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-common.h"
  19 #include "block/block.h"
  20 #include "qemu/error-report.h"
  21 #include "qemu/main-loop.h"
  22 #include "hw/hw.h"
  23 #include "qemu/cutils.h"
  24 #include "qemu/queue.h"
  25 #include "qemu/timer.h"
  26 #include "migration/block.h"
  27 #include "migration/migration.h"
  28 #include "sysemu/blockdev.h"
  29 #include "sysemu/block-backend.h"
  30
  31 #define BLOCK_SIZE                       (1 << 20)
  32 #define BDRV_SECTORS_PER_DIRTY_CHUNK     (BLOCK_SIZE >> BDRV_SECTOR_BITS)
  33
  34 #define BLK_MIG_FLAG_DEVICE_BLOCK       0x01
  35 #define BLK_MIG_FLAG_EOS                0x02
  36 #define BLK_MIG_FLAG_PROGRESS           0x04
  37 #define BLK_MIG_FLAG_ZERO_BLOCK         0x08
  38
  39 #define MAX_IS_ALLOCATED_SEARCH 65536
  40
  41 #define MAX_INFLIGHT_IO 512
  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     struct iovec iov;
  88     QEMUIOVector qiov;
  89     BlockAIOCB *aiocb;
  90
  91     /* Protected by block migration lock.  */
  92     int ret;
  93     QSIMPLEQ_ENTRY(BlkMigBlock) entry;
  94 } BlkMigBlock;
  95
  96 typedef struct BlkMigState {
  97     /* Written during setup phase.  Can be read without a lock.  */
  98     int blk_enable;
  99     int shared_base;
 100     QSIMPLEQ_HEAD(bmds_list, BlkMigDevState) bmds_list;
 101     int64_t total_sector_sum;
 102     bool zero_blocks;
 103
 104     /* Protected by lock.  */
 105     QSIMPLEQ_HEAD(blk_list, BlkMigBlock) blk_list;
 106     int submitted;
 107     int read_done;
 108
 109     /* Only used by migration thread.  Does not need a lock.  */
 110     int transferred;
 111     int prev_progress;
 112     int bulk_completed;
 113
 114     /* Lock must be taken _inside_ the iothread lock and any AioContexts.  */
 115     QemuMutex lock;
 116 } BlkMigState;
 117
 118 static BlkMigState block_mig_state;
 119
 120 static void blk_mig_lock(void)
 121 {
 122     qemu_mutex_lock(&block_mig_state.lock);
 123 }
 124
 125 static void blk_mig_unlock(void)
 126 {
 127     qemu_mutex_unlock(&block_mig_state.lock);
 128 }
 129
 130 /* Must run outside of the iothread lock during the bulk phase,
 131  * or the VM will stall.
 132  */
 133
 134 static void blk_send(QEMUFile *f, BlkMigBlock * blk)
 135 {
 136     int len;
 137     uint64_t flags = BLK_MIG_FLAG_DEVICE_BLOCK;
 138
 139     if (block_mig_state.zero_blocks &&
 140         buffer_is_zero(blk->buf, BLOCK_SIZE)) {
 141         flags |= BLK_MIG_FLAG_ZERO_BLOCK;
 142     }
 143
 144     /* sector number and flags */
 145     qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
 146                      | flags);
 147
 148     /* device name */
 149     len = strlen(blk->bmds->blk_name);
 150     qemu_put_byte(f, len);
 151     qemu_put_buffer(f, (uint8_t *) blk->bmds->blk_name, len);
 152
 153     /* if a block is zero we need to flush here since the network
 154      * bandwidth is now a lot higher than the storage device bandwidth.
