migration/block.c

   1 /*
   2  * QEMU live block migration
   3  *
   4  * Copyright IBM, Corp. 2009
   5  *
   6  * Authors:
   7  *  Liran Schour   <[email protected]>
   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 "block.h"
  27 #include "migration/misc.h"
  28 #include "migration.h"
  29 #include "migration/register.h"
  30 #include "sysemu/blockdev.h"
  31 #include "qemu-file.h"
  32 #include "migration/vmstate.h"
  33 #include "sysemu/block-backend.h"
  34
  35 #define BLOCK_SIZE                       (1 << 20)
  36 #define BDRV_SECTORS_PER_DIRTY_CHUNK     (BLOCK_SIZE >> BDRV_SECTOR_BITS)
  37
  38 #define BLK_MIG_FLAG_DEVICE_BLOCK       0x01
  39 #define BLK_MIG_FLAG_EOS                0x02
  40 #define BLK_MIG_FLAG_PROGRESS           0x04
  41 #define BLK_MIG_FLAG_ZERO_BLOCK         0x08
  42
  43 #define MAX_IS_ALLOCATED_SEARCH 65536
  44
  45 #define MAX_INFLIGHT_IO 512
  46
  47 //#define DEBUG_BLK_MIGRATION
  48
  49 #ifdef DEBUG_BLK_MIGRATION
  50 #define DPRINTF(fmt, ...) \
  51     do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
  52 #else
  53 #define DPRINTF(fmt, ...) \
  54     do { } while (0)
  55 #endif
  56
  57 typedef struct BlkMigDevState {
  58     /* Written during setup phase.  Can be read without a lock.  */
  59     BlockBackend *blk;
  60     char *blk_name;
  61     int shared_base;
  62     int64_t total_sectors;
  63     QSIMPLEQ_ENTRY(BlkMigDevState) entry;
  64     Error *blocker;
  65
  66     /* Only used by migration thread.  Does not need a lock.  */
  67     int bulk_completed;
  68     int64_t cur_sector;
  69     int64_t cur_dirty;
  70
  71     /* Data in the aio_bitmap is protected by block migration lock.
  72      * Allocation and free happen during setup and cleanup respectively.
  73      */
  74     unsigned long *aio_bitmap;
  75
  76     /* Protected by block migration lock.  */
  77     int64_t completed_sectors;
  78
  79     /* During migration this is protected by iothread lock / AioContext.
  80      * Allocation and free happen during setup and cleanup respectively.
  81      */
  82     BdrvDirtyBitmap *dirty_bitmap;
  83 } BlkMigDevState;
  84
  85 typedef struct BlkMigBlock {
  86     /* Only used by migration thread.  */
  87     uint8_t *buf;
  88     BlkMigDevState *bmds;
  89     int64_t sector;
  90     int nr_sectors;
  91     struct iovec iov;
  92     QEMUIOVector qiov;
  93     BlockAIOCB *aiocb;
  94
  95     /* Protected by block migration lock.  */
  96     int ret;
  97     QSIMPLEQ_ENTRY(BlkMigBlock) entry;
  98 } BlkMigBlock;
  99
 100 typedef struct BlkMigState {
 101     QSIMPLEQ_HEAD(bmds_list, BlkMigDevState) bmds_list;
 102     int64_t total_sector_sum;
 103     bool zero_blocks;
 104
 105     /* Protected by lock.  */
 106     QSIMPLEQ_HEAD(blk_list, BlkMigBlock) blk_list;
 107     int submitted;
 108     int read_done;
 109
 110     /* Only used by migration thread.  Does not need a lock.  */
 111     int transferred;
 112     int prev_progress;
 113     int bulk_completed;
 114
 115     /* Lock must be taken _inside_ the iothread lock and any AioContexts.  */
 116     QemuMutex lock;
 117 } BlkMigState;
 118
 119 static BlkMigState block_mig_state;
 120
 121 static void blk_mig_lock(void)
 122 {
 123     qemu_mutex_lock(&block_mig_state.lock);
 124 }
 125
 126 static void blk_mig_unlock(void)
 127 {
 128     qemu_mutex_unlock(&block_mig_state.lock);
 129 }
 130
 131 /* Must run outside of the iothread lock during the bulk phase,
 132  * or the VM will stall.
