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/main-loop.h"
  20 #include "qemu/cutils.h"
  21 #include "qemu/queue.h"
  22 #include "block.h"
  23 #include "migration/misc.h"
  24 #include "migration.h"
  25 #include "migration/register.h"
  26 #include "qemu-file.h"
  27 #include "migration/vmstate.h"
  28 #include "sysemu/block-backend.h"
  29
  30 #define BLK_MIG_BLOCK_SIZE           (1 << 20)
  31 #define BDRV_SECTORS_PER_DIRTY_CHUNK (BLK_MIG_BLOCK_SIZE >> BDRV_SECTOR_BITS)
  32
  33 #define BLK_MIG_FLAG_DEVICE_BLOCK       0x01
  34 #define BLK_MIG_FLAG_EOS                0x02
  35 #define BLK_MIG_FLAG_PROGRESS           0x04
  36 #define BLK_MIG_FLAG_ZERO_BLOCK         0x08
  37
  38 #define MAX_IS_ALLOCATED_SEARCH (65536 * BDRV_SECTOR_SIZE)
  39
  40 #define MAX_IO_BUFFERS 512
  41 #define MAX_PARALLEL_IO 16
  42
  43 //#define DEBUG_BLK_MIGRATION
  44
  45 #ifdef DEBUG_BLK_MIGRATION
  46 #define DPRINTF(fmt, ...) \
  47     do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
  48 #else
  49 #define DPRINTF(fmt, ...) \
  50     do { } while (0)
  51 #endif
  52
  53 typedef struct BlkMigDevState {
  54     /* Written during setup phase.  Can be read without a lock.  */
  55     BlockBackend *blk;
  56     char *blk_name;
  57     int shared_base;
  58     int64_t total_sectors;
  59     QSIMPLEQ_ENTRY(BlkMigDevState) entry;
  60     Error *blocker;
  61
  62     /* Only used by migration thread.  Does not need a lock.  */
  63     int bulk_completed;
  64     int64_t cur_sector;
  65     int64_t cur_dirty;
  66
  67     /* Data in the aio_bitmap is protected by block migration lock.
  68      * Allocation and free happen during setup and cleanup respectively.
  69      */
  70     unsigned long *aio_bitmap;
  71
  72     /* Protected by block migration lock.  */
  73     int64_t completed_sectors;
  74
  75     /* During migration this is protected by iothread lock / AioContext.
  76      * Allocation and free happen during setup and cleanup respectively.
  77      */
  78     BdrvDirtyBitmap *dirty_bitmap;
  79 } BlkMigDevState;
  80
  81 typedef struct BlkMigBlock {
  82     /* Only used by migration thread.  */
  83     uint8_t *buf;
  84     BlkMigDevState *bmds;
  85     int64_t sector;
  86     int nr_sectors;
  87     QEMUIOVector qiov;
  88     BlockAIOCB *aiocb;
  89
  90     /* Protected by block migration lock.  */
  91     int ret;
  92     QSIMPLEQ_ENTRY(BlkMigBlock) entry;
  93 } BlkMigBlock;
  94
  95 typedef struct BlkMigState {
  96     QSIMPLEQ_HEAD(, BlkMigDevState) bmds_list;
  97     int64_t total_sector_sum;
  98     bool zero_blocks;
  99
 100     /* Protected by lock.  */
 101     QSIMPLEQ_HEAD(, BlkMigBlock) blk_list;
 102     int submitted;
 103     int read_done;
 104
 105     /* Only used by migration thread.  Does not need a lock.  */
 106     int transferred;
 107     int prev_progress;
 108     int bulk_completed;
 109
 110     /* Lock must be taken _inside_ the iothread lock and any AioContexts.  */
 111     QemuMutex lock;
 112 } BlkMigState;
 113
 114 static BlkMigState block_mig_state;
 115
 116 static void blk_mig_lock(void)
 117 {
 118     qemu_mutex_lock(&block_mig_state.lock);
 119 }
 120
 121 static void blk_mig_unlock(void)
 122 {
 123     qemu_mutex_unlock(&block_mig_state.lock);
 124 }
 125
 126 /* Must run outside of the iothread lock during the bulk phase,
 127  * or the VM will stall.
