1 /* 2 * Image mirroring 3 * 4 * Copyright Red Hat, Inc. 2012 5 * 6 * Authors: 7 * Paolo Bonzini <pbonzini@redhat.com> 8 * 9 * This work is licensed under the terms of the GNU LGPL, version 2 or later. 10 * See the COPYING.LIB file in the top-level directory. 11 * 12 */ 13 14 #include "qemu/osdep.h" 15 #include "trace.h" 16 #include "block/blockjob_int.h" 17 #include "block/block_int.h" 18 #include "sysemu/block-backend.h" 19 #include "qapi/error.h" 20 #include "qapi/qmp/qerror.h" 21 #include "qemu/ratelimit.h" 22 #include "qemu/bitmap.h" 23 24 #define SLICE_TIME 100000000ULL /* ns */ 25 #define MAX_IN_FLIGHT 16 26 #define MAX_IO_SECTORS ((1 << 20) >> BDRV_SECTOR_BITS) /* 1 Mb */ 27 #define DEFAULT_MIRROR_BUF_SIZE \ 28 (MAX_IN_FLIGHT * MAX_IO_SECTORS * BDRV_SECTOR_SIZE) 29 30 /* The mirroring buffer is a list of granularity-sized chunks. 31 * Free chunks are organized in a list. 32 */ 33 typedef struct MirrorBuffer { 34 QSIMPLEQ_ENTRY(MirrorBuffer) next; 35 } MirrorBuffer; 36 37 typedef struct MirrorBlockJob { 38 BlockJob common; 39 RateLimit limit; 40 BlockBackend *target; 41 BlockDriverState *base; 42 /* The name of the graph node to replace */ 43 char *replaces; 44 /* The BDS to replace */ 45 BlockDriverState *to_replace; 46 /* Used to block operations on the drive-mirror-replace target */ 47 Error *replace_blocker; 48 bool is_none_mode; 49 BlockMirrorBackingMode backing_mode; 50 BlockdevOnError on_source_error, on_target_error; 51 bool synced; 52 bool should_complete; 53 int64_t granularity; 54 size_t buf_size; 55 int64_t bdev_length; 56 unsigned long *cow_bitmap; 57 BdrvDirtyBitmap *dirty_bitmap; 58 BdrvDirtyBitmapIter *dbi; 59 uint8_t *buf; 60 QSIMPLEQ_HEAD(, MirrorBuffer) buf_free; 61 int buf_free_count; 62 63 uint64_t last_pause_ns; 64 unsigned long *in_flight_bitmap; 65 int in_flight; 66 int64_t sectors_in_flight; 67 int ret; 68 bool unmap; 69 bool waiting_for_io; 70 int target_cluster_sectors; 71 int max_iov; 72 } MirrorBlockJob; 73 74 typedef struct MirrorOp { 75 MirrorBlockJob *s; 76 QEMUIOVector qiov; 77 int64_t sector_num; 78 int nb_sectors; 79 } MirrorOp; 80 81 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read, 82 int error) 83 { 84 s->synced = false; 85 if (read) { 86 return block_job_error_action(&s->common, s->on_source_error, 87 true, error); 88 } else { 89 return block_job_error_action(&s->common, s->on_target_error, 90 false, error); 91 } 92 } 93 94 static void mirror_iteration_done(MirrorOp *op, int ret) 95 { 96 MirrorBlockJob *s = op->s; 97 struct iovec *iov; 98 int64_t chunk_num; 99 int i, nb_chunks, sectors_per_chunk; 100 101 trace_mirror_iteration_done(s, op->sector_num, op->nb_sectors, ret); 102 103 s->in_flight--; 104 s->sectors_in_flight -= op->nb_sectors; 105 iov = op->qiov.iov; 106 for (i = 0; i < op->qiov.niov; i++) { 107 MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base; 108 QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next); 109 s->buf_free_count++; 110 } 111 112 sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS; 113 chunk_num = op->sector_num / sectors_per_chunk; 114 nb_chunks = DIV_ROUND_UP(op->nb_sectors, sectors_per_chunk); 115 bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks); 116 if (ret >= 0) { 117 if (s->cow_bitmap) { 118 bitmap_set(s->cow_bitmap, chunk_num, nb_chunks); 119 } 120 s->common.offset += (uint64_t)op->nb_sectors * BDRV_SECTOR_SIZE; 121 } 122 123 qemu_iovec_destroy(&op->qiov); 124 g_free(op); 125 126 if (s->waiting_for_io) { 127 qemu_coroutine_enter(s->common.co); 128 } 129 } 130 131 static void mirror_write_complete(void *opaque, int ret) 132 { 133 MirrorOp *op = opaque; 134 MirrorBlockJob *s = op->s; 135 if (ret < 0) { 136 BlockErrorAction action; 137 138 bdrv_set_dirty_bitmap(s->dirty_bitmap, op->sector_num, op->nb_sectors); 139 action = mirror_error_action(s, false, -ret); 140 if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) { 141 s->ret = ret; 142 } 143 } 144 mirror_iteration_done(op, ret); 145 } 146 147 static void mirror_read_complete(void *opaque, int ret) 148 { 149 MirrorOp *op = opaque; 150 MirrorBlockJob *s = op->s; 151 if (ret < 0) { 152 BlockErrorAction action; 153 154 bdrv_set_dirty_bitmap(s->dirty_bitmap, op->sector_num, op->nb_sectors); 