block-migration.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  */
  13
  14 #include "qemu-common.h"
  15 #include "block_int.h"
  16 #include "hw/hw.h"
  17 #include "qemu-queue.h"
  18 #include "qemu-timer.h"
  19 #include "monitor.h"
  20 #include "block-migration.h"
  21 #include "migration.h"
  22 #include <assert.h>
  23
  24 #define BLOCK_SIZE (BDRV_SECTORS_PER_DIRTY_CHUNK << BDRV_SECTOR_BITS)
  25
  26 #define BLK_MIG_FLAG_DEVICE_BLOCK       0x01
  27 #define BLK_MIG_FLAG_EOS                0x02
  28 #define BLK_MIG_FLAG_PROGRESS           0x04
  29
  30 #define MAX_IS_ALLOCATED_SEARCH 65536
  31
  32 //#define DEBUG_BLK_MIGRATION
  33
  34 #ifdef DEBUG_BLK_MIGRATION
  35 #define DPRINTF(fmt, ...) \
  36     do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
  37 #else
  38 #define DPRINTF(fmt, ...) \
  39     do { } while (0)
  40 #endif
  41
  42 typedef struct BlkMigDevState {
  43     BlockDriverState *bs;
  44     int bulk_completed;
  45     int shared_base;
  46     int64_t cur_sector;
  47     int64_t cur_dirty;
  48     int64_t completed_sectors;
  49     int64_t total_sectors;
  50     int64_t dirty;
  51     QSIMPLEQ_ENTRY(BlkMigDevState) entry;
  52 } BlkMigDevState;
  53
  54 typedef struct BlkMigBlock {
  55     uint8_t *buf;
  56     BlkMigDevState *bmds;
  57     int64_t sector;
  58     struct iovec iov;
  59     QEMUIOVector qiov;
  60     BlockDriverAIOCB *aiocb;
  61     int ret;
  62     int64_t time;
  63     QSIMPLEQ_ENTRY(BlkMigBlock) entry;
  64 } BlkMigBlock;
  65
  66 typedef struct BlkMigState {
  67     int blk_enable;
  68     int shared_base;
  69     QSIMPLEQ_HEAD(bmds_list, BlkMigDevState) bmds_list;
  70     QSIMPLEQ_HEAD(blk_list, BlkMigBlock) blk_list;
  71     int submitted;
  72     int read_done;
  73     int transferred;
  74     int64_t total_sector_sum;
  75     int prev_progress;
  76     int bulk_completed;
  77     long double total_time;
  78     int reads;
  79 } BlkMigState;
  80
  81 static BlkMigState block_mig_state;
  82
  83 static void blk_send(QEMUFile *f, BlkMigBlock * blk)
  84 {
  85     int len;
  86
  87     /* sector number and flags */
  88     qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
  89                      | BLK_MIG_FLAG_DEVICE_BLOCK);
  90
  91     /* device name */
  92     len = strlen(blk->bmds->bs->device_name);
  93     qemu_put_byte(f, len);
  94     qemu_put_buffer(f, (uint8_t *)blk->bmds->bs->device_name, len);
  95
  96     qemu_put_buffer(f, blk->buf, BLOCK_SIZE);
  97 }
  98
  99 int blk_mig_active(void)
 100 {
 101     return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
 102 }
 103
 104 uint64_t blk_mig_bytes_transferred(void)
 105 {
 106     BlkMigDevState *bmds;
 107     uint64_t sum = 0;
 108
 109     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 110         sum += bmds->completed_sectors;
 111     }
 112     return sum << BDRV_SECTOR_BITS;
 113 }
 114
 115 uint64_t blk_mig_bytes_remaining(void)
 116 {
 117     return blk_mig_bytes_total() - blk_mig_bytes_transferred();
 118 }
 119
 120 uint64_t blk_mig_bytes_total(void)
 121 {
 122     BlkMigDevState *bmds;
 123     uint64_t sum = 0;
 124
 125     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 126         sum += bmds->total_sectors;
 127     }
 128     return sum << BDRV_SECTOR_BITS;
 129 }
 130
 131 static inline void add_avg_read_time(int64_t time)
 132 {
 133     block_mig_state.reads++;
 134     block_mig_state.total_time += time;
 135 }
 136
 137 static inline long double compute_read_bwidth(void)
 138 {
 139     assert(block_mig_state.total_time != 0);
 140     return  (block_mig_state.reads * BLOCK_SIZE)/ block_mig_state.total_time;
 141 }
 142
 143 static void blk_mig_read_cb(void *opaque, int ret)
 144 {
 145     BlkMigBlock *blk = opaque;
 146
 147     blk->ret = ret;
 148
 149     blk->time = qemu_get_clock_ns(rt_clock) - blk->time;
 150
 151     add_avg_read_time(blk->time);
 152
 153     QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry);
