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