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 "block/dirty-bitmap.h"
  24 #include "migration/misc.h"
  25 #include "migration.h"
  26 #include "migration/register.h"
  27 #include "qemu-file.h"
  28 #include "migration/vmstate.h"
  29 #include "sysemu/block-backend.h"
  30 #include "trace.h"
  31
  32 #define BLK_MIG_BLOCK_SIZE           (1ULL << 20)
  33 #define BDRV_SECTORS_PER_DIRTY_CHUNK (BLK_MIG_BLOCK_SIZE >> BDRV_SECTOR_BITS)
  34
  35 #define BLK_MIG_FLAG_DEVICE_BLOCK       0x01
  36 #define BLK_MIG_FLAG_EOS                0x02
  37 #define BLK_MIG_FLAG_PROGRESS           0x04
  38 #define BLK_MIG_FLAG_ZERO_BLOCK         0x08
  39
  40 #define MAX_IS_ALLOCATED_SEARCH (65536 * BDRV_SECTOR_SIZE)
  41
  42 #define MAX_IO_BUFFERS 512
  43 #define MAX_PARALLEL_IO 16
  44
  45 /* #define DEBUG_BLK_MIGRATION */
  46
  47 #ifdef DEBUG_BLK_MIGRATION
  48 #define DPRINTF(fmt, ...) \
  49     do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
  50 #else
  51 #define DPRINTF(fmt, ...) \
  52     do { } while (0)
  53 #endif
  54
  55 typedef struct BlkMigDevState {
  56     /* Written during setup phase.  Can be read without a lock.  */
  57     BlockBackend *blk;
  58     char *blk_name;
  59     int shared_base;
  60     int64_t total_sectors;
  61     QSIMPLEQ_ENTRY(BlkMigDevState) entry;
  62     Error *blocker;
  63
  64     /* Only used by migration thread.  Does not need a lock.  */
  65     int bulk_completed;
  66     int64_t cur_sector;
  67     int64_t cur_dirty;
  68
  69     /* Data in the aio_bitmap is protected by block migration lock.
  70      * Allocation and free happen during setup and cleanup respectively.
  71      */
  72     unsigned long *aio_bitmap;
  73
  74     /* Protected by block migration lock.  */
  75     int64_t completed_sectors;
  76
  77     /* During migration this is protected by iothread lock / AioContext.
  78      * Allocation and free happen during setup and cleanup respectively.
  79      */
  80     BdrvDirtyBitmap *dirty_bitmap;
  81 } BlkMigDevState;
  82
  83 typedef struct BlkMigBlock {
  84     /* Only used by migration thread.  */
  85     uint8_t *buf;
  86     BlkMigDevState *bmds;
  87     int64_t sector;
  88     int nr_sectors;
  89     QEMUIOVector qiov;
  90     BlockAIOCB *aiocb;
  91
  92     /* Protected by block migration lock.  */
  93     int ret;
  94     QSIMPLEQ_ENTRY(BlkMigBlock) entry;
  95 } BlkMigBlock;
  96
  97 typedef struct BlkMigState {
  98     QSIMPLEQ_HEAD(, BlkMigDevState) bmds_list;
  99     int64_t total_sector_sum;
 100     bool zero_blocks;
 101
 102     /* Protected by lock.  */
 103     QSIMPLEQ_HEAD(, BlkMigBlock) blk_list;
 104     int submitted;
 105     int read_done;
 106
 107     /* Only used by migration thread.  Does not need a lock.  */
 108     int transferred;
 109     int prev_progress;
 110     int bulk_completed;
 111
 112     /* Lock must be taken _inside_ the iothread lock and any AioContexts.  */
 113     QemuMutex lock;
 114 } BlkMigState;
 115
 116 static BlkMigState block_mig_state;
 117
 118 static void blk_mig_lock(void)
 119 {
 120     qemu_mutex_lock(&block_mig_state.lock);
 121 }
 122
 123 static void blk_mig_unlock(void)
 124 {
 125     qemu_mutex_unlock(&block_mig_state.lock);
 126 }
 127
 128 /* Must run outside of the iothread lock during the bulk phase,
 129  * or the VM will stall.