 154
 155     block_mig_state.submitted--;
 156     block_mig_state.read_done++;
 157     assert(block_mig_state.submitted >= 0);
 158 }
 159
 160 static int mig_save_device_bulk(Monitor *mon, QEMUFile *f,
 161                                 BlkMigDevState *bmds)
 162 {
 163     int64_t total_sectors = bmds->total_sectors;
 164     int64_t cur_sector = bmds->cur_sector;
 165     BlockDriverState *bs = bmds->bs;
 166     BlkMigBlock *blk;
 167     int nr_sectors;
 168
 169     if (bmds->shared_base) {
 170         while (cur_sector < total_sectors &&
 171                !bdrv_is_allocated(bs, cur_sector, MAX_IS_ALLOCATED_SEARCH,
 172                                   &nr_sectors)) {
 173             cur_sector += nr_sectors;
 174         }
 175     }
 176
 177     if (cur_sector >= total_sectors) {
 178         bmds->cur_sector = bmds->completed_sectors = total_sectors;
 179         return 1;
 180     }
 181
 182     bmds->completed_sectors = cur_sector;
 183
 184     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
 185
 186     /* we are going to transfer a full block even if it is not allocated */
 187     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 188
 189     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 190         nr_sectors = total_sectors - cur_sector;
 191     }
 192
 193     blk = qemu_malloc(sizeof(BlkMigBlock));
 194     blk->buf = qemu_malloc(BLOCK_SIZE);
 195     blk->bmds = bmds;
 196     blk->sector = cur_sector;
 197
 198     blk->iov.iov_base = blk->buf;
 199     blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 200     qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 201
 202     blk->time = qemu_get_clock_ns(rt_clock);
 203
 204     blk->aiocb = bdrv_aio_readv(bs, cur_sector, &blk->qiov,
 205                                 nr_sectors, blk_mig_read_cb, blk);
 206     if (!blk->aiocb) {
 207         goto error;
 208     }
 209     block_mig_state.submitted++;
 210
 211     bdrv_reset_dirty(bs, cur_sector, nr_sectors);
 212     bmds->cur_sector = cur_sector + nr_sectors;
 213
 214     return (bmds->cur_sector >= total_sectors);
 215
 216 error:
 217     monitor_printf(mon, "Error reading sector %" PRId64 "\n", cur_sector);
 218     qemu_file_set_error(f);
 219     qemu_free(blk->buf);
 220     qemu_free(blk);
 221     return 0;
 222 }
 223
 224 static void set_dirty_tracking(int enable)
 225 {
 226     BlkMigDevState *bmds;
 227
 228     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 229         bdrv_set_dirty_tracking(bmds->bs, enable);
 230     }
 231 }
 232
 233 static void init_blk_migration(Monitor *mon, QEMUFile *f)
 234 {
 235     BlkMigDevState *bmds;
 236     BlockDriverState *bs;
 237     int64_t sectors;
 238
 239     block_mig_state.submitted = 0;
 240     block_mig_state.read_done = 0;
 241     block_mig_state.transferred = 0;
 242     block_mig_state.total_sector_sum = 0;
 243     block_mig_state.prev_progress = -1;
 244     block_mig_state.bulk_completed = 0;
 245     block_mig_state.total_time = 0;
 246     block_mig_state.reads = 0;
 247
 248     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
 249         if (bs->type == BDRV_TYPE_HD) {
 250             sectors = bdrv_getlength(bs) >> BDRV_SECTOR_BITS;
 251             if (sectors == 0) {
 252                 continue;
 253             }
 254
 255             bmds = qemu_mallocz(sizeof(BlkMigDevState));
 256             bmds->bs = bs;
 257             bmds->bulk_completed = 0;
 258             bmds->total_sectors = sectors;
 259             bmds->completed_sectors = 0;
 260             bmds->shared_base = block_mig_state.shared_base;
 261
 262             block_mig_state.total_sector_sum += sectors;
 263
 264             if (bmds->shared_base) {
 265                 monitor_printf(mon, "Start migration for %s with shared base "
 266                                     "image\n",
 267                                bs->device_name);
 268             } else {
