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         /* Skip unallocated sectors; intentionally treats failure as
 280          * an allocated sector */
 281         while (cur_sector < total_sectors &&
 282                !bdrv_is_allocated(blk_bs(bb), cur_sector,
 283                                   MAX_IS_ALLOCATED_SEARCH, &nr_sectors)) {
 284             cur_sector += nr_sectors;
 285         }
 286         aio_context_release(blk_get_aio_context(bb));
 287         qemu_mutex_unlock_iothread();
 288     }
 289
 290     if (cur_sector >= total_sectors) {
 291         bmds->cur_sector = bmds->completed_sectors = total_sectors;
 292         return 1;
 293     }
 294
 295     bmds->completed_sectors = cur_sector;
 296
 297     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
 298
 299     /* we are going to transfer a full block even if it is not allocated */
 300     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 301
 302     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 303         nr_sectors = total_sectors - cur_sector;
 304     }
 305
 306     blk = g_new(BlkMigBlock, 1);
 307     blk->buf = g_malloc(BLOCK_SIZE);
 308     blk->bmds = bmds;
 309     blk->sector = cur_sector;
 310     blk->nr_sectors = nr_sectors;
 311
 312     blk->iov.iov_base = blk->buf;
 313     blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 314     qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 315
 316     blk_mig_lock();
 317     block_mig_state.submitted++;
 318     blk_mig_unlock();
 319
 320     /* We do not know if bs is under the main thread (and thus does
 321      * not acquire the AioContext when doing AIO) or rather under
 322      * dataplane.  Thus acquire both the iothread mutex and the
 323      * AioContext.
 324      *
 325      * This is ugly and will disappear when we make bdrv_* thread-safe,
 326      * without the need to acquire the AioContext.
 327      */
 328     qemu_mutex_lock_iothread();
 329     aio_context_acquire(blk_get_aio_context(bmds->blk));
 330     blk->aiocb = blk_aio_preadv(bb, cur_sector * BDRV_SECTOR_SIZE, &blk->qiov,
 331                                 0, blk_mig_read_cb, blk);
 332
 333     bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, cur_sector, nr_sectors);
 334     aio_context_release(blk_get_aio_context(bmds->blk));
 335     qemu_mutex_unlock_iothread();
 336
 337     bmds->cur_sector = cur_sector + nr_sectors;
 338     return (bmds->cur_sector >= total_sectors);
 339 }
 340
 341 /* Called with iothread lock taken.  */
 342
 343 static int set_dirty_tracking(void)
 344 {
 345     BlkMigDevState *bmds;
 346     int ret;
 347
 348     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 349         aio_context_acquire(blk_get_aio_context(bmds->blk));
 350         bmds->dirty_bitmap = bdrv_create_dirty_bitmap(blk_bs(bmds->blk),
 351                                                       BLOCK_SIZE, NULL, NULL);
 352         aio_context_release(blk_get_aio_context(bmds->blk));
 353         if (!bmds->dirty_bitmap) {
 354             ret = -errno;
 355             goto fail;
 356         }
 357     }
 358     return 0;
 359
 360 fail:
 361     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 362         if (bmds->dirty_bitmap) {
 363             aio_context_acquire(blk_get_aio_context(bmds->blk));
 364             bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
 365             aio_context_release(blk_get_aio_context(bmds->blk));
 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         aio_context_acquire(blk_get_aio_context(bmds->blk));
 379         bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
 380         aio_context_release(blk_get_aio_context(bmds->blk));
 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             goto out;
 420         }
 421
 422         bmds = g_new0(BlkMigDevState, 1);
 423         bmds->blk = blk_new(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 = block_mig_state.shared_base;
