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 "migration/misc.h"
  24 #include "migration.h"
  25 #include "migration/register.h"
  26 #include "qemu-file.h"
  27 #include "migration/vmstate.h"
  28 #include "sysemu/block-backend.h"
  29
  30 #define BLOCK_SIZE                       (1 << 20)
  31 #define BDRV_SECTORS_PER_DIRTY_CHUNK     (BLOCK_SIZE >> BDRV_SECTOR_BITS)
  32
  33 #define BLK_MIG_FLAG_DEVICE_BLOCK       0x01
  34 #define BLK_MIG_FLAG_EOS                0x02
  35 #define BLK_MIG_FLAG_PROGRESS           0x04
  36 #define BLK_MIG_FLAG_ZERO_BLOCK         0x08
  37
  38 #define MAX_IS_ALLOCATED_SEARCH (65536 * BDRV_SECTOR_SIZE)
  39
  40 #define MAX_IO_BUFFERS 512
  41 #define MAX_PARALLEL_IO 16
  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     QEMUIOVector qiov;
  88     BlockAIOCB *aiocb;
  89
  90     /* Protected by block migration lock.  */
  91     int ret;
  92     QSIMPLEQ_ENTRY(BlkMigBlock) entry;
  93 } BlkMigBlock;
  94
  95 typedef struct BlkMigState {
  96     QSIMPLEQ_HEAD(, BlkMigDevState) bmds_list;
  97     int64_t total_sector_sum;
  98     bool zero_blocks;
  99
 100     /* Protected by lock.  */
 101     QSIMPLEQ_HEAD(, BlkMigBlock) blk_list;
 102     int submitted;
 103     int read_done;
 104
 105     /* Only used by migration thread.  Does not need a lock.  */
 106     int transferred;
 107     int prev_progress;
 108     int bulk_completed;
 109
 110     /* Lock must be taken _inside_ the iothread lock and any AioContexts.  */
 111     QemuMutex lock;
 112 } BlkMigState;
 113
 114 static BlkMigState block_mig_state;
 115
 116 static void blk_mig_lock(void)
 117 {
 118     qemu_mutex_lock(&block_mig_state.lock);
 119 }
 120
 121 static void blk_mig_unlock(void)
 122 {
 123     qemu_mutex_unlock(&block_mig_state.lock);
 124 }
 125
 126 /* Must run outside of the iothread lock during the bulk phase,
 127  * or the VM will stall.
 128  */
 129
 130 static void blk_send(QEMUFile *f, BlkMigBlock * blk)
 131 {
 132     int len;
 133     uint64_t flags = BLK_MIG_FLAG_DEVICE_BLOCK;
 134
 135     if (block_mig_state.zero_blocks &&
 136         buffer_is_zero(blk->buf, BLOCK_SIZE)) {
 137         flags |= BLK_MIG_FLAG_ZERO_BLOCK;
 138     }
 139
 140     /* sector number and flags */
 141     qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
 142                      | flags);
 143
 144     /* device name */
 145     len = strlen(blk->bmds->blk_name);
 146     qemu_put_byte(f, len);
 147     qemu_put_buffer(f, (uint8_t *) blk->bmds->blk_name, len);
 148
 149     /* if a block is zero we need to flush here since the network
 150      * bandwidth is now a lot higher than the storage device bandwidth.