 128  */
 129
 130 static void blk_send(QEMUFile *f, BlkMigBlock * blk)
 131 {
 132     int len;
 133     uint64_t flags = BLK_MIG_FLAG_DEVICE_BLOCK;
 134
 135     if (block_mig_state.zero_blocks &&
 136         buffer_is_zero(blk->buf, BLK_MIG_BLOCK_SIZE)) {
 137         flags |= BLK_MIG_FLAG_ZERO_BLOCK;
 138     }
 139
 140     /* sector number and flags */
 141     qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
 142                      | flags);
 143
 144     /* device name */
 145     len = strlen(blk->bmds->blk_name);
 146     qemu_put_byte(f, len);
 147     qemu_put_buffer(f, (uint8_t *) blk->bmds->blk_name, len);
 148
 149     /* if a block is zero we need to flush here since the network
 150      * bandwidth is now a lot higher than the storage device bandwidth.
 151      * thus if we queue zero blocks we slow down the migration */
 152     if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 153         qemu_fflush(f);
 154         return;
 155     }
 156
 157     qemu_put_buffer(f, blk->buf, BLK_MIG_BLOCK_SIZE);
 158 }
 159
 160 int blk_mig_active(void)
 161 {
 162     return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
 163 }
 164
 165 int blk_mig_bulk_active(void)
 166 {
 167     return blk_mig_active() && !block_mig_state.bulk_completed;
 168 }
 169
 170 uint64_t blk_mig_bytes_transferred(void)
 171 {
 172     BlkMigDevState *bmds;
 173     uint64_t sum = 0;
 174
 175     blk_mig_lock();
 176     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 177         sum += bmds->completed_sectors;
 178     }
 179     blk_mig_unlock();
 180     return sum << BDRV_SECTOR_BITS;
 181 }
 182
 183 uint64_t blk_mig_bytes_remaining(void)
 184 {
 185     return blk_mig_bytes_total() - blk_mig_bytes_transferred();
 186 }
 187
 188 uint64_t blk_mig_bytes_total(void)
 189 {
 190     BlkMigDevState *bmds;
 191     uint64_t sum = 0;
 192
 193     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 194         sum += bmds->total_sectors;
 195     }
 196     return sum << BDRV_SECTOR_BITS;
 197 }
 198
 199
 200 /* Called with migration lock held.  */
 201
 202 static int bmds_aio_inflight(BlkMigDevState *bmds, int64_t sector)
 203 {
 204     int64_t chunk = sector / (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK;
 205
 206     if (sector < blk_nb_sectors(bmds->blk)) {
 207         return !!(bmds->aio_bitmap[chunk / (sizeof(unsigned long) * 8)] &
 208             (1UL << (chunk % (sizeof(unsigned long) * 8))));
 209     } else {
 210         return 0;
 211     }
 212 }
 213
 214 /* Called with migration lock held.  */
 215
 216 static void bmds_set_aio_inflight(BlkMigDevState *bmds, int64_t sector_num,
 217                              int nb_sectors, int set)
 218 {
 219     int64_t start, end;
 220     unsigned long val, idx, bit;
 221
 222     start = sector_num / BDRV_SECTORS_PER_DIRTY_CHUNK;
 223     end = (sector_num + nb_sectors - 1) / BDRV_SECTORS_PER_DIRTY_CHUNK;
 224
 225     for (; start <= end; start++) {
 226         idx = start / (sizeof(unsigned long) * 8);
 227         bit = start % (sizeof(unsigned long) * 8);
 228         val = bmds->aio_bitmap[idx];
 229         if (set) {
 230             val |= 1UL << bit;
 231         } else {
 232             val &= ~(1UL << bit);
 233         }
 234         bmds->aio_bitmap[idx] = val;
