migration/block.c

   1 /*
   2  * QEMU live block migration
   3  *
   4  * Copyright IBM, Corp. 2009
   5  *
   6  * Authors:
   7  *  Liran Schour   <lirans@il.ibm.com>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.  See
  10  * the COPYING file in the top-level directory.
  11  *
  12  * Contributions after 2012-01-13 are licensed under the terms of the
  13  * GNU GPL, version 2 or (at your option) any later version.
  14  */
  15
  16 #include "qemu/osdep.h"
  17 #include "qapi/error.h"
  18 #include "qemu/error-report.h"
  19 #include "qemu/cutils.h"
  20 #include "qemu/queue.h"
  21 #include "block.h"
  22 #include "migration/misc.h"
  23 #include "migration.h"
  24 #include "migration/register.h"
  25 #include "qemu-file.h"
  26 #include "migration/vmstate.h"
  27 #include "sysemu/block-backend.h"
  28
  29 #define BLOCK_SIZE                       (1 << 20)
  30 #define BDRV_SECTORS_PER_DIRTY_CHUNK     (BLOCK_SIZE >> BDRV_SECTOR_BITS)
  31
  32 #define BLK_MIG_FLAG_DEVICE_BLOCK       0x01
  33 #define BLK_MIG_FLAG_EOS                0x02
  34 #define BLK_MIG_FLAG_PROGRESS           0x04
  35 #define BLK_MIG_FLAG_ZERO_BLOCK         0x08
  36
  37 #define MAX_IS_ALLOCATED_SEARCH (65536 * BDRV_SECTOR_SIZE)
  38
  39 #define MAX_INFLIGHT_IO 512
  40
  41 //#define DEBUG_BLK_MIGRATION
  42
  43 #ifdef DEBUG_BLK_MIGRATION
  44 #define DPRINTF(fmt, ...) \
  45     do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
  46 #else
  47 #define DPRINTF(fmt, ...) \
  48     do { } while (0)
  49 #endif
  50
  51 typedef struct BlkMigDevState {
  52     /* Written during setup phase.  Can be read without a lock.  */
  53     BlockBackend *blk;
  54     char *blk_name;
  55     int shared_base;
  56     int64_t total_sectors;
  57     QSIMPLEQ_ENTRY(BlkMigDevState) entry;
  58     Error *blocker;
  59
  60     /* Only used by migration thread.  Does not need a lock.  */
  61     int bulk_completed;
  62     int64_t cur_sector;
  63     int64_t cur_dirty;
  64
  65     /* Data in the aio_bitmap is protected by block migration lock.
  66      * Allocation and free happen during setup and cleanup respectively.
  67      */
  68     unsigned long *aio_bitmap;
  69
  70     /* Protected by block migration lock.  */
  71     int64_t completed_sectors;
  72
  73     /* During migration this is protected by iothread lock / AioContext.
  74      * Allocation and free happen during setup and cleanup respectively.
  75      */
  76     BdrvDirtyBitmap *dirty_bitmap;
  77 } BlkMigDevState;
  78
  79 typedef struct BlkMigBlock {
  80     /* Only used by migration thread.  */
  81     uint8_t *buf;
  82     BlkMigDevState *bmds;
  83     int64_t sector;
  84     int nr_sectors;
  85     struct iovec iov;
  86     QEMUIOVector qiov;
  87     BlockAIOCB *aiocb;
  88
  89     /* Protected by block migration lock.  */
  90     int ret;
  91     QSIMPLEQ_ENTRY(BlkMigBlock) entry;
  92 } BlkMigBlock;
  93
  94 typedef struct BlkMigState {
  95     QSIMPLEQ_HEAD(bmds_list, BlkMigDevState) bmds_list;
  96     int64_t total_sector_sum;
  97     bool zero_blocks;
  98
  99     /* Protected by lock.  */
 100     QSIMPLEQ_HEAD(blk_list, BlkMigBlock) blk_list;
 101     int submitted;
 102     int read_done;
 103
 104     /* Only used by migration thread.  Does not need a lock.  */
 105     int transferred;
 106     int prev_progress;
 107     int bulk_completed;
 108
 109     /* Lock must be taken _inside_ the iothread lock and any AioContexts.  */
 110     QemuMutex lock;
 111 } BlkMigState;
 112
 113 static BlkMigState block_mig_state;
 114
 115 static void blk_mig_lock(void)
 116 {
 117     qemu_mutex_lock(&block_mig_state.lock);
 118 }
 119
 120 static void blk_mig_unlock(void)
 121 {
 122     qemu_mutex_unlock(&block_mig_state.lock);
 123 }
 124
 125 /* Must run outside of the iothread lock during the bulk phase,
 126  * or the VM will stall.