 151      * thus if we queue zero blocks we slow down the migration */
 152     if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 153         qemu_fflush(f);
 154         return;
 155     }
 156
 157     qemu_put_buffer(f, blk->buf, BLOCK_SIZE);
 158 }
 159
 160 int blk_mig_active(void)
 161 {
 162     return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
 163 }
 164
 165 int blk_mig_bulk_active(void)
 166 {
 167     return blk_mig_active() && !block_mig_state.bulk_completed;
 168 }
 169
 170 uint64_t blk_mig_bytes_transferred(void)
 171 {
 172     BlkMigDevState *bmds;
 173     uint64_t sum = 0;
 174
 175     blk_mig_lock();
 176     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 177         sum += bmds->completed_sectors;
 178     }
 179     blk_mig_unlock();
 180     return sum << BDRV_SECTOR_BITS;
 181 }
 182
 183 uint64_t blk_mig_bytes_remaining(void)
 184 {
 185     return blk_mig_bytes_total() - blk_mig_bytes_transferred();
 186 }
 187
 188 uint64_t blk_mig_bytes_total(void)
 189 {
 190     BlkMigDevState *bmds;
 191     uint64_t sum = 0;
 192
 193     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 194         sum += bmds->total_sectors;
 195     }
 196     return sum << BDRV_SECTOR_BITS;
 197 }
 198
 199
 200 /* Called with migration lock held.  */
 201
 202 static int bmds_aio_inflight(BlkMigDevState *bmds, int64_t sector)
 203 {
 204     int64_t chunk = sector / (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK;
 205
 206     if (sector < blk_nb_sectors(bmds->blk)) {
 207         return !!(bmds->aio_bitmap[chunk / (sizeof(unsigned long) * 8)] &
 208             (1UL << (chunk % (sizeof(unsigned long) * 8))));
 209     } else {
 210         return 0;
 211     }
 212 }
 213
 214 /* Called with migration lock held.  */
 215
 216 static void bmds_set_aio_inflight(BlkMigDevState *bmds, int64_t sector_num,
 217                              int nb_sectors, int set)
 218 {
 219     int64_t start, end;
 220     unsigned long val, idx, bit;
 221
 222     start = sector_num / BDRV_SECTORS_PER_DIRTY_CHUNK;
 223     end = (sector_num + nb_sectors - 1) / BDRV_SECTORS_PER_DIRTY_CHUNK;
 224
 225     for (; start <= end; start++) {
 226         idx = start / (sizeof(unsigned long) * 8);
 227         bit = start % (sizeof(unsigned long) * 8);
 228         val = bmds->aio_bitmap[idx];
 229         if (set) {
 230             val |= 1UL << bit;
 231         } else {
 232             val &= ~(1UL << bit);
 233         }
 234         bmds->aio_bitmap[idx] = val;
 235     }
 236 }
 237
 238 static void alloc_aio_bitmap(BlkMigDevState *bmds)
 239 {
 240     BlockBackend *bb = bmds->blk;
 241     int64_t bitmap_size;
 242
 243     bitmap_size = blk_nb_sectors(bb) + BDRV_SECTORS_PER_DIRTY_CHUNK * 8 - 1;
 244     bitmap_size /= BDRV_SECTORS_PER_DIRTY_CHUNK * 8;
 245
 246     bmds->aio_bitmap = g_malloc0(bitmap_size);
 247 }
 248
 249 /* Never hold migration lock when yielding to the main loop!  */
 250
 251 static void blk_mig_read_cb(void *opaque, int ret)
 252 {
 253     BlkMigBlock *blk = opaque;
 254
 255     blk_mig_lock();
 256     blk->ret = ret;
 257
 258     QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry);
 259     bmds_set_aio_inflight(blk->bmds, blk->sector, blk->nr_sectors, 0);
 260
 261     block_mig_state.submitted--;
 262     block_mig_state.read_done++;
 263     assert(block_mig_state.submitted >= 0);
 264     blk_mig_unlock();
 265 }
 266
 267 /* Called with no lock taken.  */
 268
 269 static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds)
 270 {
 271     int64_t total_sectors = bmds->total_sectors;
 272     int64_t cur_sector = bmds->cur_sector;
 273     BlockBackend *bb = bmds->blk;
 274     BlkMigBlock *blk;
 275     int nr_sectors;
 276     int64_t count;
 277
 278     if (bmds->shared_base) {
 279         qemu_mutex_lock_iothread();
 280         aio_context_acquire(blk_get_aio_context(bb));
 281         /* Skip unallocated sectors; intentionally treats failure or
 282          * partial sector as an allocated sector */
 283         while (cur_sector < total_sectors &&
 284                !bdrv_is_allocated(blk_bs(bb), cur_sector * BDRV_SECTOR_SIZE,
 285                                   MAX_IS_ALLOCATED_SEARCH, &count)) {
