migration/block.c

   1 /*
   2  * QEMU live block migration
   3  *
   4  * Copyright IBM, Corp. 2009
   5  *
   6  * Authors:
   7  *  Liran Schour   <lirans@il.ibm.com>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.  See
  10  * the COPYING file in the top-level directory.
  11  *
  12  * Contributions after 2012-01-13 are licensed under the terms of the
  13  * GNU GPL, version 2 or (at your option) any later version.
  14  */
  15
  16 #include "qemu/osdep.h"
  17 #include "qapi/error.h"
  18 #include "qemu/error-report.h"
  19 #include "qemu/cutils.h"
  20 #include "qemu/queue.h"
  21 #include "block.h"
  22 #include "migration/misc.h"
  23 #include "migration.h"
  24 #include "migration/register.h"
  25 #include "qemu-file.h"
  26 #include "migration/vmstate.h"
  27 #include "sysemu/block-backend.h"
  28
  29 #define BLOCK_SIZE                       (1 << 20)
  30 #define BDRV_SECTORS_PER_DIRTY_CHUNK     (BLOCK_SIZE >> BDRV_SECTOR_BITS)
  31
  32 #define BLK_MIG_FLAG_DEVICE_BLOCK       0x01
  33 #define BLK_MIG_FLAG_EOS                0x02
  34 #define BLK_MIG_FLAG_PROGRESS           0x04
  35 #define BLK_MIG_FLAG_ZERO_BLOCK         0x08
  36
  37 #define MAX_IS_ALLOCATED_SEARCH (65536 * BDRV_SECTOR_SIZE)
  38
  39 #define MAX_INFLIGHT_IO 512
  40
  41 //#define DEBUG_BLK_MIGRATION
  42
  43 #ifdef DEBUG_BLK_MIGRATION
  44 #define DPRINTF(fmt, ...) \
  45     do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
  46 #else
  47 #define DPRINTF(fmt, ...) \
  48     do { } while (0)
  49 #endif
  50
  51 typedef struct BlkMigDevState {
  52     /* Written during setup phase.  Can be read without a lock.  */
  53     BlockBackend *blk;
  54     char *blk_name;
  55     int shared_base;
  56     int64_t total_sectors;
  57     QSIMPLEQ_ENTRY(BlkMigDevState) entry;
  58     Error *blocker;
  59
  60     /* Only used by migration thread.  Does not need a lock.  */
  61     int bulk_completed;
  62     int64_t cur_sector;
  63     int64_t cur_dirty;
  64
  65     /* Data in the aio_bitmap is protected by block migration lock.
  66      * Allocation and free happen during setup and cleanup respectively.
  67      */
  68     unsigned long *aio_bitmap;
  69
  70     /* Protected by block migration lock.  */
  71     int64_t completed_sectors;
  72
  73     /* During migration this is protected by iothread lock / AioContext.
  74      * Allocation and free happen during setup and cleanup respectively.
  75      */
  76     BdrvDirtyBitmap *dirty_bitmap;
  77 } BlkMigDevState;
  78
  79 typedef struct BlkMigBlock {
  80     /* Only used by migration thread.  */
  81     uint8_t *buf;
  82     BlkMigDevState *bmds;
  83     int64_t sector;
  84     int nr_sectors;
  85     struct iovec iov;
  86     QEMUIOVector qiov;
  87     BlockAIOCB *aiocb;
  88
  89     /* Protected by block migration lock.  */
  90     int ret;
  91     QSIMPLEQ_ENTRY(BlkMigBlock) entry;
  92 } BlkMigBlock;
  93
  94 typedef struct BlkMigState {
  95     QSIMPLEQ_HEAD(bmds_list, BlkMigDevState) bmds_list;
  96     int64_t total_sector_sum;
  97     bool zero_blocks;
  98
  99     /* Protected by lock.  */
 100     QSIMPLEQ_HEAD(blk_list, BlkMigBlock) blk_list;
 101     int submitted;
 102     int read_done;
 103
 104     /* Only used by migration thread.  Does not need a lock.  */
 105     int transferred;
 106     int prev_progress;
 107     int bulk_completed;
 108
 109     /* Lock must be taken _inside_ the iothread lock and any AioContexts.  */
 110     QemuMutex lock;
 111 } BlkMigState;
 112
 113 static BlkMigState block_mig_state;
 114
 115 static void blk_mig_lock(void)
 116 {
 117     qemu_mutex_lock(&block_mig_state.lock);
 118 }
 119
 120 static void blk_mig_unlock(void)
 121 {
 122     qemu_mutex_unlock(&block_mig_state.lock);
 123 }
 124
 125 /* Must run outside of the iothread lock during the bulk phase,
 126  * or the VM will stall.