 133  */
 134
 135 static void blk_send(QEMUFile *f, BlkMigBlock * blk)
 136 {
 137     int len;
 138     uint64_t flags = BLK_MIG_FLAG_DEVICE_BLOCK;
 139
 140     if (block_mig_state.zero_blocks &&
 141         buffer_is_zero(blk->buf, BLOCK_SIZE)) {
 142         flags |= BLK_MIG_FLAG_ZERO_BLOCK;
 143     }
 144
 145     /* sector number and flags */
 146     qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
 147                      | flags);
 148
 149     /* device name */
 150     len = strlen(blk->bmds->blk_name);
 151     qemu_put_byte(f, len);
 152     qemu_put_buffer(f, (uint8_t *) blk->bmds->blk_name, len);
 153
 154     /* if a block is zero we need to flush here since the network
 155      * bandwidth is now a lot higher than the storage device bandwidth.
 156      * thus if we queue zero blocks we slow down the migration */
 157     if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 158         qemu_fflush(f);
 159         return;
 160     }
 161
 162     qemu_put_buffer(f, blk->buf, BLOCK_SIZE);
 163 }
 164
 165 int blk_mig_active(void)
 166 {
 167     return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
 168 }
 169
 170 uint64_t blk_mig_bytes_transferred(void)
 171 {
 172     BlkMigDevState *bmds;
 173     uint64_t sum = 0;
 174
 175     blk_mig_lock();
 176     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 177         sum += bmds->completed_sectors;
 178     }
 179     blk_mig_unlock();
 180     return sum << BDRV_SECTOR_BITS;
 181 }
 182
 183 uint64_t blk_mig_bytes_remaining(void)
 184 {
 185     return blk_mig_bytes_total() - blk_mig_bytes_transferred();
 186 }
 187
 188 uint64_t blk_mig_bytes_total(void)
 189 {
 190     BlkMigDevState *bmds;
 191     uint64_t sum = 0;
 192
 193     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 194         sum += bmds->total_sectors;
 195     }
 196     return sum << BDRV_SECTOR_BITS;
 197 }
 198
 199
 200 /* Called with migration lock held.  */
 201
 202 static int bmds_aio_inflight(BlkMigDevState *bmds, int64_t sector)
 203 {
 204     int64_t chunk = sector / (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK;
 205
 206     if (sector < blk_nb_sectors(bmds->blk)) {
 207         return !!(bmds->aio_bitmap[chunk / (sizeof(unsigned long) * 8)] &
 208             (1UL << (chunk % (sizeof(unsigned long) * 8))));
 209     } else {
 210         return 0;
 211     }
 212 }
 213
 214 /* Called with migration lock held.  */
 215
 216 static void bmds_set_aio_inflight(BlkMigDevState *bmds, int64_t sector_num,
 217                              int nb_sectors, int set)
 218 {
 219     int64_t start, end;
 220     unsigned long val, idx, bit;
 221
 222     start = sector_num / BDRV_SECTORS_PER_DIRTY_CHUNK;
 223     end = (sector_num + nb_sectors - 1) / BDRV_SECTORS_PER_DIRTY_CHUNK;
 224
 225     for (; start <= end; start++) {
 226         idx = start / (sizeof(unsigned long) * 8);
 227         bit = start % (sizeof(unsigned long) * 8);
 228         val = bmds->aio_bitmap[idx];
 229         if (set) {
 230             val |= 1UL << bit;
 231         } else {
 232             val &= ~(1UL << bit);