 286             if (count < BDRV_SECTOR_SIZE) {
 287                 break;
 288             }
 289             cur_sector += count >> BDRV_SECTOR_BITS;
 290         }
 291         aio_context_release(blk_get_aio_context(bb));
 292         qemu_mutex_unlock_iothread();
 293     }
 294
 295     if (cur_sector >= total_sectors) {
 296         bmds->cur_sector = bmds->completed_sectors = total_sectors;
 297         return 1;
 298     }
 299
 300     bmds->completed_sectors = cur_sector;
 301
 302     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
 303
 304     /* we are going to transfer a full block even if it is not allocated */
 305     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 306
 307     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 308         nr_sectors = total_sectors - cur_sector;
 309     }
 310
 311     blk = g_new(BlkMigBlock, 1);
 312     blk->buf = g_malloc(BLOCK_SIZE);
 313     blk->bmds = bmds;
 314     blk->sector = cur_sector;
 315     blk->nr_sectors = nr_sectors;
 316
 317     qemu_iovec_init_buf(&blk->qiov, blk->buf, nr_sectors * BDRV_SECTOR_SIZE);
 318
 319     blk_mig_lock();
 320     block_mig_state.submitted++;
 321     blk_mig_unlock();
 322
 323     /* We do not know if bs is under the main thread (and thus does
 324      * not acquire the AioContext when doing AIO) or rather under
 325      * dataplane.  Thus acquire both the iothread mutex and the
 326      * AioContext.
 327      *
 328      * This is ugly and will disappear when we make bdrv_* thread-safe,
 329      * without the need to acquire the AioContext.
 330      */
 331     qemu_mutex_lock_iothread();
 332     aio_context_acquire(blk_get_aio_context(bmds->blk));
 333     bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, cur_sector * BDRV_SECTOR_SIZE,
 334                             nr_sectors * BDRV_SECTOR_SIZE);
 335     blk->aiocb = blk_aio_preadv(bb, cur_sector * BDRV_SECTOR_SIZE, &blk->qiov,
 336                                 0, blk_mig_read_cb, blk);
 337     aio_context_release(blk_get_aio_context(bmds->blk));
 338     qemu_mutex_unlock_iothread();
 339
 340     bmds->cur_sector = cur_sector + nr_sectors;
 341     return (bmds->cur_sector >= total_sectors);
 342 }
 343
 344 /* Called with iothread lock taken.  */
 345
 346 static int set_dirty_tracking(void)
 347 {
 348     BlkMigDevState *bmds;
 349     int ret;
 350
 351     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 352         bmds->dirty_bitmap = bdrv_create_dirty_bitmap(blk_bs(bmds->blk),
 353                                                       BLOCK_SIZE, NULL, NULL);
 354         if (!bmds->dirty_bitmap) {
 355             ret = -errno;
 356             goto fail;
 357         }
 358     }
 359     return 0;
 360
 361 fail:
 362     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 363         if (bmds->dirty_bitmap) {
 364             bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
 365         }
 366     }
 367     return ret;
 368 }
 369
 370 /* Called with iothread lock taken.  */
 371
 372 static void unset_dirty_tracking(void)
 373 {
 374     BlkMigDevState *bmds;
 375
 376     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 377         bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
 378     }
 379 }
 380
 381 static int init_blk_migration(QEMUFile *f)
 382 {
 383     BlockDriverState *bs;
 384     BlkMigDevState *bmds;
 385     int64_t sectors;
 386     BdrvNextIterator it;
 387     int i, num_bs = 0;
 388     struct {
 389         BlkMigDevState *bmds;
 390         BlockDriverState *bs;
 391     } *bmds_bs;
 392     Error *local_err = NULL;
 393     int ret;
 394
 395     block_mig_state.submitted = 0;
 396     block_mig_state.read_done = 0;
 397     block_mig_state.transferred = 0;
 398     block_mig_state.total_sector_sum = 0;
 399     block_mig_state.prev_progress = -1;
 400     block_mig_state.bulk_completed = 0;
 401     block_mig_state.zero_blocks = migrate_zero_blocks();
 402
 403     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
 404         num_bs++;
 405     }
 406     bmds_bs = g_malloc0(num_bs * sizeof(*bmds_bs));
 407
 408     for (i = 0, bs = bdrv_first(&it); bs; bs = bdrv_next(&it), i++) {
