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/cutils.h"
  20 #include "qemu/queue.h"
  21 #include "block.h"
  22 #include "migration/misc.h"
  23 #include "migration.h"
  24 #include "migration/register.h"
  25 #include "qemu-file.h"
  26 #include "migration/vmstate.h"
  27 #include "sysemu/block-backend.h"
  28
  29 #define BLOCK_SIZE                       (1 << 20)
  30 #define BDRV_SECTORS_PER_DIRTY_CHUNK     (BLOCK_SIZE >> BDRV_SECTOR_BITS)
  31
  32 #define BLK_MIG_FLAG_DEVICE_BLOCK       0x01
  33 #define BLK_MIG_FLAG_EOS                0x02
  34 #define BLK_MIG_FLAG_PROGRESS           0x04
  35 #define BLK_MIG_FLAG_ZERO_BLOCK         0x08
  36
  37 #define MAX_IS_ALLOCATED_SEARCH (65536 * BDRV_SECTOR_SIZE)
  38
  39 #define MAX_INFLIGHT_IO 512
  40
  41 //#define DEBUG_BLK_MIGRATION
  42
  43 #ifdef DEBUG_BLK_MIGRATION
  44 #define DPRINTF(fmt, ...) \
  45     do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
  46 #else
  47 #define DPRINTF(fmt, ...) \
  48     do { } while (0)
  49 #endif
  50
  51 typedef struct BlkMigDevState {
  52     /* Written during setup phase.  Can be read without a lock.  */
  53     BlockBackend *blk;
  54     char *blk_name;
  55     int shared_base;
  56     int64_t total_sectors;
  57     QSIMPLEQ_ENTRY(BlkMigDevState) entry;
  58     Error *blocker;
  59
  60     /* Only used by migration thread.  Does not need a lock.  */
  61     int bulk_completed;
  62     int64_t cur_sector;
  63     int64_t cur_dirty;
  64
  65     /* Data in the aio_bitmap is protected by block migration lock.
  66      * Allocation and free happen during setup and cleanup respectively.
  67      */
  68     unsigned long *aio_bitmap;
  69
  70     /* Protected by block migration lock.  */
  71     int64_t completed_sectors;
  72
  73     /* During migration this is protected by iothread lock / AioContext.
  74      * Allocation and free happen during setup and cleanup respectively.
  75      */
  76     BdrvDirtyBitmap *dirty_bitmap;
  77 } BlkMigDevState;
  78
  79 typedef struct BlkMigBlock {
  80     /* Only used by migration thread.  */
  81     uint8_t *buf;
  82     BlkMigDevState *bmds;
  83     int64_t sector;
  84     int nr_sectors;
  85     struct iovec iov;
  86     QEMUIOVector qiov;
  87     BlockAIOCB *aiocb;
  88
  89     /* Protected by block migration lock.  */
  90     int ret;
  91     QSIMPLEQ_ENTRY(BlkMigBlock) entry;
  92 } BlkMigBlock;
  93
  94 typedef struct BlkMigState {
  95     QSIMPLEQ_HEAD(bmds_list, BlkMigDevState) bmds_list;
  96     int64_t total_sector_sum;
  97     bool zero_blocks;
  98
  99     /* Protected by lock.  */
 100     QSIMPLEQ_HEAD(blk_list, BlkMigBlock) blk_list;
 101     int submitted;
 102     int read_done;
 103
 104     /* Only used by migration thread.  Does not need a lock.  */
 105     int transferred;
 106     int prev_progress;
 107     int bulk_completed;
 108
 109     /* Lock must be taken _inside_ the iothread lock and any AioContexts.  */
 110     QemuMutex lock;
 111 } BlkMigState;
 112
 113 static BlkMigState block_mig_state;
 114
 115 static void blk_mig_lock(void)
 116 {
 117     qemu_mutex_lock(&block_mig_state.lock);
 118 }
 119
 120 static void blk_mig_unlock(void)
 121 {
 122     qemu_mutex_unlock(&block_mig_state.lock);
 123 }
 124
 125 /* Must run outside of the iothread lock during the bulk phase,
 126  * or the VM will stall.
