migration/block.c

   1 /*
   2  * QEMU live block migration
   3  *
   4  * Copyright IBM, Corp. 2009
   5  *
   6  * Authors:
   7  *  Liran Schour   <lirans@il.ibm.com>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.  See
  10  * the COPYING file in the top-level directory.
  11  *
  12  * Contributions after 2012-01-13 are licensed under the terms of the
  13  * GNU GPL, version 2 or (at your option) any later version.
  14  */
  15
  16 #include "qemu/osdep.h"
  17 #include "qapi/error.h"
  18 #include "qemu/error-report.h"
  19 #include "qemu/main-loop.h"
  20 #include "qemu/cutils.h"
  21 #include "qemu/queue.h"
  22 #include "block.h"
  23 #include "block/dirty-bitmap.h"
  24 #include "migration/misc.h"
  25 #include "migration.h"
  26 #include "migration/register.h"
  27 #include "qemu-file.h"
  28 #include "migration/vmstate.h"
  29 #include "sysemu/block-backend.h"
  30 #include "trace.h"
  31 #include "options.h"
  32
  33 #define BLK_MIG_BLOCK_SIZE           (1ULL << 20)
  34 #define BDRV_SECTORS_PER_DIRTY_CHUNK (BLK_MIG_BLOCK_SIZE >> BDRV_SECTOR_BITS)
  35
  36 #define BLK_MIG_FLAG_DEVICE_BLOCK       0x01
  37 #define BLK_MIG_FLAG_EOS                0x02
  38 #define BLK_MIG_FLAG_PROGRESS           0x04
  39 #define BLK_MIG_FLAG_ZERO_BLOCK         0x08
  40
  41 #define MAX_IS_ALLOCATED_SEARCH (65536 * BDRV_SECTOR_SIZE)
  42
  43 #define MAX_IO_BUFFERS 512
  44 #define MAX_PARALLEL_IO 16
  45
  46 typedef struct BlkMigDevState {
  47     /* Written during setup phase.  Can be read without a lock.  */
  48     BlockBackend *blk;
  49     char *blk_name;
  50     int shared_base;
  51     int64_t total_sectors;
  52     QSIMPLEQ_ENTRY(BlkMigDevState) entry;
  53     Error *blocker;
  54
  55     /* Only used by migration thread.  Does not need a lock.  */
  56     int bulk_completed;
  57     int64_t cur_sector;
  58     int64_t cur_dirty;
  59
  60     /* Data in the aio_bitmap is protected by block migration lock.
  61      * Allocation and free happen during setup and cleanup respectively.
  62      */
  63     unsigned long *aio_bitmap;
  64
  65     /* Protected by block migration lock.  */
  66     int64_t completed_sectors;
  67
  68     /* During migration this is protected by iothread lock / AioContext.
  69      * Allocation and free happen during setup and cleanup respectively.
  70      */
  71     BdrvDirtyBitmap *dirty_bitmap;
  72 } BlkMigDevState;
  73
  74 typedef struct BlkMigBlock {
  75     /* Only used by migration thread.  */
  76     uint8_t *buf;
  77     BlkMigDevState *bmds;
  78     int64_t sector;
  79     int nr_sectors;
  80     QEMUIOVector qiov;
  81     BlockAIOCB *aiocb;
  82
  83     /* Protected by block migration lock.  */
  84     int ret;
  85     QSIMPLEQ_ENTRY(BlkMigBlock) entry;
  86 } BlkMigBlock;
  87
  88 typedef struct BlkMigState {
  89     QSIMPLEQ_HEAD(, BlkMigDevState) bmds_list;
  90     int64_t total_sector_sum;
  91     bool zero_blocks;
  92
  93     /* Protected by lock.  */
  94     QSIMPLEQ_HEAD(, BlkMigBlock) blk_list;
  95     int submitted;
  96     int read_done;
  97
  98     /* Only used by migration thread.  Does not need a lock.  */
  99     int transferred;
 100     int prev_progress;
 101     int bulk_completed;
 102
 103     /* Lock must be taken _inside_ the iothread lock and any AioContexts.  */
 104     QemuMutex lock;
 105 } BlkMigState;
 106
 107 static BlkMigState block_mig_state;
 108
 109 static void blk_mig_lock(void)
 110 {
 111     qemu_mutex_lock(&block_mig_state.lock);
 112 }
 113
 114 static void blk_mig_unlock(void)
 115 {
 116     qemu_mutex_unlock(&block_mig_state.lock);
 117 }
 118
 119 /* Must run outside of the iothread lock during the bulk phase,
 120  * or the VM will stall.