 409         if (bdrv_is_read_only(bs)) {
 410             continue;
 411         }
 412
 413         sectors = bdrv_nb_sectors(bs);
 414         if (sectors <= 0) {
 415             ret = sectors;
 416             bdrv_next_cleanup(&it);
 417             goto out;
 418         }
 419
 420         bmds = g_new0(BlkMigDevState, 1);
 421         bmds->blk = blk_new(qemu_get_aio_context(),
 422                             BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 423         bmds->blk_name = g_strdup(bdrv_get_device_name(bs));
 424         bmds->bulk_completed = 0;
 425         bmds->total_sectors = sectors;
 426         bmds->completed_sectors = 0;
 427         bmds->shared_base = migrate_use_block_incremental();
 428
 429         assert(i < num_bs);
 430         bmds_bs[i].bmds = bmds;
 431         bmds_bs[i].bs = bs;
 432
 433         block_mig_state.total_sector_sum += sectors;
 434
 435         if (bmds->shared_base) {
 436             DPRINTF("Start migration for %s with shared base image\n",
 437                     bdrv_get_device_name(bs));
 438         } else {
 439             DPRINTF("Start full migration for %s\n", bdrv_get_device_name(bs));
 440         }
 441
 442         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 443     }
 444
 445     /* Can only insert new BDSes now because doing so while iterating block
 446      * devices may end up in a deadlock (iterating the new BDSes, too). */
 447     for (i = 0; i < num_bs; i++) {
 448         BlkMigDevState *bmds = bmds_bs[i].bmds;
 449         BlockDriverState *bs = bmds_bs[i].bs;
 450
 451         if (bmds) {
 452             ret = blk_insert_bs(bmds->blk, bs, &local_err);
 453             if (ret < 0) {
 454                 error_report_err(local_err);
 455                 goto out;
 456             }
 457
 458             alloc_aio_bitmap(bmds);
 459             error_setg(&bmds->blocker, "block device is in use by migration");
 460             bdrv_op_block_all(bs, bmds->blocker);
 461         }
 462     }
 463
 464     ret = 0;
 465 out:
 466     g_free(bmds_bs);
 467     return ret;
 468 }
 469
 470 /* Called with no lock taken.  */
 471
 472 static int blk_mig_save_bulked_block(QEMUFile *f)
 473 {
 474     int64_t completed_sector_sum = 0;
 475     BlkMigDevState *bmds;
 476     int progress;
 477     int ret = 0;
 478
 479     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 480         if (bmds->bulk_completed == 0) {
 481             if (mig_save_device_bulk(f, bmds) == 1) {
 482                 /* completed bulk section for this device */
 483                 bmds->bulk_completed = 1;
 484             }
 485             completed_sector_sum += bmds->completed_sectors;
 486             ret = 1;
 487             break;
 488         } else {
 489             completed_sector_sum += bmds->completed_sectors;
 490         }
 491     }
 492
 493     if (block_mig_state.total_sector_sum != 0) {
 494         progress = completed_sector_sum * 100 /
 495                    block_mig_state.total_sector_sum;
 496     } else {
 497         progress = 100;
 498     }
 499     if (progress != block_mig_state.prev_progress) {
 500         block_mig_state.prev_progress = progress;
 501         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 502                          | BLK_MIG_FLAG_PROGRESS);
 503         DPRINTF("Completed %d %%\r", progress);
 504     }
 505
 506     return ret;
 507 }
 508
 509 static void blk_mig_reset_dirty_cursor(void)
 510 {
 511     BlkMigDevState *bmds;
 512
 513     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 514         bmds->cur_dirty = 0;
 515     }
 516 }
 517
 518 /* Called with iothread lock and AioContext taken.  */
 519
 520 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
 521                                  int is_async)
 522 {
 523     BlkMigBlock *blk;
 524     int64_t total_sectors = bmds->total_sectors;
 525     int64_t sector;
 526     int nr_sectors;
 527     int ret = -EIO;
 528
 529     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
 530         blk_mig_lock();
 531         if (bmds_aio_inflight(bmds, sector)) {
 532             blk_mig_unlock();
