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 "qemu/cutils.h" 16 #include "qemu/coroutine.h" 17 #include "qemu/range.h" 18 #include "trace.h" 19 #include "block/blockjob_int.h" 20 #include "block/block_int.h" 21 #include "sysemu/block-backend.h" 22 #include "qapi/error.h" 23 #include "qapi/qmp/qerror.h" 24 #include "qemu/ratelimit.h" 25 #include "qemu/bitmap.h" 26 27 #define MAX_IN_FLIGHT 16 28 #define MAX_IO_BYTES (1 << 20) /* 1 Mb */ 29 #define DEFAULT_MIRROR_BUF_SIZE (MAX_IN_FLIGHT * MAX_IO_BYTES) 30 31 /* The mirroring buffer is a list of granularity-sized chunks. 32 * Free chunks are organized in a list. 33 */ 34 typedef struct MirrorBuffer { 35 QSIMPLEQ_ENTRY(MirrorBuffer) next; 36 } MirrorBuffer; 37 38 typedef struct MirrorOp MirrorOp; 39 40 typedef struct MirrorBlockJob { 41 BlockJob common; 42 BlockBackend *target; 43 BlockDriverState *mirror_top_bs; 44 BlockDriverState *base; 45 BlockDriverState *base_overlay; 46 47 /* The name of the graph node to replace */ 48 char *replaces; 49 /* The BDS to replace */ 50 BlockDriverState *to_replace; 51 /* Used to block operations on the drive-mirror-replace target */ 52 Error *replace_blocker; 53 bool is_none_mode; 54 BlockMirrorBackingMode backing_mode; 55 /* Whether the target image requires explicit zero-initialization */ 56 bool zero_target; 57 MirrorCopyMode copy_mode; 58 BlockdevOnError on_source_error, on_target_error; 59 bool synced; 60 /* Set when the target is synced (dirty bitmap is clean, nothing 61 * in flight) and the job is running in active mode */ 62 bool actively_synced; 63 bool should_complete; 64 int64_t granularity; 65 size_t buf_size; 66 int64_t bdev_length; 67 unsigned long *cow_bitmap; 68 BdrvDirtyBitmap *dirty_bitmap; 69 BdrvDirtyBitmapIter *dbi; 70 uint8_t *buf; 71 QSIMPLEQ_HEAD(, MirrorBuffer) buf_free; 72 int buf_free_count; 73 74 uint64_t last_pause_ns; 75 unsigned long *in_flight_bitmap; 76 int in_flight; 77 int64_t bytes_in_flight; 78 QTAILQ_HEAD(, MirrorOp) ops_in_flight; 79 int ret; 80 bool unmap; 81 int target_cluster_size; 82 int max_iov; 83 bool initial_zeroing_ongoing; 84 int in_active_write_counter; 85 bool prepared; 86 bool in_drain; 87 } MirrorBlockJob; 88 89 typedef struct MirrorBDSOpaque { 90 MirrorBlockJob *job; 91 bool stop; 92 bool is_commit; 93 } MirrorBDSOpaque; 94 95 struct MirrorOp { 96 MirrorBlockJob *s; 97 QEMUIOVector qiov; 98 int64_t offset; 99 uint64_t bytes; 100 101 /* The pointee is set by mirror_co_read(), mirror_co_zero(), and 102 * mirror_co_discard() before yielding for the first time */ 103 int64_t *bytes_handled; 104 105 bool is_pseudo_op; 106 bool is_active_write; 107 bool is_in_flight; 108 CoQueue waiting_requests; 109 Coroutine *co; 110 111 QTAILQ_ENTRY(MirrorOp) next; 112 }; 113 114 typedef enum MirrorMethod { 115 MIRROR_METHOD_COPY, 116 MIRROR_METHOD_ZERO, 117 MIRROR_METHOD_DISCARD, 118 } MirrorMethod; 119 120 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read, 121 int error) 122 { 123 s->synced = false; 124 s->actively_synced = false; 125 if (read) { 126 return block_job_error_action(&s->common, s->on_source_error, 127 true, error); 128 } else { 129 return block_job_error_action(&s->common, s->on_target_error, 130 false, error); 131 } 132 } 133 134 static void coroutine_fn mirror_wait_on_conflicts(MirrorOp *self, 135 MirrorBlockJob *s, 136 uint64_t offset, 137 uint64_t bytes) 138 { 139 uint64_t self_start_chunk = offset / s->granularity; 140 uint64_t self_end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity); 141 uint64_t self_nb_chunks = self_end_chunk - self_start_chunk; 142 143 while (find_next_bit(s->in_flight_bitmap, self_end_chunk, 144 self_start_chunk) < self_end_chunk && 145 s->ret >= 0) 146 { 