 127  */
 128
 129 static void blk_send(QEMUFile *f, BlkMigBlock * blk)
 130 {
 131     int len;
 132     uint64_t flags = BLK_MIG_FLAG_DEVICE_BLOCK;
 133
 134     if (block_mig_state.zero_blocks &&
 135         buffer_is_zero(blk->buf, BLOCK_SIZE)) {
 136         flags |= BLK_MIG_FLAG_ZERO_BLOCK;
 137     }
 138
 139     /* sector number and flags */
 140     qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
 141                      | flags);
 142
 143     /* device name */
 144     len = strlen(blk->bmds->blk_name);
 145     qemu_put_byte(f, len);
 146     qemu_put_buffer(f, (uint8_t *) blk->bmds->blk_name, len);
 147
 148     /* if a block is zero we need to flush here since the network
 149      * bandwidth is now a lot higher than the storage device bandwidth.
 150      * thus if we queue zero blocks we slow down the migration */
 151     if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 152         qemu_fflush(f);
 153         return;
 154     }
 155
 156     qemu_put_buffer(f, blk->buf, BLOCK_SIZE);
 157 }
 158
 159 int blk_mig_active(void)
 160 {
 161     return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
 162 }
 163
 164 uint64_t blk_mig_bytes_transferred(void)
 165 {
 166     BlkMigDevState *bmds;
 167     uint64_t sum = 0;
 168
 169     blk_mig_lock();
 170     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 171         sum += bmds->completed_sectors;
 172     }
 173     blk_mig_unlock();
 174     return sum << BDRV_SECTOR_BITS;
 175 }
 176
 177 uint64_t blk_mig_bytes_remaining(void)
 178 {
 179     return blk_mig_bytes_total() - blk_mig_bytes_transferred();
 180 }
 181
 182 uint64_t blk_mig_bytes_total(void)
 183 {
 184     BlkMigDevState *bmds;
 185     uint64_t sum = 0;
 186
 187     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 188         sum += bmds->total_sectors;
 189     }
 190     return sum << BDRV_SECTOR_BITS;
 191 }
 192
 193
 194 /* Called with migration lock held.  */
 195
 196 static int bmds_aio_inflight(BlkMigDevState *bmds, int64_t sector)
 197 {
 198     int64_t chunk = sector / (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK;
 199
 200     if (sector < blk_nb_sectors(bmds->blk)) {
 201         return !!(bmds->aio_bitmap[chunk / (sizeof(unsigned long) * 8)] &
 202             (1UL << (chunk % (sizeof(unsigned long) * 8))));
 203     } else {
 204         return 0;
 205     }
 206 }
 207
 208 /* Called with migration lock held.  */
 209
 210 static void bmds_set_aio_inflight(BlkMigDevState *bmds, int64_t sector_num,
 211                              int nb_sectors, int set)
 212 {
 213     int64_t start, end;
 214     unsigned long val, idx, bit;
 215
 216     start = sector_num / BDRV_SECTORS_PER_DIRTY_CHUNK;
 217     end = (sector_num + nb_sectors - 1) / BDRV_SECTORS_PER_DIRTY_CHUNK;
 218
 219     for (; start <= end; start++) {
 220         idx = start / (sizeof(unsigned long) * 8);
 221         bit = start % (sizeof(unsigned long) * 8);
 222         val = bmds->aio_bitmap[idx];
 223         if (set) {
 224             val |= 1UL << bit;
 225         } else {
 226             val &= ~(1UL << bit);
 227         }
 228         bmds->aio_bitmap[idx] = val;
 229     }
 230 }
 231
 232 static void alloc_aio_bitmap(BlkMigDevState *bmds)
 233 {
 234     BlockBackend *bb = bmds->blk;
 235     int64_t bitmap_size;
 236
 237     bitmap_size = blk_nb_sectors(bb) + BDRV_SECTORS_PER_DIRTY_CHUNK * 8 - 1;
 238     bitmap_size /= BDRV_SECTORS_PER_DIRTY_CHUNK * 8;
 239
 240     bmds->aio_bitmap = g_malloc0(bitmap_size);
 241 }
 242
 243 /* Never hold migration lock when yielding to the main loop!  */
 244
 245 static void blk_mig_read_cb(void *opaque, int ret)
 246 {
 247     BlkMigBlock *blk = opaque;
 248
 249     blk_mig_lock();
 250     blk->ret = ret;
 251
 252     QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry);
 253     bmds_set_aio_inflight(blk->bmds, blk->sector, blk->nr_sectors, 0);
 254
 255     block_mig_state.submitted--;
