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