 233         }
 234         bmds->aio_bitmap[idx] = val;
 235     }
 236 }
 237
 238 static void alloc_aio_bitmap(BlkMigDevState *bmds)
 239 {
 240     BlockBackend *bb = bmds->blk;
 241     int64_t bitmap_size;
 242
 243     bitmap_size = blk_nb_sectors(bb) + BDRV_SECTORS_PER_DIRTY_CHUNK * 8 - 1;
 244     bitmap_size /= BDRV_SECTORS_PER_DIRTY_CHUNK * 8;
 245
 246     bmds->aio_bitmap = g_malloc0(bitmap_size);
 247 }
 248
 249 /* Never hold migration lock when yielding to the main loop!  */
 250
 251 static void blk_mig_read_cb(void *opaque, int ret)
 252 {
 253     BlkMigBlock *blk = opaque;
 254
 255     blk_mig_lock();
 256     blk->ret = ret;
 257
 258     QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry);
 259     bmds_set_aio_inflight(blk->bmds, blk->sector, blk->nr_sectors, 0);
 260
 261     block_mig_state.submitted--;
 262     block_mig_state.read_done++;
 263     assert(block_mig_state.submitted >= 0);
 264     blk_mig_unlock();
 265 }
 266
 267 /* Called with no lock taken.  */
 268
 269 static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds)
 270 {
 271     int64_t total_sectors = bmds->total_sectors;
 272     int64_t cur_sector = bmds->cur_sector;
 273     BlockBackend *bb = bmds->blk;
 274     BlkMigBlock *blk;
 275     int nr_sectors;
 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 as
 281          * an allocated sector */
 282         while (cur_sector < total_sectors &&
 283                !bdrv_is_allocated(blk_bs(bb), cur_sector,
 284                                   MAX_IS_ALLOCATED_SEARCH, &nr_sectors)) {
 285             cur_sector += nr_sectors;
 286         }
 287         aio_context_release(blk_get_aio_context(bb));
 288         qemu_mutex_unlock_iothread();
 289     }
 290
 291     if (cur_sector >= total_sectors) {
 292         bmds->cur_sector = bmds->completed_sectors = total_sectors;
 293         return 1;
 294     }
 295
 296     bmds->completed_sectors = cur_sector;
 297
 298     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
 299
 300     /* we are going to transfer a full block even if it is not allocated */
 301     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 302
 303     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 304         nr_sectors = total_sectors - cur_sector;
 305     }
 306
 307     blk = g_new(BlkMigBlock, 1);
 308     blk->buf = g_malloc(BLOCK_SIZE);
 309     blk->bmds = bmds;
 310     blk->sector = cur_sector;
 311     blk->nr_sectors = nr_sectors;
 312
 313     blk->iov.iov_base = blk->buf;
 314     blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 315     qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 316
 317     blk_mig_lock();
 318     block_mig_state.submitted++;
 319     blk_mig_unlock();
 320
 321     /* We do not know if bs is under the main thread (and thus does
 322      * not acquire the AioContext when doing AIO) or rather under
 323      * dataplane.  Thus acquire both the iothread mutex and the
 324      * AioContext.
 325      *
 326      * This is ugly and will disappear when we make bdrv_* thread-safe,
 327      * without the need to acquire the AioContext.