 121  */
 122
 123 static void blk_send(QEMUFile *f, BlkMigBlock * blk)
 124 {
 125     int len;
 126     uint64_t flags = BLK_MIG_FLAG_DEVICE_BLOCK;
 127
 128     if (block_mig_state.zero_blocks &&
 129         buffer_is_zero(blk->buf, BLK_MIG_BLOCK_SIZE)) {
 130         flags |= BLK_MIG_FLAG_ZERO_BLOCK;
 131     }
 132
 133     /* sector number and flags */
 134     qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
 135                      | flags);
 136
 137     /* device name */
 138     len = strlen(blk->bmds->blk_name);
 139     qemu_put_byte(f, len);
 140     qemu_put_buffer(f, (uint8_t *) blk->bmds->blk_name, len);
 141
 142     /* if a block is zero we need to flush here since the network
 143      * bandwidth is now a lot higher than the storage device bandwidth.
 144      * thus if we queue zero blocks we slow down the migration */
 145     if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 146         qemu_fflush(f);
 147         return;
 148     }
 149
 150     qemu_put_buffer(f, blk->buf, BLK_MIG_BLOCK_SIZE);
 151 }
 152
 153 int blk_mig_active(void)
 154 {
 155     return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
 156 }
 157
 158 int blk_mig_bulk_active(void)
 159 {
 160     return blk_mig_active() && !block_mig_state.bulk_completed;
 161 }
 162
 163 uint64_t blk_mig_bytes_transferred(void)
 164 {
 165     BlkMigDevState *bmds;
 166     uint64_t sum = 0;
 167
 168     blk_mig_lock();
 169     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 170         sum += bmds->completed_sectors;
 171     }
 172     blk_mig_unlock();
 173     return sum << BDRV_SECTOR_BITS;
 174 }
 175
 176 uint64_t blk_mig_bytes_remaining(void)
 177 {
 178     return blk_mig_bytes_total() - blk_mig_bytes_transferred();
 179 }
 180
 181 uint64_t blk_mig_bytes_total(void)
 182 {
 183     BlkMigDevState *bmds;
 184     uint64_t sum = 0;
 185
 186     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 187         sum += bmds->total_sectors;
 188     }
 189     return sum << BDRV_SECTOR_BITS;
 190 }
 191
 192
 193 /* Called with migration lock held.  */
 194
 195 static int bmds_aio_inflight(BlkMigDevState *bmds, int64_t sector)
 196 {
 197     int64_t chunk = sector / (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK;
 198
 199     if (sector < bmds->total_sectors) {
 200         return !!(bmds->aio_bitmap[chunk / (sizeof(unsigned long) * 8)] &
 201             (1UL << (chunk % (sizeof(unsigned long) * 8))));
 202     } else {
 203         return 0;
 204     }
 205 }
 206
 207 /* Called with migration lock held.  */
 208
 209 static void bmds_set_aio_inflight(BlkMigDevState *bmds, int64_t sector_num,
 210                              int nb_sectors, int set)
 211 {
 212     int64_t start, end;
 213     unsigned long val, idx, bit;
 214
 215     start = sector_num / BDRV_SECTORS_PER_DIRTY_CHUNK;
 216     end = (sector_num + nb_sectors - 1) / BDRV_SECTORS_PER_DIRTY_CHUNK;
 217
 218     for (; start <= end; start++) {
 219         idx = start / (sizeof(unsigned long) * 8);
 220         bit = start % (sizeof(unsigned long) * 8);
 221         val = bmds->aio_bitmap[idx];
 222         if (set) {
 223             val |= 1UL << bit;
 224         } else {
 225             val &= ~(1UL << bit);
 226         }
 227         bmds->aio_bitmap[idx] = val;