 269                 monitor_printf(mon, "Start full migration for %s\n",
 270                                bs->device_name);
 271             }
 272
 273             QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 274         }
 275     }
 276 }
 277
 278 static int blk_mig_save_bulked_block(Monitor *mon, QEMUFile *f)
 279 {
 280     int64_t completed_sector_sum = 0;
 281     BlkMigDevState *bmds;
 282     int progress;
 283     int ret = 0;
 284
 285     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 286         if (bmds->bulk_completed == 0) {
 287             if (mig_save_device_bulk(mon, f, bmds) == 1) {
 288                 /* completed bulk section for this device */
 289                 bmds->bulk_completed = 1;
 290             }
 291             completed_sector_sum += bmds->completed_sectors;
 292             ret = 1;
 293             break;
 294         } else {
 295             completed_sector_sum += bmds->completed_sectors;
 296         }
 297     }
 298
 299     progress = completed_sector_sum * 100 / block_mig_state.total_sector_sum;
 300     if (progress != block_mig_state.prev_progress) {
 301         block_mig_state.prev_progress = progress;
 302         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 303                          | BLK_MIG_FLAG_PROGRESS);
 304         monitor_printf(mon, "Completed %d %%\r", progress);
 305         monitor_flush(mon);
 306     }
 307
 308     return ret;
 309 }
 310
 311 static void blk_mig_reset_dirty_cursor(void)
 312 {
 313     BlkMigDevState *bmds;
 314
 315     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 316         bmds->cur_dirty = 0;
 317     }
 318 }
 319
 320 static int mig_save_device_dirty(Monitor *mon, QEMUFile *f,
 321                                  BlkMigDevState *bmds, int is_async)
 322 {
 323     BlkMigBlock *blk;
 324     int64_t total_sectors = bmds->total_sectors;
 325     int64_t sector;
 326     int nr_sectors;
 327
 328     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
 329         if (bdrv_get_dirty(bmds->bs, sector)) {
 330
 331             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 332                 nr_sectors = total_sectors - sector;
 333             } else {
 334                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 335             }
 336             blk = qemu_malloc(sizeof(BlkMigBlock));
 337             blk->buf = qemu_malloc(BLOCK_SIZE);
 338             blk->bmds = bmds;
 339             blk->sector = sector;
 340
 341             if (is_async) {
 342                 blk->iov.iov_base = blk->buf;
 343                 blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 344                 qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 345
 346                 blk->time = qemu_get_clock_ns(rt_clock);
 347
 348                 blk->aiocb = bdrv_aio_readv(bmds->bs, sector, &blk->qiov,
 349                                             nr_sectors, blk_mig_read_cb, blk);
 350                 if (!blk->aiocb) {
 351                     goto error;
 352                 }
 353                 block_mig_state.submitted++;
 354             } else {
 355                 if (bdrv_read(bmds->bs, sector, blk->buf,
 356                               nr_sectors) < 0) {
 357                     goto error;
 358                 }
 359                 blk_send(f, blk);
 360
 361                 qemu_free(blk->buf);
 362                 qemu_free(blk);
 363             }
 364
 365             bdrv_reset_dirty(bmds->bs, sector, nr_sectors);
 366             break;
 367         }
 368         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 369         bmds->cur_dirty = sector;
 370     }
 371
 372     return (bmds->cur_dirty >= bmds->total_sectors);
 373
 374 error:
 375     monitor_printf(mon, "Error reading sector %" PRId64 "\n", sector);
 376     qemu_file_set_error(f);
 377     qemu_free(blk->buf);
 378     qemu_free(blk);
 379     return 0;
 380 }
 381
 382 static int blk_mig_save_dirty_block(Monitor *mon, QEMUFile *f, int is_async)