155 action = mirror_error_action(s, true, -ret); 156 if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) { 157 s->ret = ret; 158 } 159 160 mirror_iteration_done(op, ret); 161 return; 162 } 163 blk_aio_pwritev(s->target, op->sector_num * BDRV_SECTOR_SIZE, &op->qiov, 164 0, mirror_write_complete, op); 165 } 166 167 static inline void mirror_clip_sectors(MirrorBlockJob *s, 168 int64_t sector_num, 169 int *nb_sectors) 170 { 171 *nb_sectors = MIN(*nb_sectors, 172 s->bdev_length / BDRV_SECTOR_SIZE - sector_num); 173 } 174 175 /* Round sector_num and/or nb_sectors to target cluster if COW is needed, and 176 * return the offset of the adjusted tail sector against original. */ 177 static int mirror_cow_align(MirrorBlockJob *s, 178 int64_t *sector_num, 179 int *nb_sectors) 180 { 181 bool need_cow; 182 int ret = 0; 183 int chunk_sectors = s->granularity >> BDRV_SECTOR_BITS; 184 int64_t align_sector_num = *sector_num; 185 int align_nb_sectors = *nb_sectors; 186 int max_sectors = chunk_sectors * s->max_iov; 187 188 need_cow = !test_bit(*sector_num / chunk_sectors, s->cow_bitmap); 189 need_cow |= !test_bit((*sector_num + *nb_sectors - 1) / chunk_sectors, 190 s->cow_bitmap); 191 if (need_cow) { 192 bdrv_round_sectors_to_clusters(blk_bs(s->target), *sector_num, 193 *nb_sectors, &align_sector_num, 194 &align_nb_sectors); 195 } 196 197 if (align_nb_sectors > max_sectors) { 198 align_nb_sectors = max_sectors; 199 if (need_cow) { 200 align_nb_sectors = QEMU_ALIGN_DOWN(align_nb_sectors, 201 s->target_cluster_sectors); 202 } 203 } 204 /* Clipping may result in align_nb_sectors unaligned to chunk boundary, but 205 * that doesn't matter because it's already the end of source image. */ 206 mirror_clip_sectors(s, align_sector_num, &align_nb_sectors); 207 208 ret = align_sector_num + align_nb_sectors - (*sector_num + *nb_sectors); 209 *sector_num = align_sector_num; 210 *nb_sectors = align_nb_sectors; 211 assert(ret >= 0); 212 return ret; 213 } 214 215 static inline void mirror_wait_for_io(MirrorBlockJob *s) 216 { 217 assert(!s->waiting_for_io); 218 s->waiting_for_io = true; 219 qemu_coroutine_yield(); 220 s->waiting_for_io = false; 221 } 222 223 /* Submit async read while handling COW. 224 * Returns: The number of sectors copied after and including sector_num, 225 * excluding any sectors copied prior to sector_num due to alignment. 226 * This will be nb_sectors if no alignment is necessary, or 227 * (new_end - sector_num) if tail is rounded up or down due to 228 * alignment or buffer limit. 229 */ 230 static int mirror_do_read(MirrorBlockJob *s, int64_t sector_num, 231 int nb_sectors) 232 { 233 BlockBackend *source = s->common.blk; 234 int sectors_per_chunk, nb_chunks; 235 int ret; 236 MirrorOp *op; 237 int max_sectors; 238 239 sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS; 240 max_sectors = sectors_per_chunk * s->max_iov; 241 242 /* We can only handle as much as buf_size at a time. */ 243 nb_sectors = MIN(s->buf_size >> BDRV_SECTOR_BITS, nb_sectors); 244 nb_sectors = MIN(max_sectors, nb_sectors); 245 assert(nb_sectors); 246 ret = nb_sectors; 247 248 if (s->cow_bitmap) { 249 ret += mirror_cow_align(s, §or_num, &nb_sectors); 250 } 251 assert(nb_sectors << BDRV_SECTOR_BITS <= s->buf_size); 252 /* The sector range must meet granularity because: 253 * 1) Caller passes in aligned values; 254 * 2) mirror_cow_align is used only when target cluster is larger. */ 255 assert(!(sector_num % sectors_per_chunk)); 256 nb_chunks = DIV_ROUND_UP(nb_sectors, sectors_per_chunk); 257 258 while (s->buf_free_count < nb_chunks) { 259 trace_mirror_yield_in_flight(s, sector_num, s->in_flight); 260 mirror_wait_for_io(s); 261 } 262 263 /* Allocate a MirrorOp that is used as an AIO callback. */ 264 op = g_new(MirrorOp, 1); 265 op->s = s; 266 op->sector_num = sector_num; 267 op->nb_sectors = nb_sectors; 268 269 /* Now make a QEMUIOVector taking enough granularity-sized chunks 270 * from s->buf_free. 271 */ 272 qemu_iovec_init(&op->qiov, nb_chunks); 273 while (nb_chunks-- > 0) { 274 MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free); 275 size_t remaining = nb_sectors * BDRV_SECTOR_SIZE - op->qiov.size; 276 277 QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next); 