 130  */
 131
 132 static void blk_send(QEMUFile *f, BlkMigBlock * blk)
 133 {
 134     int len;
 135     uint64_t flags = BLK_MIG_FLAG_DEVICE_BLOCK;
 136
 137     if (block_mig_state.zero_blocks &&
 138         buffer_is_zero(blk->buf, BLK_MIG_BLOCK_SIZE)) {
 139         flags |= BLK_MIG_FLAG_ZERO_BLOCK;
 140     }
 141
 142     /* sector number and flags */
 143     qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
 144                      | flags);
 145
 146     /* device name */
 147     len = strlen(blk->bmds->blk_name);
 148     qemu_put_byte(f, len);
 149     qemu_put_buffer(f, (uint8_t *) blk->bmds->blk_name, len);
 150
 151     /* if a block is zero we need to flush here since the network
 152      * bandwidth is now a lot higher than the storage device bandwidth.
 153      * thus if we queue zero blocks we slow down the migration */
 154     if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 155         qemu_fflush(f);
 156         return;
 157     }
 158
 159     qemu_put_buffer(f, blk->buf, BLK_MIG_BLOCK_SIZE);
 160 }
 161
 162 int blk_mig_active(void)
 163 {
 164     return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
 165 }
 166
 167 int blk_mig_bulk_active(void)
 168 {
 169     return blk_mig_active() && !block_mig_state.bulk_completed;
 170 }
 171
 172 uint64_t blk_mig_bytes_transferred(void)
 173 {
 174     BlkMigDevState *bmds;
 175     uint64_t sum = 0;
 176
 177     blk_mig_lock();
 178     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 179         sum += bmds->completed_sectors;
 180     }
 181     blk_mig_unlock();
 182     return sum << BDRV_SECTOR_BITS;
 183 }
 184
 185 uint64_t blk_mig_bytes_remaining(void)
 186 {
 187     return blk_mig_bytes_total() - blk_mig_bytes_transferred();
 188 }
 189
 190 uint64_t blk_mig_bytes_total(void)
 191 {
 192     BlkMigDevState *bmds;
 193     uint64_t sum = 0;
 194
 195     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 196         sum += bmds->total_sectors;
 197     }
 198     return sum << BDRV_SECTOR_BITS;
 199 }
 200
 201
 202 /* Called with migration lock held.  */
 203
 204 static int bmds_aio_inflight(BlkMigDevState *bmds, int64_t sector)
 205 {
 206     int64_t chunk = sector / (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK;
 207
 208     if (sector < blk_nb_sectors(bmds->blk)) {
 209         return !!(bmds->aio_bitmap[chunk / (sizeof(unsigned long) * 8)] &
 210             (1UL << (chunk % (sizeof(unsigned long) * 8))));
 211     } else {
 212         return 0;
 213     }
 214 }
 215
 216 /* Called with migration lock held.  */
 217
 218 static void bmds_set_aio_inflight(BlkMigDevState *bmds, int64_t sector_num,
 219                              int nb_sectors, int set)
 220 {
 221     int64_t start, end;
 222     unsigned long val, idx, bit;
 223
 224     start = sector_num / BDRV_SECTORS_PER_DIRTY_CHUNK;
 225     end = (sector_num + nb_sectors - 1) / BDRV_SECTORS_PER_DIRTY_CHUNK;
 226
 227     for (; start <= end; start++) {
 228         idx = start / (sizeof(unsigned long) * 8);
 229         bit = start % (sizeof(unsigned long) * 8);
 230         val = bmds->aio_bitmap[idx];
 231         if (set) {
 232             val |= 1UL << bit;
 233         } else {
 234             val &= ~(1UL << bit);
 235         }
 236         bmds->aio_bitmap[idx] = val;
 237     }
 238 }
 239
 240 static void alloc_aio_bitmap(BlkMigDevState *bmds)
 241 {
 242     BlockBackend *bb = bmds->blk;
 243     int64_t bitmap_size;
 244
 245     bitmap_size = blk_nb_sectors(bb) + BDRV_SECTORS_PER_DIRTY_CHUNK * 8 - 1;
 246     bitmap_size /= BDRV_SECTORS_PER_DIRTY_CHUNK * 8;
 247
 248     bmds->aio_bitmap = g_malloc0(bitmap_size);
 249 }
 250
 251 /* Never hold migration lock when yielding to the main loop!  */
 252
 253 static void blk_mig_read_cb(void *opaque, int ret)
 254 {
 255     BlkMigBlock *blk = opaque;
 256
 257     blk_mig_lock();
 258     blk->ret = ret;
 259
 260     QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry);
 261     bmds_set_aio_inflight(blk->bmds, blk->sector, blk->nr_sectors, 0);
