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 void 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
 394     block_mig_state.submitted = 0;
 395     block_mig_state.read_done = 0;
 396     block_mig_state.transferred = 0;
 397     block_mig_state.total_sector_sum = 0;
 398     block_mig_state.prev_progress = -1;
 399     block_mig_state.bulk_completed = 0;
 400     block_mig_state.zero_blocks = migrate_zero_blocks();
 401
 402     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
 403         num_bs++;
 404     }
 405     bmds_bs = g_malloc0(num_bs * sizeof(*bmds_bs));
 406
 407     for (i = 0, bs = bdrv_first(&it); bs; bs = bdrv_next(&it), i++) {
 408         if (bdrv_is_read_only(bs)) {
 409             continue;
 410         }
 411
 412         sectors = bdrv_nb_sectors(bs);
 413         if (sectors <= 0) {
 414             goto out;
 415         }
 416
 417         bmds = g_new0(BlkMigDevState, 1);
 418         bmds->blk = blk_new();
 419         bmds->blk_name = g_strdup(bdrv_get_device_name(bs));
 420         bmds->bulk_completed = 0;
 421         bmds->total_sectors = sectors;
 422         bmds->completed_sectors = 0;
 423         bmds->shared_base = block_mig_state.shared_base;
 424
 425         assert(i < num_bs);
 426         bmds_bs[i].bmds = bmds;
 427         bmds_bs[i].bs = bs;
 428
 429         block_mig_state.total_sector_sum += sectors;
 430
 431         if (bmds->shared_base) {
 432             DPRINTF("Start migration for %s with shared base image\n",
 433                     bdrv_get_device_name(bs));
 434         } else {
 435             DPRINTF("Start full migration for %s\n", bdrv_get_device_name(bs));
 436         }
 437
 438         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 439     }
 440
 441     /* Can only insert new BDSes now because doing so while iterating block
 442      * devices may end up in a deadlock (iterating the new BDSes, too). */
 443     for (i = 0; i < num_bs; i++) {
 444         BlkMigDevState *bmds = bmds_bs[i].bmds;
 445         BlockDriverState *bs = bmds_bs[i].bs;
 446
 447         if (bmds) {
 448             blk_insert_bs(bmds->blk, bs);
 449
 450             alloc_aio_bitmap(bmds);
 451             error_setg(&bmds->blocker, "block device is in use by migration");
 452             bdrv_op_block_all(bs, bmds->blocker);
 453         }
 454     }
 455
 456 out:
 457     g_free(bmds_bs);
 458 }
 459
 460 /* Called with no lock taken.  */
 461
 462 static int blk_mig_save_bulked_block(QEMUFile *f)
 463 {
 464     int64_t completed_sector_sum = 0;
 465     BlkMigDevState *bmds;
 466     int progress;
 467     int ret = 0;
 468
 469     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 470         if (bmds->bulk_completed == 0) {
 471             if (mig_save_device_bulk(f, bmds) == 1) {
 472                 /* completed bulk section for this device */
 473                 bmds->bulk_completed = 1;
 474             }
 475             completed_sector_sum += bmds->completed_sectors;
 476             ret = 1;
 477             break;
 478         } else {
 479             completed_sector_sum += bmds->completed_sectors;
 480         }
 481     }
 482
 483     if (block_mig_state.total_sector_sum != 0) {
 484         progress = completed_sector_sum * 100 /
 485                    block_mig_state.total_sector_sum;
 486     } else {
 487         progress = 100;
 488     }
 489     if (progress != block_mig_state.prev_progress) {
 490         block_mig_state.prev_progress = progress;
 491         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 492                          | BLK_MIG_FLAG_PROGRESS);
 493         DPRINTF("Completed %d %%\r", progress);
 494     }
 495
 496     return ret;
 497 }
 498
 499 static void blk_mig_reset_dirty_cursor(void)
 500 {
 501     BlkMigDevState *bmds;
 502
 503     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 504         bmds->cur_dirty = 0;
 505     }
 506 }
 507
 508 /* Called with iothread lock and AioContext taken.  */
 509
 510 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
 511                                  int is_async)
 512 {
 513     BlkMigBlock *blk;
 514     BlockDriverState *bs = blk_bs(bmds->blk);
 515     int64_t total_sectors = bmds->total_sectors;
 516     int64_t sector;
 517     int nr_sectors;
 518     int ret = -EIO;
 519
 520     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
 521         blk_mig_lock();
 522         if (bmds_aio_inflight(bmds, sector)) {
 523             blk_mig_unlock();
 524             blk_drain(bmds->blk);
 525         } else {
 526             blk_mig_unlock();
 527         }
 528         if (bdrv_get_dirty(bs, bmds->dirty_bitmap, sector)) {
 529
 530             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 531                 nr_sectors = total_sectors - sector;
 532             } else {
 533                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 534             }