 228     }
 229 }
 230
 231 static void alloc_aio_bitmap(BlkMigDevState *bmds)
 232 {
 233     int64_t bitmap_size;
 234
 235     bitmap_size = bmds->total_sectors + BDRV_SECTORS_PER_DIRTY_CHUNK * 8 - 1;
 236     bitmap_size /= BDRV_SECTORS_PER_DIRTY_CHUNK * 8;
 237
 238     bmds->aio_bitmap = g_malloc0(bitmap_size);
 239 }
 240
 241 /* Never hold migration lock when yielding to the main loop!  */
 242
 243 static void blk_mig_read_cb(void *opaque, int ret)
 244 {
 245     BlkMigBlock *blk = opaque;
 246
 247     blk_mig_lock();
 248     blk->ret = ret;
 249
 250     QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry);
 251     bmds_set_aio_inflight(blk->bmds, blk->sector, blk->nr_sectors, 0);
 252
 253     block_mig_state.submitted--;
 254     block_mig_state.read_done++;
 255     assert(block_mig_state.submitted >= 0);
 256     blk_mig_unlock();
 257 }
 258
 259 /* Called with no lock taken.  */
 260
 261 static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds)
 262 {
 263     int64_t total_sectors = bmds->total_sectors;
 264     int64_t cur_sector = bmds->cur_sector;
 265     BlockBackend *bb = bmds->blk;
 266     BlkMigBlock *blk;
 267     int nr_sectors;
 268     int64_t count;
 269
 270     if (bmds->shared_base) {
 271         qemu_mutex_lock_iothread();
 272         aio_context_acquire(blk_get_aio_context(bb));
 273         /* Skip unallocated sectors; intentionally treats failure or
 274          * partial sector as an allocated sector */
 275         while (cur_sector < total_sectors &&
 276                !bdrv_is_allocated(blk_bs(bb), cur_sector * BDRV_SECTOR_SIZE,
 277                                   MAX_IS_ALLOCATED_SEARCH, &count)) {
 278             if (count < BDRV_SECTOR_SIZE) {
 279                 break;
 280             }
 281             cur_sector += count >> BDRV_SECTOR_BITS;
 282         }
 283         aio_context_release(blk_get_aio_context(bb));
 284         qemu_mutex_unlock_iothread();
 285     }
 286
 287     if (cur_sector >= total_sectors) {
 288         bmds->cur_sector = bmds->completed_sectors = total_sectors;
 289         return 1;
 290     }
 291
 292     bmds->completed_sectors = cur_sector;
 293
 294     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
 295
 296     /* we are going to transfer a full block even if it is not allocated */
 297     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 298
 299     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 300         nr_sectors = total_sectors - cur_sector;
 301     }
 302
 303     blk = g_new(BlkMigBlock, 1);
 304     blk->buf = g_malloc(BLK_MIG_BLOCK_SIZE);
 305     blk->bmds = bmds;
 306     blk->sector = cur_sector;
 307     blk->nr_sectors = nr_sectors;
 308
 309     qemu_iovec_init_buf(&blk->qiov, blk->buf, nr_sectors * BDRV_SECTOR_SIZE);
 310
 311     blk_mig_lock();
 312     block_mig_state.submitted++;
 313     blk_mig_unlock();
 314
 315     /* We do not know if bs is under the main thread (and thus does
 316      * not acquire the AioContext when doing AIO) or rather under
 317      * dataplane.  Thus acquire both the iothread mutex and the
 318      * AioContext.
 319      *
 320      * This is ugly and will disappear when we make bdrv_* thread-safe,
 321      * without the need to acquire the AioContext.