147 MirrorOp *op; 148 149 QTAILQ_FOREACH(op, &s->ops_in_flight, next) { 150 uint64_t op_start_chunk = op->offset / s->granularity; 151 uint64_t op_nb_chunks = DIV_ROUND_UP(op->offset + op->bytes, 152 s->granularity) - 153 op_start_chunk; 154 155 if (op == self) { 156 continue; 157 } 158 159 if (ranges_overlap(self_start_chunk, self_nb_chunks, 160 op_start_chunk, op_nb_chunks)) 161 { 162 qemu_co_queue_wait(&op->waiting_requests, NULL); 163 break; 164 } 165 } 166 } 167 } 168 169 static void coroutine_fn mirror_iteration_done(MirrorOp *op, int ret) 170 { 171 MirrorBlockJob *s = op->s; 172 struct iovec *iov; 173 int64_t chunk_num; 174 int i, nb_chunks; 175 176 trace_mirror_iteration_done(s, op->offset, op->bytes, ret); 177 178 s->in_flight--; 179 s->bytes_in_flight -= op->bytes; 180 iov = op->qiov.iov; 181 for (i = 0; i < op->qiov.niov; i++) { 182 MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base; 183 QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next); 184 s->buf_free_count++; 185 } 186 187 chunk_num = op->offset / s->granularity; 188 nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity); 189 190 bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks); 191 QTAILQ_REMOVE(&s->ops_in_flight, op, next); 192 if (ret >= 0) { 193 if (s->cow_bitmap) { 194 bitmap_set(s->cow_bitmap, chunk_num, nb_chunks); 195 } 196 if (!s->initial_zeroing_ongoing) { 197 job_progress_update(&s->common.job, op->bytes); 198 } 199 } 200 qemu_iovec_destroy(&op->qiov); 201 202 qemu_co_queue_restart_all(&op->waiting_requests); 203 g_free(op); 204 } 205 206 static void coroutine_fn mirror_write_complete(MirrorOp *op, int ret) 207 { 208 MirrorBlockJob *s = op->s; 209 210 if (ret < 0) { 211 BlockErrorAction action; 212 213 bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes); 214 action = mirror_error_action(s, false, -ret); 215 if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) { 216 s->ret = ret; 217 } 218 } 219 220 mirror_iteration_done(op, ret); 221 } 222 223 static void coroutine_fn mirror_read_complete(MirrorOp *op, int ret) 224 { 225 MirrorBlockJob *s = op->s; 226 227 if (ret < 0) { 228 BlockErrorAction action; 229 230 bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes); 231 action = mirror_error_action(s, true, -ret); 232 if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) { 233 s->ret = ret; 234 } 235 236 mirror_iteration_done(op, ret); 237 return; 238 } 239 240 ret = blk_co_pwritev(s->target, op->offset, op->qiov.size, &op->qiov, 0); 241 mirror_write_complete(op, ret); 242 } 243 244 /* Clip bytes relative to offset to not exceed end-of-file */ 245 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s, 246 int64_t offset, 247 int64_t bytes) 248 { 249 return MIN(bytes, s->bdev_length - offset); 250 } 251 252 /* Round offset and/or bytes to target cluster if COW is needed, and 253 * return the offset of the adjusted tail against original. */ 254 static int mirror_cow_align(MirrorBlockJob *s, int64_t *offset, 255 uint64_t *bytes) 256 { 257 bool need_cow; 258 int ret = 0; 259 int64_t align_offset = *offset; 260 int64_t align_bytes = *bytes; 261 int max_bytes = s->granularity * s->max_iov; 262 263 need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap); 264 need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity, 265 s->cow_bitmap); 266 if (need_cow) { 267 bdrv_round_to_clusters(blk_bs(s->target), *offset, *bytes, 268 &align_offset, &align_bytes); 269 } 270 271 if (align_bytes > max_bytes) { 272 align_bytes = max_bytes; 273 if (need_cow) { 274 align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size); 275 } 276 } 277 /* Clipping may result in align_bytes unaligned to chunk boundary, but 278 * that doesn't matter