 235     }
 236 }
 237
 238 static void alloc_aio_bitmap(BlkMigDevState *bmds)
 239 {
 240     BlockBackend *bb = bmds->blk;
 241     int64_t bitmap_size;
 242
 243     bitmap_size = blk_nb_sectors(bb) + BDRV_SECTORS_PER_DIRTY_CHUNK * 8 - 1;
 244     bitmap_size /= BDRV_SECTORS_PER_DIRTY_CHUNK * 8;
 245
 246     bmds->aio_bitmap = g_malloc0(bitmap_size);
 247 }
 248
 249 /* Never hold migration lock when yielding to the main loop!  */
 250
 251 static void blk_mig_read_cb(void *opaque, int ret)
 252 {
 253     BlkMigBlock *blk = opaque;
 254
 255     blk_mig_lock();
 256     blk->ret = ret;
 257
 258     QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry);
 259     bmds_set_aio_inflight(blk->bmds, blk->sector, blk->nr_sectors, 0);
 260
 261     block_mig_state.submitted--;
 262     block_mig_state.read_done++;
 263     assert(block_mig_state.submitted >= 0);
 264     blk_mig_unlock();
 265 }
 266
 267 /* Called with no lock taken.  */
 268
 269 static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds)
 270 {
 271     int64_t total_sectors = bmds->total_sectors;
 272     int64_t cur_sector = bmds->cur_sector;
 273     BlockBackend *bb = bmds->blk;
 274     BlkMigBlock *blk;
 275     int nr_sectors;
 276     int64_t count;
 277
 278     if (bmds->shared_base) {
 279         qemu_mutex_lock_iothread();
 280         aio_context_acquire(blk_get_aio_context(bb));
 281         /* Skip unallocated sectors; intentionally treats failure or
 282          * partial sector as an allocated sector */
 283         while (cur_sector < total_sectors &&
 284                !bdrv_is_allocated(blk_bs(bb), cur_sector * BDRV_SECTOR_SIZE,
 285                                   MAX_IS_ALLOCATED_SEARCH, &count)) {
 286             if (count < BDRV_SECTOR_SIZE) {
 287                 break;
 288             }
 289             cur_sector += count >> BDRV_SECTOR_BITS;
 290         }
 291         aio_context_release(blk_get_aio_context(bb));
 292         qemu_mutex_unlock_iothread();
 293     }
 294
 295     if (cur_sector >= total_sectors) {
 296         bmds->cur_sector = bmds->completed_sectors = total_sectors;
 297         return 1;
 298     }
 299
 300     bmds->completed_sectors = cur_sector;
 301
 302     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
 303
 304     /* we are going to transfer a full block even if it is not allocated */
 305     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 306
 307     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 308         nr_sectors = total_sectors - cur_sector;
 309     }
 310
 311     blk = g_new(BlkMigBlock, 1);
 312     blk->buf = g_malloc(BLK_MIG_BLOCK_SIZE);
 313     blk->bmds = bmds;
 314     blk->sector = cur_sector;
 315     blk->nr_sectors = nr_sectors;
 316
 317     qemu_iovec_init_buf(&blk->qiov, blk->buf, nr_sectors * BDRV_SECTOR_SIZE);
 318
 319     blk_mig_lock();
 320     block_mig_state.submitted++;
 321     blk_mig_unlock();
 322
 323     /* We do not know if bs is under the main thread (and thus does
 324      * not acquire the AioContext when doing AIO) or rather under
 325      * dataplane.  Thus acquire both the iothread mutex and the
 326      * AioContext.
 327      *
 328      * This is ugly and will disappear when we make bdrv_* thread-safe,
 329      * without the need to acquire the AioContext.