 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     BlockDriverState *bs = blk_bs(bmds->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         if (bdrv_get_dirty(bs, bmds->dirty_bitmap, sector)) {
 540
 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             blk = g_new(BlkMigBlock, 1);
 547             blk->buf = g_malloc(BLOCK_SIZE);
 548             blk->bmds = bmds;
 549             blk->sector = sector;
 550             blk->nr_sectors = nr_sectors;
 551
 552             if (is_async) {
 553                 blk->iov.iov_base = blk->buf;
 554                 blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 555                 qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 556
 557                 blk->aiocb = blk_aio_preadv(bmds->blk,
 558                                             sector * BDRV_SECTOR_SIZE,
 559                                             &blk->qiov, 0, blk_mig_read_cb,
 560                                             blk);
 561
 562                 blk_mig_lock();
 563                 block_mig_state.submitted++;
 564                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 565                 blk_mig_unlock();
 566             } else {
 567                 ret = blk_pread(bmds->blk, sector * BDRV_SECTOR_SIZE, blk->buf,
 568                                 nr_sectors * BDRV_SECTOR_SIZE);
 569                 if (ret < 0) {
 570                     goto error;
 571                 }
 572                 blk_send(f, blk);
 573
 574                 g_free(blk->buf);
 575                 g_free(blk);
 576             }
 577
 578             bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, sector, nr_sectors);
 579             sector += nr_sectors;
 580             bmds->cur_dirty = sector;
 581
 582             break;
 583         }
 584         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 585         bmds->cur_dirty = sector;
 586     }
 587
 588     return (bmds->cur_dirty >= bmds->total_sectors);
 589
 590 error:
 591     DPRINTF("Error reading sector %" PRId64 "\n", sector);
 592     g_free(blk->buf);
 593     g_free(blk);
 594     return ret;
 595 }
 596
 597 /* Called with iothread lock taken.
 598  *
 599  * return value:
 600  * 0: too much data for max_downtime
 601  * 1: few enough data for max_downtime
 602 */
 603 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 604 {
 605     BlkMigDevState *bmds;
 606     int ret = 1;
 607
 608     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 609         aio_context_acquire(blk_get_aio_context(bmds->blk));
 610         ret = mig_save_device_dirty(f, bmds, is_async);
 611         aio_context_release(blk_get_aio_context(bmds->blk));
 612         if (ret <= 0) {
 613             break;
 614         }
 615     }
 616
 617     return ret;
 618 }
 619
 620 /* Called with no locks taken.  */
 621
 622 static int flush_blks(QEMUFile *f)
 623 {
 624     BlkMigBlock *blk;
 625     int ret = 0;
 626
 627     DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
 628             __FUNCTION__, block_mig_state.submitted, block_mig_state.read_done,
 629             block_mig_state.transferred);
 630
 631     blk_mig_lock();
 632     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 633         if (qemu_file_rate_limit(f)) {
 634             break;
 635         }
 636         if (blk->ret < 0) {
 637             ret = blk->ret;
 638             break;
 639         }
 640
 641         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 642         blk_mig_unlock();
 643         blk_send(f, blk);
 644         blk_mig_lock();
 645
 646         g_free(blk->buf);
 647         g_free(blk);
 648
 649         block_mig_state.read_done--;
 650         block_mig_state.transferred++;
 651         assert(block_mig_state.read_done >= 0);
 652     }
 653     blk_mig_unlock();
 654