 127  */
 128
 129 static void blk_send(QEMUFile *f, BlkMigBlock * blk)
 130 {
 131     int len;
 132     uint64_t flags = BLK_MIG_FLAG_DEVICE_BLOCK;
 133
 134     if (block_mig_state.zero_blocks &&
 135         buffer_is_zero(blk->buf, BLOCK_SIZE)) {
 136         flags |= BLK_MIG_FLAG_ZERO_BLOCK;
 137     }
 138
 139     /* sector number and flags */
 140     qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
 141                      | flags);
 142
 143     /* device name */
 144     len = strlen(blk->bmds->blk_name);
 145     qemu_put_byte(f, len);
 146     qemu_put_buffer(f, (uint8_t *) blk->bmds->blk_name, len);
 147
 148     /* if a block is zero we need to flush here since the network
 149      * bandwidth is now a lot higher than the storage device bandwidth.
 150      * thus if we queue zero blocks we slow down the migration */
 151     if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 152         qemu_fflush(f);
 153         return;
 154     }
 155
 156     qemu_put_buffer(f, blk->buf, BLOCK_SIZE);
 157 }
 158
 159 int blk_mig_active(void)
 160 {
 161     return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
 162 }
 163
 164 int blk_mig_bulk_active(void)
 165 {
 166     return blk_mig_active() && !block_mig_state.bulk_completed;
 167 }
 168
 169 uint64_t blk_mig_bytes_transferred(void)
 170 {
 171     BlkMigDevState *bmds;
 172     uint64_t sum = 0;
 173
 174     blk_mig_lock();
 175     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 176         sum += bmds->completed_sectors;
 177     }
 178     blk_mig_unlock();
 179     return sum << BDRV_SECTOR_BITS;
 180 }
 181
 182 uint64_t blk_mig_bytes_remaining(void)
 183 {
 184     return blk_mig_bytes_total() - blk_mig_bytes_transferred();
 185 }
 186
 187 uint64_t blk_mig_bytes_total(void)
 188 {
 189     BlkMigDevState *bmds;
 190     uint64_t sum = 0;
 191
 192     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 193         sum += bmds->total_sectors;
 194     }
 195     return sum << BDRV_SECTOR_BITS;
 196 }
 197
 198
 199 /* Called with migration lock held.  */
 200
 201 static int bmds_aio_inflight(BlkMigDevState *bmds, int64_t sector)
 202 {
 203     int64_t chunk = sector / (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK;
 204
 205     if (sector < blk_nb_sectors(bmds->blk)) {
 206         return !!(bmds->aio_bitmap[chunk / (sizeof(unsigned long) * 8)] &
 207             (1UL << (chunk % (sizeof(unsigned long) * 8))));
 208     } else {
 209         return 0;
 210     }
 211 }
 212
 213 /* Called with migration lock held.  */
 214
 215 static void bmds_set_aio_inflight(BlkMigDevState *bmds, int64_t sector_num,
 216                              int nb_sectors, int set)
 217 {
 218     int64_t start, end;
 219     unsigned long val, idx, bit;
 220
 221     start = sector_num / BDRV_SECTORS_PER_DIRTY_CHUNK;
 222     end = (sector_num + nb_sectors - 1) / BDRV_SECTORS_PER_DIRTY_CHUNK;
 223
 224     for (; start <= end; start++) {
 225         idx = start / (sizeof(unsigned long) * 8);
 226         bit = start % (sizeof(unsigned long) * 8);
 227         val = bmds->aio_bitmap[idx];
 228         if (set) {
 229             val |= 1UL << bit;
 230         } else {
 231             val &= ~(1UL << bit);
 232         }
 233         bmds->aio_bitmap[idx] = val;
 234     }
 235 }
 236
