735             /* no more dirty blocks */
 736             break;
 737         }
 738     }
 739     blk_mig_unlock();
 740
 741     ret = flush_blks(f);
 742     if (ret) {
 743         return ret;
 744     }
 745
 746     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 747     delta_ftell = qemu_ftell(f) - last_ftell;
 748     if (delta_ftell > 0) {
 749         return 1;
 750     } else if (delta_ftell < 0) {
 751         return -1;
 752     } else {
 753         return 0;
 754     }
 755 }
 756
 757 /* Called with iothread lock taken.  */
 758
 759 static int block_save_complete(QEMUFile *f, void *opaque)
 760 {
 761     int ret;
 762
 763     DPRINTF("Enter save live complete submitted %d transferred %d\n",
 764             block_mig_state.submitted, block_mig_state.transferred);
 765
 766     ret = flush_blks(f);
 767     if (ret) {
 768         return ret;
 769     }
 770
 771     blk_mig_reset_dirty_cursor();
 772
 773     /* we know for sure that save bulk is completed and
 774        all async read completed */
 775     blk_mig_lock();
 776     assert(block_mig_state.submitted == 0);
 777     blk_mig_unlock();
 778
 779     do {
 780         ret = blk_mig_save_dirty_block(f, 0);
 781         if (ret < 0) {
 782             return ret;
 783         }
 784     } while (ret == 0);
 785
 786     /* report completion */
 787     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 788
 789     DPRINTF("Block migration completed\n");
 790
 791     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 792
 793     return 0;
 794 }
 795
 796 static void block_save_pending(QEMUFile *f, void *opaque, uint64_t max_size,
 797                                uint64_t *non_postcopiable_pending,
 798                                uint64_t *postcopiable_pending)
 799 {
 800     /* Estimate pending number of bytes to send */
 801     uint64_t pending;
 802
 803     qemu_mutex_lock_iothread();
 804     pending = get_remaining_dirty();
 805     qemu_mutex_unlock_iothread();
 806
 807     blk_mig_lock();
 808     pending += block_mig_state.submitted * BLOCK_SIZE +
 809                block_mig_state.read_done * BLOCK_SIZE;
 810     blk_mig_unlock();
 811
 812     /* Report at least one block pending during bulk phase */
 813     if (pending <= max_size && !block_mig_state.bulk_completed) {
 814         pending = max_size + BLOCK_SIZE;
 815     }
 816
 817     DPRINTF("Enter save live pending  %" PRIu64 "\n", pending);
 818     /* We don't do postcopy */
 819     *non_postcopiable_pending += pending;
 820 }
 821
 822 static int block_load(QEMUFile *f, void *opaque, int version_id)
 823 {
 824     static int banner_printed;
 825     int len, flags;
 826     char device_name[256];
 827     int64_t addr;
 828     BlockDriverState *bs, *bs_prev = NULL;
 829     BlockBackend *blk;
 830     Error *local_err = NULL;
 831     uint8_t *buf;
 832     int64_t total_sectors = 0;
 833     int nr_sectors;
 834     int ret;
 835
 836     do {
 837         addr = qemu_get_be64(f);
 838
 839         flags = addr & ~BDRV_SECTOR_MASK;
 840         addr >>= BDRV_SECTOR_BITS;
 841
 842         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 843             /* get device name */
 844             len = qemu_get_byte(f);
 845             qemu_get_buffer(f, (uint8_t *)device_name, len);
 846             device_name[len] = '\0';
 847
 848             blk = blk_by_name(device_name);
 849             if (!blk) {
 850                 fprintf(stderr, "Error unknown block device %s\n",
 851                         device_name);
 852                 return -EINVAL;
 853             }
 854             bs = blk_bs(blk);
 855             if (!bs) {
 856                 fprintf(stderr, "Block device %s has no medium\n",
 857                         device_name);
 858                 return -EINVAL;
 859             }
 860
 861             if (bs != bs_prev) {
 862                 bs_prev = bs;
 863                 total_sectors = bdrv_nb_sectors(bs);
 864                 if (total_sectors <= 0) {
 865                     error_report("Error getting length of block device %s",
 866                                  device_name);
 867                     return -EINVAL;
 868                 }
 869
 870                 bdrv_invalidate_cache(bs, &local_err);
 871                 if (local_err) {
 872                     error_report_err(local_err);
 873                     return -EINVAL;
 874                 }
 875             }
 876
 877             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 878                 nr_sectors = total_sectors - addr;
 879             } else {
 880                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 881             }
 882
 883             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 884                 ret = bdrv_write_zeroes(bs, addr, nr_sectors,
 885                                         BDRV_REQ_MAY_UNMAP);
 886             } else {
 887                 buf = g_malloc(BLOCK_SIZE);
 888                 qemu_get_buffer(f, buf, BLOCK_SIZE);
 889                 ret = bdrv_write(bs, addr, buf, nr_sectors);
 890                 g_free(buf);
 891             }
 892
 893             if (ret < 0) {
 894                 return ret;
 895             }
 896         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 897             if (!banner_printed) {
 898                 printf("Receiving block device images\n");
 899                 banner_printed = 1;
 900             }
 901             printf("Completed %d %%%c", (int)addr,
 902                    (addr == 100) ? '\n' : '\r');
 903             fflush(stdout);
 904         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
 905             fprintf(stderr, "Unknown block migration flags: %#x\n", flags);
 906             return -EINVAL;
 907         }
 908         ret = qemu_file_get_error(f);
 909         if (ret != 0) {
 910             return ret;
 911         }
 912     } while (!(flags & BLK_MIG_FLAG_EOS));
 913
 914     return 0;
 915 }
 916
 917 static void block_set_params(const MigrationParams *params, void *opaque)
 918 {
 919     block_mig_state.blk_enable = params->blk;
 920     block_mig_state.shared_base = params->shared;
 921
 922     /* shared base means that blk_enable = 1 */
 923     block_mig_state.blk_enable |= params->shared;
 924 }
 925
 926 static bool block_is_active(void *opaque)
 927 {
 928     return block_mig_state.blk_enable == 1;
 929 }
 930
 931 static SaveVMHandlers savevm_block_handlers = {
 932     .set_params = block_set_params,
 933     .save_live_setup = block_save_setup,
 934     .save_live_iterate = block_save_iterate,
 935     .save_live_complete_precopy = block_save_complete,
 936     .save_live_pending = block_save_pending,
 937     .load_state = block_load,
 938     .cleanup = block_migration_cleanup,
 939     .is_active = block_is_active,
 940 };
 941
 942 void blk_mig_init(void)
 943 {
 944     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
 945     QSIMPLEQ_INIT(&block_mig_state.blk_list);
 946     qemu_mutex_init(&block_mig_state.lock);
 947
 948     register_savevm_live(NULL, "block", 0, 1, &savevm_block_handlers,
 949                          &block_mig_state);
 950 }