 330      */
 331     qemu_mutex_lock_iothread();
 332     aio_context_acquire(blk_get_aio_context(bmds->blk));
 333     bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, cur_sector * BDRV_SECTOR_SIZE,
 334                             nr_sectors * BDRV_SECTOR_SIZE);
 335     blk->aiocb = blk_aio_preadv(bb, cur_sector * BDRV_SECTOR_SIZE, &blk->qiov,
 336                                 0, blk_mig_read_cb, blk);
 337     aio_context_release(blk_get_aio_context(bmds->blk));
 338     qemu_mutex_unlock_iothread();
 339
 340     bmds->cur_sector = cur_sector + nr_sectors;
 341     return (bmds->cur_sector >= total_sectors);
 342 }
 343
 344 /* Called with iothread lock taken.  */
 345
 346 static int set_dirty_tracking(void)
 347 {
 348     BlkMigDevState *bmds;
 349     int ret;
 350
 351     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 352         bmds->dirty_bitmap = bdrv_create_dirty_bitmap(blk_bs(bmds->blk),
 353                                                       BLK_MIG_BLOCK_SIZE,
 354                                                       NULL, NULL);
 355         if (!bmds->dirty_bitmap) {
 356             ret = -errno;
 357             goto fail;
 358         }
 359     }
 360     return 0;
 361
 362 fail:
 363     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 364         if (bmds->dirty_bitmap) {
 365             bdrv_release_dirty_bitmap(bmds->dirty_bitmap);
 366         }
 367     }
 368     return ret;
 369 }
 370
 371 /* Called with iothread lock taken.  */
 372
 373 static void unset_dirty_tracking(void)
 374 {
 375     BlkMigDevState *bmds;
 376
 377     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 378         bdrv_release_dirty_bitmap(bmds->dirty_bitmap);
 379     }
 380 }
 381
 382 static int init_blk_migration(QEMUFile *f)
 383 {
 384     BlockDriverState *bs;
 385     BlkMigDevState *bmds;
 386     int64_t sectors;
 387     BdrvNextIterator it;
 388     int i, num_bs = 0;
 389     struct {
 390         BlkMigDevState *bmds;
 391         BlockDriverState *bs;
 392     } *bmds_bs;
 393     Error *local_err = NULL;
 394     int ret;
 395
 396     block_mig_state.submitted = 0;
 397     block_mig_state.read_done = 0;
 398     block_mig_state.transferred = 0;
 399     block_mig_state.total_sector_sum = 0;
 400     block_mig_state.prev_progress = -1;
 401     block_mig_state.bulk_completed = 0;
 402     block_mig_state.zero_blocks = migrate_zero_blocks();
 403
 404     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
 405         num_bs++;
 406     }
 407     bmds_bs = g_malloc0(num_bs * sizeof(*bmds_bs));
 408
 409     for (i = 0, bs = bdrv_first(&it); bs; bs = bdrv_next(&it), i++) {
 410         if (bdrv_is_read_only(bs)) {
 411             continue;
 412         }
 413
 414         sectors = bdrv_nb_sectors(bs);
 415         if (sectors <= 0) {
 416             ret = sectors;
 417             bdrv_next_cleanup(&it);
 418             goto out;
 419         }
 420
 421         bmds = g_new0(BlkMigDevState, 1);
 422         bmds->blk = blk_new(qemu_get_aio_context(),
 423                             BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 424         bmds->blk_name = g_strdup(bdrv_get_device_name(bs));
 425         bmds->bulk_completed = 0;
 426         bmds->total_sectors = sectors;
 427         bmds->completed_sectors = 0;
 428         bmds->shared_base = migrate_use_block_incremental();
 429
 430         assert(i < num_bs);
 431         bmds_bs[i].bmds = bmds;
 432         bmds_bs[i].bs = bs;
 433
 434         block_mig_state.total_sector_sum += sectors;
 435
 436         if (bmds->shared_base) {
 437             DPRINTF("Start migration for %s with shared base image\n",
 438                     bdrv_get_device_name(bs));
 439         } else {
 440             DPRINTF("Start full migration for %s\n", bdrv_get_device_name(bs));
 441         }
 442
 443         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 444     }
 445
 446     /* Can only insert new BDSes now because doing so while iterating block
 447      * devices may end up in a deadlock (iterating the new BDSes, too). */
 448     for (i = 0; i < num_bs; i++) {
 449         BlkMigDevState *bmds = bmds_bs[i].bmds;
 450         BlockDriverState *bs = bmds_bs[i].bs;
 451
 452         if (bmds) {
 453             ret = blk_insert_bs(bmds->blk, bs, &local_err);
 454             if (ret < 0) {
 455                 error_report_err(local_err);
 456                 goto out;
 457             }
 458
 459             alloc_aio_bitmap(bmds);
 460             error_setg(&bmds->blocker, "block device is in use by migration");
 461             bdrv_op_block_all(bs, bmds->blocker);
 462         }
 463     }
 464
 465     ret = 0;
 466 out:
 467     g_free(bmds_bs);
 468     return ret;
 469 }
 470
 471 /* Called with no lock taken.  */
 472
 473 static int blk_mig_save_bulked_block(QEMUFile *f)
 474 {
 475     int64_t completed_sector_sum = 0;
 476     BlkMigDevState *bmds;
 477     int progress;
 478     int ret = 0;
 479
 480     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 481         if (bmds->bulk_completed == 0) {
 482             if (mig_save_device_bulk(f, bmds) == 1) {
 483                 /* completed bulk section for this device */
 484                 bmds->bulk_completed = 1;
 485             }
 486             completed_sector_sum += bmds->completed_sectors;
 487             ret = 1;
 488             break;
 489         } else {
 490             completed_sector_sum += bmds->completed_sectors;
 491         }
 492     }
 493
 494     if (block_mig_state.total_sector_sum != 0) {
 495         progress = completed_sector_sum * 100 /
 496                    block_mig_state.total_sector_sum;
 497     } else {
 498         progress = 100;
 499     }
 500     if (progress != block_mig_state.prev_progress) {
 501         block_mig_state.prev_progress = progress;
 502         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 503                          | BLK_MIG_FLAG_PROGRESS);
 504         DPRINTF("Completed %d %%\r", progress);
 505     }
 506
 507     return ret;
 508 }
 509
 510 static void blk_mig_reset_dirty_cursor(void)
 511 {
 512     BlkMigDevState *bmds;
 513
 514     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 515         bmds->cur_dirty = 0;
 516     }
 517 }
 518
 519 /* Called with iothread lock and AioContext taken.  */
 520
 521 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
 522                                  int is_async)
 523 {
 524     BlkMigBlock *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_dirty_bitmap_get_locked(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(BLK_MIG_BLOCK_SIZE);
 553             blk->bmds = bmds;
 554             blk->sector = sector;
 555             blk->nr_sectors = nr_sectors;
 556
 557             if (is_async) {
 558                 qemu_iovec_init_buf(&blk->qiov, blk->buf,
 559                                     nr_sectors * BDRV_SECTOR_SIZE);
 560
 561                 blk->aiocb = blk_aio_preadv(bmds->blk,
 562                                             sector * BDRV_SECTOR_SIZE,
 563                                             &blk->qiov, 0, blk_mig_read_cb,
 564                                             blk);
 565
 566                 blk_mig_lock();
 567                 block_mig_state.submitted++;
 568                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 569                 blk_mig_unlock();
 570             } else {
 571                 ret = blk_pread(bmds->blk, sector * BDRV_SECTOR_SIZE, blk->buf,
 572                                 nr_sectors * BDRV_SECTOR_SIZE);
 573                 if (ret < 0) {
 574                     goto error;
 575                 }
 576                 blk_send(f, blk);
 577
 578                 g_free(blk->buf);
 579                 g_free(blk);
 580             }
 581
 582             sector += nr_sectors;
 583             bmds->cur_dirty = sector;
 584             break;
 585         }
 586
 587         bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 588         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 589         bmds->cur_dirty = sector;
 590     }
 591
 592     return (bmds->cur_dirty >= bmds->total_sectors);
 593
 594 error:
 595     DPRINTF("Error reading sector %" PRId64 "\n", sector);
 596     g_free(blk->buf);
 597     g_free(blk);
 598     return ret;
 599 }
 600
 601 /* Called with iothread lock taken.
 602  *
 603  * return value:
 604  * 0: too much data for max_downtime
 605  * 1: few enough data for max_downtime
 606 */
 607 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 608 {
 609     BlkMigDevState *bmds;
 610     int ret = 1;
 611
 612     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 613         aio_context_acquire(blk_get_aio_context(bmds->blk));
 614         ret = mig_save_device_dirty(f, bmds, is_async);
 615         aio_context_release(blk_get_aio_context(bmds->blk));
 616         if (ret <= 0) {
 617             break;
 618         }
 619     }
 620
 621     return ret;
 622 }
 623
 624 /* Called with no locks taken.  */
 625
 626 static int flush_blks(QEMUFile *f)
 627 {
 628     BlkMigBlock *blk;
 629     int ret = 0;
 630
 631     DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
 632             __func__, block_mig_state.submitted, block_mig_state.read_done,
 633             block_mig_state.transferred);
 634
 635     blk_mig_lock();