 237 static void alloc_aio_bitmap(BlkMigDevState *bmds)
 238 {
 239     BlockBackend *bb = bmds->blk;
 240     int64_t bitmap_size;
 241
 242     bitmap_size = blk_nb_sectors(bb) + BDRV_SECTORS_PER_DIRTY_CHUNK * 8 - 1;
 243     bitmap_size /= BDRV_SECTORS_PER_DIRTY_CHUNK * 8;
 244
 245     bmds->aio_bitmap = g_malloc0(bitmap_size);
 246 }
 247
 248 /* Never hold migration lock when yielding to the main loop!  */
 249
 250 static void blk_mig_read_cb(void *opaque, int ret)
 251 {
 252     BlkMigBlock *blk = opaque;
 253
 254     blk_mig_lock();
 255     blk->ret = ret;
 256
 257     QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry);
 258     bmds_set_aio_inflight(blk->bmds, blk->sector, blk->nr_sectors, 0);
 259
 260     block_mig_state.submitted--;
 261     block_mig_state.read_done++;
 262     assert(block_mig_state.submitted >= 0);
 263     blk_mig_unlock();
 264 }
 265
 266 /* Called with no lock taken.  */
 267
 268 static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds)
 269 {
 270     int64_t total_sectors = bmds->total_sectors;
 271     int64_t cur_sector = bmds->cur_sector;
 272     BlockBackend *bb = bmds->blk;
 273     BlkMigBlock *blk;
 274     int nr_sectors;
 275     int64_t count;
 276
 277     if (bmds->shared_base) {
 278         qemu_mutex_lock_iothread();
 279         aio_context_acquire(blk_get_aio_context(bb));
 280         /* Skip unallocated sectors; intentionally treats failure or
 281          * partial sector as an allocated sector */
 282         while (cur_sector < total_sectors &&
 283                !bdrv_is_allocated(blk_bs(bb), cur_sector * BDRV_SECTOR_SIZE,
 284                                   MAX_IS_ALLOCATED_SEARCH, &count)) {
 285             if (count < BDRV_SECTOR_SIZE) {
 286                 break;
 287             }
 288             cur_sector += count >> BDRV_SECTOR_BITS;
 289         }
 290         aio_context_release(blk_get_aio_context(bb));
 291         qemu_mutex_unlock_iothread();
 292     }
 293
 294     if (cur_sector >= total_sectors) {
 295         bmds->cur_sector = bmds->completed_sectors = total_sectors;
 296         return 1;
 297     }
 298
 299     bmds->completed_sectors = cur_sector;
 300
 301     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
 302
 303     /* we are going to transfer a full block even if it is not allocated */
 304     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 305
 306     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 307         nr_sectors = total_sectors - cur_sector;
 308     }
 309
 310     blk = g_new(BlkMigBlock, 1);
 311     blk->buf = g_malloc(BLOCK_SIZE);
 312     blk->bmds = bmds;
 313     blk->sector = cur_sector;
 314     blk->nr_sectors = nr_sectors;
 315
 316     blk->iov.iov_base = blk->buf;
 317     blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 318     qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 319
 320     blk_mig_lock();
 321     block_mig_state.submitted++;
 322     blk_mig_unlock();
 323
 324     /* We do not know if bs is under the main thread (and thus does
 325      * not acquire the AioContext when doing AIO) or rather under
 326      * dataplane.  Thus acquire both the iothread mutex and the
 327      * AioContext.
 328      *
 329      * This is ugly and will disappear when we make bdrv_* thread-safe,
 330      * without the need to acquire the AioContext.