 262
 263     block_mig_state.submitted--;
 264     block_mig_state.read_done++;
 265     assert(block_mig_state.submitted >= 0);
 266     blk_mig_unlock();
 267 }
 268
 269 /* Called with no lock taken.  */
 270
 271 static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds)
 272 {
 273     int64_t total_sectors = bmds->total_sectors;
 274     int64_t cur_sector = bmds->cur_sector;
 275     BlockBackend *bb = bmds->blk;
 276     BlkMigBlock *blk;
 277     int nr_sectors;
 278     int64_t count;
 279
 280     if (bmds->shared_base) {
 281         qemu_mutex_lock_iothread();
 282         aio_context_acquire(blk_get_aio_context(bb));
 283         /* Skip unallocated sectors; intentionally treats failure or
 284          * partial sector as an allocated sector */
 285         while (cur_sector < total_sectors &&
 286                !bdrv_is_allocated(blk_bs(bb), cur_sector * BDRV_SECTOR_SIZE,
 287                                   MAX_IS_ALLOCATED_SEARCH, &count)) {
 288             if (count < BDRV_SECTOR_SIZE) {
 289                 break;
 290             }
 291             cur_sector += count >> BDRV_SECTOR_BITS;
 292         }
 293         aio_context_release(blk_get_aio_context(bb));
 294         qemu_mutex_unlock_iothread();
 295     }
 296
 297     if (cur_sector >= total_sectors) {
 298         bmds->cur_sector = bmds->completed_sectors = total_sectors;
 299         return 1;
 300     }
 301
 302     bmds->completed_sectors = cur_sector;
 303
 304     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
 305
 306     /* we are going to transfer a full block even if it is not allocated */
 307     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 308
 309     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 310         nr_sectors = total_sectors - cur_sector;
 311     }
 312
 313     blk = g_new(BlkMigBlock, 1);
 314     blk->buf = g_malloc(BLK_MIG_BLOCK_SIZE);
 315     blk->bmds = bmds;
 316     blk->sector = cur_sector;
 317     blk->nr_sectors = nr_sectors;
 318
 319     qemu_iovec_init_buf(&blk->qiov, blk->buf, nr_sectors * BDRV_SECTOR_SIZE);
 320
 321     blk_mig_lock();
 322     block_mig_state.submitted++;
 323     blk_mig_unlock();
 324
 325     /* We do not know if bs is under the main thread (and thus does
 326      * not acquire the AioContext when doing AIO) or rather under
 327      * dataplane.  Thus acquire both the iothread mutex and the
 328      * AioContext.
 329      *
 330      * This is ugly and will disappear when we make bdrv_* thread-safe,
 331      * without the need to acquire the AioContext.
 332      */
 333     qemu_mutex_lock_iothread();
 334     aio_context_acquire(blk_get_aio_context(bmds->blk));
 335     bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, cur_sector * BDRV_SECTOR_SIZE,
 336                             nr_sectors * BDRV_SECTOR_SIZE);
 337     blk->aiocb = blk_aio_preadv(bb, cur_sector * BDRV_SECTOR_SIZE, &blk->qiov,
 338                                 0, blk_mig_read_cb, blk);
 339     aio_context_release(blk_get_aio_context(bmds->blk));
 340     qemu_mutex_unlock_iothread();
 341
 342     bmds->cur_sector = cur_sector + nr_sectors;
 343     return (bmds->cur_sector >= total_sectors);
 344 }
 345
 346 /* Called with iothread lock taken.  */
 347
 348 static int set_dirty_tracking(void)
 349 {
 350     BlkMigDevState *bmds;
 351     int ret;
 352
 353     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 354         bmds->dirty_bitmap = bdrv_create_dirty_bitmap(blk_bs(bmds->blk),
 355                                                       BLK_MIG_BLOCK_SIZE,
 356                                                       NULL, NULL);
 357         if (!bmds->dirty_bitmap) {
 358             ret = -errno;
 359             goto fail;
 360         }
 361     }
 362     return 0;
 363
 364 fail:
 365     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 366         if (bmds->dirty_bitmap) {
 367             bdrv_release_dirty_bitmap(bmds->dirty_bitmap);
 368         }
 369     }