 533             blk_drain(bmds->blk);
 534         } else {
 535             blk_mig_unlock();
 536         }
 537         bdrv_dirty_bitmap_lock(bmds->dirty_bitmap);
 538         if (bdrv_dirty_bitmap_get_locked(bmds->dirty_bitmap,
 539                                          sector * BDRV_SECTOR_SIZE)) {
 540             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 541                 nr_sectors = total_sectors - sector;
 542             } else {
 543                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 544             }
 545             bdrv_reset_dirty_bitmap_locked(bmds->dirty_bitmap,
 546                                            sector * BDRV_SECTOR_SIZE,
 547                                            nr_sectors * BDRV_SECTOR_SIZE);
 548             bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 549
 550             blk = g_new(BlkMigBlock, 1);
 551             blk->buf = g_malloc(BLOCK_SIZE);
 552             blk->bmds = bmds;
 553             blk->sector = sector;
 554             blk->nr_sectors = nr_sectors;
 555
 556             if (is_async) {
 557                 qemu_iovec_init_buf(&blk->qiov, blk->buf,
 558                                     nr_sectors * BDRV_SECTOR_SIZE);
 559
 560                 blk->aiocb = blk_aio_preadv(bmds->blk,
 561                                             sector * BDRV_SECTOR_SIZE,
 562                                             &blk->qiov, 0, blk_mig_read_cb,
 563                                             blk);
 564
 565                 blk_mig_lock();
 566                 block_mig_state.submitted++;
 567                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 568                 blk_mig_unlock();
 569             } else {
 570                 ret = blk_pread(bmds->blk, sector * BDRV_SECTOR_SIZE, blk->buf,
 571                                 nr_sectors * BDRV_SECTOR_SIZE);
 572                 if (ret < 0) {
 573                     goto error;
 574                 }
 575                 blk_send(f, blk);
 576
 577                 g_free(blk->buf);
 578                 g_free(blk);
 579             }
 580
 581             sector += nr_sectors;
 582             bmds->cur_dirty = sector;
 583             break;
 584         }
 585
 586         bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 587         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 588         bmds->cur_dirty = sector;
 589     }
 590
 591     return (bmds->cur_dirty >= bmds->total_sectors);
 592
 593 error:
 594     DPRINTF("Error reading sector %" PRId64 "\n", sector);
 595     g_free(blk->buf);
 596     g_free(blk);
 597     return ret;
 598 }
 599
 600 /* Called with iothread lock taken.
 601  *
 602  * return value:
 603  * 0: too much data for max_downtime
 604  * 1: few enough data for max_downtime
 605 */
 606 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 607 {
 608     BlkMigDevState *bmds;
 609     int ret = 1;
 610
 611     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 612         aio_context_acquire(blk_get_aio_context(bmds->blk));
 613         ret = mig_save_device_dirty(f, bmds, is_async);
 614         aio_context_release(blk_get_aio_context(bmds->blk));
 615         if (ret <= 0) {
 616             break;
 617         }
 618     }
 619
 620     return ret;
 621 }
 622
 623 /* Called with no locks taken.  */
 624
 625 static int flush_blks(QEMUFile *f)
 626 {
 627     BlkMigBlock *blk;
 628     int ret = 0;
 629
 630     DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
 631             __func__, block_mig_state.submitted, block_mig_state.read_done,
 632             block_mig_state.transferred);
 633
 634     blk_mig_lock();
 635     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 636         if (qemu_file_rate_limit(f)) {
 637             break;
 638         }
 639         if (blk->ret < 0) {
 640             ret = blk->ret;
 641             break;
 642         }
 643
 644         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 645         blk_mig_unlock();
 646         blk_send(f, blk);
 647         blk_mig_lock();
 648
 649         g_free(blk->buf);
 650         g_free(blk);
 651
 652         block_mig_state.read_done--;