 331      */
 332     qemu_mutex_lock_iothread();
 333     aio_context_acquire(blk_get_aio_context(bmds->blk));
 334     blk->aiocb = blk_aio_preadv(bb, cur_sector * BDRV_SECTOR_SIZE, &blk->qiov,
 335                                 0, blk_mig_read_cb, blk);
 336
 337     bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, cur_sector * BDRV_SECTOR_SIZE,
 338                             nr_sectors * BDRV_SECTOR_SIZE);
 339     aio_context_release(blk_get_aio_context(bmds->blk));
 340     qemu_mutex_unlock_iothread();
 341
 342     bmds->cur_sector = cur_sector + nr_sectors;
 343     return (bmds->cur_sector >= total_sectors);
 344 }
 345
 346 /* Called with iothread lock taken.  */
 347
 348 static int set_dirty_tracking(void)
 349 {
 350     BlkMigDevState *bmds;
 351     int ret;
 352
 353     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 354         bmds->dirty_bitmap = bdrv_create_dirty_bitmap(blk_bs(bmds->blk),
 355                                                       BLOCK_SIZE, NULL, NULL);
 356         if (!bmds->dirty_bitmap) {
 357             ret = -errno;
 358             goto fail;
 359         }
 360     }
 361     return 0;
 362
 363 fail:
 364     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 365         if (bmds->dirty_bitmap) {
 366             bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
 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         bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
 380     }
 381 }
 382
 383 static int init_blk_migration(QEMUFile *f)
 384 {
 385     BlockDriverState *bs;
 386     BlkMigDevState *bmds;
 387     int64_t sectors;
 388     BdrvNextIterator it;
 389     int i, num_bs = 0;
 390     struct {
 391         BlkMigDevState *bmds;
 392         BlockDriverState *bs;
 393     } *bmds_bs;
 394     Error *local_err = NULL;
 395     int ret;
 396
 397     block_mig_state.submitted = 0;
 398     block_mig_state.read_done = 0;
 399     block_mig_state.transferred = 0;
 400     block_mig_state.total_sector_sum = 0;
 401     block_mig_state.prev_progress = -1;
 402     block_mig_state.bulk_completed = 0;
 403     block_mig_state.zero_blocks = migrate_zero_blocks();
 404
 405     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
 406         num_bs++;
 407     }
 408     bmds_bs = g_malloc0(num_bs * sizeof(*bmds_bs));
 409
 410     for (i = 0, bs = bdrv_first(&it); bs; bs = bdrv_next(&it), i++) {
 411         if (bdrv_is_read_only(bs)) {
 412             continue;
 413         }
 414
 415         sectors = bdrv_nb_sectors(bs);
 416         if (sectors <= 0) {
 417             ret = sectors;
 418             goto out;
 419         }
 420
 421         bmds = g_new0(BlkMigDevState, 1);
 422         bmds->blk = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 423         bmds->blk_name = g_strdup(bdrv_get_device_name(bs));
 424         bmds->bulk_completed = 0;
 425         bmds->total_sectors = sectors;
 426         bmds->completed_sectors = 0;
 427         bmds->shared_base = migrate_use_block_incremental();
 428
 429         assert(i < num_bs);
 430         bmds_bs[i].bmds = bmds;
 431         bmds_bs[i].bs = bs;
 432
 433         block_mig_state.total_sector_sum += sectors;
 434
 435         if (bmds->shared_base) {
 436             DPRINTF("Start migration for %s with shared base image\n",
 437                     bdrv_get_device_name(bs));
 438         } else {
 439             DPRINTF("Start full migration for %s\n", bdrv_get_device_name(bs));
 440         }
 441
 442         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 443     }
 444
 445     /* Can only insert new BDSes now because doing so while iterating block
 446      * devices may end up in a deadlock (iterating the new BDSes, too). */
 447     for (i = 0; i < num_bs; i++) {
 448         BlkMigDevState *bmds = bmds_bs[i].bmds;
 449         BlockDriverState *bs = bmds_bs[i].bs;
 450
 451         if (bmds) {
 452             ret = blk_insert_bs(bmds->blk, bs, &local_err);
 453             if (ret < 0) {
 454                 error_report_err(local_err);
 455                 goto out;
 456             }
 457
 458             alloc_aio_bitmap(bmds);
 459             error_setg(&bmds->blocker, "block device is in use by migration");
 460             bdrv_op_block_all(bs, bmds->blocker);
 461         }
 462     }
 463
 464     ret = 0;
 465 out:
 466     g_free(bmds_bs);
 467     return ret;
 468 }
 469
 470 /* Called with no lock taken.  */
 471
 472 static int blk_mig_save_bulked_block(QEMUFile *f)
 473 {
 474     int64_t completed_sector_sum = 0;
 475     BlkMigDevState *bmds;
 476     int progress;
 477     int ret = 0;