 127  */
 128
 129 static void blk_send(QEMUFile *f, BlkMigBlock * blk)
 130 {
 131     int len;
 132     uint64_t flags = BLK_MIG_FLAG_DEVICE_BLOCK;
 133
 134     if (block_mig_state.zero_blocks &&
 135         buffer_is_zero(blk->buf, BLOCK_SIZE)) {
 136         flags |= BLK_MIG_FLAG_ZERO_BLOCK;
 137     }
 138
 139     /* sector number and flags */
 140     qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
 141                      | flags);
 142
 143     /* device name */
 144     len = strlen(blk->bmds->blk_name);
 145     qemu_put_byte(f, len);
 146     qemu_put_buffer(f, (uint8_t *) blk->bmds->blk_name, len);
 147
 148     /* if a block is zero we need to flush here since the network
 149      * bandwidth is now a lot higher than the storage device bandwidth.
 150      * thus if we queue zero blocks we slow down the migration */
 151     if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 152         qemu_fflush(f);
 153         return;
 154     }
 155
 156     qemu_put_buffer(f, blk->buf, BLOCK_SIZE);
 157 }
 158
 159 int blk_mig_active(void)
 160 {
 161     return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
 162 }
 163
 164 int blk_mig_bulk_active(void)
 165 {
 166     return blk_mig_active() && !block_mig_state.bulk_completed;
 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     int64_t count;
 276
 277     if (bmds->shared_base) {
 278         qemu_mutex_lock_iothread();
 279         aio_context_acquire(blk_get_aio_context(bb));
 280         /* Skip unallocated sectors; intentionally treats failure or
 281          * partial sector as an allocated sector */
 282         while (cur_sector < total_sectors &&
 283                !bdrv_is_allocated(blk_bs(bb), cur_sector * BDRV_SECTOR_SIZE,
 284                                   MAX_IS_ALLOCATED_SEARCH, &count)) {
 285             if (count < BDRV_SECTOR_SIZE) {
 286                 break;
 287             }
 288             cur_sector += count >> BDRV_SECTOR_BITS;
 289         }
 290         aio_context_release(blk_get_aio_context(bb));
 291         qemu_mutex_unlock_iothread();
 292     }
 293
 294     if (cur_sector >= total_sectors) {
 295         bmds->cur_sector = bmds->completed_sectors = total_sectors;
 296         return 1;
 297     }
 298
 299     bmds->completed_sectors = cur_sector;
 300
 301     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
 302
 303     /* we are going to transfer a full block even if it is not allocated */
 304     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 305
 306     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 307         nr_sectors = total_sectors - cur_sector;
 308     }
 309
 310     blk = g_new(BlkMigBlock, 1);
 311     blk->buf = g_malloc(BLOCK_SIZE);
 312     blk->bmds = bmds;
 313     blk->sector = cur_sector;
 314     blk->nr_sectors = nr_sectors;
 315
 316     blk->iov.iov_base = blk->buf;
 317     blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 318     qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 319
 320     blk_mig_lock();
 321     block_mig_state.submitted++;
 322     blk_mig_unlock();
 323
 324     /* We do not know if bs is under the main thread (and thus does
 325      * not acquire the AioContext when doing AIO) or rather under
 326      * dataplane.  Thus acquire both the iothread mutex and the
 327      * AioContext.
 328      *
 329      * This is ugly and will disappear when we make bdrv_* thread-safe,
 330      * without the need to acquire the AioContext.