 370     return ret;
 371 }
 372
 373 /* Called with iothread lock taken.  */
 374
 375 static void unset_dirty_tracking(void)
 376 {
 377     BlkMigDevState *bmds;
 378
 379     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 380         bdrv_release_dirty_bitmap(bmds->dirty_bitmap);
 381     }
 382 }
 383
 384 static int init_blk_migration(QEMUFile *f)
 385 {
 386     BlockDriverState *bs;
 387     BlkMigDevState *bmds;
 388     int64_t sectors;
 389     BdrvNextIterator it;
 390     int i, num_bs = 0;
 391     struct {
 392         BlkMigDevState *bmds;
 393         BlockDriverState *bs;
 394     } *bmds_bs;
 395     Error *local_err = NULL;
 396     int ret;
 397
 398     block_mig_state.submitted = 0;
 399     block_mig_state.read_done = 0;
 400     block_mig_state.transferred = 0;
 401     block_mig_state.total_sector_sum = 0;
 402     block_mig_state.prev_progress = -1;
 403     block_mig_state.bulk_completed = 0;
 404     block_mig_state.zero_blocks = migrate_zero_blocks();
 405
 406     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
 407         num_bs++;
 408     }
 409     bmds_bs = g_malloc0(num_bs * sizeof(*bmds_bs));
 410
 411     for (i = 0, bs = bdrv_first(&it); bs; bs = bdrv_next(&it), i++) {
 412         if (bdrv_is_read_only(bs)) {
 413             continue;
 414         }
 415
 416         sectors = bdrv_nb_sectors(bs);
 417         if (sectors <= 0) {
 418             ret = sectors;
 419             bdrv_next_cleanup(&it);
 420             goto out;
 421         }
 422
 423         bmds = g_new0(BlkMigDevState, 1);
 424         bmds->blk = blk_new(qemu_get_aio_context(),
 425                             BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 426         bmds->blk_name = g_strdup(bdrv_get_device_name(bs));
 427         bmds->bulk_completed = 0;
 428         bmds->total_sectors = sectors;
 429         bmds->completed_sectors = 0;
 430         bmds->shared_base = migrate_use_block_incremental();
 431
 432         assert(i < num_bs);
 433         bmds_bs[i].bmds = bmds;
 434         bmds_bs[i].bs = bs;
 435
 436         block_mig_state.total_sector_sum += sectors;
 437
 438         if (bmds->shared_base) {
 439             trace_migration_block_init_shared(bdrv_get_device_name(bs));
 440         } else {
 441             trace_migration_block_init_full(bdrv_get_device_name(bs));
 442         }
 443
 444         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 445     }
 446
 447     /* Can only insert new BDSes now because doing so while iterating block
 448      * devices may end up in a deadlock (iterating the new BDSes, too). */
 449     for (i = 0; i < num_bs; i++) {
 450         BlkMigDevState *bmds = bmds_bs[i].bmds;
 451         BlockDriverState *bs = bmds_bs[i].bs;
 452
 453         if (bmds) {
 454             ret = blk_insert_bs(bmds->blk, bs, &local_err);
 455             if (ret < 0) {
 456                 error_report_err(local_err);
 457                 goto out;
 458             }
 459
 460             alloc_aio_bitmap(bmds);
 461             error_setg(&bmds->blocker, "block device is in use by migration");
 462             bdrv_op_block_all(bs, bmds->blocker);
 463         }
 464     }
 465
 466     ret = 0;
 467 out:
 468     g_free(bmds_bs);
 469     return ret;
 470 }
 471
 472 /* Called with no lock taken.  */
 473
 474 static int blk_mig_save_bulked_block(QEMUFile *f)
 475 {
 476     int64_t completed_sector_sum = 0;
 477     BlkMigDevState *bmds;
 478     int progress;
 479     int ret = 0;
 480
 481     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 482         if (bmds->bulk_completed == 0) {
 483             if (mig_save_device_bulk(f, bmds) == 1) {
 484                 /* completed bulk section for this device */
 485                 bmds->bulk_completed = 1;
 486             }
 487             completed_sector_sum += bmds->completed_sectors;
 488             ret = 1;
 489             break;
 490         } else {
 491             completed_sector_sum += bmds->completed_sectors;