 655     DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__,
 656             block_mig_state.submitted, block_mig_state.read_done,
 657             block_mig_state.transferred);
 658     return ret;
 659 }
 660
 661 /* Called with iothread lock taken.  */
 662
 663 static int64_t get_remaining_dirty(void)
 664 {
 665     BlkMigDevState *bmds;
 666     int64_t dirty = 0;
 667
 668     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 669         aio_context_acquire(blk_get_aio_context(bmds->blk));
 670         dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
 671         aio_context_release(blk_get_aio_context(bmds->blk));
 672     }
 673
 674     return dirty << BDRV_SECTOR_BITS;
 675 }
 676
 677 /* Called with iothread lock taken.  */
 678
 679 static void block_migration_cleanup(void *opaque)
 680 {
 681     BlkMigDevState *bmds;
 682     BlkMigBlock *blk;
 683     AioContext *ctx;
 684
 685     bdrv_drain_all();
 686
 687     unset_dirty_tracking();
 688
 689     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 690         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 691         bdrv_op_unblock_all(blk_bs(bmds->blk), bmds->blocker);
 692         error_free(bmds->blocker);
 693
 694         /* Save ctx, because bmds->blk can disappear during blk_unref.  */
 695         ctx = blk_get_aio_context(bmds->blk);
 696         aio_context_acquire(ctx);
 697         blk_unref(bmds->blk);
 698         aio_context_release(ctx);
 699
 700         g_free(bmds->blk_name);
 701         g_free(bmds->aio_bitmap);
 702         g_free(bmds);
 703     }
 704
 705     blk_mig_lock();
 706     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 707         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 708         g_free(blk->buf);
 709         g_free(blk);
 710     }
 711     blk_mig_unlock();
 712 }
 713
 714 static int block_save_setup(QEMUFile *f, void *opaque)
 715 {
 716     int ret;
 717
 718     DPRINTF("Enter save live setup submitted %d transferred %d\n",
 719             block_mig_state.submitted, block_mig_state.transferred);
 720
 721     qemu_mutex_lock_iothread();
 722     ret = init_blk_migration(f);
 723     if (ret < 0) {
 724         qemu_mutex_unlock_iothread();
 725         return ret;
 726     }
 727
 728     /* start track dirty blocks */
 729     ret = set_dirty_tracking();
 730
 731     qemu_mutex_unlock_iothread();
 732
 733     if (ret) {
 734         return ret;
 735     }
 736
 737     ret = flush_blks(f);
 738     blk_mig_reset_dirty_cursor();
 739     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 740
 741     return ret;
 742 }
 743
 744 static int block_save_iterate(QEMUFile *f, void *opaque)
 745 {
 746     int ret;
 747     int64_t last_ftell = qemu_ftell(f);
 748     int64_t delta_ftell;
 749
 750     DPRINTF("Enter save live iterate submitted %d transferred %d\n",
 751             block_mig_state.submitted, block_mig_state.transferred);
 752
 753     ret = flush_blks(f);
 754     if (ret) {
 755         return ret;
 756     }
 757
 758     blk_mig_reset_dirty_cursor();
 759
 760     /* control the rate of transfer */
 761     blk_mig_lock();
 762     while ((block_mig_state.submitted +
 763             block_mig_state.read_done) * BLOCK_SIZE <
 764            qemu_file_get_rate_limit(f) &&
 765            (block_mig_state.submitted +
 766             block_mig_state.read_done) <
 767            MAX_INFLIGHT_IO) {
 768         blk_mig_unlock();
 769         if (block_mig_state.bulk_completed == 0) {
 770             /* first finish the bulk phase */
 771             if (blk_mig_save_bulked_block(f) == 0) {
 772                 /* finished saving bulk on all devices */
 773                 block_mig_state.bulk_completed = 1;
 774             }
 775             ret = 0;
 776         } else {
 777             /* Always called with iothread lock taken for
 778              * simplicity, block_save_complete also calls it.