 256     block_mig_state.read_done++;
 257     assert(block_mig_state.submitted >= 0);
 258     blk_mig_unlock();
 259 }
 260
 261 /* Called with no lock taken.  */
 262
 263 static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds)
 264 {
 265     int64_t total_sectors = bmds->total_sectors;
 266     int64_t cur_sector = bmds->cur_sector;
 267     BlockBackend *bb = bmds->blk;
 268     BlkMigBlock *blk;
 269     int nr_sectors;
 270     int64_t count;
 271
 272     if (bmds->shared_base) {
 273         qemu_mutex_lock_iothread();
 274         aio_context_acquire(blk_get_aio_context(bb));
 275         /* Skip unallocated sectors; intentionally treats failure or
 276          * partial sector as an allocated sector */
 277         while (cur_sector < total_sectors &&
 278                !bdrv_is_allocated(blk_bs(bb), cur_sector * BDRV_SECTOR_SIZE,
 279                                   MAX_IS_ALLOCATED_SEARCH, &count)) {
 280             if (count < BDRV_SECTOR_SIZE) {
 281                 break;
 282             }
 283             cur_sector += count >> BDRV_SECTOR_BITS;
 284         }
 285         aio_context_release(blk_get_aio_context(bb));
 286         qemu_mutex_unlock_iothread();
 287     }
 288
 289     if (cur_sector >= total_sectors) {
 290         bmds->cur_sector = bmds->completed_sectors = total_sectors;
 291         return 1;
 292     }
 293
 294     bmds->completed_sectors = cur_sector;
 295
 296     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
 297
 298     /* we are going to transfer a full block even if it is not allocated */
 299     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 300
 301     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 302         nr_sectors = total_sectors - cur_sector;
 303     }
 304
 305     blk = g_new(BlkMigBlock, 1);
 306     blk->buf = g_malloc(BLOCK_SIZE);
 307     blk->bmds = bmds;
 308     blk->sector = cur_sector;
 309     blk->nr_sectors = nr_sectors;
 310
 311     blk->iov.iov_base = blk->buf;
 312     blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 313     qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 314
 315     blk_mig_lock();
 316     block_mig_state.submitted++;
 317     blk_mig_unlock();
 318
 319     /* We do not know if bs is under the main thread (and thus does
 320      * not acquire the AioContext when doing AIO) or rather under
 321      * dataplane.  Thus acquire both the iothread mutex and the
 322      * AioContext.
 323      *
 324      * This is ugly and will disappear when we make bdrv_* thread-safe,
 325      * without the need to acquire the AioContext.
 326      */
 327     qemu_mutex_lock_iothread();
 328     aio_context_acquire(blk_get_aio_context(bmds->blk));
 329     blk->aiocb = blk_aio_preadv(bb, cur_sector * BDRV_SECTOR_SIZE, &blk->qiov,
 330                                 0, blk_mig_read_cb, blk);
 331
 332     bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, cur_sector, nr_sectors);
 333     aio_context_release(blk_get_aio_context(bmds->blk));
 334     qemu_mutex_unlock_iothread();
 335
 336     bmds->cur_sector = cur_sector + nr_sectors;
 337     return (bmds->cur_sector >= total_sectors);
 338 }
 339
 340 /* Called with iothread lock taken.  */
 341
 342 static int set_dirty_tracking(void)
 343 {
 344     BlkMigDevState *bmds;
 345     int ret;
 346
 347     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 348         bmds->dirty_bitmap = bdrv_create_dirty_bitmap(blk_bs(bmds->blk),
 349                                                       BLOCK_SIZE, NULL, NULL);
 350         if (!bmds->dirty_bitmap) {
 351             ret = -errno;
 352             goto fail;
 353         }
 354     }
 355     return 0;
 356
 357 fail:
 358     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 359         if (bmds->dirty_bitmap) {
 360             bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
 361         }
 362     }
 363     return ret;
 364 }
