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