 331      */
 332     qemu_mutex_lock_iothread();
 333     aio_context_acquire(blk_get_aio_context(bmds->blk));
 334     blk->aiocb = blk_aio_preadv(bb, cur_sector * BDRV_SECTOR_SIZE, &blk->qiov,
 335                                 0, blk_mig_read_cb, blk);
 336
 337     bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, cur_sector, nr_sectors);
 338     aio_context_release(blk_get_aio_context(bmds->blk));
 339     qemu_mutex_unlock_iothread();
 340
 341     bmds->cur_sector = cur_sector + nr_sectors;
 342     return (bmds->cur_sector >= total_sectors);
 343 }
 344
 345 /* Called with iothread lock taken.  */
 346
 347 static int set_dirty_tracking(void)
 348 {
 349     BlkMigDevState *bmds;
 350     int ret;
 351
 352     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 353         bmds->dirty_bitmap = bdrv_create_dirty_bitmap(blk_bs(bmds->blk),
 354                                                       BLOCK_SIZE, NULL, NULL);
 355         if (!bmds->dirty_bitmap) {
 356             ret = -errno;
 357             goto fail;
 358         }
 359     }
 360     return 0;
 361
 362 fail:
 363     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 364         if (bmds->dirty_bitmap) {
 365             bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
 366         }
 367     }
 368     return ret;
 369 }
 370
 371 /* Called with iothread lock taken.  */
 372
 373 static void unset_dirty_tracking(void)
 374 {
 375     BlkMigDevState *bmds;
 376
 377     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 378         bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
 379     }
 380 }
 381
 382 static int 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     Error *local_err = NULL;
 394     int ret;
 395
 396     block_mig_state.submitted = 0;
 397     block_mig_state.read_done = 0;
 398     block_mig_state.transferred = 0;
 399     block_mig_state.total_sector_sum = 0;
 400     block_mig_state.prev_progress = -1;
 401     block_mig_state.bulk_completed = 0;
 402     block_mig_state.zero_blocks = migrate_zero_blocks();
 403
 404     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
 405         num_bs++;
 406     }
 407     bmds_bs = g_malloc0(num_bs * sizeof(*bmds_bs));
 408
 409     for (i = 0, bs = bdrv_first(&it); bs; bs = bdrv_next(&it), i++) {
 410         if (bdrv_is_read_only(bs)) {
 411             continue;
 412         }
 413
 414         sectors = bdrv_nb_sectors(bs);
 415         if (sectors <= 0) {
 416             ret = sectors;
 417             goto out;
 418         }
 419
 420         bmds = g_new0(BlkMigDevState, 1);
 421         bmds->blk = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 422         bmds->blk_name = g_strdup(bdrv_get_device_name(bs));
 423         bmds->bulk_completed = 0;
 424         bmds->total_sectors = sectors;
 425         bmds->completed_sectors = 0;
 426         bmds->shared_base = migrate_use_block_incremental();
 427
 428         assert(i < num_bs);
 429         bmds_bs[i].bmds = bmds;
 430         bmds_bs[i].bs = bs;
 431
 432         block_mig_state.total_sector_sum += sectors;
 433
 434         if (bmds->shared_base) {
 435             DPRINTF("Start migration for %s with shared base image\n",
 436                     bdrv_get_device_name(bs));
 437         } else {
 438             DPRINTF("Start full migration for %s\n", bdrv_get_device_name(bs));
 439         }
 440
 441         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 442     }
 443
 444     /* Can only insert new BDSes now because doing so while iterating block
 445      * devices may end up in a deadlock (iterating the new BDSes, too). */
 446     for (i = 0; i < num_bs; i++) {
 447         BlkMigDevState *bmds = bmds_bs[i].bmds;
 448         BlockDriverState *bs = bmds_bs[i].bs;
 449
 450         if (bmds) {
 451             ret = blk_insert_bs(bmds->blk, bs, &local_err);
 452             if (ret < 0) {
 453                 error_report_err(local_err);
 454                 goto out;
 455             }
 456
 457             alloc_aio_bitmap(bmds);
 458             error_setg(&bmds->blocker, "block device is in use by migration");
 459             bdrv_op_block_all(bs, bmds->blocker);
 460         }
 461     }
 462
 463     ret = 0;
 464 out:
 465     g_free(bmds_bs);