 492         }
 493     }
 494
 495     if (block_mig_state.total_sector_sum != 0) {
 496         progress = completed_sector_sum * 100 /
 497                    block_mig_state.total_sector_sum;
 498     } else {
 499         progress = 100;
 500     }
 501     if (progress != block_mig_state.prev_progress) {
 502         block_mig_state.prev_progress = progress;
 503         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 504                          | BLK_MIG_FLAG_PROGRESS);
 505         DPRINTF("Completed %d %%\r", progress);
 506     }
 507
 508     return ret;
 509 }
 510
 511 static void blk_mig_reset_dirty_cursor(void)
 512 {
 513     BlkMigDevState *bmds;
 514
 515     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 516         bmds->cur_dirty = 0;
 517     }
 518 }
 519
 520 /* Called with iothread lock and AioContext taken.  */
 521
 522 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
 523                                  int is_async)
 524 {
 525     BlkMigBlock *blk;
 526     int64_t total_sectors = bmds->total_sectors;
 527     int64_t sector;
 528     int nr_sectors;
 529     int ret = -EIO;
 530
 531     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
 532         blk_mig_lock();
 533         if (bmds_aio_inflight(bmds, sector)) {
 534             blk_mig_unlock();
 535             blk_drain(bmds->blk);
 536         } else {
 537             blk_mig_unlock();
 538         }
 539         bdrv_dirty_bitmap_lock(bmds->dirty_bitmap);
 540         if (bdrv_dirty_bitmap_get_locked(bmds->dirty_bitmap,
 541                                          sector * BDRV_SECTOR_SIZE)) {
 542             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 543                 nr_sectors = total_sectors - sector;
 544             } else {
 545                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 546             }
 547             bdrv_reset_dirty_bitmap_locked(bmds->dirty_bitmap,
 548                                            sector * BDRV_SECTOR_SIZE,
 549                                            nr_sectors * BDRV_SECTOR_SIZE);
 550             bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 551
 552             blk = g_new(BlkMigBlock, 1);
 553             blk->buf = g_malloc(BLK_MIG_BLOCK_SIZE);
 554             blk->bmds = bmds;
 555             blk->sector = sector;
 556             blk->nr_sectors = nr_sectors;
 557
 558             if (is_async) {
 559                 qemu_iovec_init_buf(&blk->qiov, blk->buf,
 560                                     nr_sectors * BDRV_SECTOR_SIZE);
 561
 562                 blk->aiocb = blk_aio_preadv(bmds->blk,
 563                                             sector * BDRV_SECTOR_SIZE,
 564                                             &blk->qiov, 0, blk_mig_read_cb,
 565                                             blk);
 566
 567                 blk_mig_lock();
 568                 block_mig_state.submitted++;
 569                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 570                 blk_mig_unlock();
 571             } else {
 572                 ret = blk_pread(bmds->blk, sector * BDRV_SECTOR_SIZE,
 573                                 nr_sectors * BDRV_SECTOR_SIZE, blk->buf, 0);
 574                 if (ret < 0) {
 575                     goto error;
 576                 }
 577                 blk_send(f, blk);
 578
 579                 g_free(blk->buf);
 580                 g_free(blk);
 581             }
 582
 583             sector += nr_sectors;
 584             bmds->cur_dirty = sector;
 585             break;
 586         }
 587
 588         bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 589         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 590         bmds->cur_dirty = sector;
 591     }
 592
 593     return (bmds->cur_dirty >= bmds->total_sectors);
 594
 595 error:
 596     trace_migration_block_save_device_dirty(sector);
 597     g_free(blk->buf);
 598     g_free(blk);
 599     return ret;
 600 }
 601
 602 /* Called with iothread lock taken.
