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