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 "block-migration.h"
  18 #include <assert.h>
  19 #include <pthread.h>
  20
  21 #define SECTOR_BITS 9
  22 #define SECTOR_SIZE (1 << SECTOR_BITS)
  23 #define SECTOR_MASK ~(SECTOR_SIZE - 1);
  24
  25 #define BLOCK_SIZE (block_mig_state->sectors_per_block << SECTOR_BITS)
  26
  27 #define BLK_MIG_FLAG_DEVICE_BLOCK       0x01
  28 #define BLK_MIG_FLAG_EOS                0x02
  29
  30 #define MAX_IS_ALLOCATED_SEARCH 65536
  31 #define MAX_BLOCKS_READ 10000
  32 #define BLOCKS_READ_CHANGE 100
  33 #define INITIAL_BLOCKS_READ 100
  34
  35 //#define DEBUG_BLK_MIGRATION
  36
  37 #ifdef DEBUG_BLK_MIGRATION
  38 #define dprintf(fmt, ...)                                               \
  39     do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
  40 #else
  41 #define dprintf(fmt, ...)                       \
  42     do { } while (0)
  43 #endif
  44
  45 typedef struct BlkMigBlock {
  46     uint8_t *buf;
  47     BlkMigDevState *bmds;
  48     int64_t sector;
  49     struct iovec iov;
  50     QEMUIOVector qiov;
  51     BlockDriverAIOCB *aiocb;
  52     int ret;
  53     struct BlkMigBlock *next;
  54 } BlkMigBlock;
  55
  56 typedef struct BlkMigState {
  57     int bulk_completed;
  58     int blk_enable;
  59     int shared_base;
  60     int no_dirty;
  61     QEMUFile *load_file;
  62     BlkMigDevState *bmds_first;
  63     int sectors_per_block;
  64     BlkMigBlock *first_blk;
  65     BlkMigBlock *last_blk;
  66     int submitted;
  67     int read_done;
  68     int transferred;
  69     int64_t print_completion;
  70 } BlkMigState;
  71
  72 static BlkMigState *block_mig_state = NULL;
  73
  74 static void blk_mig_read_cb(void *opaque, int ret)
  75 {
  76     BlkMigBlock *blk = opaque;
  77
  78     blk->ret = ret;
  79
  80     /* insert at the end */
  81     if(block_mig_state->last_blk == NULL) {
  82         block_mig_state->first_blk = blk;
  83         block_mig_state->last_blk = blk;
  84     } else {
  85         block_mig_state->last_blk->next = blk;
  86         block_mig_state->last_blk = blk;
  87     }
  88
  89     block_mig_state->submitted--;
  90     block_mig_state->read_done++;
  91     assert(block_mig_state->submitted >= 0);
  92
  93     return;
  94 }
  95
  96 static int mig_read_device_bulk(QEMUFile *f, BlkMigDevState *bms)
  97 {
  98     int nr_sectors;
  99     int64_t total_sectors, cur_sector = 0;
 100     BlockDriverState *bs = bms->bs;
 101     BlkMigBlock *blk;
 102
 103     blk = qemu_malloc(sizeof(BlkMigBlock));
 104     blk->buf = qemu_malloc(BLOCK_SIZE);
 105
 106     cur_sector = bms->cur_sector;
 107     total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 108
 109     if(bms->shared_base) {
 110         while(cur_sector < bms->total_sectors &&
 111               !bdrv_is_allocated(bms->bs, cur_sector,
 112                                  MAX_IS_ALLOCATED_SEARCH, &nr_sectors)) {
 113             cur_sector += nr_sectors;
 114         }
 115     }
 116
 117     if(cur_sector >= total_sectors) {
 118         bms->cur_sector = total_sectors;
 119         qemu_free(blk->buf);
 120         qemu_free(blk);
 121         return 1;
 122     }
 123
 124     if(cur_sector >= block_mig_state->print_completion) {
 125         printf("Completed %" PRId64 " %%\r", cur_sector * 100 / total_sectors);
 126         fflush(stdout);
 127         block_mig_state->print_completion +=
 128             (block_mig_state->sectors_per_block * 10000);
 129     }
 130
 131     /* we going to transfder BLOCK_SIZE any way even if it is not allocated */
 132     nr_sectors = block_mig_state->sectors_per_block;
 133
 134     cur_sector &= ~((int64_t)block_mig_state->sectors_per_block -1);
 135
 136     if(total_sectors - cur_sector < block_mig_state->sectors_per_block) {
 137         nr_sectors = (total_sectors - cur_sector);
 138     }
 139
 140     bms->cur_sector = cur_sector + nr_sectors;
 141     blk->sector = cur_sector;
 142     blk->bmds = bms;
 143     blk->next = NULL;
 144
 145     blk->iov.iov_base = blk->buf;
 146     blk->iov.iov_len = nr_sectors * SECTOR_SIZE;
 147     qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 148