 155      * thus if we queue zero blocks we slow down the migration */
 156     if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 157         qemu_fflush(f);
 158         return;
 159     }
 160
 161     qemu_put_buffer(f, blk->buf, BLOCK_SIZE);
 162 }
 163
 164 int blk_mig_active(void)
 165 {
 166     return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
 167 }
 168
 169 uint64_t blk_mig_bytes_transferred(void)
 170 {
 171     BlkMigDevState *bmds;
 172     uint64_t sum = 0;
 173
 174     blk_mig_lock();
 175     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 176         sum += bmds->completed_sectors;
 177     }
 178     blk_mig_unlock();
 179     return sum << BDRV_SECTOR_BITS;
 180 }
 181
 182 uint64_t blk_mig_bytes_remaining(void)
 183 {
 184     return blk_mig_bytes_total() - blk_mig_bytes_transferred();
 185 }
 186
 187 uint64_t blk_mig_bytes_total(void)
 188 {
 189     BlkMigDevState *bmds;
 190     uint64_t sum = 0;
 191
 192     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 193         sum += bmds->total_sectors;
 194     }
 195     return sum << BDRV_SECTOR_BITS;
 196 }
 197
 198
 199 /* Called with migration lock held.  */
 200
 201 static int bmds_aio_inflight(BlkMigDevState *bmds, int64_t sector)
 202 {
 203     int64_t chunk = sector / (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK;
 204
 205     if (sector < blk_nb_sectors(bmds->blk)) {
 206         return !!(bmds->aio_bitmap[chunk / (sizeof(unsigned long) * 8)] &
 207             (1UL << (chunk % (sizeof(unsigned long) * 8))));
 208     } else {
 209         return 0;
 210     }
 211 }
 212
 213 /* Called with migration lock held.  */
 214
 215 static void bmds_set_aio_inflight(BlkMigDevState *bmds, int64_t sector_num,
 216                              int nb_sectors, int set)
 217 {
 218     int64_t start, end;
 219     unsigned long val, idx, bit;
 220
 221     start = sector_num / BDRV_SECTORS_PER_DIRTY_CHUNK;
 222     end = (sector_num + nb_sectors - 1) / BDRV_SECTORS_PER_DIRTY_CHUNK;
 223
 224     for (; start <= end; start++) {
 225         idx = start / (sizeof(unsigned long) * 8);
 226         bit = start % (sizeof(unsigned long) * 8);
 227         val = bmds->aio_bitmap[idx];
 228         if (set) {
 229             val |= 1UL << bit;
 230         } else {
 231             val &= ~(1UL << bit);
 232         }
 233         bmds->aio_bitmap[idx] = val;
 234     }
 235 }
 236
 237 static void alloc_aio_bitmap(BlkMigDevState *bmds)
 238 {
 239     BlockBackend *bb = bmds->blk;
 240     int64_t bitmap_size;
 241
 242     bitmap_size = blk_nb_sectors(bb) + BDRV_SECTORS_PER_DIRTY_CHUNK * 8 - 1;
 243     bitmap_size /= BDRV_SECTORS_PER_DIRTY_CHUNK * 8;
 244
 245     bmds->aio_bitmap = g_malloc0(bitmap_size);
 246 }
 247
 248 /* Never hold migration lock when yielding to the main loop!  */
 249
 250 static void blk_mig_read_cb(void *opaque, int ret)
 251 {
 252     BlkMigBlock *blk = opaque;
 253
 254     blk_mig_lock();
 255     blk->ret = ret;
 256
 257     QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry);
 258     bmds_set_aio_inflight(blk->bmds, blk->sector, blk->nr_sectors, 0);
 259
 260     block_mig_state.submitted--;
 261     block_mig_state.read_done++;
 262     assert(block_mig_state.submitted >= 0);
 263     blk_mig_unlock();
 264 }
 265
 266 /* Called with no lock taken.  */
 267
 268 static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds)
 269 {
 270     int64_t total_sectors = bmds->total_sectors;
 271     int64_t cur_sector = bmds->cur_sector;
 272     BlockBackend *bb = bmds->blk;
 273     BlkMigBlock *blk;
 274     int nr_sectors;
 275
 276     if (bmds->shared_base) {
 277         qemu_mutex_lock_iothread();
 278         aio_context_acquire(blk_get_aio_context(bb));
 279         while (cur_sector < total_sectors &&
 280                !bdrv_is_allocated(blk_bs(bb), cur_sector,
 281                                   MAX_IS_ALLOCATED_SEARCH, &nr_sectors)) {
 282             cur_sector += nr_sectors;
 283         }
 284         aio_context_release(blk_get_aio_context(bb));
 285         qemu_mutex_unlock_iothread();
 286     }
 287
 288     if (cur_sector >= total_sectors) {
 289         bmds->cur_sector = bmds->completed_sectors = total_sectors;
 290         return 1;
 291     }
 292
 293     bmds->completed_sectors = cur_sector;
 294
 295     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
 296
 297     /* we are going to transfer a full block even if it is not allocated */
 298     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 299
 300     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 301         nr_sectors = total_sectors - cur_sector;
 302     }
 303
 304     blk = g_new(BlkMigBlock, 1);
 305     blk->buf = g_malloc(BLOCK_SIZE);
 306     blk->bmds = bmds;
 307     blk->sector = cur_sector;
 308     blk->nr_sectors = nr_sectors;
 309
 310     blk->iov.iov_base = blk->buf;
 311     blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 312     qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 313
 314     blk_mig_lock();
 315     block_mig_state.submitted++;
 316     blk_mig_unlock();
 317
 318     /* We do not know if bs is under the main thread (and thus does
 319      * not acquire the AioContext when doing AIO) or rather under
 320      * dataplane.  Thus acquire both the iothread mutex and the
 321      * AioContext.