because it's already the end of source image. */ 279 align_bytes = mirror_clip_bytes(s, align_offset, align_bytes); 280 281 ret = align_offset + align_bytes - (*offset + *bytes); 282 *offset = align_offset; 283 *bytes = align_bytes; 284 assert(ret >= 0); 285 return ret; 286 } 287 288 static inline void coroutine_fn 289 mirror_wait_for_any_operation(MirrorBlockJob *s, bool active) 290 { 291 MirrorOp *op; 292 293 QTAILQ_FOREACH(op, &s->ops_in_flight, next) { 294 /* Do not wait on pseudo ops, because it may in turn wait on 295 * some other operation to start, which may in fact be the 296 * caller of this function. Since there is only one pseudo op 297 * at any given time, we will always find some real operation 298 * to wait on. */ 299 if (!op->is_pseudo_op && op->is_in_flight && 300 op->is_active_write == active) 301 { 302 qemu_co_queue_wait(&op->waiting_requests, NULL); 303 return; 304 } 305 } 306 abort(); 307 } 308 309 static inline void coroutine_fn 310 mirror_wait_for_free_in_flight_slot(MirrorBlockJob *s) 311 { 312 /* Only non-active operations use up in-flight slots */ 313 mirror_wait_for_any_operation(s, false); 314 } 315 316 /* Perform a mirror copy operation. 317 * 318 * *op->bytes_handled is set to the number of bytes copied after and 319 * including offset, excluding any bytes copied prior to offset due 320 * to alignment. This will be op->bytes if no alignment is necessary, 321 * or (new_end - op->offset) if the tail is rounded up or down due to 322 * alignment or buffer limit. 323 */ 324 static void coroutine_fn mirror_co_read(void *opaque) 325 { 326 MirrorOp *op = opaque; 327 MirrorBlockJob *s = op->s; 328 int nb_chunks; 329 uint64_t ret; 330 uint64_t max_bytes; 331 332 max_bytes = s->granularity * s->max_iov; 333 334 /* We can only handle as much as buf_size at a time. */ 335 op->bytes = MIN(s->buf_size, MIN(max_bytes, op->bytes)); 336 assert(op->bytes); 337 assert(op->bytes < BDRV_REQUEST_MAX_BYTES); 338 *op->bytes_handled = op->bytes; 339 340 if (s->cow_bitmap) { 341 *op->bytes_handled += mirror_cow_align(s, &op->offset, &op->bytes); 342 } 343 /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */ 344 assert(*op->bytes_handled <= UINT_MAX); 345 assert(op->bytes <= s->buf_size); 346 /* The offset is granularity-aligned because: 347 * 1) Caller passes in aligned values; 348 * 2) mirror_cow_align is used only when target cluster is larger. */ 349 assert(QEMU_IS_ALIGNED(op->offset, s->granularity)); 350 /* The range is sector-aligned, since bdrv_getlength() rounds up. */ 351 assert(QEMU_IS_ALIGNED(op->bytes, BDRV_SECTOR_SIZE)); 352 nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity); 353 354 while (s->buf_free_count < nb_chunks) { 355 trace_mirror_yield_in_flight(s, op->offset, s->in_flight); 356 mirror_wait_for_free_in_flight_slot(s); 357 } 358 359 /* Now make a QEMUIOVector taking enough granularity-sized chunks 360 * from s->buf_free. 361 */ 362 qemu_iovec_init(&op->qiov, nb_chunks); 363 while (nb_chunks-- > 0) { 364 MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free); 365 size_t remaining = op->bytes - op->qiov.size; 366 367 QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next); 368 s->buf_free_count--; 369 qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining)); 370 } 371 372 /* Copy the dirty cluster. */ 373 s->in_flight++; 374 s->bytes_in_flight += op->bytes; 375 op->is_in_flight = true; 376 trace_mirror_one_iteration(s, op->offset, op->bytes); 377 378 ret = bdrv_co_preadv(s->mirror_top_bs->backing, op->offset, op->bytes, 379 &op->qiov, 0); 380 mirror_read_complete(op, ret); 381 } 382 383 static void coroutine_fn mirror_co_zero(void *opaque) 384 { 385 MirrorOp *op = opaque; 386 int ret; 387 388 