 322      */
 323     qemu_mutex_lock_iothread();
 324     aio_context_acquire(blk_get_aio_context(bmds->blk));
 325     bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, cur_sector * BDRV_SECTOR_SIZE,
 326                             nr_sectors * BDRV_SECTOR_SIZE);
 327     blk->aiocb = blk_aio_preadv(bb, cur_sector * BDRV_SECTOR_SIZE, &blk->qiov,
 328                                 0, blk_mig_read_cb, blk);
 329     aio_context_release(blk_get_aio_context(bmds->blk));
 330     qemu_mutex_unlock_iothread();
 331
 332     bmds->cur_sector = cur_sector + nr_sectors;
 333     return (bmds->cur_sector >= total_sectors);
 334 }
 335
 336 /* Called with iothread lock taken.  */
 337
 338 static int set_dirty_tracking(void)
 339 {
 340     BlkMigDevState *bmds;
 341     int ret;
 342
 343     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 344         bmds->dirty_bitmap = bdrv_create_dirty_bitmap(blk_bs(bmds->blk),
 345                                                       BLK_MIG_BLOCK_SIZE,
 346                                                       NULL, NULL);
 347         if (!bmds->dirty_bitmap) {
 348             ret = -errno;
 349             goto fail;
 350         }
 351     }
 352     return 0;
 353
 354 fail:
 355     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 356         if (bmds->dirty_bitmap) {
 357             bdrv_release_dirty_bitmap(bmds->dirty_bitmap);
 358         }
 359     }
 360     return ret;
 361 }
 362
 363 /* Called with iothread lock taken.  */
 364
 365 static void unset_dirty_tracking(void)
 366 {
 367     BlkMigDevState *bmds;
 368
 369     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 370         bdrv_release_dirty_bitmap(bmds->dirty_bitmap);
 371     }
 372 }
 373
 374 static int init_blk_migration(QEMUFile *f)
 375 {
 376     BlockDriverState *bs;
 377     BlkMigDevState *bmds;
 378     int64_t sectors;
 379     BdrvNextIterator it;
 380     int i, num_bs = 0;
 381     struct {
 382         BlkMigDevState *bmds;
 383         BlockDriverState *bs;
 384     } *bmds_bs;
 385     Error *local_err = NULL;
 386     int ret;
 387
 388     block_mig_state.submitted = 0;
 389     block_mig_state.read_done = 0;
 390     block_mig_state.transferred = 0;
 391     block_mig_state.total_sector_sum = 0;
 392     block_mig_state.prev_progress = -1;
 393     block_mig_state.bulk_completed = 0;
 394     block_mig_state.zero_blocks = migrate_zero_blocks();
 395
 396     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
 397         num_bs++;
 398     }
 399     bmds_bs = g_malloc0(num_bs * sizeof(*bmds_bs));
 400
 401     for (i = 0, bs = bdrv_first(&it); bs; bs = bdrv_next(&it), i++) {
 402         if (bdrv_is_read_only(bs)) {
 403             continue;
 404         }
 405
 406         sectors = bdrv_nb_sectors(bs);
 407         if (sectors <= 0) {
 408             ret = sectors;
 409             bdrv_next_cleanup(&it);
 410             goto out;
 411         }
 412
 413         bmds = g_new0(BlkMigDevState, 1);
 414         bmds->blk = blk_new(qemu_get_aio_context(),
 415                             BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 416         bmds->blk_name = g_strdup(bdrv_get_device_name(bs));
 417         bmds->bulk_completed = 0;
 418         bmds->total_sectors = sectors;
 419         bmds->completed_sectors = 0;
 420         bmds->shared_base = migrate_block_incremental();
 421
 422         assert(i < num_bs);
 423         bmds_bs[i].bmds = bmds;
 424         bmds_bs[i].bs = bs;
 425
 426         block_mig_state.total_sector_sum += sectors;
 427
 428         if (bmds->shared_base) {
 429             trace_migration_block_init_shared(bdrv_get_device_name(bs));
 430         } else {
 431             trace_migration_block_init_full(bdrv_get_device_name(bs));
 432         }
 433
 434         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 435     }
 436
 437     /* Can only insert new BDSes now because doing so while iterating block
 438      * devices may end up in a deadlock (iterating the new BDSes, too). */
 439     for (i = 0; i < num_bs; i++) {
 440         BlkMigDevState *bmds = bmds_bs[i].bmds;
 441         BlockDriverState *bs = bmds_bs[i].bs;
 442
 443         if (bmds) {
 444             ret = blk_insert_bs(bmds->blk, bs, &local_err);
 445             if (ret < 0) {
 446                 error_report_err(local_err);
 447                 goto out;
 448             }
 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     ret = 0;
 457 out:
 458     g_free(bmds_bs);
 459     return ret;
 460 }
 461
 462 /* Called with no lock taken.  */
 463
 464 static int blk_mig_save_bulked_block(QEMUFile *f)
 465 {