 322      *
 323      * This is ugly and will disappear when we make bdrv_* thread-safe,
 324      * without the need to acquire the AioContext.
 325      */
 326     qemu_mutex_lock_iothread();
 327     aio_context_acquire(blk_get_aio_context(bmds->blk));
 328     blk->aiocb = blk_aio_preadv(bb, cur_sector * BDRV_SECTOR_SIZE, &blk->qiov,
 329                                 0, blk_mig_read_cb, blk);
 330
 331     bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, cur_sector, nr_sectors);
 332     aio_context_release(blk_get_aio_context(bmds->blk));
 333     qemu_mutex_unlock_iothread();
 334
 335     bmds->cur_sector = cur_sector + nr_sectors;
 336     return (bmds->cur_sector >= total_sectors);
 337 }
 338
 339 /* Called with iothread lock taken.  */
 340
 341 static int set_dirty_tracking(void)
 342 {
 343     BlkMigDevState *bmds;
 344     int ret;
 345
 346     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 347         aio_context_acquire(blk_get_aio_context(bmds->blk));
 348         bmds->dirty_bitmap = bdrv_create_dirty_bitmap(blk_bs(bmds->blk),
 349                                                       BLOCK_SIZE, NULL, NULL);
 350         aio_context_release(blk_get_aio_context(bmds->blk));
 351         if (!bmds->dirty_bitmap) {
 352             ret = -errno;
 353             goto fail;
 354         }
 355     }
 356     return 0;
 357
 358 fail:
 359     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 360         if (bmds->dirty_bitmap) {
 361             aio_context_acquire(blk_get_aio_context(bmds->blk));
 362             bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
 363             aio_context_release(blk_get_aio_context(bmds->blk));
 364         }
 365     }
 366     return ret;
 367 }
 368
 369 /* Called with iothread lock taken.  */
 370
 371 static void unset_dirty_tracking(void)
 372 {
 373     BlkMigDevState *bmds;
 374
 375     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 376         aio_context_acquire(blk_get_aio_context(bmds->blk));
 377         bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
 378         aio_context_release(blk_get_aio_context(bmds->blk));
 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             goto out;
 418         }
 419
 420         bmds = g_new0(BlkMigDevState, 1);
 421         bmds->blk = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 422         bmds->blk_name = g_strdup(bdrv_get_device_name(bs));
 423         bmds->bulk_completed = 0;
 424         bmds->total_sectors = sectors;
 425         bmds->completed_sectors = 0;
 426         bmds->shared_base = block_mig_state.shared_base;
 427
 428         assert(i < num_bs);
 429         bmds_bs[i].bmds = bmds;
 430         bmds_bs[i].bs = bs;
 431
 432         block_mig_state.total_sector_sum += sectors;
 433
 434         if (bmds->shared_base) {
 435             DPRINTF("Start migration for %s with shared base image\n",
 436                     bdrv_get_device_name(bs));
 437         } else {
 438             DPRINTF("Start full migration for %s\n", bdrv_get_device_name(bs));
 439         }
 440
 441         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 442     }
 443
 444     /* Can only insert new BDSes now because doing so while iterating block
 445      * devices may end up in a deadlock (iterating the new BDSes, too). */
 446     for (i = 0; i < num_bs; i++) {
 447         BlkMigDevState *bmds = bmds_bs[i].bmds;
 448         BlockDriverState *bs = bmds_bs[i].bs;
 449
 450         if (bmds) {
 451             ret = blk_insert_bs(bmds->blk, bs, &local_err);
 452             if (ret < 0) {
 453                 error_report_err(local_err);
 454                 goto out;
 455             }
 456
 457             alloc_aio_bitmap(bmds);
 458             error_setg(&bmds->blocker, "block device is in use by migration");
 459             bdrv_op_block_all(bs, bmds->blocker);
 460         }
 461     }
 462
 463     ret = 0;
 464 out:
 465     g_free(bmds_bs);
 466     return ret;
 467 }
 468
 469 /* Called with no lock taken.  */
 470
 471 static int blk_mig_save_bulked_block(QEMUFile *f)
 472 {
 473     int64_t completed_sector_sum = 0;
 474     BlkMigDevState *bmds;
 475     int progress;