 328      */
 329     qemu_mutex_lock_iothread();
 330     aio_context_acquire(blk_get_aio_context(bmds->blk));
 331     blk->aiocb = blk_aio_preadv(bb, cur_sector * BDRV_SECTOR_SIZE, &blk->qiov,
 332                                 0, blk_mig_read_cb, blk);
 333
 334     bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, cur_sector, nr_sectors);
 335     aio_context_release(blk_get_aio_context(bmds->blk));
 336     qemu_mutex_unlock_iothread();
 337
 338     bmds->cur_sector = cur_sector + nr_sectors;
 339     return (bmds->cur_sector >= total_sectors);
 340 }
 341
 342 /* Called with iothread lock taken.  */
 343
 344 static int set_dirty_tracking(void)
 345 {
 346     BlkMigDevState *bmds;
 347     int ret;
 348
 349     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 350         aio_context_acquire(blk_get_aio_context(bmds->blk));
 351         bmds->dirty_bitmap = bdrv_create_dirty_bitmap(blk_bs(bmds->blk),
 352                                                       BLOCK_SIZE, NULL, NULL);
 353         aio_context_release(blk_get_aio_context(bmds->blk));
 354         if (!bmds->dirty_bitmap) {
 355             ret = -errno;
 356             goto fail;
 357         }
 358     }
 359     return 0;
 360
 361 fail:
 362     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 363         if (bmds->dirty_bitmap) {
 364             aio_context_acquire(blk_get_aio_context(bmds->blk));
 365             bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
 366             aio_context_release(blk_get_aio_context(bmds->blk));
 367         }
 368     }
 369     return ret;
 370 }
 371
 372 /* Called with iothread lock taken.  */
 373
 374 static void unset_dirty_tracking(void)
 375 {
 376     BlkMigDevState *bmds;
 377
 378     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 379         aio_context_acquire(blk_get_aio_context(bmds->blk));
 380         bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
 381         aio_context_release(blk_get_aio_context(bmds->blk));
 382     }
 383 }
 384
 385 static int init_blk_migration(QEMUFile *f)
 386 {
 387     BlockDriverState *bs;
 388     BlkMigDevState *bmds;
 389     int64_t sectors;
 390     BdrvNextIterator it;
 391     int i, num_bs = 0;
 392     struct {
 393         BlkMigDevState *bmds;
 394         BlockDriverState *bs;
 395     } *bmds_bs;
 396     Error *local_err = NULL;
 397     int ret;
 398
 399     block_mig_state.submitted = 0;
 400     block_mig_state.read_done = 0;
 401     block_mig_state.transferred = 0;
 402     block_mig_state.total_sector_sum = 0;
 403     block_mig_state.prev_progress = -1;
 404     block_mig_state.bulk_completed = 0;
 405     block_mig_state.zero_blocks = migrate_zero_blocks();
 406
 407     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
 408         num_bs++;
 409     }
 410     bmds_bs = g_malloc0(num_bs * sizeof(*bmds_bs));
 411
 412     for (i = 0, bs = bdrv_first(&it); bs; bs = bdrv_next(&it), i++) {
 413         if (bdrv_is_read_only(bs)) {
 414             continue;
 415         }
 416
 417         sectors = bdrv_nb_sectors(bs);
 418         if (sectors <= 0) {
 419             ret = sectors;
 420             goto out;
 421         }
 422
 423         bmds = g_new0(BlkMigDevState, 1);
 424         bmds->blk = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 425         bmds->blk_name = g_strdup(bdrv_get_device_name(bs));
 426         bmds->bulk_completed = 0;
 427         bmds->total_sectors = sectors;
 428         bmds->completed_sectors = 0;
 429         bmds->shared_base = migrate_use_block_incremental();
 430
 431         assert(i < num_bs);
 432         bmds_bs[i].bmds = bmds;
 433         bmds_bs[i].bs = bs;
 434
 435         block_mig_state.total_sector_sum += sectors;
 436
 437         if (bmds->shared_base) {
 438             DPRINTF("Start migration for %s with shared base image\n",
 439                     bdrv_get_device_name(bs));
 440         } else {
 441             DPRINTF("Start full migration for %s\n", bdrv_get_device_name(bs));
 442         }
 443
 444         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 445     }
 446
 447     /* Can only insert new BDSes now because doing so while iterating block
 448      * devices may end up in a deadlock (iterating the new BDSes, too). */
 449     for (i = 0; i < num_bs; i++) {
 450         BlkMigDevState *bmds = bmds_bs[i].bmds;
 451         BlockDriverState *bs = bmds_bs[i].bs;
 452
 453         if (bmds) {
 454             ret = blk_insert_bs(bmds->blk, bs, &local_err);
 455             if (ret < 0) {
 456                 error_report_err(local_err);
 457                 goto out;
 458             }
 459
 460             alloc_aio_bitmap(bmds);