 535             blk = g_new(BlkMigBlock, 1);
 536             blk->buf = g_malloc(BLOCK_SIZE);
 537             blk->bmds = bmds;
 538             blk->sector = sector;
 539             blk->nr_sectors = nr_sectors;
 540
 541             if (is_async) {
 542                 blk->iov.iov_base = blk->buf;
 543                 blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 544                 qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 545
 546                 blk->aiocb = blk_aio_preadv(bmds->blk,
 547                                             sector * BDRV_SECTOR_SIZE,
 548                                             &blk->qiov, 0, blk_mig_read_cb,
 549                                             blk);
 550
 551                 blk_mig_lock();
 552                 block_mig_state.submitted++;
 553                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 554                 blk_mig_unlock();
 555             } else {
 556                 ret = blk_pread(bmds->blk, sector * BDRV_SECTOR_SIZE, blk->buf,
 557                                 nr_sectors * BDRV_SECTOR_SIZE);
 558                 if (ret < 0) {
 559                     goto error;
 560                 }
 561                 blk_send(f, blk);
 562
 563                 g_free(blk->buf);
 564                 g_free(blk);
 565             }
 566
 567             bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, sector, nr_sectors);
 568             break;
 569         }
 570         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 571         bmds->cur_dirty = sector;
 572     }
 573
 574     return (bmds->cur_dirty >= bmds->total_sectors);
 575
 576 error:
 577     DPRINTF("Error reading sector %" PRId64 "\n", sector);
 578     g_free(blk->buf);
 579     g_free(blk);
 580     return ret;
 581 }
 582
 583 /* Called with iothread lock taken.
 584  *
 585  * return value:
 586  * 0: too much data for max_downtime
 587  * 1: few enough data for max_downtime
 588 */
 589 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 590 {
 591     BlkMigDevState *bmds;
 592     int ret = 1;
 593
 594     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 595         aio_context_acquire(blk_get_aio_context(bmds->blk));
 596         ret = mig_save_device_dirty(f, bmds, is_async);
 597         aio_context_release(blk_get_aio_context(bmds->blk));
 598         if (ret <= 0) {
 599             break;
 600         }
 601     }
 602
 603     return ret;
 604 }
 605
 606 /* Called with no locks taken.  */
 607
 608 static int flush_blks(QEMUFile *f)
 609 {
 610     BlkMigBlock *blk;
 611     int ret = 0;
 612
 613     DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
 614             __FUNCTION__, block_mig_state.submitted, block_mig_state.read_done,
 615             block_mig_state.transferred);
 616
 617     blk_mig_lock();
 618     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 619         if (qemu_file_rate_limit(f)) {
 620             break;
 621         }
 622         if (blk->ret < 0) {
 623             ret = blk->ret;
 624             break;
 625         }
 626
 627         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 628         blk_mig_unlock();
 629         blk_send(f, blk);
 630         blk_mig_lock();
 631
 632         g_free(blk->buf);
 633         g_free(blk);
 634
 635         block_mig_state.read_done--;
 636         block_mig_state.transferred++;
 637         assert(block_mig_state.read_done >= 0);
 638     }
 639     blk_mig_unlock();
 640
 641     DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__,
 642             block_mig_state.submitted, block_mig_state.read_done,
 643             block_mig_state.transferred);
 644     return ret;
 645 }
 646
 647 /* Called with iothread lock taken.  */
 648
 649 static int64_t get_remaining_dirty(void)
 650 {
 651     BlkMigDevState *bmds;
 652     int64_t dirty = 0;
 653
 654     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 655         aio_context_acquire(blk_get_aio_context(bmds->blk));
 656         dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
 657         aio_context_release(blk_get_aio_context(bmds->blk));
 658     }
 659
 660     return dirty << BDRV_SECTOR_BITS;
 661 }
 662
 663 /* Called with iothread lock taken.  */
 664
 665 static void block_migration_cleanup(void *opaque)
 666 {
 667     BlkMigDevState *bmds;
 668     BlkMigBlock *blk;
 669     AioContext *ctx;
 670
 671     bdrv_drain_all();
 672
 673     unset_dirty_tracking();
 674
 675     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 676         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 677         bdrv_op_unblock_all(blk_bs(bmds->blk), bmds->blocker);
 678         error_free(bmds->blocker);
 679
 680         /* Save ctx, because bmds->blk can disappear during blk_unref.  */
 681         ctx = blk_get_aio_context(bmds->blk);
 682         aio_context_acquire(ctx);