 365
 366 /* Called with iothread lock taken.  */
 367
 368 static void unset_dirty_tracking(void)
 369 {
 370     BlkMigDevState *bmds;
 371
 372     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 373         bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
 374     }
 375 }
 376
 377 static int init_blk_migration(QEMUFile *f)
 378 {
 379     BlockDriverState *bs;
 380     BlkMigDevState *bmds;
 381     int64_t sectors;
 382     BdrvNextIterator it;
 383     int i, num_bs = 0;
 384     struct {
 385         BlkMigDevState *bmds;
 386         BlockDriverState *bs;
 387     } *bmds_bs;
 388     Error *local_err = NULL;
 389     int ret;
 390
 391     block_mig_state.submitted = 0;
 392     block_mig_state.read_done = 0;
 393     block_mig_state.transferred = 0;
 394     block_mig_state.total_sector_sum = 0;
 395     block_mig_state.prev_progress = -1;
 396     block_mig_state.bulk_completed = 0;
 397     block_mig_state.zero_blocks = migrate_zero_blocks();
 398
 399     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
 400         num_bs++;
 401     }
 402     bmds_bs = g_malloc0(num_bs * sizeof(*bmds_bs));
 403
 404     for (i = 0, bs = bdrv_first(&it); bs; bs = bdrv_next(&it), i++) {
 405         if (bdrv_is_read_only(bs)) {
 406             continue;
 407         }
 408
 409         sectors = bdrv_nb_sectors(bs);
 410         if (sectors <= 0) {
 411             ret = sectors;
 412             goto out;
 413         }
 414
 415         bmds = g_new0(BlkMigDevState, 1);
 416         bmds->blk = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 417         bmds->blk_name = g_strdup(bdrv_get_device_name(bs));
 418         bmds->bulk_completed = 0;
 419         bmds->total_sectors = sectors;
 420         bmds->completed_sectors = 0;
 421         bmds->shared_base = migrate_use_block_incremental();
 422
 423         assert(i < num_bs);
 424         bmds_bs[i].bmds = bmds;
 425         bmds_bs[i].bs = bs;
 426
 427         block_mig_state.total_sector_sum += sectors;
 428
 429         if (bmds->shared_base) {
 430             DPRINTF("Start migration for %s with shared base image\n",
 431                     bdrv_get_device_name(bs));
 432         } else {
 433             DPRINTF("Start full migration for %s\n", bdrv_get_device_name(bs));
 434         }
 435
 436         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 437     }
 438
 439     /* Can only insert new BDSes now because doing so while iterating block
 440      * devices may end up in a deadlock (iterating the new BDSes, too). */
 441     for (i = 0; i < num_bs; i++) {
 442         BlkMigDevState *bmds = bmds_bs[i].bmds;
 443         BlockDriverState *bs = bmds_bs[i].bs;
 444
 445         if (bmds) {
 446             ret = blk_insert_bs(bmds->blk, bs, &local_err);
 447             if (ret < 0) {
 448                 error_report_err(local_err);
 449                 goto out;
 450             }
 451
 452             alloc_aio_bitmap(bmds);
 453             error_setg(&bmds->blocker, "block device is in use by migration");
 454             bdrv_op_block_all(bs, bmds->blocker);
 455         }
 456     }
 457
 458     ret = 0;
 459 out:
 460     g_free(bmds_bs);
 461     return ret;
 462 }
 463
 464 /* Called with no lock taken.  */
 465
 466 static int blk_mig_save_bulked_block(QEMUFile *f)
 467 {
 468     int64_t completed_sector_sum = 0;
 469     BlkMigDevState *bmds;
 470     int progress;
 471     int ret = 0;
 472
 473     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 474         if (bmds->bulk_completed == 0) {
 475             if (mig_save_device_bulk(f, bmds) == 1) {
 476                 /* completed bulk section for this device */
 477                 bmds->bulk_completed = 1;
 478             }
 479             completed_sector_sum += bmds->completed_sectors;
 480             ret = 1;
 481             break;
 482         } else {
 483             completed_sector_sum += bmds->completed_sectors;