 461             error_setg(&bmds->blocker, "block device is in use by migration");
 462             bdrv_op_block_all(bs, bmds->blocker);
 463         }
 464     }
 465
 466     ret = 0;
 467 out:
 468     g_free(bmds_bs);
 469     return ret;
 470 }
 471
 472 /* Called with no lock taken.  */
 473
 474 static int blk_mig_save_bulked_block(QEMUFile *f)
 475 {
 476     int64_t completed_sector_sum = 0;
 477     BlkMigDevState *bmds;
 478     int progress;
 479     int ret = 0;
 480
 481     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 482         if (bmds->bulk_completed == 0) {
 483             if (mig_save_device_bulk(f, bmds) == 1) {
 484                 /* completed bulk section for this device */
 485                 bmds->bulk_completed = 1;
 486             }
 487             completed_sector_sum += bmds->completed_sectors;
 488             ret = 1;
 489             break;
 490         } else {
 491             completed_sector_sum += bmds->completed_sectors;
 492         }
 493     }
 494
 495     if (block_mig_state.total_sector_sum != 0) {
 496         progress = completed_sector_sum * 100 /
 497                    block_mig_state.total_sector_sum;
 498     } else {
 499         progress = 100;
 500     }
 501     if (progress != block_mig_state.prev_progress) {
 502         block_mig_state.prev_progress = progress;
 503         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 504                          | BLK_MIG_FLAG_PROGRESS);
 505         DPRINTF("Completed %d %%\r", progress);
 506     }
 507
 508     return ret;
 509 }
 510
 511 static void blk_mig_reset_dirty_cursor(void)
 512 {
 513     BlkMigDevState *bmds;
 514
 515     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 516         bmds->cur_dirty = 0;
 517     }
 518 }
 519
 520 /* Called with iothread lock and AioContext taken.  */
 521
 522 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
 523                                  int is_async)
 524 {
 525     BlkMigBlock *blk;
 526     BlockDriverState *bs = blk_bs(bmds->blk);
 527     int64_t total_sectors = bmds->total_sectors;
 528     int64_t sector;
 529     int nr_sectors;
 530     int ret = -EIO;
 531
 532     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
 533         blk_mig_lock();
 534         if (bmds_aio_inflight(bmds, sector)) {
 535             blk_mig_unlock();
 536             blk_drain(bmds->blk);
 537         } else {
 538             blk_mig_unlock();
 539         }
 540         if (bdrv_get_dirty(bs, bmds->dirty_bitmap, sector)) {
 541
 542             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 543                 nr_sectors = total_sectors - sector;
 544             } else {
 545                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 546             }
 547             blk = g_new(BlkMigBlock, 1);
 548             blk->buf = g_malloc(BLOCK_SIZE);
 549             blk->bmds = bmds;
 550             blk->sector = sector;
 551             blk->nr_sectors = nr_sectors;
 552
 553             if (is_async) {
 554                 blk->iov.iov_base = blk->buf;
 555                 blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 556                 qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 557
 558                 blk->aiocb = blk_aio_preadv(bmds->blk,
 559                                             sector * BDRV_SECTOR_SIZE,
 560                                             &blk->qiov, 0, blk_mig_read_cb,
 561                                             blk);
 562
 563                 blk_mig_lock();
 564                 block_mig_state.submitted++;
 565                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 566                 blk_mig_unlock();
 567             } else {
 568                 ret = blk_pread(bmds->blk, sector * BDRV_SECTOR_SIZE, blk->buf,
 569                                 nr_sectors * BDRV_SECTOR_SIZE);
 570                 if (ret < 0) {
 571                     goto error;
 572                 }
 573                 blk_send(f, blk);
 574
 575                 g_free(blk->buf);
 576                 g_free(blk);
 577             }
 578
 579             bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, sector, nr_sectors);
 580             sector += nr_sectors;
 581             bmds->cur_dirty = sector;
 582
 583             break;
 584         }
 585         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 586         bmds->cur_dirty = sector;
 587     }
 588
 589     return (bmds->cur_dirty >= bmds->total_sectors);
 590
 591 error:
 592     DPRINTF("Error reading sector %" PRId64 "\n", sector);
 593     g_free(blk->buf);
 594     g_free(blk);
 595     return ret;
 596 }
 597
 598 /* Called with iothread lock taken.