 683         blk_unref(bmds->blk);
 684         aio_context_release(ctx);
 685
 686         g_free(bmds->blk_name);
 687         g_free(bmds->aio_bitmap);
 688         g_free(bmds);
 689     }
 690
 691     blk_mig_lock();
 692     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 693         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 694         g_free(blk->buf);
 695         g_free(blk);
 696     }
 697     blk_mig_unlock();
 698 }
 699
 700 static int block_save_setup(QEMUFile *f, void *opaque)
 701 {
 702     int ret;
 703
 704     DPRINTF("Enter save live setup submitted %d transferred %d\n",
 705             block_mig_state.submitted, block_mig_state.transferred);
 706
 707     qemu_mutex_lock_iothread();
 708     init_blk_migration(f);
 709
 710     /* start track dirty blocks */
 711     ret = set_dirty_tracking();
 712
 713     qemu_mutex_unlock_iothread();
 714
 715     if (ret) {
 716         return ret;
 717     }
 718
 719     ret = flush_blks(f);
 720     blk_mig_reset_dirty_cursor();
 721     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 722
 723     return ret;
 724 }
 725
 726 static int block_save_iterate(QEMUFile *f, void *opaque)
 727 {
 728     int ret;
 729     int64_t last_ftell = qemu_ftell(f);
 730     int64_t delta_ftell;
 731
 732     DPRINTF("Enter save live iterate submitted %d transferred %d\n",
 733             block_mig_state.submitted, block_mig_state.transferred);
 734
 735     ret = flush_blks(f);
 736     if (ret) {
 737         return ret;
 738     }
 739
 740     blk_mig_reset_dirty_cursor();
 741
 742     /* control the rate of transfer */
 743     blk_mig_lock();
 744     while ((block_mig_state.submitted +
 745             block_mig_state.read_done) * BLOCK_SIZE <
 746            qemu_file_get_rate_limit(f) &&
 747            (block_mig_state.submitted +
 748             block_mig_state.read_done) <
 749            MAX_INFLIGHT_IO) {
 750         blk_mig_unlock();
 751         if (block_mig_state.bulk_completed == 0) {
 752             /* first finish the bulk phase */
 753             if (blk_mig_save_bulked_block(f) == 0) {
 754                 /* finished saving bulk on all devices */
 755                 block_mig_state.bulk_completed = 1;
 756             }
 757             ret = 0;
 758         } else {
 759             /* Always called with iothread lock taken for
 760              * simplicity, block_save_complete also calls it.
 761              */
 762             qemu_mutex_lock_iothread();
 763             ret = blk_mig_save_dirty_block(f, 1);
 764             qemu_mutex_unlock_iothread();
 765         }
 766         if (ret < 0) {
 767             return ret;
 768         }
 769         blk_mig_lock();
 770         if (ret != 0) {
 771             /* no more dirty blocks */
 772             break;
 773         }
 774     }
 775     blk_mig_unlock();
 776
 777     ret = flush_blks(f);
 778     if (ret) {
 779         return ret;
 780     }
 781
 782     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 783     delta_ftell = qemu_ftell(f) - last_ftell;
 784     if (delta_ftell > 0) {
 785         return 1;
 786     } else if (delta_ftell < 0) {
 787         return -1;
 788     } else {
 789         return 0;
 790     }
 791 }
 792
 793 /* Called with iothread lock taken.  */
 794
 795 static int block_save_complete(QEMUFile *f, void *opaque)
 796 {
 797     int ret;
 798
 799     DPRINTF("Enter save live complete submitted %d transferred %d\n",
 800             block_mig_state.submitted, block_mig_state.transferred);
 801
 802     ret = flush_blks(f);
 803     if (ret) {
 804         return ret;
 805     }
 806
 807     blk_mig_reset_dirty_cursor();
 808
 809     /* we know for sure that save bulk is completed and
 810        all async read completed */
 811     blk_mig_lock();
 812     assert(block_mig_state.submitted == 0);
 813     blk_mig_unlock();
 814
 815     do {
 816         ret = blk_mig_save_dirty_block(f, 0);
 817         if (ret < 0) {
 818             return ret;
 819         }
 820     } while (ret == 0);
 821
 822     /* report completion */
 823     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 824
 825     DPRINTF("Block migration completed\n");
 826
 827     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 828
 829     return 0;
 830 }
 831
 832 static void block_save_pending(QEMUFile *f, void *opaque, uint64_t max_size,
 833                                uint64_t *non_postcopiable_pending,
 834                                uint64_t *postcopiable_pending)
 835 {
 836     /* Estimate pending number of bytes to send */
 837     uint64_t pending;
 838
 839     qemu_mutex_lock_iothread();
 840     pending = get_remaining_dirty();
 841     qemu_mutex_unlock_iothread();
 842
 843     blk_mig_lock();
 844     pending += block_mig_state.submitted * BLOCK_SIZE +
 845                block_mig_state.read_done * BLOCK_SIZE;
 846     blk_mig_unlock();
 847