 476     int ret = 0;
 477
 478     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 479         if (bmds->bulk_completed == 0) {
 480             if (mig_save_device_bulk(f, bmds) == 1) {
 481                 /* completed bulk section for this device */
 482                 bmds->bulk_completed = 1;
 483             }
 484             completed_sector_sum += bmds->completed_sectors;
 485             ret = 1;
 486             break;
 487         } else {
 488             completed_sector_sum += bmds->completed_sectors;
 489         }
 490     }
 491
 492     if (block_mig_state.total_sector_sum != 0) {
 493         progress = completed_sector_sum * 100 /
 494                    block_mig_state.total_sector_sum;
 495     } else {
 496         progress = 100;
 497     }
 498     if (progress != block_mig_state.prev_progress) {
 499         block_mig_state.prev_progress = progress;
 500         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 501                          | BLK_MIG_FLAG_PROGRESS);
 502         DPRINTF("Completed %d %%\r", progress);
 503     }
 504
 505     return ret;
 506 }
 507
 508 static void blk_mig_reset_dirty_cursor(void)
 509 {
 510     BlkMigDevState *bmds;
 511
 512     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 513         bmds->cur_dirty = 0;
 514     }
 515 }
 516
 517 /* Called with iothread lock and AioContext taken.  */
 518
 519 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
 520                                  int is_async)
 521 {
 522     BlkMigBlock *blk;
 523     BlockDriverState *bs = blk_bs(bmds->blk);
 524     int64_t total_sectors = bmds->total_sectors;
 525     int64_t sector;
 526     int nr_sectors;
 527     int ret = -EIO;
 528
 529     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
 530         blk_mig_lock();
 531         if (bmds_aio_inflight(bmds, sector)) {
 532             blk_mig_unlock();
 533             blk_drain(bmds->blk);
 534         } else {
 535             blk_mig_unlock();
 536         }
 537         if (bdrv_get_dirty(bs, bmds->dirty_bitmap, sector)) {
 538
 539             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 540                 nr_sectors = total_sectors - sector;
 541             } else {
 542                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 543             }
 544             blk = g_new(BlkMigBlock, 1);
 545             blk->buf = g_malloc(BLOCK_SIZE);
 546             blk->bmds = bmds;
 547             blk->sector = sector;
 548             blk->nr_sectors = nr_sectors;
 549
 550             if (is_async) {
 551                 blk->iov.iov_base = blk->buf;
 552                 blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 553                 qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 554
 555                 blk->aiocb = blk_aio_preadv(bmds->blk,
 556                                             sector * BDRV_SECTOR_SIZE,
 557                                             &blk->qiov, 0, blk_mig_read_cb,
 558                                             blk);
 559
 560                 blk_mig_lock();
 561                 block_mig_state.submitted++;
 562                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 563                 blk_mig_unlock();
 564             } else {
 565                 ret = blk_pread(bmds->blk, sector * BDRV_SECTOR_SIZE, blk->buf,
 566                                 nr_sectors * BDRV_SECTOR_SIZE);
 567                 if (ret < 0) {
 568                     goto error;
 569                 }
 570                 blk_send(f, blk);
 571
 572                 g_free(blk->buf);
 573                 g_free(blk);
 574             }
 575
 576             bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, sector, nr_sectors);
 577             break;
 578         }
 579         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 580         bmds->cur_dirty = sector;
 581     }
 582
 583     return (bmds->cur_dirty >= bmds->total_sectors);
 584
 585 error:
 586     DPRINTF("Error reading sector %" PRId64 "\n", sector);
 587     g_free(blk->buf);
 588     g_free(blk);
 589     return ret;
 590 }
 591
 592 /* Called with iothread lock taken.