 466     return ret;
 467 }
 468
 469 /* Called with no lock taken.  */
 470
 471 static int blk_mig_save_bulked_block(QEMUFile *f)
 472 {
 473     int64_t completed_sector_sum = 0;
 474     BlkMigDevState *bmds;
 475     int progress;
 476     int ret = 0;
 477
 478     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 479         if (bmds->bulk_completed == 0) {
 480             if (mig_save_device_bulk(f, bmds) == 1) {
 481                 /* completed bulk section for this device */
 482                 bmds->bulk_completed = 1;
 483             }
 484             completed_sector_sum += bmds->completed_sectors;
 485             ret = 1;
 486             break;
 487         } else {
 488             completed_sector_sum += bmds->completed_sectors;
 489         }
 490     }
 491
 492     if (block_mig_state.total_sector_sum != 0) {
 493         progress = completed_sector_sum * 100 /
 494                    block_mig_state.total_sector_sum;
 495     } else {
 496         progress = 100;
 497     }
 498     if (progress != block_mig_state.prev_progress) {
 499         block_mig_state.prev_progress = progress;
 500         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 501                          | BLK_MIG_FLAG_PROGRESS);
 502         DPRINTF("Completed %d %%\r", progress);
 503     }
 504
 505     return ret;
 506 }
 507
 508 static void blk_mig_reset_dirty_cursor(void)
 509 {
 510     BlkMigDevState *bmds;
 511
 512     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 513         bmds->cur_dirty = 0;
 514     }
 515 }
 516
 517 /* Called with iothread lock and AioContext taken.  */
 518
 519 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
 520                                  int is_async)
 521 {
 522     BlkMigBlock *blk;
 523     BlockDriverState *bs = blk_bs(bmds->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_get_dirty_locked(bs, bmds->dirty_bitmap, sector)) {
 539             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 540                 nr_sectors = total_sectors - sector;
 541             } else {
 542                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 543             }
 544             bdrv_reset_dirty_bitmap_locked(bmds->dirty_bitmap, sector, nr_sectors);
 545             bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 546
 547             blk = g_new(BlkMigBlock, 1);
 548             blk->buf = g_malloc(BLOCK_SIZE);
 549             blk->bmds = bmds;
 550             blk->sector = sector;
 551             blk->nr_sectors = nr_sectors;
 552
 553             if (is_async) {
 554                 blk->iov.iov_base = blk->buf;
 555                 blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 556                 qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 557
 558                 blk->aiocb = blk_aio_preadv(bmds->blk,
 559                                             sector * BDRV_SECTOR_SIZE,
 560                                             &blk->qiov, 0, blk_mig_read_cb,
 561                                             blk);
 562
 563                 blk_mig_lock();
 564                 block_mig_state.submitted++;
 565                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 566                 blk_mig_unlock();
 567             } else {
 568                 ret = blk_pread(bmds->blk, sector * BDRV_SECTOR_SIZE, blk->buf,
 569                                 nr_sectors * BDRV_SECTOR_SIZE);
 570                 if (ret < 0) {
 571                     goto error;
 572                 }
 573                 blk_send(f, blk);
 574
 575                 g_free(blk->buf);
 576                 g_free(blk);
 577             }
 578
 579             sector += nr_sectors;
 580             bmds->cur_dirty = sector;
 581             break;
 582         }
 583
 584         bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 585         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 586         bmds->cur_dirty = sector;
 587     }
 588
 589     return (bmds->cur_dirty >= bmds->total_sectors);
 590
 591 error:
 592     DPRINTF("Error reading sector %" PRId64 "\n", sector);
 593     g_free(blk->buf);
 594     g_free(blk);
 595     return ret;
 596 }
 597
 598 /* Called with iothread lock taken.