 603  *
 604  * return value:
 605  * 0: too much data for max_downtime
 606  * 1: few enough data for max_downtime
 607 */
 608 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 609 {
 610     BlkMigDevState *bmds;
 611     int ret = 1;
 612
 613     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 614         aio_context_acquire(blk_get_aio_context(bmds->blk));
 615         ret = mig_save_device_dirty(f, bmds, is_async);
 616         aio_context_release(blk_get_aio_context(bmds->blk));
 617         if (ret <= 0) {
 618             break;
 619         }
 620     }
 621
 622     return ret;
 623 }
 624
 625 /* Called with no locks taken.  */
 626
 627 static int flush_blks(QEMUFile *f)
 628 {
 629     BlkMigBlock *blk;
 630     int ret = 0;
 631
 632     trace_migration_block_flush_blks("Enter", block_mig_state.submitted,
 633                                      block_mig_state.read_done,
 634                                      block_mig_state.transferred);
 635
 636     blk_mig_lock();
 637     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 638         if (qemu_file_rate_limit(f)) {
 639             break;
 640         }
 641         if (blk->ret < 0) {
 642             ret = blk->ret;
 643             break;
 644         }
 645
 646         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 647         blk_mig_unlock();
 648         blk_send(f, blk);
 649         blk_mig_lock();
 650
 651         g_free(blk->buf);
 652         g_free(blk);
 653
 654         block_mig_state.read_done--;
 655         block_mig_state.transferred++;
 656         assert(block_mig_state.read_done >= 0);
 657     }
 658     blk_mig_unlock();
 659
 660     trace_migration_block_flush_blks("Exit", block_mig_state.submitted,
 661                                      block_mig_state.read_done,
 662                                      block_mig_state.transferred);
 663     return ret;
 664 }
 665
 666 /* Called with iothread lock taken.  */
 667
 668 static int64_t get_remaining_dirty(void)
 669 {
 670     BlkMigDevState *bmds;
 671     int64_t dirty = 0;
 672
 673     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 674         aio_context_acquire(blk_get_aio_context(bmds->blk));
 675         dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
 676         aio_context_release(blk_get_aio_context(bmds->blk));
 677     }
 678
 679     return dirty;
 680 }
 681
 682
 683
 684 /* Called with iothread lock taken.  */
 685 static void block_migration_cleanup_bmds(void)
 686 {
 687     BlkMigDevState *bmds;
 688     AioContext *ctx;
 689
 690     unset_dirty_tracking();
 691
 692     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 693         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 694         bdrv_op_unblock_all(blk_bs(bmds->blk), bmds->blocker);
 695         error_free(bmds->blocker);
 696
 697         /* Save ctx, because bmds->blk can disappear during blk_unref.  */
 698         ctx = blk_get_aio_context(bmds->blk);
 699         aio_context_acquire(ctx);
 700         blk_unref(bmds->blk);
 701         aio_context_release(ctx);
 702
 703         g_free(bmds->blk_name);
 704         g_free(bmds->aio_bitmap);
 705         g_free(bmds);
 706     }
 707 }
 708
 709 /* Called with iothread lock taken.  */
 710 static void block_migration_cleanup(void *opaque)
 711 {
 712     BlkMigBlock *blk;
 713
 714     bdrv_drain_all();
 715
 716     block_migration_cleanup_bmds();
 717
 718     blk_mig_lock();
 719     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 720         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 721         g_free(blk->buf);
 722         g_free(blk);
 723     }
 724     blk_mig_unlock();
 725 }
 726
 727 static int block_save_setup(QEMUFile *f, void *opaque)
 728 {
 729     int ret;
 730
 731     trace_migration_block_save("setup", block_mig_state.submitted,
 732                                block_mig_state.transferred);
 733
 734     qemu_mutex_lock_iothread();
 735     ret = init_blk_migration(f);
 736     if (ret < 0) {
 737         qemu_mutex_unlock_iothread();
 738         return ret;
 739     }
 740
 741     /* start track dirty blocks */
 742     ret = set_dirty_tracking();
 743
 744     qemu_mutex_unlock_iothread();
 745
 746     if (ret) {
 747         return ret;
 748     }
 749
 750     ret = flush_blks(f);
 751     blk_mig_reset_dirty_cursor();
 752     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 753
 754     return ret;
 755 }
 756
 757 static int block_save_iterate(QEMUFile *f, void *opaque)
 758 {
 759     int ret;
 760     int64_t last_bytes = qemu_file_total_transferred(f);
 761     int64_t delta_bytes;
 762
 763     trace_migration_block_save("iterate", block_mig_state.submitted,
 764                                block_mig_state.transferred);
 765
 766     ret = flush_blks(f);
 767     if (ret) {
 768         return ret;
 769     }
 770
 771     blk_mig_reset_dirty_cursor();
 772
 773     /* control the rate of transfer */
 774     blk_mig_lock();
 775     while (block_mig_state.read_done * BLK_MIG_BLOCK_SIZE <
 776            qemu_file_get_rate_limit(f) &&
 777            block_mig_state.submitted < MAX_PARALLEL_IO &&
 778            (block_mig_state.submitted + block_mig_state.read_done) <
 779            MAX_IO_BUFFERS) {
 780         blk_mig_unlock();
 781         if (block_mig_state.bulk_completed == 0) {
 782             /* first finish the bulk phase */
 783             if (blk_mig_save_bulked_block(f) == 0) {
 784                 /* finished saving bulk on all devices */
 785                 block_mig_state.bulk_completed = 1;
 786             }
 787             ret = 0;
 788         } else {
 789             /* Always called with iothread lock taken for
 790              * simplicity, block_save_complete also calls it.