 478
 479     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 480         if (bmds->bulk_completed == 0) {
 481             if (mig_save_device_bulk(f, bmds) == 1) {
 482                 /* completed bulk section for this device */
 483                 bmds->bulk_completed = 1;
 484             }
 485             completed_sector_sum += bmds->completed_sectors;
 486             ret = 1;
 487             break;
 488         } else {
 489             completed_sector_sum += bmds->completed_sectors;
 490         }
 491     }
 492
 493     if (block_mig_state.total_sector_sum != 0) {
 494         progress = completed_sector_sum * 100 /
 495                    block_mig_state.total_sector_sum;
 496     } else {
 497         progress = 100;
 498     }
 499     if (progress != block_mig_state.prev_progress) {
 500         block_mig_state.prev_progress = progress;
 501         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 502                          | BLK_MIG_FLAG_PROGRESS);
 503         DPRINTF("Completed %d %%\r", progress);
 504     }
 505
 506     return ret;
 507 }
 508
 509 static void blk_mig_reset_dirty_cursor(void)
 510 {
 511     BlkMigDevState *bmds;
 512
 513     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 514         bmds->cur_dirty = 0;
 515     }
 516 }
 517
 518 /* Called with iothread lock and AioContext taken.  */
 519
 520 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
 521                                  int is_async)
 522 {
 523     BlkMigBlock *blk;
 524     BlockDriverState *bs = blk_bs(bmds->blk);
 525     int64_t total_sectors = bmds->total_sectors;
 526     int64_t sector;
 527     int nr_sectors;
 528     int ret = -EIO;
 529
 530     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
 531         blk_mig_lock();
 532         if (bmds_aio_inflight(bmds, sector)) {
 533             blk_mig_unlock();
 534             blk_drain(bmds->blk);
 535         } else {
 536             blk_mig_unlock();
 537         }
 538         bdrv_dirty_bitmap_lock(bmds->dirty_bitmap);
 539         if (bdrv_get_dirty_locked(bs, bmds->dirty_bitmap,
 540                                   sector * BDRV_SECTOR_SIZE)) {
 541             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 542                 nr_sectors = total_sectors - sector;
 543             } else {
 544                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 545             }
 546             bdrv_reset_dirty_bitmap_locked(bmds->dirty_bitmap,
 547                                            sector * BDRV_SECTOR_SIZE,
 548                                            nr_sectors * BDRV_SECTOR_SIZE);
 549             bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 550
 551             blk = g_new(BlkMigBlock, 1);
 552             blk->buf = g_malloc(BLOCK_SIZE);
 553             blk->bmds = bmds;
 554             blk->sector = sector;
 555             blk->nr_sectors = nr_sectors;
 556
 557             if (is_async) {
 558                 blk->iov.iov_base = blk->buf;
 559                 blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 560                 qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 561
 562                 blk->aiocb = blk_aio_preadv(bmds->blk,
 563                                             sector * BDRV_SECTOR_SIZE,
 564                                             &blk->qiov, 0, blk_mig_read_cb,
 565                                             blk);
 566
 567                 blk_mig_lock();
 568                 block_mig_state.submitted++;
 569                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 570                 blk_mig_unlock();
 571             } else {
 572                 ret = blk_pread(bmds->blk, sector * BDRV_SECTOR_SIZE, blk->buf,
 573                                 nr_sectors * BDRV_SECTOR_SIZE);
 574                 if (ret < 0) {
 575                     goto error;
 576                 }
 577                 blk_send(f, blk);
 578
 579                 g_free(blk->buf);
 580                 g_free(blk);
 581             }
 582
 583             sector += nr_sectors;
 584             bmds->cur_dirty = sector;
 585             break;
 586         }
 587
 588         bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 589         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 590         bmds->cur_dirty = sector;
 591     }
 592
 593     return (bmds->cur_dirty >= bmds->total_sectors);
 594
 595 error:
 596     DPRINTF("Error reading sector %" PRId64 "\n", sector);
 597     g_free(blk->buf);
 598     g_free(blk);
 599     return ret;
 600 }
 601
 602 /* Called with iothread lock taken.