 636     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 637         if (qemu_file_rate_limit(f)) {
 638             break;
 639         }
 640         if (blk->ret < 0) {
 641             ret = blk->ret;
 642             break;
 643         }
 644
 645         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 646         blk_mig_unlock();
 647         blk_send(f, blk);
 648         blk_mig_lock();
 649
 650         g_free(blk->buf);
 651         g_free(blk);
 652
 653         block_mig_state.read_done--;
 654         block_mig_state.transferred++;
 655         assert(block_mig_state.read_done >= 0);
 656     }
 657     blk_mig_unlock();
 658
 659     DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __func__,
 660             block_mig_state.submitted, block_mig_state.read_done,
 661             block_mig_state.transferred);
 662     return ret;
 663 }
 664
 665 /* Called with iothread lock taken.  */
 666
 667 static int64_t get_remaining_dirty(void)
 668 {
 669     BlkMigDevState *bmds;
 670     int64_t dirty = 0;
 671
 672     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 673         aio_context_acquire(blk_get_aio_context(bmds->blk));
 674         dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
 675         aio_context_release(blk_get_aio_context(bmds->blk));
 676     }
 677
 678     return dirty;
 679 }
 680
 681
 682
 683 /* Called with iothread lock taken.  */
 684 static void block_migration_cleanup_bmds(void)
 685 {
 686     BlkMigDevState *bmds;
 687     AioContext *ctx;
 688
 689     unset_dirty_tracking();
 690
 691     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 692         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 693         bdrv_op_unblock_all(blk_bs(bmds->blk), bmds->blocker);
 694         error_free(bmds->blocker);
 695
 696         /* Save ctx, because bmds->blk can disappear during blk_unref.  */
 697         ctx = blk_get_aio_context(bmds->blk);
 698         aio_context_acquire(ctx);
 699         blk_unref(bmds->blk);
 700         aio_context_release(ctx);
 701
 702         g_free(bmds->blk_name);
 703         g_free(bmds->aio_bitmap);
 704         g_free(bmds);
 705     }
 706 }
 707
 708 /* Called with iothread lock taken.  */
 709 static void block_migration_cleanup(void *opaque)
 710 {
 711     BlkMigBlock *blk;
 712
 713     bdrv_drain_all();
 714
 715     block_migration_cleanup_bmds();
 716
 717     blk_mig_lock();
 718     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 719         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 720         g_free(blk->buf);
 721         g_free(blk);
 722     }
 723     blk_mig_unlock();
 724 }
 725
 726 static int block_save_setup(QEMUFile *f, void *opaque)
 727 {
 728     int ret;
 729
 730     DPRINTF("Enter save live setup submitted %d transferred %d\n",
 731             block_mig_state.submitted, block_mig_state.transferred);
 732
 733     qemu_mutex_lock_iothread();
 734     ret = init_blk_migration(f);
 735     if (ret < 0) {
 736         qemu_mutex_unlock_iothread();
 737         return ret;
 738     }
 739
 740     /* start track dirty blocks */
 741     ret = set_dirty_tracking();
 742
 743     qemu_mutex_unlock_iothread();
 744
 745     if (ret) {
 746         return ret;
 747     }
 748
 749     ret = flush_blks(f);
 750     blk_mig_reset_dirty_cursor();
 751     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 752
 753     return ret;
 754 }
 755
 756 static int block_save_iterate(QEMUFile *f, void *opaque)
 757 {
 758     int ret;
 759     int64_t last_ftell = qemu_ftell(f);
 760     int64_t delta_ftell;
 761
 762     DPRINTF("Enter save live iterate submitted %d transferred %d\n",
 763             block_mig_state.submitted, block_mig_state.transferred);
 764
 765     ret = flush_blks(f);
 766     if (ret) {
 767         return ret;
 768     }
 769
 770     blk_mig_reset_dirty_cursor();
 771
 772     /* control the rate of transfer */
 773     blk_mig_lock();
 774     while (block_mig_state.read_done * BLK_MIG_BLOCK_SIZE <
 775            qemu_file_get_rate_limit(f) &&
 776            block_mig_state.submitted < MAX_PARALLEL_IO &&
 777            (block_mig_state.submitted + block_mig_state.read_done) <
 778            MAX_IO_BUFFERS) {
 779         blk_mig_unlock();
 780         if (block_mig_state.bulk_completed == 0) {
 781             /* first finish the bulk phase */
 782             if (blk_mig_save_bulked_block(f) == 0) {
 783                 /* finished saving bulk on all devices */
 784                 block_mig_state.bulk_completed = 1;
 785             }
 786             ret = 0;
 787         } else {
 788             /* Always called with iothread lock taken for
 789              * simplicity, block_save_complete also calls it.