 848     /* Report at least one block pending during bulk phase */
 849     if (pending <= max_size && !block_mig_state.bulk_completed) {
 850         pending = max_size + BLOCK_SIZE;
 851     }
 852
 853     DPRINTF("Enter save live pending  %" PRIu64 "\n", pending);
 854     /* We don't do postcopy */
 855     *non_postcopiable_pending += pending;
 856 }
 857
 858 static int block_load(QEMUFile *f, void *opaque, int version_id)
 859 {
 860     static int banner_printed;
 861     int len, flags;
 862     char device_name[256];
 863     int64_t addr;
 864     BlockBackend *blk, *blk_prev = NULL;;
 865     Error *local_err = NULL;
 866     uint8_t *buf;
 867     int64_t total_sectors = 0;
 868     int nr_sectors;
 869     int ret;
 870
 871     do {
 872         addr = qemu_get_be64(f);
 873
 874         flags = addr & ~BDRV_SECTOR_MASK;
 875         addr >>= BDRV_SECTOR_BITS;
 876
 877         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 878             /* get device name */
 879             len = qemu_get_byte(f);
 880             qemu_get_buffer(f, (uint8_t *)device_name, len);
 881             device_name[len] = '\0';
 882
 883             blk = blk_by_name(device_name);
 884             if (!blk) {
 885                 fprintf(stderr, "Error unknown block device %s\n",
 886                         device_name);
 887                 return -EINVAL;
 888             }
 889
 890             if (blk != blk_prev) {
 891                 blk_prev = blk;
 892                 total_sectors = blk_nb_sectors(blk);
 893                 if (total_sectors <= 0) {
 894                     error_report("Error getting length of block device %s",
 895                                  device_name);
 896                     return -EINVAL;
 897                 }
 898
 899                 blk_invalidate_cache(blk, &local_err);
 900                 if (local_err) {
 901                     error_report_err(local_err);
 902                     return -EINVAL;
 903                 }
 904             }
 905
 906             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 907                 nr_sectors = total_sectors - addr;
 908             } else {
 909                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 910             }
 911
 912             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 913                 ret = blk_pwrite_zeroes(blk, addr * BDRV_SECTOR_SIZE,
 914                                         nr_sectors * BDRV_SECTOR_SIZE,
 915                                         BDRV_REQ_MAY_UNMAP);
 916             } else {
 917                 buf = g_malloc(BLOCK_SIZE);
 918                 qemu_get_buffer(f, buf, BLOCK_SIZE);
 919                 ret = blk_pwrite(blk, addr * BDRV_SECTOR_SIZE, buf,
 920                                  nr_sectors * BDRV_SECTOR_SIZE, 0);
 921                 g_free(buf);
 922             }
 923
 924             if (ret < 0) {
 925                 return ret;
 926             }
 927         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 928             if (!banner_printed) {
 929                 printf("Receiving block device images\n");
 930                 banner_printed = 1;
 931             }
 932             printf("Completed %d %%%c", (int)addr,
 933                    (addr == 100) ? '\n' : '\r');
 934             fflush(stdout);
 935         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
 936             fprintf(stderr, "Unknown block migration flags: %#x\n", flags);
 937             return -EINVAL;
 938         }
 939         ret = qemu_file_get_error(f);
 940         if (ret != 0) {
 941             return ret;
 942         }
 943     } while (!(flags & BLK_MIG_FLAG_EOS));
 944
 945     return 0;
 946 }
 947
 948 static void block_set_params(const MigrationParams *params, void *opaque)
 949 {
 950     block_mig_state.blk_enable = params->blk;
 951     block_mig_state.shared_base = params->shared;
 952
 953     /* shared base means that blk_enable = 1 */
 954     block_mig_state.blk_enable |= params->shared;
 955 }
 956
 957 static bool block_is_active(void *opaque)
 958 {
 959     return block_mig_state.blk_enable == 1;
 960 }
 961
 962 static SaveVMHandlers savevm_block_handlers = {
 963     .set_params = block_set_params,
 964     .save_live_setup = block_save_setup,
 965     .save_live_iterate = block_save_iterate,
 966     .save_live_complete_precopy = block_save_complete,
 967     .save_live_pending = block_save_pending,
 968     .load_state = block_load,
 969     .cleanup = block_migration_cleanup,
 970     .is_active = block_is_active,
 971 };
 972
 973 void blk_mig_init(void)
 974 {
 975     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
 976     QSIMPLEQ_INIT(&block_mig_state.blk_list);
 977     qemu_mutex_init(&block_mig_state.lock);
 978
 979     register_savevm_live(NULL, "block", 0, 1, &savevm_block_handlers,
 980                          &block_mig_state);
 981 }