 599  *
 600  * return value:
 601  * 0: too much data for max_downtime
 602  * 1: few enough data for max_downtime
 603 */
 604 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 605 {
 606     BlkMigDevState *bmds;
 607     int ret = 1;
 608
 609     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 610         aio_context_acquire(blk_get_aio_context(bmds->blk));
 611         ret = mig_save_device_dirty(f, bmds, is_async);
 612         aio_context_release(blk_get_aio_context(bmds->blk));
 613         if (ret <= 0) {
 614             break;
 615         }
 616     }
 617
 618     return ret;
 619 }
 620
 621 /* Called with no locks taken.  */
 622
 623 static int flush_blks(QEMUFile *f)
 624 {
 625     BlkMigBlock *blk;
 626     int ret = 0;
 627
 628     DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
 629             __FUNCTION__, block_mig_state.submitted, block_mig_state.read_done,
 630             block_mig_state.transferred);
 631
 632     blk_mig_lock();
 633     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 634         if (qemu_file_rate_limit(f)) {
 635             break;
 636         }
 637         if (blk->ret < 0) {
 638             ret = blk->ret;
 639             break;
 640         }
 641
 642         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 643         blk_mig_unlock();
 644         blk_send(f, blk);
 645         blk_mig_lock();
 646
 647         g_free(blk->buf);
 648         g_free(blk);
 649
 650         block_mig_state.read_done--;
 651         block_mig_state.transferred++;
 652         assert(block_mig_state.read_done >= 0);
 653     }
 654     blk_mig_unlock();
 655
 656     DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__,
 657             block_mig_state.submitted, block_mig_state.read_done,
 658             block_mig_state.transferred);
 659     return ret;
 660 }
 661
 662 /* Called with iothread lock taken.  */
 663
 664 static int64_t get_remaining_dirty(void)
 665 {
 666     BlkMigDevState *bmds;
 667     int64_t dirty = 0;
 668
 669     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 670         aio_context_acquire(blk_get_aio_context(bmds->blk));
 671         dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
 672         aio_context_release(blk_get_aio_context(bmds->blk));
 673     }
 674
 675     return dirty << BDRV_SECTOR_BITS;
 676 }
 677
 678
 679
 680 /* Called with iothread lock taken.  */
 681 static void block_migration_cleanup_bmds(void)
 682 {
 683     BlkMigDevState *bmds;
 684     AioContext *ctx;
 685
 686     unset_dirty_tracking();
 687
 688     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 689         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 690         bdrv_op_unblock_all(blk_bs(bmds->blk), bmds->blocker);
 691         error_free(bmds->blocker);
 692
 693         /* Save ctx, because bmds->blk can disappear during blk_unref.  */
 694         ctx = blk_get_aio_context(bmds->blk);
 695         aio_context_acquire(ctx);
 696         blk_unref(bmds->blk);
 697         aio_context_release(ctx);
 698
 699         g_free(bmds->blk_name);
 700         g_free(bmds->aio_bitmap);
 701         g_free(bmds);
 702     }
 703 }
 704
 705 /* Called with iothread lock taken.  */
 706 static void block_migration_cleanup(void *opaque)
 707 {
 708     BlkMigBlock *blk;
 709
 710     bdrv_drain_all();
 711
 712     block_migration_cleanup_bmds();
 713
 714     blk_mig_lock();
 715     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 716         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 717         g_free(blk->buf);
 718         g_free(blk);
 719     }
 720     blk_mig_unlock();
 721 }
 722
 723 static int block_save_setup(QEMUFile *f, void *opaque)
 724 {
 725     int ret;
 726
 727     DPRINTF("Enter save live setup submitted %d transferred %d\n",
 728             block_mig_state.submitted, block_mig_state.transferred);
 729
 730     qemu_mutex_lock_iothread();
 731     ret = init_blk_migration(f);
 732     if (ret < 0) {
 733         qemu_mutex_unlock_iothread();
 734         return ret;
 735     }
 736
 737     /* start track dirty blocks */
 738     ret = set_dirty_tracking();
 739
 740     qemu_mutex_unlock_iothread();
 741
 742     if (ret) {
 743         return ret;
 744     }
 745
 746     ret = flush_blks(f);
 747     blk_mig_reset_dirty_cursor();
 748     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 749
 750     return ret;
 751 }
 752
 753 static int block_save_iterate(QEMUFile *f, void *opaque)
 754 {
 755     int ret;
 756     int64_t last_ftell = qemu_ftell(f);
 757     int64_t delta_ftell;
 758
 759     DPRINTF("Enter save live iterate submitted %d transferred %d\n",
 760             block_mig_state.submitted, block_mig_state.transferred);
 761
 762     ret = flush_blks(f);
 763     if (ret) {
 764         return ret;
 765     }
 766
 767     blk_mig_reset_dirty_cursor();
 768
 769     /* control the rate of transfer */
 770     blk_mig_lock();
 771     while ((block_mig_state.submitted +
 772             block_mig_state.read_done) * BLOCK_SIZE <
 773            qemu_file_get_rate_limit(f) &&
 774            (block_mig_state.submitted +
 775             block_mig_state.read_done) <
 776            MAX_INFLIGHT_IO) {
 777         blk_mig_unlock();
 778         if (block_mig_state.bulk_completed == 0) {
 779             /* first finish the bulk phase */
 780             if (blk_mig_save_bulked_block(f) == 0) {
 781                 /* finished saving bulk on all devices */
 782                 block_mig_state.bulk_completed = 1;
 783             }
 784             ret = 0;
 785         } else {
 786             /* Always called with iothread lock taken for
 787              * simplicity, block_save_complete also calls it.