 466     int64_t completed_sector_sum = 0;
 467     BlkMigDevState *bmds;
 468     int progress;
 469     int ret = 0;
 470
 471     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 472         if (bmds->bulk_completed == 0) {
 473             if (mig_save_device_bulk(f, bmds) == 1) {
 474                 /* completed bulk section for this device */
 475                 bmds->bulk_completed = 1;
 476             }
 477             completed_sector_sum += bmds->completed_sectors;
 478             ret = 1;
 479             break;
 480         } else {
 481             completed_sector_sum += bmds->completed_sectors;
 482         }
 483     }
 484
 485     if (block_mig_state.total_sector_sum != 0) {
 486         progress = completed_sector_sum * 100 /
 487                    block_mig_state.total_sector_sum;
 488     } else {
 489         progress = 100;
 490     }
 491     if (progress != block_mig_state.prev_progress) {
 492         block_mig_state.prev_progress = progress;
 493         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 494                          | BLK_MIG_FLAG_PROGRESS);
 495         trace_migration_block_progression(progress);
 496     }
 497
 498     return ret;
 499 }
 500
 501 static void blk_mig_reset_dirty_cursor(void)
 502 {
 503     BlkMigDevState *bmds;
 504
 505     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 506         bmds->cur_dirty = 0;
 507     }
 508 }
 509
 510 /* Called with iothread lock and AioContext taken.  */
 511
 512 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
 513                                  int is_async)
 514 {
 515     BlkMigBlock *blk;
 516     int64_t total_sectors = bmds->total_sectors;
 517     int64_t sector;
 518     int nr_sectors;
 519     int ret = -EIO;
 520
 521     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
 522         blk_mig_lock();
 523         if (bmds_aio_inflight(bmds, sector)) {
 524             blk_mig_unlock();
 525             blk_drain(bmds->blk);
 526         } else {
 527             blk_mig_unlock();
 528         }
 529         bdrv_dirty_bitmap_lock(bmds->dirty_bitmap);
 530         if (bdrv_dirty_bitmap_get_locked(bmds->dirty_bitmap,
 531                                          sector * BDRV_SECTOR_SIZE)) {
 532             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 533                 nr_sectors = total_sectors - sector;
 534             } else {
 535                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 536             }
 537             bdrv_reset_dirty_bitmap_locked(bmds->dirty_bitmap,
 538                                            sector * BDRV_SECTOR_SIZE,
 539                                            nr_sectors * BDRV_SECTOR_SIZE);
 540             bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 541
 542             blk = g_new(BlkMigBlock, 1);
 543             blk->buf = g_malloc(BLK_MIG_BLOCK_SIZE);
 544             blk->bmds = bmds;
 545             blk->sector = sector;
 546             blk->nr_sectors = nr_sectors;
 547
 548             if (is_async) {
 549                 qemu_iovec_init_buf(&blk->qiov, blk->buf,
 550                                     nr_sectors * BDRV_SECTOR_SIZE);
 551
 552                 blk->aiocb = blk_aio_preadv(bmds->blk,
 553                                             sector * BDRV_SECTOR_SIZE,
 554                                             &blk->qiov, 0, blk_mig_read_cb,
 555                                             blk);
 556
 557                 blk_mig_lock();
 558                 block_mig_state.submitted++;
 559                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 560                 blk_mig_unlock();
 561             } else {
 562                 ret = blk_pread(bmds->blk, sector * BDRV_SECTOR_SIZE,
 563                                 nr_sectors * BDRV_SECTOR_SIZE, blk->buf, 0);
 564                 if (ret < 0) {
 565                     goto error;
 566                 }
 567                 blk_send(f, blk);
 568
 569                 g_free(blk->buf);
 570                 g_free(blk);
 571             }
 572
 573             sector += nr_sectors;
 574             bmds->cur_dirty = sector;
 575             break;
 576         }
 577
 578         bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 579         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 580         bmds->cur_dirty = sector;
 581     }
 582
 583     return (bmds->cur_dirty >= bmds->total_sectors);
 584
 585 error:
 586     trace_migration_block_save_device_dirty(sector);
 587     g_free(blk->buf);
 588     g_free(blk);
 589     return ret;
 590 }
 591
 592 /* Called with iothread lock taken.