 149     blk->aiocb = bdrv_aio_readv(bs, cur_sector, &blk->qiov,
 150                                 nr_sectors, blk_mig_read_cb, blk);
 151
 152     if(!blk->aiocb) {
 153         printf("Error reading sector %" PRId64 "\n", cur_sector);
 154         qemu_free(blk->buf);
 155         qemu_free(blk);
 156         return 0;
 157     }
 158
 159     bdrv_reset_dirty(bms->bs, cur_sector, nr_sectors);
 160     block_mig_state->submitted++;
 161
 162     return (bms->cur_sector >= total_sectors);
 163 }
 164
 165 static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds)
 166 {
 167     int len, nr_sectors;
 168     int64_t total_sectors = bmds->total_sectors, cur_sector = 0;
 169     uint8_t *tmp_buf = NULL;
 170     BlockDriverState *bs = bmds->bs;
 171
 172     tmp_buf = qemu_malloc(BLOCK_SIZE);
 173
 174     cur_sector = bmds->cur_sector;
 175
 176     if(bmds->shared_base) {
 177         while(cur_sector < bmds->total_sectors &&
 178               !bdrv_is_allocated(bmds->bs, cur_sector,
 179                                  MAX_IS_ALLOCATED_SEARCH, &nr_sectors)) {
 180             cur_sector += nr_sectors;
 181         }
 182     }
 183
 184     if(cur_sector >= total_sectors) {
 185         bmds->cur_sector = total_sectors;
 186         qemu_free(tmp_buf);
 187         return 1;
 188     }
 189
 190     if(cur_sector >= block_mig_state->print_completion) {
 191         printf("Completed %" PRId64 " %%\r", cur_sector * 100 / total_sectors);
 192         fflush(stdout);
 193         block_mig_state->print_completion +=
 194             (block_mig_state->sectors_per_block * 10000);
 195     }
 196
 197     cur_sector &= ~((int64_t)block_mig_state->sectors_per_block -1);
 198
 199     /* we going to transfer
 200        BLOCK_SIZE
 201        any way even if it is not allocated */
 202     nr_sectors = block_mig_state->sectors_per_block;
 203
 204     if(total_sectors - cur_sector < block_mig_state->sectors_per_block) {
 205         nr_sectors = (total_sectors - cur_sector);
 206     }
 207
 208     if(bdrv_read(bs, cur_sector, tmp_buf, nr_sectors) < 0) {
 209         printf("Error reading sector %" PRId64 "\n", cur_sector);
 210     }
 211
 212     bdrv_reset_dirty(bs, cur_sector, nr_sectors);
 213
 214     /* Device name */
 215     qemu_put_be64(f,(cur_sector << SECTOR_BITS) | BLK_MIG_FLAG_DEVICE_BLOCK);
 216
 217     len = strlen(bs->device_name);
 218     qemu_put_byte(f, len);
 219     qemu_put_buffer(f, (uint8_t *)bs->device_name, len);
 220
 221     qemu_put_buffer(f, tmp_buf,
 222                     BLOCK_SIZE);
 223
 224     bmds->cur_sector = cur_sector + block_mig_state->sectors_per_block;
 225
 226     qemu_free(tmp_buf);
 227
 228     return (bmds->cur_sector >= total_sectors);
 229 }
 230
 231 static void send_blk(QEMUFile *f, BlkMigBlock * blk)
 232 {
 233     int len;
 234
 235     /* Device name */
 236     qemu_put_be64(f,(blk->sector << SECTOR_BITS) | BLK_MIG_FLAG_DEVICE_BLOCK);
 237
 238     len = strlen(blk->bmds->bs->device_name);
 239     qemu_put_byte(f, len);
 240     qemu_put_buffer(f, (uint8_t *)blk->bmds->bs->device_name, len);
 241
 242     qemu_put_buffer(f, blk->buf,
 243                     BLOCK_SIZE);
 244
 245     return;
 246 }
 247
 248 static void blk_mig_save_dev_info(QEMUFile *f, BlkMigDevState *bmds)
 249 {
 250 }
 251
 252 static void set_dirty_tracking(int enable)
 253 {
 254     BlkMigDevState *bmds;
 255     for(bmds = block_mig_state->bmds_first; bmds != NULL; bmds = bmds->next) {
 256         bdrv_set_dirty_tracking(bmds->bs,enable);
 257     }
 258
 259     return;
 260 }
 261
 262 static void init_blk_migration(QEMUFile *f)
 263 {
 264     BlkMigDevState **pbmds, *bmds;
 265     BlockDriverState *bs;
 266
 267     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
 268         if(bs->type == BDRV_TYPE_HD) {
 269             bmds = qemu_mallocz(sizeof(BlkMigDevState));
 270             bmds->bs = bs;
 271             bmds->bulk_completed = 0;
 272             bmds->total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 273             bmds->shared_base = block_mig_state->shared_base;
 274
 275             if(bmds->shared_base) {
 276                 printf("Start migration for %s with shared base image\n",