 603  *
 604  * return value:
 605  * 0: too much data for max_downtime
 606  * 1: few enough data for max_downtime
 607 */
 608 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 609 {
 610     BlkMigDevState *bmds;
 611     int ret = 1;
 612
 613     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 614         aio_context_acquire(blk_get_aio_context(bmds->blk));
 615         ret = mig_save_device_dirty(f, bmds, is_async);
 616         aio_context_release(blk_get_aio_context(bmds->blk));
 617         if (ret <= 0) {
 618             break;
 619         }
 620     }
 621
 622     return ret;
 623 }
 624
 625 /* Called with no locks taken.  */
 626
 627 static int flush_blks(QEMUFile *f)
 628 {
 629     BlkMigBlock *blk;
 630     int ret = 0;
 631
 632     DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
 633             __FUNCTION__, block_mig_state.submitted, block_mig_state.read_done,
 634             block_mig_state.transferred);
 635
 636     blk_mig_lock();
 637     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 638         if (qemu_file_rate_limit(f)) {
 639             break;
 640         }
 641         if (blk->ret < 0) {
 642             ret = blk->ret;
 643             break;
 644         }
 645
 646         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 647         blk_mig_unlock();
 648         blk_send(f, blk);
 649         blk_mig_lock();
 650
 651         g_free(blk->buf);
 652         g_free(blk);
 653
 654         block_mig_state.read_done--;
 655         block_mig_state.transferred++;
 656         assert(block_mig_state.read_done >= 0);
 657     }
 658     blk_mig_unlock();
 659
 660     DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__,
 661             block_mig_state.submitted, block_mig_state.read_done,
 662             block_mig_state.transferred);
 663     return ret;
 664 }
 665
 666 /* Called with iothread lock taken.  */
 667
 668 static int64_t get_remaining_dirty(void)
 669 {
 670     BlkMigDevState *bmds;
 671     int64_t dirty = 0;
 672
 673     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 674         aio_context_acquire(blk_get_aio_context(bmds->blk));
 675         dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
 676         aio_context_release(blk_get_aio_context(bmds->blk));
 677     }
 678
 679     return dirty;
 680 }
 681
 682
 683
 684 /* Called with iothread lock taken.  */
 685 static void block_migration_cleanup_bmds(void)
 686 {
 687     BlkMigDevState *bmds;
 688     AioContext *ctx;
 689
 690     unset_dirty_tracking();
 691
 692     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 693         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 694         bdrv_op_unblock_all(blk_bs(bmds->blk), bmds->blocker);
 695         error_free(bmds->blocker);
 696
 697         /* Save ctx, because bmds->blk can disappear during blk_unref.  */
 698         ctx = blk_get_aio_context(bmds->blk);
 699         aio_context_acquire(ctx);
 700         blk_unref(bmds->blk);
 701         aio_context_release(ctx);
 702
 703         g_free(bmds->blk_name);
 704         g_free(bmds->aio_bitmap);
 705         g_free(bmds);
 706     }
 707 }
 708
 709 /* Called with iothread lock taken.  */
 710 static void block_migration_cleanup(void *opaque)
 711 {
 712     BlkMigBlock *blk;
 713
 714     bdrv_drain_all();
 715
 716     block_migration_cleanup_bmds();
 717
 718     blk_mig_lock();
 719     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 720         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 721         g_free(blk->buf);
 722         g_free(blk);
 723     }
 724     blk_mig_unlock();
 725 }
 726
 727 static int block_save_setup(QEMUFile *f, void *opaque)
 728 {
 729     int ret;
 730
 731     DPRINTF("Enter save live setup submitted %d transferred %d\n",
 732             block_mig_state.submitted, block_mig_state.transferred);
 733
 734     qemu_mutex_lock_iothread();
 735     ret = init_blk_migration(f);
 736     if (ret < 0) {
 737         qemu_mutex_unlock_iothread();
 738         return ret;
 739     }
 740
 741     /* start track dirty blocks */
 742     ret = set_dirty_tracking();
 743
 744     qemu_mutex_unlock_iothread();
 745
 746     if (ret) {
 747         return ret;
 748     }
 749
 750     ret = flush_blks(f);
 751     blk_mig_reset_dirty_cursor();
 752     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 753
 754     return ret;
 755 }
 756
 757 static int block_save_iterate(QEMUFile *f, void *opaque)
 758 {
 759     int ret;
 760     int64_t last_ftell = qemu_ftell(f);
 761     int64_t delta_ftell;
 762
 763     DPRINTF("Enter save live iterate 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     /* control the rate of transfer */
 774     blk_mig_lock();
 775     while ((block_mig_state.submitted +
 776             block_mig_state.read_done) * BLOCK_SIZE <
 777            qemu_file_get_rate_limit(f) &&
 778            (block_mig_state.submitted +
 779             block_mig_state.read_done) <
 780            MAX_INFLIGHT_IO) {
 781         blk_mig_unlock();
 782         if (block_mig_state.bulk_completed == 0) {
 783             /* first finish the bulk phase */
 784             if (blk_mig_save_bulked_block(f) == 0) {
 785                 /* finished saving bulk on all devices */
 786                 block_mig_state.bulk_completed = 1;
 787             }
 788             ret = 0;
 789         } else {
 790             /* Always called with iothread lock taken for
 791              * simplicity, block_save_complete also calls it.