 653         block_mig_state.transferred++;
 654         assert(block_mig_state.read_done >= 0);
 655     }
 656     blk_mig_unlock();
 657
 658     DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __func__,
 659             block_mig_state.submitted, block_mig_state.read_done,
 660             block_mig_state.transferred);
 661     return ret;
 662 }
 663
 664 /* Called with iothread lock taken.  */
 665
 666 static int64_t get_remaining_dirty(void)
 667 {
 668     BlkMigDevState *bmds;
 669     int64_t dirty = 0;
 670
 671     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 672         aio_context_acquire(blk_get_aio_context(bmds->blk));
 673         dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
 674         aio_context_release(blk_get_aio_context(bmds->blk));
 675     }
 676
 677     return dirty;
 678 }
 679
 680
 681
 682 /* Called with iothread lock taken.  */
 683 static void block_migration_cleanup_bmds(void)
 684 {
 685     BlkMigDevState *bmds;
 686     AioContext *ctx;
 687
 688     unset_dirty_tracking();
 689
 690     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 691         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 692         bdrv_op_unblock_all(blk_bs(bmds->blk), bmds->blocker);
 693         error_free(bmds->blocker);
 694
 695         /* Save ctx, because bmds->blk can disappear during blk_unref.  */
 696         ctx = blk_get_aio_context(bmds->blk);
 697         aio_context_acquire(ctx);
 698         blk_unref(bmds->blk);
 699         aio_context_release(ctx);
 700
 701         g_free(bmds->blk_name);
 702         g_free(bmds->aio_bitmap);
 703         g_free(bmds);
 704     }
 705 }
 706
 707 /* Called with iothread lock taken.  */
 708 static void block_migration_cleanup(void *opaque)
 709 {
 710     BlkMigBlock *blk;
 711
 712     bdrv_drain_all();
 713
 714     block_migration_cleanup_bmds();
 715
 716     blk_mig_lock();
 717     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 718         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 719         g_free(blk->buf);
 720         g_free(blk);
 721     }
 722     blk_mig_unlock();
 723 }
 724
 725 static int block_save_setup(QEMUFile *f, void *opaque)
 726 {
 727     int ret;
 728
 729     DPRINTF("Enter save live setup submitted %d transferred %d\n",
 730             block_mig_state.submitted, block_mig_state.transferred);
 731
 732     qemu_mutex_lock_iothread();
 733     ret = init_blk_migration(f);
 734     if (ret < 0) {
 735         qemu_mutex_unlock_iothread();
 736         return ret;
 737     }
 738
 739     /* start track dirty blocks */
 740     ret = set_dirty_tracking();
 741
 742     qemu_mutex_unlock_iothread();
 743
 744     if (ret) {
 745         return ret;
 746     }
 747
 748     ret = flush_blks(f);
 749     blk_mig_reset_dirty_cursor();
 750     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 751
 752     return ret;
 753 }
 754
 755 static int block_save_iterate(QEMUFile *f, void *opaque)
 756 {
 757     int ret;
 758     int64_t last_ftell = qemu_ftell(f);
 759     int64_t delta_ftell;
 760
 761     DPRINTF("Enter save live iterate submitted %d transferred %d\n",
 762             block_mig_state.submitted, block_mig_state.transferred);
 763
 764     ret = flush_blks(f);
 765     if (ret) {
 766         return ret;
 767     }
 768
 769     blk_mig_reset_dirty_cursor();
 770
 771     /* control the rate of transfer */
 772     blk_mig_lock();
 773     while (block_mig_state.read_done * BLOCK_SIZE <
 774            qemu_file_get_rate_limit(f) &&
 775            block_mig_state.submitted < MAX_PARALLEL_IO &&
 776            (block_mig_state.submitted + block_mig_state.read_done) <
 777            MAX_IO_BUFFERS) {
 778         blk_mig_unlock();
 779         if (block_mig_state.bulk_completed == 0) {
 780             /* first finish the bulk phase */
 781             if (blk_mig_save_bulked_block(f) == 0) {
 782                 /* finished saving bulk on all devices */
 783                 block_mig_state.bulk_completed = 1;
 784             }
 785             ret = 0;
 786         } else {
 787             /* Always called with iothread lock taken for
 788              * simplicity, block_save_complete also calls it.