 484         }
 485     }
 486
 487     if (block_mig_state.total_sector_sum != 0) {
 488         progress = completed_sector_sum * 100 /
 489                    block_mig_state.total_sector_sum;
 490     } else {
 491         progress = 100;
 492     }
 493     if (progress != block_mig_state.prev_progress) {
 494         block_mig_state.prev_progress = progress;
 495         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 496                          | BLK_MIG_FLAG_PROGRESS);
 497         DPRINTF("Completed %d %%\r", progress);
 498     }
 499
 500     return ret;
 501 }
 502
 503 static void blk_mig_reset_dirty_cursor(void)
 504 {
 505     BlkMigDevState *bmds;
 506
 507     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 508         bmds->cur_dirty = 0;
 509     }
 510 }
 511
 512 /* Called with iothread lock and AioContext taken.  */
 513
 514 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
 515                                  int is_async)
 516 {
 517     BlkMigBlock *blk;
 518     BlockDriverState *bs = blk_bs(bmds->blk);
 519     int64_t total_sectors = bmds->total_sectors;
 520     int64_t sector;
 521     int nr_sectors;
 522     int ret = -EIO;
 523
 524     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
 525         blk_mig_lock();
 526         if (bmds_aio_inflight(bmds, sector)) {
 527             blk_mig_unlock();
 528             blk_drain(bmds->blk);
 529         } else {
 530             blk_mig_unlock();
 531         }
 532         bdrv_dirty_bitmap_lock(bmds->dirty_bitmap);
 533         if (bdrv_get_dirty_locked(bs, bmds->dirty_bitmap, sector)) {
 534             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 535                 nr_sectors = total_sectors - sector;
 536             } else {
 537                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 538             }
 539             bdrv_reset_dirty_bitmap_locked(bmds->dirty_bitmap, sector, nr_sectors);
 540             bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 541
 542             blk = g_new(BlkMigBlock, 1);
 543             blk->buf = g_malloc(BLOCK_SIZE);
 544             blk->bmds = bmds;
 545             blk->sector = sector;
 546             blk->nr_sectors = nr_sectors;
 547
 548             if (is_async) {
 549                 blk->iov.iov_base = blk->buf;
 550                 blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 551                 qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 552
 553                 blk->aiocb = blk_aio_preadv(bmds->blk,
 554                                             sector * BDRV_SECTOR_SIZE,
 555                                             &blk->qiov, 0, blk_mig_read_cb,
 556                                             blk);
 557
 558                 blk_mig_lock();
 559                 block_mig_state.submitted++;
 560                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 561                 blk_mig_unlock();
 562             } else {
 563                 ret = blk_pread(bmds->blk, sector * BDRV_SECTOR_SIZE, blk->buf,
 564                                 nr_sectors * BDRV_SECTOR_SIZE);
 565                 if (ret < 0) {
 566                     goto error;
 567                 }
 568                 blk_send(f, blk);
 569
 570                 g_free(blk->buf);
 571                 g_free(blk);
 572             }
 573
 574             sector += nr_sectors;
 575             bmds->cur_dirty = sector;
 576             break;
 577         }
 578
 579         bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 580         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 581         bmds->cur_dirty = sector;
 582     }
 583
 584     return (bmds->cur_dirty >= bmds->total_sectors);
 585
 586 error:
 587     DPRINTF("Error reading sector %" PRId64 "\n", sector);
 588     g_free(blk->buf);
 589     g_free(blk);
 590     return ret;
 591 }
 592
 593 /* Called with iothread lock taken.