 593  *
 594  * return value:
 595  * 0: too much data for max_downtime
 596  * 1: few enough data for max_downtime
 597 */
 598 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 599 {
 600     BlkMigDevState *bmds;
 601     int ret = 1;
 602
 603     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 604         aio_context_acquire(blk_get_aio_context(bmds->blk));
 605         ret = mig_save_device_dirty(f, bmds, is_async);
 606         aio_context_release(blk_get_aio_context(bmds->blk));
 607         if (ret <= 0) {
 608             break;
 609         }
 610     }
 611
 612     return ret;
 613 }
 614
 615 /* Called with no locks taken.  */
 616
 617 static int flush_blks(QEMUFile *f)
 618 {
 619     BlkMigBlock *blk;
 620     int ret = 0;
 621
 622     DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
 623             __FUNCTION__, block_mig_state.submitted, block_mig_state.read_done,
 624             block_mig_state.transferred);
 625
 626     blk_mig_lock();
 627     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 628         if (qemu_file_rate_limit(f)) {
 629             break;
 630         }
 631         if (blk->ret < 0) {
 632             ret = blk->ret;
 633             break;
 634         }
 635
 636         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 637         blk_mig_unlock();
 638         blk_send(f, blk);
 639         blk_mig_lock();
 640
 641         g_free(blk->buf);
 642         g_free(blk);
 643
 644         block_mig_state.read_done--;
 645         block_mig_state.transferred++;
 646         assert(block_mig_state.read_done >= 0);
 647     }
 648     blk_mig_unlock();
 649
 650     DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__,
 651             block_mig_state.submitted, block_mig_state.read_done,
 652             block_mig_state.transferred);
 653     return ret;
 654 }
 655
 656 /* Called with iothread lock taken.  */
 657
 658 static int64_t get_remaining_dirty(void)
 659 {
 660     BlkMigDevState *bmds;
 661     int64_t dirty = 0;
 662
 663     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 664         aio_context_acquire(blk_get_aio_context(bmds->blk));
 665         dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
 666         aio_context_release(blk_get_aio_context(bmds->blk));
 667     }
 668
 669     return dirty << BDRV_SECTOR_BITS;
 670 }
 671
 672 /* Called with iothread lock taken.  */
 673
 674 static void block_migration_cleanup(void *opaque)
 675 {
 676     BlkMigDevState *bmds;
 677     BlkMigBlock *blk;
 678     AioContext *ctx;
 679
 680     bdrv_drain_all();
 681
 682     unset_dirty_tracking();
 683
 684     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 685         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 686         bdrv_op_unblock_all(blk_bs(bmds->blk), bmds->blocker);
 687         error_free(bmds->blocker);
 688
 689         /* Save ctx, because bmds->blk can disappear during blk_unref.  */
 690         ctx = blk_get_aio_context(bmds->blk);
 691         aio_context_acquire(ctx);
 692         blk_unref(bmds->blk);
 693         aio_context_release(ctx);
 694
 695         g_free(bmds->blk_name);
 696         g_free(bmds->aio_bitmap);
 697         g_free(bmds);
 698     }
 699
 700     blk_mig_lock();
 701     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 702         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 703         g_free(blk->buf);
 704         g_free(blk);
 705     }
 706     blk_mig_unlock();
 707 }
 708
 709 static int block_save_setup(QEMUFile *f, void *opaque)
 710 {
 711     int ret;
 712
 713     DPRINTF("Enter save live setup submitted %d transferred %d\n",
 714             block_mig_state.submitted, block_mig_state.transferred);
 715
 716     qemu_mutex_lock_iothread();
 717     ret = init_blk_migration(f);
 718     if (ret < 0) {
 719         qemu_mutex_unlock_iothread();
 720         return ret;
 721     }
 722
 723     /* start track dirty blocks */
 724     ret = set_dirty_tracking();
 725
 726     qemu_mutex_unlock_iothread();
 727
 728     if (ret) {
 729         return ret;
 730     }
 731
 732     ret = flush_blks(f);
 733     blk_mig_reset_dirty_cursor();
 734     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 735
 736     return ret;
 737 }
 738
 739 static int block_save_iterate(QEMUFile *f, void *opaque)
