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