 792              */
 793             qemu_mutex_lock_iothread();
 794             ret = blk_mig_save_dirty_block(f, 1);
 795             qemu_mutex_unlock_iothread();
 796         }
 797         if (ret < 0) {
 798             return ret;
 799         }
 800         blk_mig_lock();
 801         if (ret != 0) {
 802             /* no more dirty blocks */
 803             break;
 804         }
 805     }
 806     blk_mig_unlock();
 807
 808     ret = flush_blks(f);
 809     if (ret) {
 810         return ret;
 811     }
 812
 813     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 814     delta_ftell = qemu_ftell(f) - last_ftell;
 815     if (delta_ftell > 0) {
 816         return 1;
 817     } else if (delta_ftell < 0) {
 818         return -1;
 819     } else {
 820         return 0;
 821     }
 822 }
 823
 824 /* Called with iothread lock taken.  */
 825
 826 static int block_save_complete(QEMUFile *f, void *opaque)
 827 {
 828     int ret;
 829
 830     DPRINTF("Enter save live complete submitted %d transferred %d\n",
 831             block_mig_state.submitted, block_mig_state.transferred);
 832
 833     ret = flush_blks(f);
 834     if (ret) {
 835         return ret;
 836     }
 837
 838     blk_mig_reset_dirty_cursor();
 839
 840     /* we know for sure that save bulk is completed and
 841        all async read completed */
 842     blk_mig_lock();
 843     assert(block_mig_state.submitted == 0);
 844     blk_mig_unlock();
 845
 846     do {
 847         ret = blk_mig_save_dirty_block(f, 0);
 848         if (ret < 0) {
 849             return ret;
 850         }
 851     } while (ret == 0);
 852
 853     /* report completion */
 854     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 855
 856     DPRINTF("Block migration completed\n");
 857
 858     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 859
 860     /* Make sure that our BlockBackends are gone, so that the block driver
 861      * nodes can be inactivated. */
 862     block_migration_cleanup_bmds();
 863
 864     return 0;
 865 }
 866
 867 static void block_save_pending(QEMUFile *f, void *opaque, uint64_t max_size,
 868                                uint64_t *non_postcopiable_pending,
 869                                uint64_t *postcopiable_pending)
 870 {
 871     /* Estimate pending number of bytes to send */
 872     uint64_t pending;
 873
 874     qemu_mutex_lock_iothread();
 875     pending = get_remaining_dirty();
 876     qemu_mutex_unlock_iothread();
 877
 878     blk_mig_lock();
 879     pending += block_mig_state.submitted * BLOCK_SIZE +
 880                block_mig_state.read_done * BLOCK_SIZE;
 881     blk_mig_unlock();
 882
 883     /* Report at least one block pending during bulk phase */
 884     if (pending <= max_size && !block_mig_state.bulk_completed) {
 885         pending = max_size + BLOCK_SIZE;
 886     }
 887
 888     DPRINTF("Enter save live pending  %" PRIu64 "\n", pending);
 889     /* We don't do postcopy */
 890     *non_postcopiable_pending += pending;
 891 }
 892
 893 static int block_load(QEMUFile *f, void *opaque, int version_id)
 894 {
 895     static int banner_printed;
 896     int len, flags;
 897     char device_name[256];
 898     int64_t addr;
 899     BlockBackend *blk, *blk_prev = NULL;;
 900     Error *local_err = NULL;
 901     uint8_t *buf;
 902     int64_t total_sectors = 0;
 903     int nr_sectors;
 904     int ret;
 905     BlockDriverInfo bdi;
 906     int cluster_size = BLOCK_SIZE;
 907
 908     do {
 909         addr = qemu_get_be64(f);
 910
 911         flags = addr & ~BDRV_SECTOR_MASK;
 912         addr >>= BDRV_SECTOR_BITS;
 913
 914         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 915             /* get device name */
 916             len = qemu_get_byte(f);
 917             qemu_get_buffer(f, (uint8_t *)device_name, len);
 918             device_name[len] = '\0';
 919
 920             blk = blk_by_name(device_name);
 921             if (!blk) {
 922                 fprintf(stderr, "Error unknown block device %s\n",
 923                         device_name);
 924                 return -EINVAL;
 925             }
 926
 927             if (blk != blk_prev) {
 928                 blk_prev = blk;
 929                 total_sectors = blk_nb_sectors(blk);
 930                 if (total_sectors <= 0) {