 790              */
 791             qemu_mutex_lock_iothread();
 792             ret = blk_mig_save_dirty_block(f, 1);
 793             qemu_mutex_unlock_iothread();
 794         }
 795         if (ret < 0) {
 796             return ret;
 797         }
 798         blk_mig_lock();
 799         if (ret != 0) {
 800             /* no more dirty blocks */
 801             break;
 802         }
 803     }
 804     blk_mig_unlock();
 805
 806     ret = flush_blks(f);
 807     if (ret) {
 808         return ret;
 809     }
 810
 811     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 812     delta_ftell = qemu_ftell(f) - last_ftell;
 813     if (delta_ftell > 0) {
 814         return 1;
 815     } else if (delta_ftell < 0) {
 816         return -1;
 817     } else {
 818         return 0;
 819     }
 820 }
 821
 822 /* Called with iothread lock taken.  */
 823
 824 static int block_save_complete(QEMUFile *f, void *opaque)
 825 {
 826     int ret;
 827
 828     DPRINTF("Enter save live complete submitted %d transferred %d\n",
 829             block_mig_state.submitted, block_mig_state.transferred);
 830
 831     ret = flush_blks(f);
 832     if (ret) {
 833         return ret;
 834     }
 835
 836     blk_mig_reset_dirty_cursor();
 837
 838     /* we know for sure that save bulk is completed and
 839        all async read completed */
 840     blk_mig_lock();
 841     assert(block_mig_state.submitted == 0);
 842     blk_mig_unlock();
 843
 844     do {
 845         ret = blk_mig_save_dirty_block(f, 0);
 846         if (ret < 0) {
 847             return ret;
 848         }
 849     } while (ret == 0);
 850
 851     /* report completion */
 852     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 853
 854     DPRINTF("Block migration completed\n");
 855
 856     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 857
 858     /* Make sure that our BlockBackends are gone, so that the block driver
 859      * nodes can be inactivated. */
 860     block_migration_cleanup_bmds();
 861
 862     return 0;
 863 }
 864
 865 static void block_save_pending(QEMUFile *f, void *opaque, uint64_t max_size,
 866                                uint64_t *res_precopy_only,
 867                                uint64_t *res_compatible,
 868                                uint64_t *res_postcopy_only)
 869 {
 870     /* Estimate pending number of bytes to send */
 871     uint64_t pending;
 872
 873     qemu_mutex_lock_iothread();
 874     pending = get_remaining_dirty();
 875     qemu_mutex_unlock_iothread();
 876
 877     blk_mig_lock();
 878     pending += block_mig_state.submitted * BLK_MIG_BLOCK_SIZE +
 879                block_mig_state.read_done * BLK_MIG_BLOCK_SIZE;
 880     blk_mig_unlock();
 881
 882     /* Report at least one block pending during bulk phase */
 883     if (pending <= max_size && !block_mig_state.bulk_completed) {
 884         pending = max_size + BLK_MIG_BLOCK_SIZE;
 885     }
 886
 887     DPRINTF("Enter save live pending  %" PRIu64 "\n", pending);
 888     /* We don't do postcopy */
 889     *res_precopy_only += pending;
 890 }
 891
 892 static int block_load(QEMUFile *f, void *opaque, int version_id)
 893 {
 894     static int banner_printed;
 895     int len, flags;
 896     char device_name[256];
 897     int64_t addr;
 898     BlockBackend *blk, *blk_prev = NULL;
 899     Error *local_err = NULL;
 900     uint8_t *buf;
 901     int64_t total_sectors = 0;
 902     int nr_sectors;
 903     int ret;
 904     BlockDriverInfo bdi;
 905     int cluster_size = BLK_MIG_BLOCK_SIZE;
 906
 907     do {
 908         addr = qemu_get_be64(f);
 909
 910         flags = addr & (BDRV_SECTOR_SIZE - 1);
 911         addr >>= BDRV_SECTOR_BITS;
 912
 913         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 914             /* get device name */
 915             len = qemu_get_byte(f);
 916             qemu_get_buffer(f, (uint8_t *)device_name, len);
 917             device_name[len] = '\0';
 918
 919             blk = blk_by_name(device_name);
 920             if (!blk) {
 921                 fprintf(stderr, "Error unknown block device %s\n",
 922                         device_name);
 923                 return -EINVAL;
 924             }
 925
 926             if (blk != blk_prev) {
 927                 blk_prev = blk;
 928                 total_sectors = blk_nb_sectors(blk);