 599  *
 600  * return value:
 601  * 0: too much data for max_downtime
 602  * 1: few enough data for max_downtime
 603 */
 604 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 605 {
 606     BlkMigDevState *bmds;
 607     int ret = 1;
 608
 609     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 610         aio_context_acquire(blk_get_aio_context(bmds->blk));
 611         ret = mig_save_device_dirty(f, bmds, is_async);
 612         aio_context_release(blk_get_aio_context(bmds->blk));
 613         if (ret <= 0) {
 614             break;
 615         }
 616     }
 617
 618     return ret;
 619 }
 620
 621 /* Called with no locks taken.  */
 622
 623 static int flush_blks(QEMUFile *f)
 624 {
 625     BlkMigBlock *blk;
 626     int ret = 0;
 627
 628     DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
 629             __FUNCTION__, block_mig_state.submitted, block_mig_state.read_done,
 630             block_mig_state.transferred);
 631
 632     blk_mig_lock();
 633     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 634         if (qemu_file_rate_limit(f)) {
 635             break;
 636         }
 637         if (blk->ret < 0) {
 638             ret = blk->ret;
 639             break;
 640         }
 641
 642         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 643         blk_mig_unlock();
 644         blk_send(f, blk);
 645         blk_mig_lock();
 646
 647         g_free(blk->buf);
 648         g_free(blk);
 649
 650         block_mig_state.read_done--;
 651         block_mig_state.transferred++;
 652         assert(block_mig_state.read_done >= 0);
 653     }
 654     blk_mig_unlock();
 655
 656     DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__,
 657             block_mig_state.submitted, block_mig_state.read_done,
 658             block_mig_state.transferred);
 659     return ret;
 660 }
 661
 662 /* Called with iothread lock taken.  */
 663
 664 static int64_t get_remaining_dirty(void)
 665 {
 666     BlkMigDevState *bmds;
 667     int64_t dirty = 0;
 668
 669     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 670         aio_context_acquire(blk_get_aio_context(bmds->blk));
 671         dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
 672         aio_context_release(blk_get_aio_context(bmds->blk));
 673     }
 674
 675     return dirty << BDRV_SECTOR_BITS;
 676 }
 677
 678
 679
 680 /* Called with iothread lock taken.  */
 681 static void block_migration_cleanup_bmds(void)
 682 {
 683     BlkMigDevState *bmds;
 684     AioContext *ctx;
 685
 686     unset_dirty_tracking();
 687
 688     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 689         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 690         bdrv_op_unblock_all(blk_bs(bmds->blk), bmds->blocker);
 691         error_free(bmds->blocker);
 692
 693         /* Save ctx, because bmds->blk can disappear during blk_unref.  */
 694         ctx = blk_get_aio_context(bmds->blk);
 695         aio_context_acquire(ctx);
 696         blk_unref(bmds->blk);
 697         aio_context_release(ctx);
 698
 699         g_free(bmds->blk_name);
 700         g_free(bmds->aio_bitmap);
 701         g_free(bmds);
 702     }
 703 }
 704
 705 /* Called with iothread lock taken.  */
 706 static void block_migration_cleanup(void *opaque)
 707 {
 708     BlkMigBlock *blk;
 709
 710     bdrv_drain_all();
 711
 712     block_migration_cleanup_bmds();
 713
 714     blk_mig_lock();
 715     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 716         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 717         g_free(blk->buf);
 718         g_free(blk);
 719     }
 720     blk_mig_unlock();
 721 }
 722
 723 static int block_save_setup(QEMUFile *f, void *opaque)
 724 {
 725     int ret;
 726
 727     DPRINTF("Enter save live setup submitted %d transferred %d\n",
 728             block_mig_state.submitted, block_mig_state.transferred);
 729