 740 {
 741     int ret;
 742     int64_t last_ftell = qemu_ftell(f);
 743     int64_t delta_ftell;
 744
 745     DPRINTF("Enter save live iterate submitted %d transferred %d\n",
 746             block_mig_state.submitted, block_mig_state.transferred);
 747
 748     ret = flush_blks(f);
 749     if (ret) {
 750         return ret;
 751     }
 752
 753     blk_mig_reset_dirty_cursor();
 754
 755     /* control the rate of transfer */
 756     blk_mig_lock();
 757     while ((block_mig_state.submitted +
 758             block_mig_state.read_done) * BLOCK_SIZE <
 759            qemu_file_get_rate_limit(f) &&
 760            (block_mig_state.submitted +
 761             block_mig_state.read_done) <
 762            MAX_INFLIGHT_IO) {
 763         blk_mig_unlock();
 764         if (block_mig_state.bulk_completed == 0) {
 765             /* first finish the bulk phase */
 766             if (blk_mig_save_bulked_block(f) == 0) {
 767                 /* finished saving bulk on all devices */
 768                 block_mig_state.bulk_completed = 1;
 769             }
 770             ret = 0;
 771         } else {
 772             /* Always called with iothread lock taken for
 773              * simplicity, block_save_complete also calls it.
 774              */
 775             qemu_mutex_lock_iothread();
 776             ret = blk_mig_save_dirty_block(f, 1);
 777             qemu_mutex_unlock_iothread();
 778         }
 779         if (ret < 0) {
 780             return ret;
 781         }
 782         blk_mig_lock();
 783         if (ret != 0) {
 784             /* no more dirty blocks */
 785             break;
 786         }
 787     }
 788     blk_mig_unlock();
 789
 790     ret = flush_blks(f);
 791     if (ret) {
 792         return ret;
 793     }
 794
 795     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 796     delta_ftell = qemu_ftell(f) - last_ftell;
 797     if (delta_ftell > 0) {
 798         return 1;
 799     } else if (delta_ftell < 0) {
 800         return -1;
 801     } else {
 802         return 0;
 803     }
 804 }
 805
 806 /* Called with iothread lock taken.  */
 807
 808 static int block_save_complete(QEMUFile *f, void *opaque)
 809 {
 810     int ret;
 811
 812     DPRINTF("Enter save live complete submitted %d transferred %d\n",
 813             block_mig_state.submitted, block_mig_state.transferred);
 814
 815     ret = flush_blks(f);
 816     if (ret) {
 817         return ret;
 818     }
 819
 820     blk_mig_reset_dirty_cursor();
 821
 822     /* we know for sure that save bulk is completed and
 823        all async read completed */
 824     blk_mig_lock();
 825     assert(block_mig_state.submitted == 0);
 826     blk_mig_unlock();
 827
 828     do {
 829         ret = blk_mig_save_dirty_block(f, 0);
 830         if (ret < 0) {
 831             return ret;
 832         }
 833     } while (ret == 0);
 834
 835     /* report completion */
 836     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 837
 838     DPRINTF("Block migration completed\n");
 839
 840     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 841
 842     return 0;
 843 }
 844
 845 static void block_save_pending(QEMUFile *f, void *opaque, uint64_t max_size,
 846                                uint64_t *non_postcopiable_pending,
 847                                uint64_t *postcopiable_pending)
 848 {
 849     /* Estimate pending number of bytes to send */
 850     uint64_t pending;
 851
 852     qemu_mutex_lock_iothread();
 853     pending = get_remaining_dirty();
 854     qemu_mutex_unlock_iothread();
 855
 856     blk_mig_lock();
 857     pending += block_mig_state.submitted * BLOCK_SIZE +
 858                block_mig_state.read_done * BLOCK_SIZE;
 859     blk_mig_unlock();
 860
 861     /* Report at least one block pending during bulk phase */
 862     if (pending <= max_size && !block_mig_state.bulk_completed) {
 863         pending = max_size + BLOCK_SIZE;
 864     }
 865
 866     DPRINTF("Enter save live pending  %" PRIu64 "\n", pending);
 867     /* We don't do postcopy */
 868     *non_postcopiable_pending += pending;
 869 }
 870
 871 static int block_load(QEMUFile *f, void *opaque, int version_id)
 872 {
 873     static int banner_printed;
 874     int len, flags;
 875     char device_name[256];
 876     int64_t addr;
 877     BlockBackend *blk, *blk_prev = NULL;;