 789              */
 790             qemu_mutex_lock_iothread();
 791             ret = blk_mig_save_dirty_block(f, 1);
 792             qemu_mutex_unlock_iothread();
 793         }
 794         if (ret < 0) {
 795             return ret;
 796         }
 797         blk_mig_lock();
 798         if (ret != 0) {
 799             /* no more dirty blocks */
 800             break;
 801         }
 802     }
 803     blk_mig_unlock();
 804
 805     ret = flush_blks(f);
 806     if (ret) {
 807         return ret;
 808     }
 809
 810     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 811     delta_ftell = qemu_ftell(f) - last_ftell;
 812     if (delta_ftell > 0) {
 813         return 1;
 814     } else if (delta_ftell < 0) {
 815         return -1;
 816     } else {
 817         return 0;
 818     }
 819 }
 820
 821 /* Called with iothread lock taken.  */
 822
 823 static int block_save_complete(QEMUFile *f, void *opaque)
 824 {
 825     int ret;
 826
 827     DPRINTF("Enter save live complete submitted %d transferred %d\n",
 828             block_mig_state.submitted, block_mig_state.transferred);
 829
 830     ret = flush_blks(f);
 831     if (ret) {
 832         return ret;
 833     }
 834
 835     blk_mig_reset_dirty_cursor();
 836
 837     /* we know for sure that save bulk is completed and
 838        all async read completed */
 839     blk_mig_lock();
 840     assert(block_mig_state.submitted == 0);
 841     blk_mig_unlock();
 842
 843     do {
 844         ret = blk_mig_save_dirty_block(f, 0);
 845         if (ret < 0) {
 846             return ret;
 847         }
 848     } while (ret == 0);
 849
 850     /* report completion */
 851     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 852
 853     DPRINTF("Block migration completed\n");
 854
 855     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 856
 857     /* Make sure that our BlockBackends are gone, so that the block driver
 858      * nodes can be inactivated. */
 859     block_migration_cleanup_bmds();
 860
 861     return 0;
 862 }
 863
 864 static void block_save_pending(QEMUFile *f, void *opaque, uint64_t max_size,
 865                                uint64_t *res_precopy_only,
 866                                uint64_t *res_compatible,
 867                                uint64_t *res_postcopy_only)
 868 {
 869     /* Estimate pending number of bytes to send */
 870     uint64_t pending;
 871
 872     qemu_mutex_lock_iothread();
 873     pending = get_remaining_dirty();
 874     qemu_mutex_unlock_iothread();
 875
 876     blk_mig_lock();
 877     pending += block_mig_state.submitted * BLOCK_SIZE +
 878                block_mig_state.read_done * BLOCK_SIZE;
 879     blk_mig_unlock();
 880
 881     /* Report at least one block pending during bulk phase */
 882     if (pending <= max_size && !block_mig_state.bulk_completed) {
 883         pending = max_size + BLOCK_SIZE;
 884     }
 885
 886     DPRINTF("Enter save live pending  %" PRIu64 "\n", pending);
 887     /* We don't do postcopy */
 888     *res_precopy_only += pending;
 889 }
 890
 891 static int block_load(QEMUFile *f, void *opaque, int version_id)
 892 {
 893     static int banner_printed;
 894     int len, flags;
 895     char device_name[256];
 896     int64_t addr;
 897     BlockBackend *blk, *blk_prev = NULL;
 898     Error *local_err = NULL;
 899     uint8_t *buf;
 900     int64_t total_sectors = 0;
 901     int nr_sectors;
 902     int ret;
 903     BlockDriverInfo bdi;
 904     int cluster_size = BLOCK_SIZE;
 905
 906     do {
 907         addr = qemu_get_be64(f);
 908
 909         flags = addr & ~BDRV_SECTOR_MASK;
 910         addr >>= BDRV_SECTOR_BITS;
 911
 912         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 913             /* get device name */
 914             len = qemu_get_byte(f);
 915             qemu_get_buffer(f, (uint8_t *)device_name, len);
 916             device_name[len] = '\0';
 917
 918             blk = blk_by_name(device_name);
 919             if (!blk) {
 920                 fprintf(stderr, "Error unknown block device %s\n",
 921                         device_name);
 922                 return -EINVAL;
 923             }
 924
 925             if (blk != blk_prev) {
 926                 blk_prev = blk;
 927                 total_sectors = blk_nb_sectors(blk);