 277                        bs->device_name);
 278             } else {
 279                 printf("Start full migration for %s\n", bs->device_name);
 280             }
 281
 282             /* insert at the end */
 283             pbmds = &block_mig_state->bmds_first;
 284             while (*pbmds != NULL)
 285                 pbmds = &(*pbmds)->next;
 286             *pbmds = bmds;
 287
 288             blk_mig_save_dev_info(f, bmds);
 289
 290         }
 291     }
 292
 293     block_mig_state->sectors_per_block = bdrv_get_sectors_per_chunk();
 294
 295     return;
 296 }
 297
 298 static int blk_mig_save_bulked_block(QEMUFile *f, int is_async)
 299 {
 300     BlkMigDevState *bmds;
 301
 302     for (bmds = block_mig_state->bmds_first; bmds != NULL; bmds = bmds->next) {
 303         if(bmds->bulk_completed == 0) {
 304             if(is_async) {
 305                 if(mig_read_device_bulk(f, bmds) == 1) {
 306                     /* completed bulk section for this device */
 307                     bmds->bulk_completed = 1;
 308                 }
 309             } else {
 310                 if(mig_save_device_bulk(f,bmds) == 1) {
 311                     /* completed bulk section for this device */
 312                     bmds->bulk_completed = 1;
 313                 }
 314             }
 315             return 1;
 316         }
 317     }
 318
 319     /* we reached here means bulk is completed */
 320     block_mig_state->bulk_completed = 1;
 321
 322     return 0;
 323
 324 }
 325
 326 #define MAX_NUM_BLOCKS 4
 327
 328 static void blk_mig_save_dirty_blocks(QEMUFile *f)
 329 {
 330     BlkMigDevState *bmds;
 331     uint8_t buf[BLOCK_SIZE];
 332     int64_t sector;
 333     int len;
 334
 335     for(bmds = block_mig_state->bmds_first; bmds != NULL; bmds = bmds->next) {
 336         for(sector = 0; sector < bmds->cur_sector;) {
 337
 338             if(bdrv_get_dirty(bmds->bs,sector)) {
 339
 340                 if(bdrv_read(bmds->bs, sector, buf,
 341                              block_mig_state->sectors_per_block) < 0) {
 342                 }
 343
 344                 /* device name */
 345                 qemu_put_be64(f,(sector << SECTOR_BITS)
 346                               | BLK_MIG_FLAG_DEVICE_BLOCK);
 347
 348                 len = strlen(bmds->bs->device_name);
 349
 350                 qemu_put_byte(f, len);
 351                 qemu_put_buffer(f, (uint8_t *)bmds->bs->device_name, len);
 352
 353                 qemu_put_buffer(f, buf,
 354                                 (block_mig_state->sectors_per_block *
 355                                  SECTOR_SIZE));
 356
 357                 bdrv_reset_dirty(bmds->bs, sector,
 358                                  block_mig_state->sectors_per_block);
 359
 360                 sector += block_mig_state->sectors_per_block;
 361             } else {
 362                 /* sector is clean */
 363                 sector += block_mig_state->sectors_per_block;
 364             }
 365         }
 366     }
 367
 368     return;
 369 }
 370
 371 static void flush_blks(QEMUFile* f)
 372 {
 373     BlkMigBlock *blk, *tmp;
 374
 375     dprintf("%s Enter submitted %d read_done %d transfered\n", __FUNCTION__,
 376             submitted, read_done, transfered);
 377
 378     for(blk = block_mig_state->first_blk;
 379         blk != NULL && !qemu_file_rate_limit(f); blk = tmp) {
 380         send_blk(f, blk);
 381
 382         tmp = blk->next;
 383         qemu_free(blk->buf);
 384         qemu_free(blk);
 385
 386         block_mig_state->read_done--;
 387         block_mig_state->transferred++;
 388         assert(block_mig_state->read_done >= 0);
 389     }
 390     block_mig_state->first_blk = blk;
 391
 392     if(block_mig_state->first_blk == NULL) {
 393         block_mig_state->last_blk = NULL;
 394     }
 395
 396     dprintf("%s Exit submitted %d read_done %d transferred%d\n", __FUNCTION__,
 397             block_mig_state->submitted, block_mig_state->read_done,
 398             block_mig_state->transferred);
 399
 400     return;
 401 }
 402
 403 static int is_stage2_completed(void)
 404 {
 405     BlkMigDevState *bmds;
