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