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 #define MAX_BLOCKS_READ 10000
  30 #define BLOCKS_READ_CHANGE 100
  31 #define INITIAL_BLOCKS_READ 100
  32
  33 //#define DEBUG_BLK_MIGRATION
  34
  35 #ifdef DEBUG_BLK_MIGRATION
  36 #define dprintf(fmt, ...) \
  37     do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
  38 #else
  39 #define dprintf(fmt, ...) \
  40     do { } while (0)
  41 #endif
  42
  43 typedef struct BlkMigDevState {
  44     BlockDriverState *bs;
  45     int bulk_completed;
  46     int shared_base;
  47     int64_t cur_sector;
  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     QSIMPLEQ_ENTRY(BlkMigBlock) entry;
  63 } BlkMigBlock;
  64
  65 typedef struct BlkMigState {
  66     int blk_enable;
  67     int shared_base;
  68     QSIMPLEQ_HEAD(bmds_list, BlkMigDevState) bmds_list;
  69     QSIMPLEQ_HEAD(blk_list, BlkMigBlock) blk_list;
  70     int submitted;
  71     int read_done;
  72     int transferred;
  73     int64_t total_sector_sum;
  74     int prev_progress;
  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, int is_async)
 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     if (is_async) {
 179         blk->iov.iov_base = blk->buf;
 180         blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 181         qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 182
 183         blk->aiocb = bdrv_aio_readv(bs, cur_sector, &blk->qiov,
 184                                     nr_sectors, blk_mig_read_cb, blk);
 185         if (!blk->aiocb) {
 186             goto error;
 187         }
 188         block_mig_state.submitted++;
 189     } else {
 190         if (bdrv_read(bs, cur_sector, blk->buf, nr_sectors) < 0) {
 191             goto error;
 192         }
 193         blk_send(f, blk);
 194
 195         qemu_free(blk->buf);
 196         qemu_free(blk);
 197     }
 198
 199     bdrv_reset_dirty(bs, cur_sector, nr_sectors);
 200     bmds->cur_sector = cur_sector + nr_sectors;
 201
 202     return (bmds->cur_sector >= total_sectors);
 203
 204 error:
 205     monitor_printf(mon, "Error reading sector %" PRId64 "\n", cur_sector);
 206     qemu_file_set_error(f);
 207     qemu_free(blk->buf);
 208     qemu_free(blk);
 209     return 0;
 210 }
 211
 212 static void set_dirty_tracking(int enable)
 213 {
 214     BlkMigDevState *bmds;
 215
 216     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 217         bdrv_set_dirty_tracking(bmds->bs, enable);
 218     }
 219 }
 220
 221 static void init_blk_migration(Monitor *mon, QEMUFile *f)
 222 {
 223     BlkMigDevState *bmds;
 224     BlockDriverState *bs;
 225
 226     block_mig_state.submitted = 0;
 227     block_mig_state.read_done = 0;
 228     block_mig_state.transferred = 0;
 229     block_mig_state.total_sector_sum = 0;
 230     block_mig_state.prev_progress = -1;
 231
 232     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
 233         if (bs->type == BDRV_TYPE_HD) {
 234             bmds = qemu_mallocz(sizeof(BlkMigDevState));
 235             bmds->bs = bs;
 236             bmds->bulk_completed = 0;
 237             bmds->total_sectors = bdrv_getlength(bs) >> BDRV_SECTOR_BITS;
 238             bmds->completed_sectors = 0;
 239             bmds->shared_base = block_mig_state.shared_base;
 240
 241             block_mig_state.total_sector_sum += bmds->total_sectors;
 242
 243             if (bmds->shared_base) {
 244                 monitor_printf(mon, "Start migration for %s with shared base "
 245                                     "image\n",
 246                                bs->device_name);
 247             } else {
 248                 monitor_printf(mon, "Start full migration for %s\n",
 249                                bs->device_name);
 250             }
 251
 252             QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 253         }
 254     }
 255 }
 256
 257 static int blk_mig_save_bulked_block(Monitor *mon, QEMUFile *f, int is_async)
 258 {
 259     int64_t completed_sector_sum = 0;
 260     BlkMigDevState *bmds;
 261     int progress;
 262     int ret = 0;
 263
 264     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 265         if (bmds->bulk_completed == 0) {
 266             if (mig_save_device_bulk(mon, f, bmds, is_async) == 1) {