 791              */
 792             qemu_mutex_lock_iothread();
 793             ret = blk_mig_save_dirty_block(f, 1);
 794             qemu_mutex_unlock_iothread();
 795         }
 796         if (ret < 0) {
 797             return ret;
 798         }
 799         blk_mig_lock();
 800         if (ret != 0) {
 801             /* no more dirty blocks */
 802             break;
 803         }
 804     }
 805     blk_mig_unlock();
 806
 807     ret = flush_blks(f);
 808     if (ret) {
 809         return ret;
 810     }
 811
 812     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 813     delta_bytes = qemu_file_total_transferred(f) - last_bytes;
 814     if (delta_bytes > 0) {
 815         return 1;
 816     } else if (delta_bytes < 0) {
 817         return -1;
 818     } else {
 819         return 0;
 820     }
 821 }
 822
 823 /* Called with iothread lock taken.  */
 824
 825 static int block_save_complete(QEMUFile *f, void *opaque)
 826 {
 827     int ret;
 828
 829     trace_migration_block_save("complete", block_mig_state.submitted,
 830                                block_mig_state.transferred);
 831
 832     ret = flush_blks(f);
 833     if (ret) {
 834         return ret;
 835     }
 836
 837     blk_mig_reset_dirty_cursor();
 838
 839     /* we know for sure that save bulk is completed and
 840        all async read completed */
 841     blk_mig_lock();
 842     assert(block_mig_state.submitted == 0);
 843     blk_mig_unlock();
 844
 845     do {
 846         ret = blk_mig_save_dirty_block(f, 0);
 847         if (ret < 0) {
 848             return ret;
 849         }
 850     } while (ret == 0);
 851
 852     /* report completion */
 853     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 854
 855     trace_migration_block_save_complete();
 856
 857     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 858
 859     /* Make sure that our BlockBackends are gone, so that the block driver
 860      * nodes can be inactivated. */
 861     block_migration_cleanup_bmds();
 862
 863     return 0;
 864 }
 865
 866 static void block_save_pending(QEMUFile *f, void *opaque, uint64_t max_size,
 867                                uint64_t *res_precopy_only,
 868                                uint64_t *res_compatible,
 869                                uint64_t *res_postcopy_only)
 870 {
 871     /* Estimate pending number of bytes to send */
 872     uint64_t pending;
 873
 874     qemu_mutex_lock_iothread();
 875     pending = get_remaining_dirty();
 876     qemu_mutex_unlock_iothread();
 877
 878     blk_mig_lock();
 879     pending += block_mig_state.submitted * BLK_MIG_BLOCK_SIZE +
 880                block_mig_state.read_done * BLK_MIG_BLOCK_SIZE;
 881     blk_mig_unlock();
 882
 883     /* Report at least one block pending during bulk phase */
 884     if (!pending && !block_mig_state.bulk_completed) {
 885         pending = BLK_MIG_BLOCK_SIZE;
 886     }
 887
 888     trace_migration_block_save_pending(pending);
 889     /* We don't do postcopy */
 890     *res_precopy_only += pending;
 891 }
 892
 893 static int block_load(QEMUFile *f, void *opaque, int version_id)
 894 {
 895     static int banner_printed;
 896     int len, flags;
 897     char device_name[256];
 898     int64_t addr;
 899     BlockBackend *blk, *blk_prev = NULL;
 900     Error *local_err = NULL;
 901     uint8_t *buf;
 902     int64_t total_sectors = 0;
 903     int nr_sectors;
 904     int ret;
 905     BlockDriverInfo bdi;
 906     int cluster_size = BLK_MIG_BLOCK_SIZE;
 907
 908     do {
 909         addr = qemu_get_be64(f);
 910
 911         flags = addr & (BDRV_SECTOR_SIZE - 1);
 912         addr >>= BDRV_SECTOR_BITS;
 913
 914         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 915             /* get device name */
 916             len = qemu_get_byte(f);
 917             qemu_get_buffer(f, (uint8_t *)device_name, len);
 918             device_name[len] = '\0';
 919
 920             blk = blk_by_name(device_name);
 921             if (!blk) {
 922                 fprintf(stderr, "Error unknown block device %s\n",
 923                         device_name);
 924                 return -EINVAL;
 925             }
 926
 927             if (blk != blk_prev) {
 928                 blk_prev = blk;
 929                 total_sectors = blk_nb_sectors(blk);
 930                 if (total_sectors <= 0) {