 593  *
 594  * return value:
 595  * 0: too much data for max_downtime
 596  * 1: few enough data for max_downtime
 597 */
 598 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 599 {
 600     BlkMigDevState *bmds;
 601     int ret = 1;
 602
 603     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 604         aio_context_acquire(blk_get_aio_context(bmds->blk));
 605         ret = mig_save_device_dirty(f, bmds, is_async);
 606         aio_context_release(blk_get_aio_context(bmds->blk));
 607         if (ret <= 0) {
 608             break;
 609         }
 610     }
 611
 612     return ret;
 613 }
 614
 615 /* Called with no locks taken.  */
 616
 617 static int flush_blks(QEMUFile *f)
 618 {
 619     BlkMigBlock *blk;
 620     int ret = 0;
 621
 622     trace_migration_block_flush_blks("Enter", block_mig_state.submitted,
 623                                      block_mig_state.read_done,
 624                                      block_mig_state.transferred);
 625
 626     blk_mig_lock();
 627     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 628         if (qemu_file_rate_limit(f)) {
 629             break;
 630         }
 631         if (blk->ret < 0) {
 632             ret = blk->ret;
 633             break;
 634         }
 635
 636         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 637         blk_mig_unlock();
 638         blk_send(f, blk);
 639         blk_mig_lock();
 640
 641         g_free(blk->buf);
 642         g_free(blk);
 643
 644         block_mig_state.read_done--;
 645         block_mig_state.transferred++;
 646         assert(block_mig_state.read_done >= 0);
 647     }
 648     blk_mig_unlock();
 649
 650     trace_migration_block_flush_blks("Exit", block_mig_state.submitted,
 651                                      block_mig_state.read_done,
 652                                      block_mig_state.transferred);
 653     return ret;
 654 }
 655
 656 /* Called with iothread lock taken.  */
 657
 658 static int64_t get_remaining_dirty(void)
 659 {
 660     BlkMigDevState *bmds;
 661     int64_t dirty = 0;
 662
 663     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 664         aio_context_acquire(blk_get_aio_context(bmds->blk));
 665         dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
 666         aio_context_release(blk_get_aio_context(bmds->blk));
 667     }
 668
 669     return dirty;
 670 }
 671
 672
 673
 674 /* Called with iothread lock taken.  */
 675 static void block_migration_cleanup_bmds(void)
 676 {
 677     BlkMigDevState *bmds;
 678     AioContext *ctx;
 679
 680     unset_dirty_tracking();
 681
 682     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 683         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 684         bdrv_op_unblock_all(blk_bs(bmds->blk), bmds->blocker);
 685         error_free(bmds->blocker);
 686
 687         /* Save ctx, because bmds->blk can disappear during blk_unref.  */
 688         ctx = blk_get_aio_context(bmds->blk);
 689         aio_context_acquire(ctx);
 690         blk_unref(bmds->blk);
 691         aio_context_release(ctx);
 692
 693         g_free(bmds->blk_name);
 694         g_free(bmds->aio_bitmap);
 695         g_free(bmds);
 696     }
 697 }
 698
 699 /* Called with iothread lock taken.  */
 700 static void block_migration_cleanup(void *opaque)
 701 {
 702     BlkMigBlock *blk;
 703
 704     bdrv_drain_all();
 705
 706     block_migration_cleanup_bmds();
 707
 708     blk_mig_lock();
 709     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 710         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 711         g_free(blk->buf);
 712         g_free(blk);
 713     }
 714     blk_mig_unlock();
 715 }
 716
 717 static int block_save_setup(QEMUFile *f, void *opaque)
 718 {
 719     int ret;
 720
 721     trace_migration_block_save("setup", block_mig_state.submitted,
 722                                block_mig_state.transferred);
 723
 724     qemu_mutex_lock_iothread();
 725     ret = init_blk_migration(f);
 726     if (ret < 0) {
 727         qemu_mutex_unlock_iothread();
 728         return ret;
 729     }
 730
 731     /* start track dirty blocks */
 732     ret = set_dirty_tracking();
 733
 734     qemu_mutex_unlock_iothread();
 735
 736     if (ret) {
 737         return ret;
 738     }
 739
 740     ret = flush_blks(f);
 741     blk_mig_reset_dirty_cursor();
 742     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 743
 744     return ret;
 745 }
 746
 747 static int block_save_iterate(QEMUFile *f, void *opaque)
 748 {
 749     int ret;
 750     int64_t last_bytes = qemu_file_total_transferred(f);
 751     int64_t delta_bytes;
 752
 753     trace_migration_block_save("iterate", block_mig_state.submitted,
 754                                block_mig_state.transferred);
 755
 756     ret = flush_blks(f);
 757     if (ret) {
 758         return ret;
 759     }
 760
 761     blk_mig_reset_dirty_cursor();
 762
 763     /* control the rate of transfer */
 764     blk_mig_lock();
 765     while (block_mig_state.read_done * BLK_MIG_BLOCK_SIZE <
 766            qemu_file_get_rate_limit(f) &&
 767            block_mig_state.submitted < MAX_PARALLEL_IO &&
 768            (block_mig_state.submitted + block_mig_state.read_done) <
 769            MAX_IO_BUFFERS) {
 770         blk_mig_unlock();
 771         if (block_mig_state.bulk_completed == 0) {
 772             /* first finish the bulk phase */
 773             if (blk_mig_save_bulked_block(f) == 0) {
 774                 /* finished saving bulk on all devices */
 775                 block_mig_state.bulk_completed = 1;
 776             }
 777             ret = 0;
 778         } else {
 779             /* Always called with iothread lock taken for
 780              * simplicity, block_save_complete also calls it.