278 s->buf_free_count--; 279 qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining)); 280 } 281 282 /* Copy the dirty cluster. */ 283 s->in_flight++; 284 s->sectors_in_flight += nb_sectors; 285 trace_mirror_one_iteration(s, sector_num, nb_sectors); 286 287 blk_aio_preadv(source, sector_num * BDRV_SECTOR_SIZE, &op->qiov, 0, 288 mirror_read_complete, op); 289 return ret; 290 } 291 292 static void mirror_do_zero_or_discard(MirrorBlockJob *s, 293 int64_t sector_num, 294 int nb_sectors, 295 bool is_discard) 296 { 297 MirrorOp *op; 298 299 /* Allocate a MirrorOp that is used as an AIO callback. The qiov is zeroed 300 * so the freeing in mirror_iteration_done is nop. */ 301 op = g_new0(MirrorOp, 1); 302 op->s = s; 303 op->sector_num = sector_num; 304 op->nb_sectors = nb_sectors; 305 306 s->in_flight++; 307 s->sectors_in_flight += nb_sectors; 308 if (is_discard) { 309 blk_aio_pdiscard(s->target, sector_num << BDRV_SECTOR_BITS, 310 op->nb_sectors << BDRV_SECTOR_BITS, 311 mirror_write_complete, op); 312 } else { 313 blk_aio_pwrite_zeroes(s->target, sector_num * BDRV_SECTOR_SIZE, 314 op->nb_sectors * BDRV_SECTOR_SIZE, 315 s->unmap ? BDRV_REQ_MAY_UNMAP : 0, 316 mirror_write_complete, op); 317 } 318 } 319 320 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s) 321 { 322 BlockDriverState *source = blk_bs(s->common.blk); 323 int64_t sector_num, first_chunk; 324 uint64_t delay_ns = 0; 325 /* At least the first dirty chunk is mirrored in one iteration. */ 326 int nb_chunks = 1; 327 int64_t end = s->bdev_length / BDRV_SECTOR_SIZE; 328 int sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS; 329 bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target)); 330 int max_io_sectors = MAX((s->buf_size >> BDRV_SECTOR_BITS) / MAX_IN_FLIGHT, 331 MAX_IO_SECTORS); 332 333 sector_num = bdrv_dirty_iter_next(s->dbi); 334 if (sector_num < 0) { 335 bdrv_set_dirty_iter(s->dbi, 0); 336 sector_num = bdrv_dirty_iter_next(s->dbi); 337 trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap)); 338 assert(sector_num >= 0); 339 } 340 341 first_chunk = sector_num / sectors_per_chunk; 342 while (test_bit(first_chunk, s->in_flight_bitmap)) { 343 trace_mirror_yield_in_flight(s, sector_num, s->in_flight); 344 mirror_wait_for_io(s); 345 } 346 347 block_job_pause_point(&s->common); 348 349 /* Find the number of consective dirty chunks following the first dirty 350 * one, and wait for in flight requests in them. */ 351 while (nb_chunks * sectors_per_chunk < (s->buf_size >> BDRV_SECTOR_BITS)) { 352 int64_t next_dirty; 353 int64_t next_sector = sector_num + nb_chunks * sectors_per_chunk; 354 int64_t next_chunk = next_sector / sectors_per_chunk; 355 if (next_sector >= end || 356 !bdrv_get_dirty(source, s->dirty_bitmap, next_sector)) { 357 break; 358 } 359 if (test_bit(next_chunk, s->in_flight_bitmap)) { 360 break; 361 } 362 363 next_dirty = bdrv_dirty_iter_next(s->dbi); 364 if (next_dirty > next_sector || next_dirty < 0) { 365 /* The bitmap iterator's cache is stale, refresh it */ 366 bdrv_set_dirty_iter(s->dbi, next_sector); 367 next_dirty = bdrv_dirty_iter_next(s->dbi); 368 } 369 assert(next_dirty == next_sector); 370 nb_chunks++; 371 } 372 373 /* Clear dirty bits before querying the block status, because 374 * calling bdrv_get_block_status_above could yield - if some blocks are 375 * marked dirty in this window, we need to know. 376 */ 377 bdrv_reset_dirty_bitmap(s->dirty_bitmap, sector_num, 378 nb_chunks * sectors_per_chunk); 379 bitmap_set(s->in_flight_bitmap, sector_num / sectors_per_chunk, nb_chunks); 380 while (nb_chunks > 0 && sector_num < end) { 381 int ret; 382 int io_sectors, io_sectors_acct; 383 BlockDriverState *file; 384 enum MirrorMethod { 385 MIRROR_METHOD_COPY, 386 MIRROR_METHOD_ZERO, 387 MIRROR_METHOD_DISCARD 388 } mirror_method = MIRROR_METHOD_COPY; 389 390 assert(!