 929                 if (total_sectors <= 0) {
 930                     error_report("Error getting length of block device %s",
 931                                  device_name);
 932                     return -EINVAL;
 933                 }
 934
 935                 blk_invalidate_cache(blk, &local_err);
 936                 if (local_err) {
 937                     error_report_err(local_err);
 938                     return -EINVAL;
 939                 }
 940
 941                 ret = bdrv_get_info(blk_bs(blk), &bdi);
 942                 if (ret == 0 && bdi.cluster_size > 0 &&
 943                     bdi.cluster_size <= BLK_MIG_BLOCK_SIZE &&
 944                     BLK_MIG_BLOCK_SIZE % bdi.cluster_size == 0) {
 945                     cluster_size = bdi.cluster_size;
 946                 } else {
 947                     cluster_size = BLK_MIG_BLOCK_SIZE;
 948                 }
 949             }
 950
 951             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 952                 nr_sectors = total_sectors - addr;
 953             } else {
 954                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 955             }
 956
 957             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 958                 ret = blk_pwrite_zeroes(blk, addr * BDRV_SECTOR_SIZE,
 959                                         nr_sectors * BDRV_SECTOR_SIZE,
 960                                         BDRV_REQ_MAY_UNMAP);
 961             } else {
 962                 int i;
 963                 int64_t cur_addr;
 964                 uint8_t *cur_buf;
 965
 966                 buf = g_malloc(BLK_MIG_BLOCK_SIZE);
 967                 qemu_get_buffer(f, buf, BLK_MIG_BLOCK_SIZE);
 968                 for (i = 0; i < BLK_MIG_BLOCK_SIZE / cluster_size; i++) {
 969                     cur_addr = addr * BDRV_SECTOR_SIZE + i * cluster_size;
 970                     cur_buf = buf + i * cluster_size;
 971
 972                     if ((!block_mig_state.zero_blocks ||
 973                         cluster_size < BLK_MIG_BLOCK_SIZE) &&
 974                         buffer_is_zero(cur_buf, cluster_size)) {
 975                         ret = blk_pwrite_zeroes(blk, cur_addr,
 976                                                 cluster_size,
 977                                                 BDRV_REQ_MAY_UNMAP);
 978                     } else {
 979                         ret = blk_pwrite(blk, cur_addr, cur_buf,
 980                                          cluster_size, 0);
 981                     }
 982                     if (ret < 0) {
 983                         break;
 984                     }
 985                 }
 986                 g_free(buf);
 987             }
 988
 989             if (ret < 0) {
 990                 return ret;
 991             }
 992         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 993             if (!banner_printed) {
 994                 printf("Receiving block device images\n");
 995                 banner_printed = 1;
 996             }
 997             printf("Completed %d %%%c", (int)addr,
 998                    (addr == 100) ? '\n' : '\r');
 999             fflush(stdout);
1000         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
1001             fprintf(stderr, "Unknown block migration flags: %#x\n", flags);
1002             return -EINVAL;
1003         }
1004         ret = qemu_file_get_error(f);
1005         if (ret != 0) {
1006             return ret;
1007         }
1008     } while (!(flags & BLK_MIG_FLAG_EOS));
1009
1010     return 0;
1011 }
1012
1013 static bool block_is_active(void *opaque)
1014 {
1015     return migrate_use_block();
1016 }
1017
1018 static SaveVMHandlers savevm_block_handlers = {
1019     .save_setup = block_save_setup,
1020     .save_live_iterate = block_save_iterate,
1021     .save_live_complete_precopy = block_save_complete,
1022     .save_live_pending = block_save_pending,
1023     .load_state = block_load,
1024     .save_cleanup = block_migration_cleanup,
1025     .is_active = block_is_active,
1026 };
1027
1028 void blk_mig_init(void)
1029 {
1030     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
1031     QSIMPLEQ_INIT(&block_mig_state.blk_list);
1032     qemu_mutex_init(&block_mig_state.lock);
1033
1034     register_savevm_live("block", 0, 1, &savevm_block_handlers,
1035                          &block_mig_state);
1036 }