 928                 if (total_sectors <= 0) {
 929                     error_report("Error getting length of block device %s",
 930                                  device_name);
 931                     return -EINVAL;
 932                 }
 933
 934                 blk_invalidate_cache(blk, &local_err);
 935                 if (local_err) {
 936                     error_report_err(local_err);
 937                     return -EINVAL;
 938                 }
 939
 940                 ret = bdrv_get_info(blk_bs(blk), &bdi);
 941                 if (ret == 0 && bdi.cluster_size > 0 &&
 942                     bdi.cluster_size <= BLOCK_SIZE &&
 943                     BLOCK_SIZE % bdi.cluster_size == 0) {
 944                     cluster_size = bdi.cluster_size;
 945                 } else {
 946                     cluster_size = BLOCK_SIZE;
 947                 }
 948             }
 949
 950             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 951                 nr_sectors = total_sectors - addr;
 952             } else {
 953                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 954             }
 955
 956             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 957                 ret = blk_pwrite_zeroes(blk, addr * BDRV_SECTOR_SIZE,
 958                                         nr_sectors * BDRV_SECTOR_SIZE,
 959                                         BDRV_REQ_MAY_UNMAP);
 960             } else {
 961                 int i;
 962                 int64_t cur_addr;
 963                 uint8_t *cur_buf;
 964
 965                 buf = g_malloc(BLOCK_SIZE);
 966                 qemu_get_buffer(f, buf, BLOCK_SIZE);
 967                 for (i = 0; i < BLOCK_SIZE / cluster_size; i++) {
 968                     cur_addr = addr * BDRV_SECTOR_SIZE + i * cluster_size;
 969                     cur_buf = buf + i * cluster_size;
 970
 971                     if ((!block_mig_state.zero_blocks ||
 972                         cluster_size < BLOCK_SIZE) &&
 973                         buffer_is_zero(cur_buf, cluster_size)) {
 974                         ret = blk_pwrite_zeroes(blk, cur_addr,
 975                                                 cluster_size,
 976                                                 BDRV_REQ_MAY_UNMAP);
 977                     } else {
 978                         ret = blk_pwrite(blk, cur_addr, cur_buf,
 979                                          cluster_size, 0);
 980                     }
 981                     if (ret < 0) {
 982                         break;
 983                     }
 984                 }
 985                 g_free(buf);
 986             }
 987
 988             if (ret < 0) {
 989                 return ret;
 990             }
 991         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 992             if (!banner_printed) {
 993                 printf("Receiving block device images\n");
 994                 banner_printed = 1;
 995             }
 996             printf("Completed %d %%%c", (int)addr,
 997                    (addr == 100) ? '\n' : '\r');
 998             fflush(stdout);
 999         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
1000             fprintf(stderr, "Unknown block migration flags: %#x\n", flags);
1001             return -EINVAL;
1002         }
1003         ret = qemu_file_get_error(f);
1004         if (ret != 0) {
1005             return ret;
1006         }
1007     } while (!(flags & BLK_MIG_FLAG_EOS));
1008
1009     return 0;
1010 }
1011
1012 static bool block_is_active(void *opaque)
1013 {
1014     return migrate_use_block();
1015 }
1016
1017 static SaveVMHandlers savevm_block_handlers = {
1018     .save_setup = block_save_setup,
1019     .save_live_iterate = block_save_iterate,
1020     .save_live_complete_precopy = block_save_complete,
1021     .save_live_pending = block_save_pending,
1022     .load_state = block_load,
1023     .save_cleanup = block_migration_cleanup,
1024     .is_active = block_is_active,
1025 };
1026
1027 void blk_mig_init(void)
1028 {
1029     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
1030     QSIMPLEQ_INIT(&block_mig_state.blk_list);
1031     qemu_mutex_init(&block_mig_state.lock);
1032
1033     register_savevm_live(NULL, "block", 0, 1, &savevm_block_handlers,
1034                          &block_mig_state);
1035 }