(sector_num % sectors_per_chunk)); 391 ret = bdrv_get_block_status_above(source, NULL, sector_num, 392 nb_chunks * sectors_per_chunk, 393 &io_sectors, &file); 394 if (ret < 0) { 395 io_sectors = MIN(nb_chunks * sectors_per_chunk, max_io_sectors); 396 } else if (ret & BDRV_BLOCK_DATA) { 397 io_sectors = MIN(io_sectors, max_io_sectors); 398 } 399 400 io_sectors -= io_sectors % sectors_per_chunk; 401 if (io_sectors < sectors_per_chunk) { 402 io_sectors = sectors_per_chunk; 403 } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) { 404 int64_t target_sector_num; 405 int target_nb_sectors; 406 bdrv_round_sectors_to_clusters(blk_bs(s->target), sector_num, 407 io_sectors, &target_sector_num, 408 &target_nb_sectors); 409 if (target_sector_num == sector_num && 410 target_nb_sectors == io_sectors) { 411 mirror_method = ret & BDRV_BLOCK_ZERO ? 412 MIRROR_METHOD_ZERO : 413 MIRROR_METHOD_DISCARD; 414 } 415 } 416 417 while (s->in_flight >= MAX_IN_FLIGHT) { 418 trace_mirror_yield_in_flight(s, sector_num, s->in_flight); 419 mirror_wait_for_io(s); 420 } 421 422 if (s->ret < 0) { 423 return 0; 424 } 425 426 mirror_clip_sectors(s, sector_num, &io_sectors); 427 switch (mirror_method) { 428 case MIRROR_METHOD_COPY: 429 io_sectors = mirror_do_read(s, sector_num, io_sectors); 430 io_sectors_acct = io_sectors; 431 break; 432 case MIRROR_METHOD_ZERO: 433 case MIRROR_METHOD_DISCARD: 434 mirror_do_zero_or_discard(s, sector_num, io_sectors, 435 mirror_method == MIRROR_METHOD_DISCARD); 436 if (write_zeroes_ok) { 437 io_sectors_acct = 0; 438 } else { 439 io_sectors_acct = io_sectors; 440 } 441 break; 442 default: 443 abort(); 444 } 445 assert(io_sectors); 446 sector_num += io_sectors; 447 nb_chunks -= DIV_ROUND_UP(io_sectors, sectors_per_chunk); 448 if (s->common.speed) { 449 delay_ns = ratelimit_calculate_delay(&s->limit, io_sectors_acct); 450 } 451 } 452 return delay_ns; 453 } 454 455 static void mirror_free_init(MirrorBlockJob *s) 456 { 457 int granularity = s->granularity; 458 size_t buf_size = s->buf_size; 459 uint8_t *buf = s->buf; 460 461 assert(s->buf_free_count == 0); 462 QSIMPLEQ_INIT(&s->buf_free); 463 while (buf_size != 0) { 464 MirrorBuffer *cur = (MirrorBuffer *)buf; 465 QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next); 466 s->buf_free_count++; 467 buf_size -= granularity; 468 buf += granularity; 469 } 470 } 471 472 /* This is also used for the .pause callback. There is no matching 473 * mirror_resume() because mirror_run() will begin iterating again 474 * when the job is resumed. 475 */ 476 static void mirror_wait_for_all_io(MirrorBlockJob *s) 477 { 478 while (s->in_flight > 0) { 479 mirror_wait_for_io(s); 480 } 481 } 482 483 typedef struct { 484 int ret; 485 } MirrorExitData; 486 487 static void mirror_exit(BlockJob *job, void *opaque) 488 { 489 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common); 490 MirrorExitData *data = opaque; 491 AioContext *replace_aio_context = NULL; 492 BlockDriverState *src = blk_bs(s->common.blk); 493 BlockDriverState *target_bs = blk_bs(s->target); 494 495 /* Make sure that the source BDS doesn't go away before we called 496 * block_job_completed(). */ 497 bdrv_ref(src); 498 499 if (s->to_replace) { 500 replace_aio_context = bdrv_get_aio_context(s->to_replace); 501 aio_context_acquire(replace_aio_context); 502 } 503 504 if (s->should_complete && data->ret == 0) { 505 BlockDriverState *to_replace = src; 506 if (s->to_replace) { 507 to_replace = s->to_replace; 508 } 509 510 if (bdrv_get_flags(target_bs) != bdrv_get_flags(to_replace)) { 511 bdrv_reopen(target_bs, bdrv_get_flags(to_replace), NULL); 512 } 513 514 /* The mirror job has no requests in flight any more, but we need to 515 * drain potential other users of the BDS before changing the graph. */ 516 bdrv_drained_begin(target_bs); 517 bdrv_replace_in_backing_chain(to_replace, target_bs); 518 bdrv_drained_end(target_bs); 519 520 /* We just changed the BDS the job BB refers to */ 521 blk_remove_bs(job->blk); 522 blk_insert_bs(job->blk, src); 523 } 524 if (s->to_replace) { 525 bdrv_op_unblock_all(s->to_replace, s->replace_blocker); 526 error_free(s->replace_blocker); 527 bdrv_unref(s->to_replace); 528 } 529 if (replace_aio_context) { 530 aio_context_release(replace_aio_context); 531 } 532 g_free(s->replaces); 533 blk_unref(s->target); 534 s->target = NULL; 535 block_job_completed(&s->common, data->ret); 536 g_free(data); 537 bdrv_drained_end(src); 538 bdrv_unref(src); 539 } 540 541 static void mirror_throttle(MirrorBlockJob *s) 542 { 543 int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); 544 545 if (now - s->last_pause_ns > SLICE_TIME) { 546 s->last_pause_ns = now; 547 block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, 0); 548 } else { 549 block_job_pause_point(&s->common); 550 } 551 } 552 553 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s) 554 { 555 int64_t sector_num, end; 556 BlockDriverState *base = s->base; 557 BlockDriverState *bs = blk_bs(s->common.blk); 558 BlockDriverState *target_bs = blk_bs(s->target); 559 int ret, n; 560 561 end = s->bdev_length / BDRV_SECTOR_SIZE; 562 563 if (base == NULL && !bdrv_has_zero_init(target_bs)) { 564 if (!bdrv_can_write_zeroes_with_unmap(target_bs)) { 565 bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, end); 566 return 0; 567 } 568 569 for (sector_num = 0; sector_num < end; ) { 570 int nb_sectors = MIN(end - sector_num, 571 QEMU_ALIGN_DOWN(INT_MAX, s->granularity) >> BDRV_SECTOR_BITS); 572 573 mirror_throttle(s); 574 575 if (block_job_is_cancelled(&s->common)) { 576 return 0; 577 } 578 579 if (s->in_flight >= MAX_IN_FLIGHT) { 580 trace_mirror_yield(s, s->in_flight, s->buf_free_count, -1); 581 mirror_wait_for_io(s); 582 continue; 583 } 584 585 mirror_do_zero_or_discard(s, sector_num, nb_sectors, false); 586 sector_num += nb_sectors; 587 } 588 589 mirror_wait_for_all_io(s); 590 } 591 592 /* First part, loop on the sectors and initialize the dirty bitmap. */ 593 for (sector_num = 0; sector_num < end; ) { 594 /* Just to make sure we are not exceeding int limit. */ 595 int nb_sectors = MIN(INT_MAX >> BDRV_SECTOR_BITS, 596 end - sector_num); 597 598 mirror_throttle(s); 599 600 if (block_job_is_cancelled(&s->common)) { 601 return 0; 602 } 603 604 ret = bdrv_is_allocated_above(bs, base, sector_num, nb_sectors, &n); 605 if (ret < 0) { 606 return ret; 607 } 608 609 assert(n > 0); 610 if (ret == 1) { 611 bdrv_set_dirty_bitmap(s->dirty_bitmap, sector_num, n); 612 } 613 sector_num += n; 614 } 615 return 0; 616 } 617 618 /* Called when going out of the streaming phase to flush the bulk of the 619 * data to the medium, or just before completing. 620 */ 621 static int mirror_flush(MirrorBlockJob *s) 622 { 623 int ret = blk_flush(s->target); 624 if (ret < 0) { 625 if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) { 626 s->ret = ret; 627 } 628 } 629 return ret; 630 } 631 632 static void coroutine_fn mirror_run(void *opaque) 633 { 634 MirrorBlockJob *s = opaque; 635 MirrorExitData *data; 636 BlockDriverState *bs = blk_bs(s->common.blk); 637 BlockDriverState *target_bs = blk_bs(s->target); 638 bool need_drain = true; 639 int64_t length; 640 BlockDriverInfo bdi; 641 char backing_filename[2]; /* we only need 2 characters because we are only 642 checking for a NULL string */ 643 int ret = 0; 644 int target_cluster_size = BDRV_SECTOR_SIZE; 645 646 if (block_job_is_cancelled(&s->common)) { 647 goto immediate_exit; 648 } 649 650 s->bdev_length = bdrv_getlength(bs); 651 if (s->bdev_length < 0) { 652 ret = s->bdev_length; 653 goto immediate_exit; 654 } else if (s->bdev_length == 0) { 655 /* Report BLOCK_JOB_READY and wait for complete. */ 656 block_job_event_ready(&s->common); 657 s->synced = true; 658 while (!block_job_is_cancelled(&s->common) && !s->should_complete) { 659 block_job_yield(&s->common); 660 } 661 s->common.cancelled = false; 662 goto immediate_exit; 663 } 664 665 length = DIV_ROUND_UP(s->bdev_length, s->granularity); 666 s->in_flight_bitmap = bitmap_new(length); 667 668 /* If we have no backing file yet in the destination, we cannot let 669 * the destination do COW. Instead, we copy sectors around the 670 * dirty data if needed. We need a bitmap to do that. 671 */ 672 bdrv_get_backing_filename(target_bs, backing_filename, 673 sizeof(backing_filename)); 674 if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) { 675 target_cluster_size = bdi.cluster_size; 676 } 677 if (backing_filename[0] && !target_bs->backing 678 && s->granularity < target_cluster_size) { 679 s->buf_size = MAX(s->buf_size, target_cluster_size); 680 s->cow_bitmap = bitmap_new(length); 681 } 682 s->target_cluster_sectors = target_cluster_size >> BDRV_SECTOR_BITS; 683 s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov); 684 685 s->buf = qemu_try_blockalign(bs, s->buf_size); 686 if (s->buf == NULL) { 687 ret = -ENOMEM; 688 