 878     Error *local_err = NULL;
 879     uint8_t *buf;
 880     int64_t total_sectors = 0;
 881     int nr_sectors;
 882     int ret;
 883
 884     do {
 885         addr = qemu_get_be64(f);
 886
 887         flags = addr & ~BDRV_SECTOR_MASK;
 888         addr >>= BDRV_SECTOR_BITS;
 889
 890         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 891             /* get device name */
 892             len = qemu_get_byte(f);
 893             qemu_get_buffer(f, (uint8_t *)device_name, len);
 894             device_name[len] = '\0';
 895
 896             blk = blk_by_name(device_name);
 897             if (!blk) {
 898                 fprintf(stderr, "Error unknown block device %s\n",
 899                         device_name);
 900                 return -EINVAL;
 901             }
 902
 903             if (blk != blk_prev) {
 904                 blk_prev = blk;
 905                 total_sectors = blk_nb_sectors(blk);
 906                 if (total_sectors <= 0) {
 907                     error_report("Error getting length of block device %s",
 908                                  device_name);
 909                     return -EINVAL;
 910                 }
 911
 912                 blk_invalidate_cache(blk, &local_err);
 913                 if (local_err) {
 914                     error_report_err(local_err);
 915                     return -EINVAL;
 916                 }
 917             }
 918
 919             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 920                 nr_sectors = total_sectors - addr;
 921             } else {
 922                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 923             }
 924
 925             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 926                 ret = blk_pwrite_zeroes(blk, addr * BDRV_SECTOR_SIZE,
 927                                         nr_sectors * BDRV_SECTOR_SIZE,
 928                                         BDRV_REQ_MAY_UNMAP);
 929             } else {
 930                 buf = g_malloc(BLOCK_SIZE);
 931                 qemu_get_buffer(f, buf, BLOCK_SIZE);
 932                 ret = blk_pwrite(blk, addr * BDRV_SECTOR_SIZE, buf,
 933                                  nr_sectors * BDRV_SECTOR_SIZE, 0);
 934                 g_free(buf);
 935             }
 936
 937             if (ret < 0) {
 938                 return ret;
 939             }
 940         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 941             if (!banner_printed) {
 942                 printf("Receiving block device images\n");
 943                 banner_printed = 1;
 944             }
 945             printf("Completed %d %%%c", (int)addr,
 946                    (addr == 100) ? '\n' : '\r');
 947             fflush(stdout);
 948         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
 949             fprintf(stderr, "Unknown block migration flags: %#x\n", flags);
 950             return -EINVAL;
 951         }
 952         ret = qemu_file_get_error(f);
 953         if (ret != 0) {
 954             return ret;
 955         }
 956     } while (!(flags & BLK_MIG_FLAG_EOS));
 957
 958     return 0;
 959 }
 960
 961 static void block_set_params(const MigrationParams *params, void *opaque)
 962 {
 963     block_mig_state.blk_enable = params->blk;
 964     block_mig_state.shared_base = params->shared;
 965
 966     /* shared base means that blk_enable = 1 */
 967     block_mig_state.blk_enable |= params->shared;
 968 }
 969
 970 static bool block_is_active(void *opaque)
 971 {
 972     return block_mig_state.blk_enable == 1;
 973 }
 974
 975 static SaveVMHandlers savevm_block_handlers = {
 976     .set_params = block_set_params,
 977     .save_live_setup = block_save_setup,
 978     .save_live_iterate = block_save_iterate,
 979     .save_live_complete_precopy = block_save_complete,
 980     .save_live_pending = block_save_pending,
 981     .load_state = block_load,
 982     .cleanup = block_migration_cleanup,
 983     .is_active = block_is_active,
 984 };
 985
 986 void blk_mig_init(void)
 987 {
 988     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
 989     QSIMPLEQ_INIT(&block_mig_state.blk_list);
 990     qemu_mutex_init(&block_mig_state.lock);
 991
 992     register_savevm_live(NULL, "block", 0, 1, &savevm_block_handlers,
 993                          &block_mig_state);
 994 }