 594  *
 595  * return value:
 596  * 0: too much data for max_downtime
 597  * 1: few enough data for max_downtime
 598 */
 599 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 600 {
 601     BlkMigDevState *bmds;
 602     int ret = 1;
 603
 604     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 605         aio_context_acquire(blk_get_aio_context(bmds->blk));
 606         ret = mig_save_device_dirty(f, bmds, is_async);
 607         aio_context_release(blk_get_aio_context(bmds->blk));
 608         if (ret <= 0) {
 609             break;
 610         }
 611     }
 612
 613     return ret;
 614 }
 615
 616 /* Called with no locks taken.  */
 617
 618 static int flush_blks(QEMUFile *f)
 619 {
 620     BlkMigBlock *blk;
 621     int ret = 0;
 622
 623     DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
 624             __FUNCTION__, block_mig_state.submitted, block_mig_state.read_done,
 625             block_mig_state.transferred);
 626
 627     blk_mig_lock();
 628     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 629         if (qemu_file_rate_limit(f)) {
 630             break;
 631         }
 632         if (blk->ret < 0) {
 633             ret = blk->ret;
 634             break;
 635         }
 636
 637         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 638         blk_mig_unlock();
 639         blk_send(f, blk);
 640         blk_mig_lock();
 641
 642         g_free(blk->buf);
 643         g_free(blk);
 644
 645         block_mig_state.read_done--;
 646         block_mig_state.transferred++;
 647         assert(block_mig_state.read_done >= 0);
 648     }
 649     blk_mig_unlock();
 650
 651     DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__,
 652             block_mig_state.submitted, block_mig_state.read_done,
 653             block_mig_state.transferred);
 654     return ret;
 655 }
 656
 657 /* Called with iothread lock taken.  */
 658
 659 static int64_t get_remaining_dirty(void)
 660 {
 661     BlkMigDevState *bmds;
 662     int64_t dirty = 0;
 663
 664     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 665         aio_context_acquire(blk_get_aio_context(bmds->blk));
 666         dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
 667         aio_context_release(blk_get_aio_context(bmds->blk));
 668     }
 669
 670     return dirty << BDRV_SECTOR_BITS;
 671 }
 672
 673
 674
 675 /* Called with iothread lock taken.  */
 676 static void block_migration_cleanup_bmds(void)
 677 {
 678     BlkMigDevState *bmds;
 679     AioContext *ctx;
 680
 681     unset_dirty_tracking();
 682
 683     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 684         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 685         bdrv_op_unblock_all(blk_bs(bmds->blk), bmds->blocker);
 686         error_free(bmds->blocker);
 687
 688         /* Save ctx, because bmds->blk can disappear during blk_unref.  */
 689         ctx = blk_get_aio_context(bmds->blk);
 690         aio_context_acquire(ctx);
 691         blk_unref(bmds->blk);
 692         aio_context_release(ctx);
 693
 694         g_free(bmds->blk_name);
 695         g_free(bmds->aio_bitmap);
 696         g_free(bmds);
 697     }
 698 }
 699
 700 /* Called with iothread lock taken.  */
 701 static void block_migration_cleanup(void *opaque)
 702 {
 703     BlkMigBlock *blk;
 704
 705     bdrv_drain_all();
 706
 707     block_migration_cleanup_bmds();
 708
 709     blk_mig_lock();
 710     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 711         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 712         g_free(blk->buf);
 713         g_free(blk);
 714     }
 715     blk_mig_unlock();
 716 }
 717
 718 static int block_save_setup(QEMUFile *f, void *opaque)
 719 {
 720     int ret;
 721
 722     DPRINTF("Enter save live setup submitted %d transferred %d\n",
 723             block_mig_state.submitted, block_mig_state.transferred);
 724
 725     qemu_mutex_lock_iothread();
 726     ret = init_blk_migration(f);
 727     if (ret < 0) {
 728         qemu_mutex_unlock_iothread();
 729         return ret;
 730     }
 731
 732     /* start track dirty blocks */
 733     ret = set_dirty_tracking();
 734
 735     qemu_mutex_unlock_iothread();
 736
 737     if (ret) {
 738         return ret;
 739     }
 740
 741     ret = flush_blks(f);
 742     blk_mig_reset_dirty_cursor();
 743     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 744
 745     return ret;
 746 }
 747
 748 static int block_save_iterate(QEMUFile *f, void *opaque)
 749 {
 750     int ret;
 751     int64_t last_ftell = qemu_ftell(f);
 752     int64_t delta_ftell;
 753
 754     DPRINTF("Enter save live iterate submitted %d transferred %d\n",
 755             block_mig_state.submitted, block_mig_state.transferred);
 756
 757     ret = flush_blks(f);
 758     if (ret) {
 759         return ret;
 760     }
 761
 762     blk_mig_reset_dirty_cursor();
 763
 764     /* control the rate of transfer */
 765     blk_mig_lock();
 766     while ((block_mig_state.submitted +
 767             block_mig_state.read_done) * BLOCK_SIZE <
 768            qemu_file_get_rate_limit(f) &&
 769            (block_mig_state.submitted +
 770             block_mig_state.read_done) <
 771            MAX_INFLIGHT_IO) {
 772         blk_mig_unlock();
 773         if (block_mig_state.bulk_completed == 0) {
 774             /* first finish the bulk phase */
 775             if (blk_mig_save_bulked_block(f) == 0) {
 776                 /* finished saving bulk on all devices */
 777                 block_mig_state.bulk_completed = 1;
 778             }
 779             ret = 0;
 780         } else {
 781             /* Always called with iothread lock taken for
 782              * simplicity, block_save_complete also calls it.