goto immediate_exit; 689 } 690 691 mirror_free_init(s); 692 693 s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); 694 if (!s->is_none_mode) { 695 ret = mirror_dirty_init(s); 696 if (ret < 0 || block_job_is_cancelled(&s->common)) { 697 goto immediate_exit; 698 } 699 } 700 701 assert(!s->dbi); 702 s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap, 0); 703 for (;;) { 704 uint64_t delay_ns = 0; 705 int64_t cnt, delta; 706 bool should_complete; 707 708 if (s->ret < 0) { 709 ret = s->ret; 710 goto immediate_exit; 711 } 712 713 block_job_pause_point(&s->common); 714 715 cnt = bdrv_get_dirty_count(s->dirty_bitmap); 716 /* s->common.offset contains the number of bytes already processed so 717 * far, cnt is the number of dirty sectors remaining and 718 * s->sectors_in_flight is the number of sectors currently being 719 * processed; together those are the current total operation length */ 720 s->common.len = s->common.offset + 721 (cnt + s->sectors_in_flight) * BDRV_SECTOR_SIZE; 722 723 /* Note that even when no rate limit is applied we need to yield 724 * periodically with no pending I/O so that bdrv_drain_all() returns. 725 * We do so every SLICE_TIME nanoseconds, or when there is an error, 726 * or when the source is clean, whichever comes first. 727 */ 728 delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns; 729 if (delta < SLICE_TIME && 730 s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) { 731 if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 || 732 (cnt == 0 && s->in_flight > 0)) { 733 trace_mirror_yield(s, s->in_flight, s->buf_free_count, cnt); 734 mirror_wait_for_io(s); 735 continue; 736 } else if (cnt != 0) { 737 delay_ns = mirror_iteration(s); 738 } 739 } 740 741 should_complete = false; 742 if (s->in_flight == 0 && cnt == 0) { 743 trace_mirror_before_flush(s); 744 if (!s->synced) { 745 if (mirror_flush(s) < 0) { 746 /* Go check s->ret. */ 747 continue; 748 } 749 /* We're out of the streaming phase. From now on, if the job 750 * is cancelled we will actually complete all pending I/O and 751 * report completion. This way, block-job-cancel will leave 752 * the target in a consistent state. 753 */ 754 block_job_event_ready(&s->common); 755 s->synced = true; 756 } 757 758 should_complete = s->should_complete || 759 block_job_is_cancelled(&s->common); 760 cnt = bdrv_get_dirty_count(s->dirty_bitmap); 761 } 762 763 if (cnt == 0 && should_complete) { 764 /* The dirty bitmap is not updated while operations are pending. 765 * If we're about to exit, wait for pending operations before 766 * calling bdrv_get_dirty_count(bs), or we may exit while the 767 * source has dirty data to copy! 768 * 769 * Note that I/O can be submitted by the guest while 770 * mirror_populate runs, so pause it now. Before deciding 771 * whether to switch to target check one last time if I/O has 772 * come in the meanwhile, and if not flush the data to disk. 773 */ 774 trace_mirror_before_drain(s, cnt); 775 776 bdrv_drained_begin(bs); 777 cnt = bdrv_get_dirty_count(s->dirty_bitmap); 778 if (cnt > 0 || mirror_flush(s) < 0) { 779 bdrv_drained_end(bs); 780 continue; 781 } 782 783 /* The two disks are in sync. Exit and report successful 784 * completion. 785 */ 786 assert(QLIST_EMPTY(&bs->tracked_requests)); 787 s->common.cancelled = false; 788 need_drain = false; 789 break; 790 } 791 792 ret = 0; 793 trace_mirror_before_sleep(s, cnt, s->synced, delay_ns); 794 if (!s->synced) { 795 block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns); 796 if (block_job_is_cancelled(&s->common)) { 797 break; 798 } 799 } else if (!should_complete) { 800 delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0); 801 block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns); 802 } 803 s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); 804 } 805 806 immediate_exit: 807 if (s->in_flight > 0) { 808 /* We get here only if something went wrong. Either the job failed, 809 * or it was cancelled prematurely so that we do not guarantee that 810 * the target is a copy of the source. 