 406
 407     if(block_mig_state->submitted > 0) {
 408         return 0;
 409     }
 410
 411     for (bmds = block_mig_state->bmds_first; bmds != NULL; bmds = bmds->next) {
 412         if(bmds->bulk_completed == 0) {
 413             return 0;
 414         }
 415     }
 416
 417     return 1;
 418 }
 419
 420 static int block_save_live(QEMUFile *f, int stage, void *opaque)
 421 {
 422     int ret = 1;
 423
 424     dprintf("Enter save live stage %d submitted %d transferred %d\n", stage,
 425             submitted, transferred);
 426
 427     if(block_mig_state->blk_enable != 1) {
 428         /* no need to migrate storage */
 429
 430         qemu_put_be64(f,BLK_MIG_FLAG_EOS);
 431         return 1;
 432     }
 433
 434     if(stage == 1) {
 435         init_blk_migration(f);
 436
 437         /* start track dirty blocks */
 438         set_dirty_tracking(1);
 439
 440     }
 441
 442     flush_blks(f);
 443
 444     /* control the rate of transfer */
 445     while ((block_mig_state->submitted + block_mig_state->read_done) *
 446            (BLOCK_SIZE) <
 447            (qemu_file_get_rate_limit(f))) {
 448
 449         ret = blk_mig_save_bulked_block(f, 1);
 450
 451         if (ret == 0) /* no more bulk blocks for now*/
 452             break;
 453     }
 454
 455     flush_blks(f);
 456
 457     if(stage == 3) {
 458
 459         while(blk_mig_save_bulked_block(f, 0) != 0);
 460
 461         blk_mig_save_dirty_blocks(f);
 462
 463         /* stop track dirty blocks */
 464         set_dirty_tracking(0);;
 465
 466         printf("\nBlock migration completed\n");
 467     }
 468
 469     qemu_put_be64(f,BLK_MIG_FLAG_EOS);
 470
 471     return ((stage == 2) && is_stage2_completed());
 472 }
 473
 474 static int block_load(QEMUFile *f, void *opaque, int version_id)
 475 {
 476     int len, flags;
 477     char device_name[256];
 478     int64_t addr;
 479     BlockDriverState *bs;
 480     uint8_t *buf;
 481
 482     block_mig_state->sectors_per_block = bdrv_get_sectors_per_chunk();
 483     buf = qemu_malloc(BLOCK_SIZE);
 484
 485     do {
 486
 487         addr = qemu_get_be64(f);
 488
 489         flags = addr & ~SECTOR_MASK;
 490         addr &= SECTOR_MASK;
 491
 492         if(flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 493
 494             /* get device name */
 495             len = qemu_get_byte(f);
 496
 497             qemu_get_buffer(f, (uint8_t *)device_name, len);
 498             device_name[len] = '\0';
 499
 500             bs = bdrv_find(device_name);
 501
 502             qemu_get_buffer(f, buf,
 503                             BLOCK_SIZE);
 504             if(bs != NULL) {
 505
 506                 bdrv_write(bs, (addr >> SECTOR_BITS),
 507                            buf, block_mig_state->sectors_per_block);
 508             } else {
 509                 printf("Error unknown block device %s\n", device_name);
 510             }
 511         } else if(flags & BLK_MIG_FLAG_EOS) {
 512
 513         } else {
 514             printf("Unknown flags\n");
 515         }
 516     } while(!(flags & BLK_MIG_FLAG_EOS));
 517
 518     qemu_free(buf);
 519
 520     return 0;
 521 }
 522
 523 static void block_set_params(int blk_enable, int shared_base, void *opaque)
 524 {
 525     assert(opaque == block_mig_state);
 526
 527     block_mig_state->blk_enable = blk_enable;
 528     block_mig_state->shared_base = shared_base;
 529
 530     /* shared base means that blk_enable = 1 */
 531     block_mig_state->blk_enable |= shared_base;
 532
 533     return;
 534 }
 535
 536 void blk_mig_info(void)
 537 {
 538     BlockDriverState *bs;
 539
 540     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
 541         printf("Device %s\n", bs->device_name);
 542         if(bs->type == BDRV_TYPE_HD) {
 543             printf("device %s format %s\n",
 544                    bs->device_name, bs->drv->format_name);
 545         }
 546     }
 547 }
 548
 549 void blk_mig_init(void)
 550 {
 551
 552     block_mig_state = qemu_mallocz(sizeof(BlkMigState));
 553
 554     register_savevm_live("block", 0, 1, block_set_params, block_save_live,
 555                          NULL, block_load, block_mig_state);
 556
 557
 558 }