 931                     error_report("Error getting length of block device %s",
 932                                  device_name);
 933                     return -EINVAL;
 934                 }
 935
 936                 blk_invalidate_cache(blk, &local_err);
 937                 if (local_err) {
 938                     error_report_err(local_err);
 939                     return -EINVAL;
 940                 }
 941
 942                 ret = bdrv_get_info(blk_bs(blk), &bdi);
 943                 if (ret == 0 && bdi.cluster_size > 0 &&
 944                     bdi.cluster_size <= BLOCK_SIZE &&
 945                     BLOCK_SIZE % bdi.cluster_size == 0) {
 946                     cluster_size = bdi.cluster_size;
 947                 } else {
 948                     cluster_size = BLOCK_SIZE;
 949                 }
 950             }
 951
 952             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 953                 nr_sectors = total_sectors - addr;
 954             } else {
 955                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 956             }
 957
 958             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 959                 ret = blk_pwrite_zeroes(blk, addr * BDRV_SECTOR_SIZE,
 960                                         nr_sectors * BDRV_SECTOR_SIZE,
 961                                         BDRV_REQ_MAY_UNMAP);
 962             } else {
 963                 int i;
 964                 int64_t cur_addr;
 965                 uint8_t *cur_buf;
 966
 967                 buf = g_malloc(BLOCK_SIZE);
 968                 qemu_get_buffer(f, buf, BLOCK_SIZE);
 969                 for (i = 0; i < BLOCK_SIZE / cluster_size; i++) {
 970                     cur_addr = addr * BDRV_SECTOR_SIZE + i * cluster_size;
 971                     cur_buf = buf + i * cluster_size;
 972
 973                     if ((!block_mig_state.zero_blocks ||
 974                         cluster_size < BLOCK_SIZE) &&
 975                         buffer_is_zero(cur_buf, cluster_size)) {
 976                         ret = blk_pwrite_zeroes(blk, cur_addr,
 977                                                 cluster_size,
 978                                                 BDRV_REQ_MAY_UNMAP);
 979                     } else {
 980                         ret = blk_pwrite(blk, cur_addr, cur_buf,
 981                                          cluster_size, 0);
 982                     }
 983                     if (ret < 0) {
 984                         break;
 985                     }
 986                 }
 987                 g_free(buf);
 988             }
 989
 990             if (ret < 0) {
 991                 return ret;
 992             }
 993         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 994             if (!banner_printed) {
 995                 printf("Receiving block device images\n");
 996                 banner_printed = 1;
 997             }
 998             printf("Completed %d %%%c", (int)addr,
 999                    (addr == 100) ? '\n' : '\r');
1000             fflush(stdout);
1001         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
1002             fprintf(stderr, "Unknown block migration flags: %#x\n", flags);
1003             return -EINVAL;
1004         }
1005         ret = qemu_file_get_error(f);
1006         if (ret != 0) {
1007             return ret;
1008         }
1009     } while (!(flags & BLK_MIG_FLAG_EOS));
1010
1011     return 0;
1012 }
1013
1014 static bool block_is_active(void *opaque)
1015 {
1016     return migrate_use_block();
1017 }
1018
1019 static SaveVMHandlers savevm_block_handlers = {
1020     .save_setup = block_save_setup,
1021     .save_live_iterate = block_save_iterate,
1022     .save_live_complete_precopy = block_save_complete,
1023     .save_live_pending = block_save_pending,
1024     .load_state = block_load,
1025     .save_cleanup = block_migration_cleanup,
1026     .is_active = block_is_active,
1027 };
1028
1029 void blk_mig_init(void)
1030 {
1031     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
1032     QSIMPLEQ_INIT(&block_mig_state.blk_list);
1033     qemu_mutex_init(&block_mig_state.lock);
1034
1035     register_savevm_live(NULL, "block", 0, 1, &savevm_block_handlers,
1036                          &block_mig_state);
1037 }