 783              */
 784             qemu_mutex_lock_iothread();
 785             ret = blk_mig_save_dirty_block(f, 1);
 786             qemu_mutex_unlock_iothread();
 787         }
 788         if (ret < 0) {
 789             return ret;
 790         }
 791         blk_mig_lock();
 792         if (ret != 0) {
 793             /* no more dirty blocks */
 794             break;
 795         }
 796     }
 797     blk_mig_unlock();
 798
 799     ret = flush_blks(f);
 800     if (ret) {
 801         return ret;
 802     }
 803
 804     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 805     delta_ftell = qemu_ftell(f) - last_ftell;
 806     if (delta_ftell > 0) {
 807         return 1;
 808     } else if (delta_ftell < 0) {
 809         return -1;
 810     } else {
 811         return 0;
 812     }
 813 }
 814
 815 /* Called with iothread lock taken.  */
 816
 817 static int block_save_complete(QEMUFile *f, void *opaque)
 818 {
 819     int ret;
 820
 821     DPRINTF("Enter save live complete submitted %d transferred %d\n",
 822             block_mig_state.submitted, block_mig_state.transferred);
 823
 824     ret = flush_blks(f);
 825     if (ret) {
 826         return ret;
 827     }
 828
 829     blk_mig_reset_dirty_cursor();
 830
 831     /* we know for sure that save bulk is completed and
 832        all async read completed */
 833     blk_mig_lock();
 834     assert(block_mig_state.submitted == 0);
 835     blk_mig_unlock();
 836
 837     do {
 838         ret = blk_mig_save_dirty_block(f, 0);
 839         if (ret < 0) {
 840             return ret;
 841         }
 842     } while (ret == 0);
 843
 844     /* report completion */
 845     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 846
 847     DPRINTF("Block migration completed\n");
 848
 849     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 850
 851     /* Make sure that our BlockBackends are gone, so that the block driver
 852      * nodes can be inactivated. */
 853     block_migration_cleanup_bmds();
 854
 855     return 0;
 856 }
 857
 858 static void block_save_pending(QEMUFile *f, void *opaque, uint64_t max_size,
 859                                uint64_t *non_postcopiable_pending,
 860                                uint64_t *postcopiable_pending)
 861 {
 862     /* Estimate pending number of bytes to send */
 863     uint64_t pending;
 864
 865     qemu_mutex_lock_iothread();
 866     pending = get_remaining_dirty();
 867     qemu_mutex_unlock_iothread();
 868
 869     blk_mig_lock();
 870     pending += block_mig_state.submitted * BLOCK_SIZE +
 871                block_mig_state.read_done * BLOCK_SIZE;
 872     blk_mig_unlock();
 873
 874     /* Report at least one block pending during bulk phase */
 875     if (pending <= max_size && !block_mig_state.bulk_completed) {
 876         pending = max_size + BLOCK_SIZE;
 877     }
 878
 879     DPRINTF("Enter save live pending  %" PRIu64 "\n", pending);
 880     /* We don't do postcopy */
 881     *non_postcopiable_pending += pending;
 882 }
 883
 884 static int block_load(QEMUFile *f, void *opaque, int version_id)
 885 {
 886     static int banner_printed;
 887     int len, flags;
 888     char device_name[256];
 889     int64_t addr;
 890     BlockBackend *blk, *blk_prev = NULL;;
 891     Error *local_err = NULL;
 892     uint8_t *buf;
 893     int64_t total_sectors = 0;
 894     int nr_sectors;
 895     int ret;
 896     BlockDriverInfo bdi;
 897     int cluster_size = BLOCK_SIZE;
 898
 899     do {
 900         addr = qemu_get_be64(f);
 901
 902         flags = addr & ~BDRV_SECTOR_MASK;
 903         addr >>= BDRV_SECTOR_BITS;
 904
 905         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 906             /* get device name */
 907             len = qemu_get_byte(f);
 908             qemu_get_buffer(f, (uint8_t *)device_name, len);
 909             device_name[len] = '\0';
 910
 911             blk = blk_by_name(device_name);
 912             if (!blk) {
 913                 fprintf(stderr, "Error unknown block device %s\n",
 914                         device_name);
 915                 return -EINVAL;
 916             }
 917
 918             if (blk != blk_prev) {
 919                 blk_prev = blk;
 920                 total_sectors = blk_nb_sectors(blk);
 921                 if (total_sectors <= 0) {