 267                 /* completed bulk section for this device */
 268                 bmds->bulk_completed = 1;
 269             }
 270             completed_sector_sum += bmds->completed_sectors;
 271             ret = 1;
 272             break;
 273         } else {
 274             completed_sector_sum += bmds->completed_sectors;
 275         }
 276     }
 277
 278     progress = completed_sector_sum * 100 / block_mig_state.total_sector_sum;
 279     if (progress != block_mig_state.prev_progress) {
 280         block_mig_state.prev_progress = progress;
 281         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 282                          | BLK_MIG_FLAG_PROGRESS);
 283         monitor_printf(mon, "Completed %d %%\r", progress);
 284         monitor_flush(mon);
 285     }
 286
 287     return ret;
 288 }
 289
 290 #define MAX_NUM_BLOCKS 4
 291
 292 static void blk_mig_save_dirty_blocks(Monitor *mon, QEMUFile *f)
 293 {
 294     BlkMigDevState *bmds;
 295     BlkMigBlock blk;
 296     int64_t sector;
 297
 298     blk.buf = qemu_malloc(BLOCK_SIZE);
 299
 300     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 301         for (sector = 0; sector < bmds->cur_sector;) {
 302             if (bdrv_get_dirty(bmds->bs, sector)) {
 303                 if (bdrv_read(bmds->bs, sector, blk.buf,
 304                               BDRV_SECTORS_PER_DIRTY_CHUNK) < 0) {
 305                     monitor_printf(mon, "Error reading sector %" PRId64 "\n",
 306                                    sector);
 307                     qemu_file_set_error(f);
 308                     qemu_free(blk.buf);
 309                     return;
 310                 }
 311                 blk.bmds = bmds;
 312                 blk.sector = sector;
 313                 blk_send(f, &blk);
 314
 315                 bdrv_reset_dirty(bmds->bs, sector,
 316                                  BDRV_SECTORS_PER_DIRTY_CHUNK);
 317             }
 318             sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 319         }
 320     }
 321
 322     qemu_free(blk.buf);
 323 }
 324
 325 static void flush_blks(QEMUFile* f)
 326 {
 327     BlkMigBlock *blk;
 328
 329     dprintf("%s Enter submitted %d read_done %d transferred %d\n",
 330             __FUNCTION__, block_mig_state.submitted, block_mig_state.read_done,
 331             block_mig_state.transferred);
 332
 333     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 334         if (qemu_file_rate_limit(f)) {
 335             break;
 336         }
 337         if (blk->ret < 0) {
 338             qemu_file_set_error(f);
 339             break;
 340         }
 341         blk_send(f, blk);
 342
 343         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 344         qemu_free(blk->buf);
 345         qemu_free(blk);
 346
 347         block_mig_state.read_done--;
 348         block_mig_state.transferred++;
 349         assert(block_mig_state.read_done >= 0);
 350     }
 351
 352     dprintf("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__,
 353             block_mig_state.submitted, block_mig_state.read_done,
 354             block_mig_state.transferred);
 355 }
 356
 357 static int is_stage2_completed(void)
 358 {
 359     BlkMigDevState *bmds;
 360
 361     if (block_mig_state.submitted > 0) {
 362         return 0;
 363     }
 364
 365     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 366         if (bmds->bulk_completed == 0) {
 367             return 0;
 368         }
 369     }
 370
 371     return 1;
 372 }
 373
 374 static void blk_mig_cleanup(Monitor *mon)
 375 {
 376     BlkMigDevState *bmds;
 377     BlkMigBlock *blk;
 378
 379     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 380         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 381         qemu_free(bmds);
 382     }
 383
 384     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 385         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 386         qemu_free(blk->buf);
 387         qemu_free(blk);
 388     }
 389
 390     set_dirty_tracking(0);
 391
 392     monitor_printf(mon, "\n");
 393 }
 394
 395 static int block_save_live(Monitor *mon, QEMUFile *f, int stage, void *opaque)
 396 {
 397     dprintf("Enter save live stage %d submitted %d transferred %d\n",