 383 {
 384     BlkMigDevState *bmds;
 385     int ret = 0;
 386
 387     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 388         if (mig_save_device_dirty(mon, f, bmds, is_async) == 0) {
 389             ret = 1;
 390             break;
 391         }
 392     }
 393
 394     return ret;
 395 }
 396
 397 static void flush_blks(QEMUFile* f)
 398 {
 399     BlkMigBlock *blk;
 400
 401     DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
 402             __FUNCTION__, block_mig_state.submitted, block_mig_state.read_done,
 403             block_mig_state.transferred);
 404
 405     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 406         if (qemu_file_rate_limit(f)) {
 407             break;
 408         }
 409         if (blk->ret < 0) {
 410             qemu_file_set_error(f);
 411             break;
 412         }
 413         blk_send(f, blk);
 414
 415         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 416         qemu_free(blk->buf);
 417         qemu_free(blk);
 418
 419         block_mig_state.read_done--;
 420         block_mig_state.transferred++;
 421         assert(block_mig_state.read_done >= 0);
 422     }
 423
 424     DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__,
 425             block_mig_state.submitted, block_mig_state.read_done,
 426             block_mig_state.transferred);
 427 }
 428
 429 static int64_t get_remaining_dirty(void)
 430 {
 431     BlkMigDevState *bmds;
 432     int64_t dirty = 0;
 433
 434     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 435         dirty += bdrv_get_dirty_count(bmds->bs);
 436     }
 437
 438     return dirty * BLOCK_SIZE;
 439 }
 440
 441 static int is_stage2_completed(void)
 442 {
 443     int64_t remaining_dirty;
 444     long double bwidth;
 445
 446     if (block_mig_state.bulk_completed == 1) {
 447
 448         remaining_dirty = get_remaining_dirty();
 449         if (remaining_dirty == 0) {
 450             return 1;
 451         }
 452
 453         bwidth = compute_read_bwidth();
 454
 455         if ((remaining_dirty / bwidth) <=
 456             migrate_max_downtime()) {
 457             /* finish stage2 because we think that we can finish remaing work
 458                below max_downtime */
 459
 460             return 1;
 461         }
 462     }
 463
 464     return 0;
 465 }
 466
 467 static void blk_mig_cleanup(Monitor *mon)
 468 {
 469     BlkMigDevState *bmds;
 470     BlkMigBlock *blk;
 471
 472     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 473         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 474         qemu_free(bmds);
 475     }
 476
 477     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 478         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 479         qemu_free(blk->buf);
 480         qemu_free(blk);
 481     }
 482
 483     set_dirty_tracking(0);
 484
 485     monitor_printf(mon, "\n");
 486 }
 487
 488 static int block_save_live(Monitor *mon, QEMUFile *f, int stage, void *opaque)
 489 {
 490     DPRINTF("Enter save live stage %d submitted %d transferred %d\n",
 491             stage, block_mig_state.submitted, block_mig_state.transferred);
 492
 493     if (stage < 0) {
 494         blk_mig_cleanup(mon);
 495         return 0;
 496     }
 497
 498     if (block_mig_state.blk_enable != 1) {
 499         /* no need to migrate storage */
 500         qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 501         return 1;
 502     }
 503
 504     if (stage == 1) {
 505         init_blk_migration(mon, f);
 506
 507         /* start track dirty blocks */
 508         set_dirty_tracking(1);
 509     }
 510
 511     flush_blks(f);
 512
 513     if (qemu_file_has_error(f)) {
 514         blk_mig_cleanup(mon);
 515         return 0;
 516     }
 517
 518     blk_mig_reset_dirty_cursor();
 519
 520     if (stage == 2) {
 521         /* control the rate of transfer */
 522         while ((block_mig_state.submitted +