 922                     error_report("Error getting length of block device %s",
 923                                  device_name);
 924                     return -EINVAL;
 925                 }
 926
 927                 blk_invalidate_cache(blk, &local_err);
 928                 if (local_err) {
 929                     error_report_err(local_err);
 930                     return -EINVAL;
 931                 }
 932
 933                 ret = bdrv_get_info(blk_bs(blk), &bdi);
 934                 if (ret == 0 && bdi.cluster_size > 0 &&
 935                     bdi.cluster_size <= BLOCK_SIZE &&
 936                     BLOCK_SIZE % bdi.cluster_size == 0) {
 937                     cluster_size = bdi.cluster_size;
 938                 } else {
 939                     cluster_size = BLOCK_SIZE;
 940                 }
 941             }
 942
 943             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 944                 nr_sectors = total_sectors - addr;
 945             } else {
 946                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 947             }
 948
 949             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 950                 ret = blk_pwrite_zeroes(blk, addr * BDRV_SECTOR_SIZE,
 951                                         nr_sectors * BDRV_SECTOR_SIZE,
 952                                         BDRV_REQ_MAY_UNMAP);
 953             } else {
 954                 int i;
 955                 int64_t cur_addr;
 956                 uint8_t *cur_buf;
 957
 958                 buf = g_malloc(BLOCK_SIZE);
 959                 qemu_get_buffer(f, buf, BLOCK_SIZE);
 960                 for (i = 0; i < BLOCK_SIZE / cluster_size; i++) {
 961                     cur_addr = addr * BDRV_SECTOR_SIZE + i * cluster_size;
 962                     cur_buf = buf + i * cluster_size;
 963
 964                     if ((!block_mig_state.zero_blocks ||
 965                         cluster_size < BLOCK_SIZE) &&
 966                         buffer_is_zero(cur_buf, cluster_size)) {
 967                         ret = blk_pwrite_zeroes(blk, cur_addr,
 968                                                 cluster_size,
 969                                                 BDRV_REQ_MAY_UNMAP);
 970                     } else {
 971                         ret = blk_pwrite(blk, cur_addr, cur_buf,
 972                                          cluster_size, 0);
 973                     }
 974                     if (ret < 0) {
 975                         break;
 976                     }
 977                 }
 978                 g_free(buf);
 979             }
 980
 981             if (ret < 0) {
 982                 return ret;
 983             }
 984         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 985             if (!banner_printed) {
 986                 printf("Receiving block device images\n");
 987                 banner_printed = 1;
 988             }
 989             printf("Completed %d %%%c", (int)addr,
 990                    (addr == 100) ? '\n' : '\r');
 991             fflush(stdout);
 992         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
 993             fprintf(stderr, "Unknown block migration flags: %#x\n", flags);
 994             return -EINVAL;
 995         }
 996         ret = qemu_file_get_error(f);
 997         if (ret != 0) {
 998             return ret;
 999         }
1000     } while (!(flags & BLK_MIG_FLAG_EOS));
1001
1002     return 0;
1003 }
1004
1005 static bool block_is_active(void *opaque)
1006 {
1007     return migrate_use_block();
1008 }
1009
1010 static SaveVMHandlers savevm_block_handlers = {
1011     .save_setup = block_save_setup,
1012     .save_live_iterate = block_save_iterate,
1013     .save_live_complete_precopy = block_save_complete,
1014     .save_live_pending = block_save_pending,
1015     .load_state = block_load,
1016     .save_cleanup = block_migration_cleanup,
1017     .is_active = block_is_active,
1018 };
1019
1020 void blk_mig_init(void)
1021 {
1022     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
1023     QSIMPLEQ_INIT(&block_mig_state.blk_list);
1024     qemu_mutex_init(&block_mig_state.lock);
1025
1026     register_savevm_live(NULL, "block", 0, 1, &savevm_block_handlers,
1027                          &block_mig_state);
1028 }