op->s->in_flight++; 389 op->s->bytes_in_flight += op->bytes; 390 *op->bytes_handled = op->bytes; 391 op->is_in_flight = true; 392 393 ret = blk_co_pwrite_zeroes(op->s->target, op->offset, op->bytes, 394 op->s->unmap ? BDRV_REQ_MAY_UNMAP : 0); 395 mirror_write_complete(op, ret); 396 } 397 398 static void coroutine_fn mirror_co_discard(void *opaque) 399 { 400 MirrorOp *op = opaque; 401 int ret; 402 403 op->s->in_flight++; 404 op->s->bytes_in_flight += op->bytes; 405 *op->bytes_handled = op->bytes; 406 op->is_in_flight = true; 407 408 ret = blk_co_pdiscard(op->s->target, op->offset, op->bytes); 409 mirror_write_complete(op, ret); 410 } 411 412 static unsigned mirror_perform(MirrorBlockJob *s, int64_t offset, 413 unsigned bytes, MirrorMethod mirror_method) 414 { 415 MirrorOp *op; 416 Coroutine *co; 417 int64_t bytes_handled = -1; 418 419 op = g_new(MirrorOp, 1); 420 *op = (MirrorOp){ 421 .s = s, 422 .offset = offset, 423 .bytes = bytes, 424 .bytes_handled = &bytes_handled, 425 }; 426 qemu_co_queue_init(&op->waiting_requests); 427 428 switch (mirror_method) { 429 case MIRROR_METHOD_COPY: 430 co = qemu_coroutine_create(mirror_co_read, op); 431 break; 432 case MIRROR_METHOD_ZERO: 433 co = qemu_coroutine_create(mirror_co_zero, op); 434 break; 435 case MIRROR_METHOD_DISCARD: 436 co = qemu_coroutine_create(mirror_co_discard, op); 437 break; 438 default: 439 abort(); 440 } 441 op->co = co; 442 443 QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next); 444 qemu_coroutine_enter(co); 445 /* At this point, ownership of op has been moved to the coroutine 446 * and the object may already be freed */ 447 448 /* Assert that this value has been set */ 449 assert(bytes_handled >= 0); 450 451 /* Same assertion as in mirror_co_read() (and for mirror_co_read() 452 * and mirror_co_discard(), bytes_handled == op->bytes, which 453 * is the @bytes parameter given to this function) */ 454 assert(bytes_handled <= UINT_MAX); 455 return bytes_handled; 456 } 457 458 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s) 459 { 460 BlockDriverState *source = s->mirror_top_bs->backing->bs; 461 MirrorOp *pseudo_op; 462 int64_t offset; 463 uint64_t delay_ns = 0, ret = 0; 464 /* At least the first dirty chunk is mirrored in one iteration. */ 465 int nb_chunks = 1; 466 bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target)); 467 int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES); 468 469 bdrv_dirty_bitmap_lock(s->dirty_bitmap); 470 offset = bdrv_dirty_iter_next(s->dbi); 471 if (offset < 0) { 472 bdrv_set_dirty_iter(s->dbi, 0); 473 offset = bdrv_dirty_iter_next(s->dbi); 474 trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap)); 475 assert(offset >= 0); 476 } 477 bdrv_dirty_bitmap_unlock(s->dirty_bitmap); 478 479 mirror_wait_on_conflicts(NULL, s, offset, 1); 480 481 job_pause_point(&s->common.job); 482 483 /* Find the number of consective dirty chunks following the first dirty 484 * one, and wait for in flight requests in them. */ 485 bdrv_dirty_bitmap_lock(s->dirty_bitmap); 486 while (nb_chunks * s->granularity < s->buf_size) { 487 int64_t next_dirty; 488 int64_t next_offset = offset + nb_chunks * s->granularity; 489 int64_t next_chunk = next_offset / s->granularity; 490 if (next_offset >= s->bdev_length || 491 !bdrv_dirty_bitmap_get_locked(s->dirty_bitmap, next_offset)) { 492 break; 493 } 494 if (test_bit(next_chunk, s->in_flight_bitmap)) { 495 break; 496 } 497 498 next_dirty = bdrv_dirty_iter_next(s->dbi); 499 if (next_dirty > next_offset || next_dirty < 0) { 500 /* The bitmap iterator's cache is stale, refresh it */ 501 bdrv_set_dirty_iter(s->dbi, next_offset); 502 next_dirty = bdrv_dirty_iter_next(s->dbi); 503 } 504 assert(next_dirty == next_offset); 