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_it(void *opaque, BlockDriverState *bs)
 234 {
 235     Monitor *mon = opaque;
 236     BlkMigDevState *bmds;
 237     int64_t sectors;
 238
 239     if (bs->type == BDRV_TYPE_HD) {
 240         sectors = bdrv_getlength(bs) >> BDRV_SECTOR_BITS;
 241         if (sectors == 0) {
 242             return;
 243         }
 244
 245         bmds = qemu_mallocz(sizeof(BlkMigDevState));
 246         bmds->bs = bs;
 247         bmds->bulk_completed = 0;
 248         bmds->total_sectors = sectors;
 249         bmds->completed_sectors = 0;
 250         bmds->shared_base = block_mig_state.shared_base;
 251
 252         block_mig_state.total_sector_sum += sectors;
 253
 254         if (bmds->shared_base) {
 255             monitor_printf(mon, "Start migration for %s with shared base "
 256                                 "image\n",
 257                            bs->device_name);
 258         } else {
 259             monitor_printf(mon, "Start full migration for %s\n",
 260                            bs->device_name);
 261         }
 262
 263         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 264     }
 265 }
 266
 267 static void init_blk_migration(Monitor *mon, QEMUFile *f)
 268 {
 269     block_mig_state.submitted = 0;
 270     block_mig_state.read_done = 0;
 271     block_mig_state.transferred = 0;
 272     block_mig_state.total_sector_sum = 0;
 273     block_mig_state.prev_progress = -1;
 274     block_mig_state.bulk_completed = 0;
 275     block_mig_state.total_time = 0;
 276     block_mig_state.reads = 0;
 277
 278     bdrv_iterate(init_blk_migration_it, mon);
 279 }
 280
 281 static int blk_mig_save_bulked_block(Monitor *mon, QEMUFile *f)
 282 {
 283     int64_t completed_sector_sum = 0;
 284     BlkMigDevState *bmds;
 285     int progress;
 286     int ret = 0;
 287
 288     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 289         if (bmds->bulk_completed == 0) {
 290             if (mig_save_device_bulk(mon, f, bmds) == 1) {
 291                 /* completed bulk section for this device */
 292                 bmds->bulk_completed = 1;
 293             }
 294             completed_sector_sum += bmds->completed_sectors;
 295             ret = 1;
 296             break;
 297         } else {
 298             completed_sector_sum += bmds->completed_sectors;
 299         }
 300     }
 301
 302     progress = completed_sector_sum * 100 / block_mig_state.total_sector_sum;
 303     if (progress != block_mig_state.prev_progress) {
 304         block_mig_state.prev_progress = progress;
 305         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 306                          | BLK_MIG_FLAG_PROGRESS);
 307         monitor_printf(mon, "Completed %d %%\r", progress);
 308         monitor_flush(mon);
 309     }
 310
 311     return ret;
 312 }
 313
 314 static void blk_mig_reset_dirty_cursor(void)
 315 {
 316     BlkMigDevState *bmds;
 317
 318     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 319         bmds->cur_dirty = 0;
 320     }
 321 }
 322
 323 static int mig_save_device_dirty(Monitor *mon, QEMUFile *f,
 324                                  BlkMigDevState *bmds, int is_async)
 325 {
 326     BlkMigBlock *blk;
 327     int64_t total_sectors = bmds->total_sectors;
 328     int64_t sector;
 329     int nr_sectors;
 330
 331     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
 332         if (bdrv_get_dirty(bmds->bs, sector)) {
 333
 334             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 335                 nr_sectors = total_sectors - sector;
 336             } else {
 337                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 338             }
 339             blk = qemu_malloc(sizeof(BlkMigBlock));
 340             blk->buf = qemu_malloc(BLOCK_SIZE);
 341             blk->bmds = bmds;
 342             blk->sector = sector;
 343
 344             if (is_async) {
 345                 blk->iov.iov_base = blk->buf;
 346                 blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 347                 qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 348
 349                 blk->time = qemu_get_clock_ns(rt_clock);
 350
 351                 blk->aiocb = bdrv_aio_readv(bmds->bs, sector, &blk->qiov,
 352                                             nr_sectors, blk_mig_read_cb, blk);
 353                 if (!blk->aiocb) {
 354                     goto error;
 355                 }