 779              */
 780             qemu_mutex_lock_iothread();
 781             ret = blk_mig_save_dirty_block(f, 1);
 782             qemu_mutex_unlock_iothread();
 783         }
 784         if (ret < 0) {
 785             return ret;
 786         }
 787         blk_mig_lock();
 788         if (ret != 0) {
 789             /* no more dirty blocks */
 790             break;
 791         }
 792     }
 793     blk_mig_unlock();
 794
 795     ret = flush_blks(f);
 796     if (ret) {
 797         return ret;
 798     }
 799
 800     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 801     delta_ftell = qemu_ftell(f) - last_ftell;
 802     if (delta_ftell > 0) {
 803         return 1;
 804     } else if (delta_ftell < 0) {
 805         return -1;
 806     } else {
 807         return 0;
 808     }
 809 }
 810
 811 /* Called with iothread lock taken.  */
 812
 813 static int block_save_complete(QEMUFile *f, void *opaque)
 814 {
 815     int ret;
 816
 817     DPRINTF("Enter save live complete submitted %d transferred %d\n",
 818             block_mig_state.submitted, block_mig_state.transferred);
 819
 820     ret = flush_blks(f);
 821     if (ret) {
 822         return ret;
 823     }
 824
 825     blk_mig_reset_dirty_cursor();
 826
 827     /* we know for sure that save bulk is completed and
 828        all async read completed */
 829     blk_mig_lock();
 830     assert(block_mig_state.submitted == 0);
 831     blk_mig_unlock();
 832
 833     do {
 834         ret = blk_mig_save_dirty_block(f, 0);
 835         if (ret < 0) {
 836             return ret;
 837         }
 838     } while (ret == 0);
 839
 840     /* report completion */
 841     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 842
 843     DPRINTF("Block migration completed\n");
 844
 845     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 846
 847     return 0;
 848 }
 849
 850 static void block_save_pending(QEMUFile *f, void *opaque, uint64_t max_size,
 851                                uint64_t *non_postcopiable_pending,
 852                                uint64_t *postcopiable_pending)
 853 {
 854     /* Estimate pending number of bytes to send */
 855     uint64_t pending;
 856
 857     qemu_mutex_lock_iothread();
 858     pending = get_remaining_dirty();
 859     qemu_mutex_unlock_iothread();
 860
 861     blk_mig_lock();
 862     pending += block_mig_state.submitted * BLOCK_SIZE +
 863                block_mig_state.read_done * BLOCK_SIZE;
 864     blk_mig_unlock();
 865
 866     /* Report at least one block pending during bulk phase */
 867     if (pending <= max_size && !block_mig_state.bulk_completed) {
 868         pending = max_size + BLOCK_SIZE;
 869     }
 870
 871     DPRINTF("Enter save live pending  %" PRIu64 "\n", pending);
 872     /* We don't do postcopy */
 873     *non_postcopiable_pending += pending;
 874 }
 875
 876 static int block_load(QEMUFile *f, void *opaque, int version_id)
 877 {
 878     static int banner_printed;
 879     int len, flags;
 880     char device_name[256];
 881     int64_t addr;
 882     BlockBackend *blk, *blk_prev = NULL;;
 883     Error *local_err = NULL;
 884     uint8_t *buf;
 885     int64_t total_sectors = 0;
 886     int nr_sectors;
 887     int ret;
 888     BlockDriverInfo bdi;
 889     int cluster_size = BLOCK_SIZE;
 890
 891     do {
 892         addr = qemu_get_be64(f);
 893
 894         flags = addr & ~BDRV_SECTOR_MASK;
 895         addr >>= BDRV_SECTOR_BITS;
 896
 897         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 898             /* get device name */
 899             len = qemu_get_byte(f);
 900             qemu_get_buffer(f, (uint8_t *)device_name, len);
 901             device_name[len] = '\0';
 902
 903             blk = blk_by_name(device_name);
 904             if (!blk) {
 905                 fprintf(stderr, "Error unknown block device %s\n",
 906                         device_name);
 907                 return -EINVAL;
 908             }
 909
 910             if (blk != blk_prev) {
 911                 blk_prev = blk;
 912                 total_sectors = blk_nb_sectors(blk);
 913                 if (total_sectors <= 0) {
 914                     error_report("Error getting length of block device %s",
 915                                  device_name);
 916                     return -EINVAL;
 917                 }
 918
 919                 blk_invalidate_cache(blk, &local_err);