 931                     error_report("Error getting length of block device %s",
 932                                  device_name);
 933                     return -EINVAL;
 934                 }
 935
 936                 blk_activate(blk, &local_err);
 937                 if (local_err) {
 938                     error_report_err(local_err);
 939                     return -EINVAL;
 940                 }
 941
 942                 ret = bdrv_get_info(blk_bs(blk), &bdi);
 943                 if (ret == 0 && bdi.cluster_size > 0 &&
 944                     bdi.cluster_size <= BLK_MIG_BLOCK_SIZE &&
 945                     BLK_MIG_BLOCK_SIZE % bdi.cluster_size == 0) {
 946                     cluster_size = bdi.cluster_size;
 947                 } else {
 948                     cluster_size = BLK_MIG_BLOCK_SIZE;
 949                 }
 950             }
 951
 952             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 953                 nr_sectors = total_sectors - addr;
 954             } else {
 955                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 956             }
 957
 958             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 959                 ret = blk_pwrite_zeroes(blk, addr * BDRV_SECTOR_SIZE,
 960                                         nr_sectors * BDRV_SECTOR_SIZE,
 961                                         BDRV_REQ_MAY_UNMAP);
 962             } else {
 963                 int i;
 964                 int64_t cur_addr;
 965                 uint8_t *cur_buf;
 966
 967                 buf = g_malloc(BLK_MIG_BLOCK_SIZE);
 968                 qemu_get_buffer(f, buf, BLK_MIG_BLOCK_SIZE);
 969                 for (i = 0; i < BLK_MIG_BLOCK_SIZE / cluster_size; i++) {
 970                     cur_addr = addr * BDRV_SECTOR_SIZE + i * cluster_size;
 971                     cur_buf = buf + i * cluster_size;
 972
 973                     if ((!block_mig_state.zero_blocks ||
 974                         cluster_size < BLK_MIG_BLOCK_SIZE) &&
 975                         buffer_is_zero(cur_buf, cluster_size)) {
 976                         ret = blk_pwrite_zeroes(blk, cur_addr,
 977                                                 cluster_size,
 978                                                 BDRV_REQ_MAY_UNMAP);
 979                     } else {
 980                         ret = blk_pwrite(blk, cur_addr, cluster_size, cur_buf,
 981                                          0);
 982                     }
 983                     if (ret < 0) {
 984                         break;
 985                     }
 986                 }
 987                 g_free(buf);
 988             }
 989
 990             if (ret < 0) {
 991                 return ret;
 992             }
 993         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 994             if (!banner_printed) {
 995                 printf("Receiving block device images\n");
 996                 banner_printed = 1;
 997             }
 998             printf("Completed %d %%%c", (int)addr,
 999                    (addr == 100) ? '\n' : '\r');
1000             fflush(stdout);
1001         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
1002             fprintf(stderr, "Unknown block migration flags: 0x%x\n", flags);
1003             return -EINVAL;
1004         }
1005         ret = qemu_file_get_error(f);
1006         if (ret != 0) {
1007             return ret;
1008         }
1009     } while (!(flags & BLK_MIG_FLAG_EOS));
1010
1011     return 0;
1012 }
1013
1014 static bool block_is_active(void *opaque)
1015 {
1016     return migrate_use_block();
1017 }
1018
1019 static SaveVMHandlers savevm_block_handlers = {
1020     .save_setup = block_save_setup,
1021     .save_live_iterate = block_save_iterate,
1022     .save_live_complete_precopy = block_save_complete,
1023     .save_live_pending = block_save_pending,
1024     .load_state = block_load,
1025     .save_cleanup = block_migration_cleanup,
1026     .is_active = block_is_active,
1027 };
1028
1029 void blk_mig_init(void)
1030 {
1031     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
1032     QSIMPLEQ_INIT(&block_mig_state.blk_list);
1033     qemu_mutex_init(&block_mig_state.lock);
1034
1035     register_savevm_live("block", 0, 1, &savevm_block_handlers,
1036                          &block_mig_state);
1037 }