 356                 block_mig_state.submitted++;
 357             } else {
 358                 if (bdrv_read(bmds->bs, sector, blk->buf,
 359                               nr_sectors) < 0) {
 360                     goto error;
 361                 }
 362                 blk_send(f, blk);
 363
 364                 qemu_free(blk->buf);
 365                 qemu_free(blk);
 366             }
 367
 368             bdrv_reset_dirty(bmds->bs, sector, nr_sectors);
 369             break;
 370         }
 371         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 372         bmds->cur_dirty = sector;
 373     }
 374
 375     return (bmds->cur_dirty >= bmds->total_sectors);
 376
 377 error:
 378     monitor_printf(mon, "Error reading sector %" PRId64 "\n", sector);
 379     qemu_file_set_error(f);
 380     qemu_free(blk->buf);
 381     qemu_free(blk);
 382     return 0;
 383 }
 384
 385 static int blk_mig_save_dirty_block(Monitor *mon, QEMUFile *f, int is_async)
 386 {
 387     BlkMigDevState *bmds;
 388     int ret = 0;
 389
 390     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 391         if (mig_save_device_dirty(mon, f, bmds, is_async) == 0) {
 392             ret = 1;
 393             break;
 394         }
 395     }
 396
 397     return ret;
 398 }
 399
 400 static void flush_blks(QEMUFile* f)
 401 {
 402     BlkMigBlock *blk;
 403
 404     DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
 405             __FUNCTION__, block_mig_state.submitted, block_mig_state.read_done,
 406             block_mig_state.transferred);
 407
 408     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 409         if (qemu_file_rate_limit(f)) {
 410             break;
 411         }
 412         if (blk->ret < 0) {
 413             qemu_file_set_error(f);
 414             break;
 415         }
 416         blk_send(f, blk);
 417
 418         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 419         qemu_free(blk->buf);
 420         qemu_free(blk);
 421
 422         block_mig_state.read_done--;
 423         block_mig_state.transferred++;
 424         assert(block_mig_state.read_done >= 0);
 425     }
 426
 427     DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__,
 428             block_mig_state.submitted, block_mig_state.read_done,
 429             block_mig_state.transferred);
 430 }
 431
 432 static int64_t get_remaining_dirty(void)
 433 {
 434     BlkMigDevState *bmds;
 435     int64_t dirty = 0;
 436
 437     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 438         dirty += bdrv_get_dirty_count(bmds->bs);
 439     }
 440
 441     return dirty * BLOCK_SIZE;
 442 }
 443
 444 static int is_stage2_completed(void)
 445 {
 446     int64_t remaining_dirty;
 447     long double bwidth;
 448
 449     if (block_mig_state.bulk_completed == 1) {
 450
 451         remaining_dirty = get_remaining_dirty();
 452         if (remaining_dirty == 0) {
 453             return 1;
 454         }
 455
 456         bwidth = compute_read_bwidth();
 457
 458         if ((remaining_dirty / bwidth) <=
 459             migrate_max_downtime()) {
 460             /* finish stage2 because we think that we can finish remaing work
 461                below max_downtime */
 462
 463             return 1;
 464         }
 465     }
 466
 467     return 0;
 468 }
 469
 470 static void blk_mig_cleanup(Monitor *mon)
 471 {
 472     BlkMigDevState *bmds;
 473     BlkMigBlock *blk;
 474
 475     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 476         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 477         qemu_free(bmds);
 478     }
 479
 480     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 481         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 482         qemu_free(blk->buf);
 483         qemu_free(blk);
 484     }
 485
 486     set_dirty_tracking(0);
 487
 488     monitor_printf(mon, "\n");
 489 }
 490
 491 static int block_save_live(Monitor *mon, QEMUFile *f, int stage, void *opaque)
 492 {
 493     DPRINTF("Enter save live stage %d submitted %d transferred %d\n",
 494             stage, block_mig_state.submitted, block_mig_state.transferred);
 495
 496     if (stage < 0) {
 497         blk_mig_cleanup(mon);
 498         return 0;
 499     }
 500
 501     if (block_mig_state.blk_enable != 1) {
 502         /* no need to migrate storage */
 503         qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 504         return 1;
 505     }
 506
 507     if (stage == 1) {