811 */ 812 assert(ret < 0 || (!s->synced && block_job_is_cancelled(&s->common))); 813 assert(need_drain); 814 mirror_wait_for_all_io(s); 815 } 816 817 assert(s->in_flight == 0); 818 qemu_vfree(s->buf); 819 g_free(s->cow_bitmap); 820 g_free(s->in_flight_bitmap); 821 bdrv_dirty_iter_free(s->dbi); 822 bdrv_release_dirty_bitmap(bs, s->dirty_bitmap); 823 824 data = g_malloc(sizeof(*data)); 825 data->ret = ret; 826 827 if (need_drain) { 828 bdrv_drained_begin(bs); 829 } 830 block_job_defer_to_main_loop(&s->common, mirror_exit, data); 831 } 832 833 static void mirror_set_speed(BlockJob *job, int64_t speed, Error **errp) 834 { 835 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common); 836 837 if (speed < 0) { 838 error_setg(errp, QERR_INVALID_PARAMETER, "speed"); 839 return; 840 } 841 ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME); 842 } 843 844 static void mirror_complete(BlockJob *job, Error **errp) 845 { 846 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common); 847 BlockDriverState *src, *target; 848 849 src = blk_bs(job->blk); 850 target = blk_bs(s->target); 851 852 if (!s->synced) { 853 error_setg(errp, "The active block job '%s' cannot be completed", 854 job->id); 855 return; 856 } 857 858 if (s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) { 859 int ret; 860 861 assert(!target->backing); 862 ret = bdrv_open_backing_file(target, NULL, "backing", errp); 863 if (ret < 0) { 864 return; 865 } 866 } 867 868 /* block all operations on to_replace bs */ 869 if (s->replaces) { 870 AioContext *replace_aio_context; 871 872 s->to_replace = bdrv_find_node(s->replaces); 873 if (!s->to_replace) { 874 error_setg(errp, "Node name '%s' not found", s->replaces); 875 return; 876 } 877 878 replace_aio_context = bdrv_get_aio_context(s->to_replace); 879 aio_context_acquire(replace_aio_context); 880 881 error_setg(&s->replace_blocker, 882 "block device is in use by block-job-complete"); 883 bdrv_op_block_all(s->to_replace, s->replace_blocker); 884 bdrv_ref(s->to_replace); 885 886 aio_context_release(replace_aio_context); 887 } 888 889 if (s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) { 890 BlockDriverState *backing = s->is_none_mode ? src : s->base; 891 if (backing_bs(target) != backing) { 892 bdrv_set_backing_hd(target, backing); 893 } 894 } 895 896 s->should_complete = true; 897 block_job_enter(&s->common); 898 } 899 900 static void mirror_pause(BlockJob *job) 901 { 902 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common); 903 904 mirror_wait_for_all_io(s); 905 } 906 907 static void mirror_attached_aio_context(BlockJob *job, AioContext *new_context) 908 { 909 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common); 910 911 blk_set_aio_context(s->target, new_context); 912 } 913 914 static void mirror_drain(BlockJob *job) 915 { 916 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common); 917 918 /* Need to keep a reference in case blk_drain triggers execution 919 * of mirror_complete... 920 */ 921 if (s->target) { 922 BlockBackend *target = s->target; 923 blk_ref(target); 924 blk_drain(target); 925 blk_unref(target); 926 } 927 } 928 929 static const BlockJobDriver mirror_job_driver = { 930 .instance_size = sizeof(MirrorBlockJob), 931 .job_type = BLOCK_JOB_TYPE_MIRROR, 932 .set_speed = mirror_set_speed, 933 .start = mirror_run, 934 .complete = mirror_complete, 935 .pause = mirror_pause, 936 .attached_aio_context = mirror_attached_aio_context, 937 .drain = mirror_drain, 938 }; 939 940 static const BlockJobDriver commit_active_job_driver = { 941 .instance_size = sizeof(MirrorBlockJob), 942 .job_type = BLOCK_JOB_TYPE_COMMIT, 943 .set_speed = mirror_set_speed, 944 .start = mirror_run, 945 .complete = mirror_complete, 946 .pause = mirror_pause, 947 .attached_aio_context = mirror_attached_aio_context, 948 .drain = mirror_drain, 949 }; 950 951 static void mirror_start_job(const char *job_id, BlockDriverState *bs, 952 int creation_flags, BlockDriverState *target, 953 const char *replaces, int64_t speed, 954 uint32_t granularity, int64_t buf_size, 955 BlockMirrorBackingMode backing_mode, 956 BlockdevOnError on_source_error, 957 BlockdevOnError on_target_error, 958 bool unmap, 959 BlockCompletionFunc *cb, 960 void *opaque, Error **errp, 961 const BlockJobDriver *driver, 962 bool is_none_mode, BlockDriverState *base, 963 bool auto_complete) 964 { 965 MirrorBlockJob *s; 966 967 if (granularity == 