 781              */
 782             qemu_mutex_lock_iothread();
 783             ret = blk_mig_save_dirty_block(f, 1);
 784             qemu_mutex_unlock_iothread();
 785         }
 786         if (ret < 0) {
 787             return ret;
 788         }
 789         blk_mig_lock();
 790         if (ret != 0) {
 791             /* no more dirty blocks */
 792             break;
 793         }
 794     }
 795     blk_mig_unlock();
 796
 797     ret = flush_blks(f);
 798     if (ret) {
 799         return ret;
 800     }
 801
 802     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 803     delta_bytes = qemu_file_total_transferred(f) - last_bytes;
 804     if (delta_bytes > 0) {
 805         return 1;
 806     } else if (delta_bytes < 0) {
 807         return -1;
 808     } else {
 809         return 0;
 810     }
 811 }
 812
 813 /* Called with iothread lock taken.  */
 814
 815 static int block_save_complete(QEMUFile *f, void *opaque)
 816 {
 817     int ret;
 818
 819     trace_migration_block_save("complete", block_mig_state.submitted,
 820                                block_mig_state.transferred);
 821
 822     ret = flush_blks(f);
 823     if (ret) {
 824         return ret;
 825     }
 826
 827     blk_mig_reset_dirty_cursor();
 828
 829     /* we know for sure that save bulk is completed and
 830        all async read completed */
 831     blk_mig_lock();
 832     assert(block_mig_state.submitted == 0);
 833     blk_mig_unlock();
 834
 835     do {
 836         ret = blk_mig_save_dirty_block(f, 0);
 837         if (ret < 0) {
 838             return ret;
 839         }
 840     } while (ret == 0);
 841
 842     /* report completion */
 843     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 844
 845     trace_migration_block_save_complete();
 846
 847     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 848
 849     /* Make sure that our BlockBackends are gone, so that the block driver
 850      * nodes can be inactivated. */
 851     block_migration_cleanup_bmds();
 852
 853     return 0;
 854 }
 855
 856 static void block_state_pending(void *opaque, uint64_t *must_precopy,
 857                                 uint64_t *can_postcopy)
 858 {
 859     /* Estimate pending number of bytes to send */
 860     uint64_t pending;
 861
 862     qemu_mutex_lock_iothread();
 863     pending = get_remaining_dirty();
 864     qemu_mutex_unlock_iothread();
 865
 866     blk_mig_lock();
 867     pending += block_mig_state.submitted * BLK_MIG_BLOCK_SIZE +
 868                block_mig_state.read_done * BLK_MIG_BLOCK_SIZE;
 869     blk_mig_unlock();
 870
 871     /* Report at least one block pending during bulk phase */
 872     if (!pending && !block_mig_state.bulk_completed) {
 873         pending = BLK_MIG_BLOCK_SIZE;
 874     }
 875
 876     trace_migration_block_state_pending(pending);
 877     /* We don't do postcopy */
 878     *must_precopy += pending;
 879 }
 880
 881 static int block_load(QEMUFile *f, void *opaque, int version_id)
 882 {
 883     static int banner_printed;
 884     int len, flags;
 885     char device_name[256];
 886     int64_t addr;
 887     BlockBackend *blk, *blk_prev = NULL;
 888     Error *local_err = NULL;
 889     uint8_t *buf;
 890     int64_t total_sectors = 0;
 891     int nr_sectors;
 892     int ret;
 893     BlockDriverInfo bdi;
 894     int cluster_size = BLK_MIG_BLOCK_SIZE;
 895
 896     do {
 897         addr = qemu_get_be64(f);
 898
 899         flags = addr & (BDRV_SECTOR_SIZE - 1);
 900         addr >>= BDRV_SECTOR_BITS;
 901
 902         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 903             /* get device name */
 904             len = qemu_get_byte(f);
 905             qemu_get_buffer(f, (uint8_t *)device_name, len);
 906             device_name[len] = '\0';
 907
 908             blk = blk_by_name(device_name);
 909             if (!blk) {
 910                 fprintf(stderr, "Error unknown block device %s\n",
 911                         device_name);
 912                 return -EINVAL;
 913             }
 914
 915             if (blk != blk_prev) {
 916                 blk_prev = blk;
 917                 total_sectors = blk_nb_sectors(blk);
 918                 if (total_sectors <= 0) {
 919                     error_report("Error getting length of block device %s",