 523                 block_mig_state.read_done) * BLOCK_SIZE <
 524                qemu_file_get_rate_limit(f)) {
 525             if (block_mig_state.bulk_completed == 0) {
 526                 /* first finish the bulk phase */
 527                 if (blk_mig_save_bulked_block(mon, f) == 0) {
 528                     /* finished saving bulk on all devices */
 529                     block_mig_state.bulk_completed = 1;
 530                 }
 531             } else {
 532                 if (blk_mig_save_dirty_block(mon, f, 1) == 0) {
 533                     /* no more dirty blocks */
 534                     break;
 535                 }
 536             }
 537         }
 538
 539         flush_blks(f);
 540
 541         if (qemu_file_has_error(f)) {
 542             blk_mig_cleanup(mon);
 543             return 0;
 544         }
 545     }
 546
 547     if (stage == 3) {
 548         /* we know for sure that save bulk is completed and
 549            all async read completed */
 550         assert(block_mig_state.submitted == 0);
 551
 552         while (blk_mig_save_dirty_block(mon, f, 0) != 0);
 553         blk_mig_cleanup(mon);
 554
 555         /* report completion */
 556         qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 557
 558         if (qemu_file_has_error(f)) {
 559             return 0;
 560         }
 561
 562         monitor_printf(mon, "Block migration completed\n");
 563     }
 564
 565     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 566
 567     return ((stage == 2) && is_stage2_completed());
 568 }
 569
 570 static int block_load(QEMUFile *f, void *opaque, int version_id)
 571 {
 572     static int banner_printed;
 573     int len, flags;
 574     char device_name[256];
 575     int64_t addr;
 576     BlockDriverState *bs;
 577     uint8_t *buf;
 578
 579     do {
 580         addr = qemu_get_be64(f);
 581
 582         flags = addr & ~BDRV_SECTOR_MASK;
 583         addr >>= BDRV_SECTOR_BITS;
 584
 585         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 586             /* get device name */
 587             len = qemu_get_byte(f);
 588             qemu_get_buffer(f, (uint8_t *)device_name, len);
 589             device_name[len] = '\0';
 590
 591             bs = bdrv_find(device_name);
 592             if (!bs) {
 593                 fprintf(stderr, "Error unknown block device %s\n",
 594                         device_name);
 595                 return -EINVAL;
 596             }
 597
 598             buf = qemu_malloc(BLOCK_SIZE);
 599
 600             qemu_get_buffer(f, buf, BLOCK_SIZE);
 601             bdrv_write(bs, addr, buf, BDRV_SECTORS_PER_DIRTY_CHUNK);
 602
 603             qemu_free(buf);
 604         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 605             if (!banner_printed) {
 606                 printf("Receiving block device images\n");
 607                 banner_printed = 1;
 608             }
 609             printf("Completed %d %%%c", (int)addr,
 610                    (addr == 100) ? '\n' : '\r');
 611             fflush(stdout);
 612         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
 613             fprintf(stderr, "Unknown flags\n");
 614             return -EINVAL;
 615         }
 616         if (qemu_file_has_error(f)) {
 617             return -EIO;
 618         }
 619     } while (!(flags & BLK_MIG_FLAG_EOS));
 620
 621     return 0;
 622 }
 623
 624 static void block_set_params(int blk_enable, int shared_base, void *opaque)
 625 {
 626     block_mig_state.blk_enable = blk_enable;
 627     block_mig_state.shared_base = shared_base;
 628
 629     /* shared base means that blk_enable = 1 */
 630     block_mig_state.blk_enable |= shared_base;
 631 }
 632
 633 void blk_mig_init(void)
 634 {
 635     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
 636     QSIMPLEQ_INIT(&block_mig_state.blk_list);
 637
 638     register_savevm_live("block", 0, 1, block_set_params, block_save_live,
 639                          NULL, block_load, &block_mig_state);
 640 }