 730     qemu_mutex_lock_iothread();
 731     ret = init_blk_migration(f);
 732     if (ret < 0) {
 733         qemu_mutex_unlock_iothread();
 734         return ret;
 735     }
 736
 737     /* start track dirty blocks */
 738     ret = set_dirty_tracking();
 739
 740     qemu_mutex_unlock_iothread();
 741
 742     if (ret) {
 743         return ret;
 744     }
 745
 746     ret = flush_blks(f);
 747     blk_mig_reset_dirty_cursor();
 748     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 749
 750     return ret;
 751 }
 752
 753 static int block_save_iterate(QEMUFile *f, void *opaque)
 754 {
 755     int ret;
 756     int64_t last_ftell = qemu_ftell(f);
 757     int64_t delta_ftell;
 758
 759     DPRINTF("Enter save live iterate submitted %d transferred %d\n",
 760             block_mig_state.submitted, block_mig_state.transferred);
 761
 762     ret = flush_blks(f);
 763     if (ret) {
 764         return ret;
 765     }
 766
 767     blk_mig_reset_dirty_cursor();
 768
 769     /* control the rate of transfer */
 770     blk_mig_lock();
 771     while ((block_mig_state.submitted +
 772             block_mig_state.read_done) * BLOCK_SIZE <
 773            qemu_file_get_rate_limit(f) &&
 774            (block_mig_state.submitted +
 775             block_mig_state.read_done) <
 776            MAX_INFLIGHT_IO) {
 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 *non_postcopiable_pending,
 865                                uint64_t *postcopiable_pending)
 866 {
 867     /* Estimate pending number of bytes to send */
 868     uint64_t pending;
 869
 870     qemu_mutex_lock_iothread();
 871     pending = get_remaining_dirty();
 872     qemu_mutex_unlock_iothread();
 873
 874     blk_mig_lock();
 875     pending += block_mig_state.submitted * BLOCK_SIZE +
 876                block_mig_state.read_done * BLOCK_SIZE;
 877     blk_mig_unlock();
 878
 879     /* Report at least one block pending during bulk phase */
 880     if (pending <= max_size && !block_mig_state.bulk_completed) {
 881         pending = max_size + BLOCK_SIZE;
 882     }
 883
 884     DPRINTF("Enter save live pending  %" PRIu64 "\n", pending);
 885     /* We don't do postcopy */
 886     *non_postcopiable_pending += pending;
 887 }
 888
 889 static int block_load(QEMUFile *f, void *opaque, int version_id)
 890 {
 891     static int banner_printed;
 892     int len, flags;
 893     char device_name[256];
 894     int64_t addr;
 895     BlockBackend *blk, *blk_prev = NULL;;
 896     Error *local_err = NULL;
 897     uint8_t *buf;
 898     int64_t total_sectors = 0;
 899     int nr_sectors;
 900     int ret;
 901     BlockDriverInfo bdi;
 902     int cluster_size = BLOCK_SIZE;
 903
 904     do {
 905         addr = qemu_get_be64(f);
 906
 907         flags = addr & ~BDRV_SECTOR_MASK;
 908         addr >>= BDRV_SECTOR_BITS;
 909
 910         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 911             /* get device name */
 912             len = qemu_get_byte(f);
 913             qemu_get_buffer(f, (uint8_t *)device_name, len);
 914             device_name[len] = '\0';
 915
 916             blk = blk_by_name(device_name);
 917             if (!blk) {
 918                 fprintf(stderr, "Error unknown block device %s\n",
 919                         device_name);
 920                 return -EINVAL;
 921             }
 922
 923             if (blk != blk_prev) {
 924                 blk_prev = blk;
 925                 total_sectors = blk_nb_sectors(blk);
 926                 if (total_sectors <= 0) {
 927                     error_report("Error getting length of block device %s",
 928                                  device_name);
 929                     return -EINVAL;
 930                 }
 931
 932                 blk_invalidate_cache(blk, &local_err);
 933                 if (local_err) {
 934                     error_report_err(local_err);
 935                     return -EINVAL;