 508         init_blk_migration(mon, f);
 509
 510         /* start track dirty blocks */
 511         set_dirty_tracking(1);
 512     }
 513
 514     flush_blks(f);
 515
 516     if (qemu_file_has_error(f)) {
 517         blk_mig_cleanup(mon);
 518         return 0;
 519     }
 520
 521     blk_mig_reset_dirty_cursor();
 522
 523     if (stage == 2) {
 524         /* control the rate of transfer */
 525         while ((block_mig_state.submitted +
 526                 block_mig_state.read_done) * BLOCK_SIZE <
 527                qemu_file_get_rate_limit(f)) {
 528             if (block_mig_state.bulk_completed == 0) {
 529                 /* first finish the bulk phase */
 530                 if (blk_mig_save_bulked_block(mon, f) == 0) {
 531                     /* finished saving bulk on all devices */
 532                     block_mig_state.bulk_completed = 1;
 533                 }
 534             } else {
 535                 if (blk_mig_save_dirty_block(mon, f, 1) == 0) {
 536                     /* no more dirty blocks */
 537                     break;
 538                 }
 539             }
 540         }
 541
 542         flush_blks(f);
 543
 544         if (qemu_file_has_error(f)) {
 545             blk_mig_cleanup(mon);
 546             return 0;
 547         }
 548     }
 549
 550     if (stage == 3) {
 551         /* we know for sure that save bulk is completed and
 552            all async read completed */
 553         assert(block_mig_state.submitted == 0);
 554
 555         while (blk_mig_save_dirty_block(mon, f, 0) != 0);
 556         blk_mig_cleanup(mon);
 557
 558         /* report completion */
 559         qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 560
 561         if (qemu_file_has_error(f)) {
 562             return 0;
 563         }
 564
 565         monitor_printf(mon, "Block migration completed\n");
 566     }
 567
 568     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 569
 570     return ((stage == 2) && is_stage2_completed());
 571 }
 572
 573 static int block_load(QEMUFile *f, void *opaque, int version_id)
 574 {
 575     static int banner_printed;
 576     int len, flags;
 577     char device_name[256];
 578     int64_t addr;
 579     BlockDriverState *bs;
 580     uint8_t *buf;
 581
 582     do {
 583         addr = qemu_get_be64(f);
 584
 585         flags = addr & ~BDRV_SECTOR_MASK;
 586         addr >>= BDRV_SECTOR_BITS;
 587
 588         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 589             /* get device name */
 590             len = qemu_get_byte(f);
 591             qemu_get_buffer(f, (uint8_t *)device_name, len);
 592             device_name[len] = '\0';
 593
 594             bs = bdrv_find(device_name);
 595             if (!bs) {
 596                 fprintf(stderr, "Error unknown block device %s\n",
 597                         device_name);
 598                 return -EINVAL;
 599             }
 600
 601             buf = qemu_malloc(BLOCK_SIZE);
 602
 603             qemu_get_buffer(f, buf, BLOCK_SIZE);
 604             bdrv_write(bs, addr, buf, BDRV_SECTORS_PER_DIRTY_CHUNK);
 605
 606             qemu_free(buf);
 607         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 608             if (!banner_printed) {
 609                 printf("Receiving block device images\n");
 610                 banner_printed = 1;
 611             }
 612             printf("Completed %d %%%c", (int)addr,
 613                    (addr == 100) ? '\n' : '\r');
 614             fflush(stdout);
 615         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
 616             fprintf(stderr, "Unknown flags\n");
 617             return -EINVAL;
 618         }
 619         if (qemu_file_has_error(f)) {
 620             return -EIO;
 621         }
 622     } while (!(flags & BLK_MIG_FLAG_EOS));
 623
 624     return 0;
 625 }
 626
 627 static void block_set_params(int blk_enable, int shared_base, void *opaque)
 628 {
 629     block_mig_state.blk_enable = blk_enable;
 630     block_mig_state.shared_base = shared_base;
 631
 632     /* shared base means that blk_enable = 1 */
 633     block_mig_state.blk_enable |= shared_base;
 634 }
 635
 636 void blk_mig_init(void)
 637 {
 638     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
 639     QSIMPLEQ_INIT(&block_mig_state.blk_list);
 640
 641     register_savevm_live("block", 0, 1, block_set_params, block_save_live,
 642                          NULL, block_load, &block_mig_state);
 643 }