 788              */
 789             qemu_mutex_lock_iothread();
 790             ret = blk_mig_save_dirty_block(f, 1);
 791             qemu_mutex_unlock_iothread();
 792         }
 793         if (ret < 0) {
 794             return ret;
 795         }
 796         blk_mig_lock();
 797         if (ret != 0) {
 798             /* no more dirty blocks */
 799             break;
 800         }
 801     }
 802     blk_mig_unlock();
 803
 804     ret = flush_blks(f);
 805     if (ret) {
 806         return ret;
 807     }
 808
 809     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 810     delta_ftell = qemu_ftell(f) - last_ftell;
 811     if (delta_ftell > 0) {
 812         return 1;
 813     } else if (delta_ftell < 0) {
 814         return -1;
 815     } else {
 816         return 0;
 817     }
 818 }
 819
 820 /* Called with iothread lock taken.  */
 821
 822 static int block_save_complete(QEMUFile *f, void *opaque)
 823 {
 824     int ret;
 825
 826     DPRINTF("Enter save live complete submitted %d transferred %d\n",
 827             block_mig_state.submitted, block_mig_state.transferred);
 828
 829     ret = flush_blks(f);
 830     if (ret) {
 831         return ret;
 832     }
 833
 834     blk_mig_reset_dirty_cursor();
 835
 836     /* we know for sure that save bulk is completed and
 837        all async read completed */
 838     blk_mig_lock();
 839     assert(block_mig_state.submitted == 0);
 840     blk_mig_unlock();
 841
 842     do {
 843         ret = blk_mig_save_dirty_block(f, 0);
 844         if (ret < 0) {
 845             return ret;
 846         }
 847     } while (ret == 0);
 848
 849     /* report completion */
 850     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 851
 852     DPRINTF("Block migration completed\n");
 853
 854     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 855
 856     /* Make sure that our BlockBackends are gone, so that the block driver
 857      * nodes can be inactivated. */
 858     block_migration_cleanup_bmds();
 859
 860     return 0;
 861 }
 862
 863 static void block_save_pending(QEMUFile *f, void *opaque, uint64_t max_size,
 864                                uint64_t *non_postcopiable_pending,
 865                                uint64_t *postcopiable_pending)
 866 {
 867     /* Estimate pending number of bytes to send */
 868     uint64_t pending;
 869
 870     qemu_mutex_lock_iothread();
 871     pending = get_remaining_dirty();
 872     qemu_mutex_unlock_iothread();
 873
 874     blk_mig_lock();
 875     pending += block_mig_state.submitted * BLOCK_SIZE +
 876                block_mig_state.read_done * BLOCK_SIZE;
 877     blk_mig_unlock();
 878
 879     /* Report at least one block pending during bulk phase */
 880     if (pending <= max_size && !block_mig_state.bulk_completed) {
 881         pending = max_size + BLOCK_SIZE;
 882     }
 883
 884     DPRINTF("Enter save live pending  %" PRIu64 "\n", pending);
 885     /* We don't do postcopy */
 886     *non_postcopiable_pending += pending;
 887 }
 888
 889 static int block_load(QEMUFile *f, void *opaque, int version_id)
 890 {
 891     static int banner_printed;
 892     int len, flags;
 893     char device_name[256];
 894     int64_t addr;
 895     BlockBackend *blk, *blk_prev = NULL;;
 896     Error *local_err = NULL;
 897     uint8_t *buf;
 898     int64_t total_sectors = 0;
 899     int nr_sectors;
 900     int ret;
 901     BlockDriverInfo bdi;
 902     int cluster_size = BLOCK_SIZE;
 903
 904     do {
 905         addr = qemu_get_be64(f);
 906
 907         flags = addr & ~BDRV_SECTOR_MASK;
 908         addr >>= BDRV_SECTOR_BITS;
 909
 910         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 911             /* get device name */
 912             len = qemu_get_byte(f);
 913             qemu_get_buffer(f, (uint8_t *)device_name, len);
 914             device_name[len] = '\0';
 915
 916             blk = blk_by_name(device_name);
 917             if (!blk) {
 918                 fprintf(stderr, "Error unknown block device %s\n",
 919                         device_name);
 920                 return -EINVAL;
 921             }
 922
 923             if (blk != blk_prev) {
 924                 blk_prev = blk;
 925                 total_sectors = blk_nb_sectors(blk);
 926                 if (total_sectors <= 0) {