 920                 if (local_err) {
 921                     error_report_err(local_err);
 922                     return -EINVAL;
 923                 }
 924
 925                 ret = bdrv_get_info(blk_bs(blk), &bdi);
 926                 if (ret == 0 && bdi.cluster_size > 0 &&
 927                     bdi.cluster_size <= BLOCK_SIZE &&
 928                     BLOCK_SIZE % bdi.cluster_size == 0) {
 929                     cluster_size = bdi.cluster_size;
 930                 } else {
 931                     cluster_size = BLOCK_SIZE;
 932                 }
 933             }
 934
 935             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 936                 nr_sectors = total_sectors - addr;
 937             } else {
 938                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 939             }
 940
 941             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 942                 ret = blk_pwrite_zeroes(blk, addr * BDRV_SECTOR_SIZE,
 943                                         nr_sectors * BDRV_SECTOR_SIZE,
 944                                         BDRV_REQ_MAY_UNMAP);
 945             } else {
 946                 int i;
 947                 int64_t cur_addr;
 948                 uint8_t *cur_buf;
 949
 950                 buf = g_malloc(BLOCK_SIZE);
 951                 qemu_get_buffer(f, buf, BLOCK_SIZE);
 952                 for (i = 0; i < BLOCK_SIZE / cluster_size; i++) {
 953                     cur_addr = addr * BDRV_SECTOR_SIZE + i * cluster_size;
 954                     cur_buf = buf + i * cluster_size;
 955
 956                     if ((!block_mig_state.zero_blocks ||
 957                         cluster_size < BLOCK_SIZE) &&
 958                         buffer_is_zero(cur_buf, cluster_size)) {
 959                         ret = blk_pwrite_zeroes(blk, cur_addr,
 960                                                 cluster_size,
 961                                                 BDRV_REQ_MAY_UNMAP);
 962                     } else {
 963                         ret = blk_pwrite(blk, cur_addr, cur_buf,
 964                                          cluster_size, 0);
 965                     }
 966                     if (ret < 0) {
 967                         break;
 968                     }
 969                 }
 970                 g_free(buf);
 971             }
 972
 973             if (ret < 0) {
 974                 return ret;
 975             }
 976         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 977             if (!banner_printed) {
 978                 printf("Receiving block device images\n");
 979                 banner_printed = 1;
 980             }
 981             printf("Completed %d %%%c", (int)addr,
 982                    (addr == 100) ? '\n' : '\r');
 983             fflush(stdout);
 984         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
 985             fprintf(stderr, "Unknown block migration flags: %#x\n", flags);
 986             return -EINVAL;
 987         }
 988         ret = qemu_file_get_error(f);
 989         if (ret != 0) {
 990             return ret;
 991         }
 992     } while (!(flags & BLK_MIG_FLAG_EOS));
 993
 994     return 0;
 995 }
 996
 997 static void block_set_params(const MigrationParams *params, void *opaque)
 998 {
 999     block_mig_state.blk_enable = params->blk;
1000     block_mig_state.shared_base = params->shared;
1001
1002     /* shared base means that blk_enable = 1 */
1003     block_mig_state.blk_enable |= params->shared;
1004 }
1005
1006 static bool block_is_active(void *opaque)
1007 {
1008     return block_mig_state.blk_enable == 1;
1009 }
1010
1011 static SaveVMHandlers savevm_block_handlers = {
1012     .set_params = block_set_params,
1013     .save_live_setup = block_save_setup,
1014     .save_live_iterate = block_save_iterate,
1015     .save_live_complete_precopy = block_save_complete,
1016     .save_live_pending = block_save_pending,
1017     .load_state = block_load,
1018     .cleanup = block_migration_cleanup,
1019     .is_active = block_is_active,
1020 };
1021
1022 void blk_mig_init(void)
1023 {
1024     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
1025     QSIMPLEQ_INIT(&block_mig_state.blk_list);
1026     qemu_mutex_init(&block_mig_state.lock);
1027
1028     register_savevm_live(NULL, "block", 0, 1, &savevm_block_handlers,
1029                          &block_mig_state);
1030 }