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