 927                     error_report("Error getting length of block device %s",
 928                                  device_name);
 929                     return -EINVAL;
 930                 }
 931
 932                 blk_invalidate_cache(blk, &local_err);
 933                 if (local_err) {
 934                     error_report_err(local_err);
 935                     return -EINVAL;
 936                 }
 937
 938                 ret = bdrv_get_info(blk_bs(blk), &bdi);
 939                 if (ret == 0 && bdi.cluster_size > 0 &&
 940                     bdi.cluster_size <= BLOCK_SIZE &&
 941                     BLOCK_SIZE % bdi.cluster_size == 0) {
 942                     cluster_size = bdi.cluster_size;
 943                 } else {
 944                     cluster_size = BLOCK_SIZE;
 945                 }
 946             }
 947
 948             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 949                 nr_sectors = total_sectors - addr;
 950             } else {
 951                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 952             }
 953
 954             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 955                 ret = blk_pwrite_zeroes(blk, addr * BDRV_SECTOR_SIZE,
 956                                         nr_sectors * BDRV_SECTOR_SIZE,
 957                                         BDRV_REQ_MAY_UNMAP);
 958             } else {
 959                 int i;
 960                 int64_t cur_addr;
 961                 uint8_t *cur_buf;
 962
 963                 buf = g_malloc(BLOCK_SIZE);
 964                 qemu_get_buffer(f, buf, BLOCK_SIZE);
 965                 for (i = 0; i < BLOCK_SIZE / cluster_size; i++) {
 966                     cur_addr = addr * BDRV_SECTOR_SIZE + i * cluster_size;
 967                     cur_buf = buf + i * cluster_size;
 968
 969                     if ((!block_mig_state.zero_blocks ||
 970                         cluster_size < BLOCK_SIZE) &&
 971                         buffer_is_zero(cur_buf, cluster_size)) {
 972                         ret = blk_pwrite_zeroes(blk, cur_addr,
 973                                                 cluster_size,
 974                                                 BDRV_REQ_MAY_UNMAP);
 975                     } else {
 976                         ret = blk_pwrite(blk, cur_addr, cur_buf,
 977                                          cluster_size, 0);
 978                     }
 979                     if (ret < 0) {
 980                         break;
 981                     }
 982                 }
 983                 g_free(buf);
 984             }
 985
 986             if (ret < 0) {
 987                 return ret;
 988             }
 989         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 990             if (!banner_printed) {
 991                 printf("Receiving block device images\n");
 992                 banner_printed = 1;
 993             }
 994             printf("Completed %d %%%c", (int)addr,
 995                    (addr == 100) ? '\n' : '\r');
 996             fflush(stdout);
 997         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
 998             fprintf(stderr, "Unknown block migration flags: %#x\n", flags);
 999             return -EINVAL;
1000         }
1001         ret = qemu_file_get_error(f);
1002         if (ret != 0) {
1003             return ret;
1004         }
1005     } while (!(flags & BLK_MIG_FLAG_EOS));
1006
1007     return 0;
1008 }
1009
1010 static bool block_is_active(void *opaque)
1011 {
1012     return migrate_use_block();
1013 }
1014
1015 static SaveVMHandlers savevm_block_handlers = {
1016     .save_setup = block_save_setup,
1017     .save_live_iterate = block_save_iterate,
1018     .save_live_complete_precopy = block_save_complete,
1019     .save_live_pending = block_save_pending,
1020     .load_state = block_load,
1021     .save_cleanup = block_migration_cleanup,
1022     .is_active = block_is_active,
1023 };
1024
1025 void blk_mig_init(void)
1026 {
1027     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
1028     QSIMPLEQ_INIT(&block_mig_state.blk_list);
1029     qemu_mutex_init(&block_mig_state.lock);
1030
1031     register_savevm_live(NULL, "block", 0, 1, &savevm_block_handlers,
1032                          &block_mig_state);
1033 }