 936                 }
 937
 938                 ret = bdrv_get_info(blk_bs(blk), &bdi);
 939                 if (ret == 0 && bdi.cluster_size > 0 &&
 940                     bdi.cluster_size <= BLOCK_SIZE &&
 941                     BLOCK_SIZE % bdi.cluster_size == 0) {
 942                     cluster_size = bdi.cluster_size;
 943                 } else {
 944                     cluster_size = BLOCK_SIZE;
 945                 }
 946             }
 947
 948             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 949                 nr_sectors = total_sectors - addr;
 950             } else {
 951                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 952             }
 953
 954             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 955                 ret = blk_pwrite_zeroes(blk, addr * BDRV_SECTOR_SIZE,
 956                                         nr_sectors * BDRV_SECTOR_SIZE,
 957                                         BDRV_REQ_MAY_UNMAP);
 958             } else {
 959                 int i;
 960                 int64_t cur_addr;
 961                 uint8_t *cur_buf;
 962
 963                 buf = g_malloc(BLOCK_SIZE);
 964                 qemu_get_buffer(f, buf, BLOCK_SIZE);
 965                 for (i = 0; i < BLOCK_SIZE / cluster_size; i++) {
 966                     cur_addr = addr * BDRV_SECTOR_SIZE + i * cluster_size;
 967                     cur_buf = buf + i * cluster_size;
 968
 969                     if ((!block_mig_state.zero_blocks ||
 970                         cluster_size < BLOCK_SIZE) &&
 971                         buffer_is_zero(cur_buf, cluster_size)) {
 972                         ret = blk_pwrite_zeroes(blk, cur_addr,
 973                                                 cluster_size,
 974                                                 BDRV_REQ_MAY_UNMAP);
 975                     } else {
 976                         ret = blk_pwrite(blk, cur_addr, cur_buf,
 977                                          cluster_size, 0);
 978                     }
 979                     if (ret < 0) {
 980                         break;
 981                     }
 982                 }
 983                 g_free(buf);
 984             }
 985
 986             if (ret < 0) {
 987                 return ret;
 988             }
 989         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 990             if (!banner_printed) {
 991                 printf("Receiving block device images\n");
 992                 banner_printed = 1;
 993             }
 994             printf("Completed %d %%%c", (int)addr,
 995                    (addr == 100) ? '\n' : '\r');
 996             fflush(stdout);
 997         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
 998             fprintf(stderr, "Unknown block migration flags: %#x\n", flags);
 999             return -EINVAL;
1000         }
1001         ret = qemu_file_get_error(f);
1002         if (ret != 0) {
1003             return ret;
1004         }
1005     } while (!(flags & BLK_MIG_FLAG_EOS));
1006
1007     return 0;
1008 }
1009
1010 static bool block_is_active(void *opaque)
1011 {
1012     return migrate_use_block();
1013 }
1014
1015 static SaveVMHandlers savevm_block_handlers = {
1016     .save_live_setup = block_save_setup,
1017     .save_live_iterate = block_save_iterate,
1018     .save_live_complete_precopy = block_save_complete,
1019     .save_live_pending = block_save_pending,
1020     .load_state = block_load,
1021     .cleanup = block_migration_cleanup,
1022     .is_active = block_is_active,
1023 };
1024
1025 void blk_mig_init(void)
1026 {
1027     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
1028     QSIMPLEQ_INIT(&block_mig_state.blk_list);
1029     qemu_mutex_init(&block_mig_state.lock);
1030
1031     register_savevm_live(NULL, "block", 0, 1, &savevm_block_handlers,
1032                          &block_mig_state);
1033 }