0) { 968 granularity = bdrv_get_default_bitmap_granularity(target); 969 } 970 971 assert ((granularity & (granularity - 1)) == 0); 972 973 if (buf_size < 0) { 974 error_setg(errp, "Invalid parameter 'buf-size'"); 975 return; 976 } 977 978 if (buf_size == 0) { 979 buf_size = DEFAULT_MIRROR_BUF_SIZE; 980 } 981 982 s = block_job_create(job_id, driver, bs, speed, creation_flags, 983 cb, opaque, errp); 984 if (!s) { 985 return; 986 } 987 988 s->target = blk_new(); 989 blk_insert_bs(s->target, target); 990 991 s->replaces = g_strdup(replaces); 992 s->on_source_error = on_source_error; 993 s->on_target_error = on_target_error; 994 s->is_none_mode = is_none_mode; 995 s->backing_mode = backing_mode; 996 s->base = base; 997 s->granularity = granularity; 998 s->buf_size = ROUND_UP(buf_size, granularity); 999 s->unmap = unmap; 1000 if (auto_complete) { 1001 s->should_complete = true; 1002 } 1003 1004 s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp); 1005 if (!s->dirty_bitmap) { 1006 g_free(s->replaces); 1007 blk_unref(s->target); 1008 block_job_unref(&s->common); 1009 return; 1010 } 1011 1012 block_job_add_bdrv(&s->common, target); 1013 /* In commit_active_start() all intermediate nodes disappear, so 1014 * any jobs in them must be blocked */ 1015 if (bdrv_chain_contains(bs, target)) { 1016 BlockDriverState *iter; 1017 for (iter = backing_bs(bs); iter != target; iter = backing_bs(iter)) { 1018 block_job_add_bdrv(&s->common, iter); 1019 } 1020 } 1021 1022 trace_mirror_start(bs, s, opaque); 1023 block_job_start(&s->common); 1024 } 1025 1026 void mirror_start(const char *job_id, BlockDriverState *bs, 1027 BlockDriverState *target, const char *replaces, 1028 int64_t speed, uint32_t granularity, int64_t buf_size, 1029 MirrorSyncMode mode, BlockMirrorBackingMode backing_mode, 1030 BlockdevOnError on_source_error, 1031 BlockdevOnError on_target_error, 1032 bool unmap, Error **errp) 1033 { 1034 bool is_none_mode; 1035 BlockDriverState *base; 1036 1037 if (mode == MIRROR_SYNC_MODE_INCREMENTAL) { 1038 error_setg(errp, "Sync mode 'incremental' not supported"); 1039 return; 1040 } 1041 is_none_mode = mode == MIRROR_SYNC_MODE_NONE; 1042 base = mode == MIRROR_SYNC_MODE_TOP ? backing_bs(bs) : NULL; 1043 mirror_start_job(job_id, bs, BLOCK_JOB_DEFAULT, target, replaces, 1044 speed, granularity, buf_size, backing_mode, 1045 on_source_error, on_target_error, unmap, NULL, NULL, errp, 1046 &mirror_job_driver, is_none_mode, base, false); 1047 } 1048 1049 void commit_active_start(const char *job_id, BlockDriverState *bs, 1050 BlockDriverState *base, int creation_flags, 1051 int64_t speed, BlockdevOnError on_error, 1052 BlockCompletionFunc *cb, void *opaque, Error **errp, 1053 bool auto_complete) 1054 { 1055 int64_t length, base_length; 1056 int orig_base_flags; 1057 int ret; 1058 Error *local_err = NULL; 1059 1060 orig_base_flags = bdrv_get_flags(base); 1061 1062 if (bdrv_reopen(base, bs->open_flags, errp)) { 1063 return; 1064 } 1065 1066 length = bdrv_getlength(bs); 1067 if (length < 0) { 1068 error_setg_errno(errp, -length, 1069 "Unable to determine length of %s", bs->filename); 1070 goto error_restore_flags; 1071 } 1072 1073 base_length = bdrv_getlength(base); 1074 if (base_length < 0) { 1075 error_setg_errno(errp, -base_length, 1076 "Unable to determine length of %s", base->filename); 1077 goto error_restore_flags; 1078 } 1079 1080 if (length > base_length) { 1081 ret = bdrv_truncate(base, length); 1082 if (ret < 0) { 1083 error_setg_errno(errp, -ret, 1084 "Top image %s is larger than base image %s, and " 1085 "resize of base image failed", 1086 bs->filename, base->filename); 1087 goto error_restore_flags; 1088 } 1089 } 1090 1091 mirror_start_job(job_id, bs, creation_flags, base, NULL, speed, 0, 0, 1092 MIRROR_LEAVE_BACKING_CHAIN, 1093 on_error, on_error, true, cb, opaque, &local_err, 1094 &commit_active_job_driver, false, base, auto_complete); 1095 if (local_err) { 1096 error_propagate(errp, local_err); 1097 goto error_restore_flags; 1098 } 1099 1100 return; 1101 1102 error_restore_flags: 1103 /* ignore error and errp for bdrv_reopen, because we want to propagate 1104 * the original error */ 1105 bdrv_reopen(base, orig_base_flags, NULL); 1106 return; 1107 } 1108