 920                                  device_name);
 921                     return -EINVAL;
 922                 }
 923
 924                 blk_activate(blk, &local_err);
 925                 if (local_err) {
 926                     error_report_err(local_err);
 927                     return -EINVAL;
 928                 }
 929
 930                 ret = bdrv_get_info(blk_bs(blk), &bdi);
 931                 if (ret == 0 && bdi.cluster_size > 0 &&
 932                     bdi.cluster_size <= BLK_MIG_BLOCK_SIZE &&
 933                     BLK_MIG_BLOCK_SIZE % bdi.cluster_size == 0) {
 934                     cluster_size = bdi.cluster_size;
 935                 } else {
 936                     cluster_size = BLK_MIG_BLOCK_SIZE;
 937                 }
 938             }
 939
 940             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 941                 nr_sectors = total_sectors - addr;
 942             } else {
 943                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 944             }
 945
 946             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 947                 ret = blk_pwrite_zeroes(blk, addr * BDRV_SECTOR_SIZE,
 948                                         nr_sectors * BDRV_SECTOR_SIZE,
 949                                         BDRV_REQ_MAY_UNMAP);
 950             } else {
 951                 int i;
 952                 int64_t cur_addr;
 953                 uint8_t *cur_buf;
 954
 955                 buf = g_malloc(BLK_MIG_BLOCK_SIZE);
 956                 qemu_get_buffer(f, buf, BLK_MIG_BLOCK_SIZE);
 957                 for (i = 0; i < BLK_MIG_BLOCK_SIZE / cluster_size; i++) {
 958                     cur_addr = addr * BDRV_SECTOR_SIZE + i * cluster_size;
 959                     cur_buf = buf + i * cluster_size;
 960
 961                     if ((!block_mig_state.zero_blocks ||
 962                         cluster_size < BLK_MIG_BLOCK_SIZE) &&
 963                         buffer_is_zero(cur_buf, cluster_size)) {
 964                         ret = blk_pwrite_zeroes(blk, cur_addr,
 965                                                 cluster_size,
 966                                                 BDRV_REQ_MAY_UNMAP);
 967                     } else {
 968                         ret = blk_pwrite(blk, cur_addr, cluster_size, cur_buf,
 969                                          0);
 970                     }
 971                     if (ret < 0) {
 972                         break;
 973                     }
 974                 }
 975                 g_free(buf);
 976             }
 977
 978             if (ret < 0) {
 979                 return ret;
 980             }
 981         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 982             if (!banner_printed) {
 983                 printf("Receiving block device images\n");
 984                 banner_printed = 1;
 985             }
 986             printf("Completed %d %%%c", (int)addr,
 987                    (addr == 100) ? '\n' : '\r');
 988             fflush(stdout);
 989         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
 990             fprintf(stderr, "Unknown block migration flags: 0x%x\n", flags);
 991             return -EINVAL;
 992         }
 993         ret = qemu_file_get_error(f);
 994         if (ret != 0) {
 995             return ret;
 996         }
 997     } while (!(flags & BLK_MIG_FLAG_EOS));
 998
 999     return 0;
1000 }
1001
1002 static bool block_is_active(void *opaque)
1003 {
1004     return migrate_block();
1005 }
1006
1007 static SaveVMHandlers savevm_block_handlers = {
1008     .save_setup = block_save_setup,
1009     .save_live_iterate = block_save_iterate,
1010     .save_live_complete_precopy = block_save_complete,
1011     .state_pending_exact = block_state_pending,
1012     .state_pending_estimate = block_state_pending,
1013     .load_state = block_load,
1014     .save_cleanup = block_migration_cleanup,
1015     .is_active = block_is_active,
1016 };
1017
1018 void blk_mig_init(void)
1019 {
1020     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
1021     QSIMPLEQ_INIT(&block_mig_state.blk_list);
1022     qemu_mutex_init(&block_mig_state.lock);
1023
1024     register_savevm_live("block", 0, 1, &savevm_block_handlers,
1025                          &block_mig_state);
1026 }