 398             stage, block_mig_state.submitted, block_mig_state.transferred);
 399
 400     if (stage < 0) {
 401         blk_mig_cleanup(mon);
 402         return 0;
 403     }
 404
 405     if (block_mig_state.blk_enable != 1) {
 406         /* no need to migrate storage */
 407         qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 408         return 1;
 409     }
 410
 411     if (stage == 1) {
 412         init_blk_migration(mon, f);
 413
 414         /* start track dirty blocks */
 415         set_dirty_tracking(1);
 416     }
 417
 418     flush_blks(f);
 419
 420     if (qemu_file_has_error(f)) {
 421         blk_mig_cleanup(mon);
 422         return 0;
 423     }
 424
 425     /* control the rate of transfer */
 426     while ((block_mig_state.submitted +
 427             block_mig_state.read_done) * BLOCK_SIZE <
 428            qemu_file_get_rate_limit(f)) {
 429         if (blk_mig_save_bulked_block(mon, f, 1) == 0) {
 430             /* no more bulk blocks for now */
 431             break;
 432         }
 433     }
 434
 435     flush_blks(f);
 436
 437     if (qemu_file_has_error(f)) {
 438         blk_mig_cleanup(mon);
 439         return 0;
 440     }
 441
 442     if (stage == 3) {
 443         while (blk_mig_save_bulked_block(mon, f, 0) != 0) {
 444             /* empty */
 445         }
 446
 447         blk_mig_save_dirty_blocks(mon, f);
 448         blk_mig_cleanup(mon);
 449
 450         /* report completion */
 451         qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 452
 453         if (qemu_file_has_error(f)) {
 454             return 0;
 455         }
 456
 457         monitor_printf(mon, "Block migration completed\n");
 458     }
 459
 460     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 461
 462     return ((stage == 2) && is_stage2_completed());
 463 }
 464
 465 static int block_load(QEMUFile *f, void *opaque, int version_id)
 466 {
 467     static int banner_printed;
 468     int len, flags;
 469     char device_name[256];
 470     int64_t addr;
 471     BlockDriverState *bs;
 472     uint8_t *buf;
 473
 474     do {
 475         addr = qemu_get_be64(f);
 476
 477         flags = addr & ~BDRV_SECTOR_MASK;
 478         addr >>= BDRV_SECTOR_BITS;
 479
 480         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 481             /* get device name */
 482             len = qemu_get_byte(f);
 483             qemu_get_buffer(f, (uint8_t *)device_name, len);
 484             device_name[len] = '\0';
 485
 486             bs = bdrv_find(device_name);
 487             if (!bs) {
 488                 fprintf(stderr, "Error unknown block device %s\n",
 489                         device_name);
 490                 return -EINVAL;
 491             }
 492
 493             buf = qemu_malloc(BLOCK_SIZE);
 494
 495             qemu_get_buffer(f, buf, BLOCK_SIZE);
 496             bdrv_write(bs, addr, buf, BDRV_SECTORS_PER_DIRTY_CHUNK);
 497
 498             qemu_free(buf);
 499         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 500             if (!banner_printed) {
 501                 printf("Receiving block device images\n");
 502                 banner_printed = 1;
 503             }
 504             printf("Completed %d %%%c", (int)addr,
 505                    (addr == 100) ? '\n' : '\r');
 506             fflush(stdout);
 507         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
 508             fprintf(stderr, "Unknown flags\n");
 509             return -EINVAL;
 510         }
 511         if (qemu_file_has_error(f)) {
 512             return -EIO;
 513         }
 514     } while (!(flags & BLK_MIG_FLAG_EOS));
 515
 516     return 0;
 517 }
 518
 519 static void block_set_params(int blk_enable, int shared_base, void *opaque)
 520 {
 521     block_mig_state.blk_enable = blk_enable;
 522     block_mig_state.shared_base = shared_base;
 523
 524     /* shared base means that blk_enable = 1 */
 525     block_mig_state.blk_enable |= shared_base;
 526 }
 527
 528 void blk_mig_init(void)
 529 {
 530     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
 531     QSIMPLEQ_INIT(&block_mig_state.blk_list);
 532
 533     register_savevm_live("block", 0, 1, block_set_params, block_save_live,
 534                          NULL, block_load, &block_mig_state);
 535 }