505 nb_chunks++; 506 } 507 508 /* Clear dirty bits before querying the block status, because 509 * calling bdrv_block_status_above could yield - if some blocks are 510 * marked dirty in this window, we need to know. 511 */ 512 bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset, 513 nb_chunks * s->granularity); 514 bdrv_dirty_bitmap_unlock(s->dirty_bitmap); 515 516 /* Before claiming an area in the in-flight bitmap, we have to 517 * create a MirrorOp for it so that conflicting requests can wait 518 * for it. mirror_perform() will create the real MirrorOps later, 519 * for now we just create a pseudo operation that will wake up all 520 * conflicting requests once all real operations have been 521 * launched. */ 522 pseudo_op = g_new(MirrorOp, 1); 523 *pseudo_op = (MirrorOp){ 524 .offset = offset, 525 .bytes = nb_chunks * s->granularity, 526 .is_pseudo_op = true, 527 }; 528 qemu_co_queue_init(&pseudo_op->waiting_requests); 529 QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next); 530 531 bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks); 532 while (nb_chunks > 0 && offset < s->bdev_length) { 533 int ret; 534 int64_t io_bytes; 535 int64_t io_bytes_acct; 536 MirrorMethod mirror_method = MIRROR_METHOD_COPY; 537 538 assert(!(offset % s->granularity)); 539 ret = bdrv_block_status_above(source, NULL, offset, 540 nb_chunks * s->granularity, 541 &io_bytes, NULL, NULL); 542 if (ret < 0) { 543 io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes); 544 } else if (ret & BDRV_BLOCK_DATA) { 545 io_bytes = MIN(io_bytes, max_io_bytes); 546 } 547 548 io_bytes -= io_bytes % s->granularity; 549 if (io_bytes < s->granularity) { 550 io_bytes = s->granularity; 551 } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) { 552 int64_t target_offset; 553 int64_t target_bytes; 554 bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes, 555 &target_offset, &target_bytes); 556 if (target_offset == offset && 557 target_bytes == io_bytes) { 558 mirror_method = ret & BDRV_BLOCK_ZERO ? 559 MIRROR_METHOD_ZERO : 560 MIRROR_METHOD_DISCARD; 561 } 562 } 563 564 while (s->in_flight >= MAX_IN_FLIGHT) { 565 trace_mirror_yield_in_flight(s, offset, s->in_flight); 566 mirror_wait_for_free_in_flight_slot(s); 567 } 568 569 if (s->ret < 0) { 570 ret = 0; 571 goto fail; 572 } 573 574 io_bytes = mirror_clip_bytes(s, offset, io_bytes); 575 io_bytes = mirror_perform(s, offset, io_bytes, mirror_method); 576 if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) { 577 io_bytes_acct = 0; 578 } else { 579 io_bytes_acct = io_bytes; 580 } 581 assert(io_bytes); 582 offset += io_bytes; 583 nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity); 584 delay_ns = block_job_ratelimit_get_delay(&s->common, io_bytes_acct); 585 } 586 587 ret = delay_ns; 588 fail: 589 QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next); 590 qemu_co_queue_restart_all(&pseudo_op->waiting_requests); 591 g_free(pseudo_op); 592 593 return ret; 594 } 595 596 static void mirror_free_init(MirrorBlockJob *s) 597 { 598 int granularity = s->granularity; 599 size_t buf_size = s->buf_size; 600 uint8_t *buf = s->buf; 601 602 assert(s->buf_free_count == 0); 603 QSIMPLEQ_INIT(&s->buf_free); 604 while (buf_size != 0) { 605 MirrorBuffer *cur = (MirrorBuffer *)buf; 606 QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next); 607 s->buf_free_count++; 608 buf_size -= granularity; 609 buf += granularity; 610 } 611 } 612 613 /* This is also used for the .pause callback. There is no matching 614 * mirror_resume() because mirror_run() will begin iterating again 615 * when the job is resumed. 616 */ 617 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s) 618 { 619 while (s->in_flight > 0) { 620 mirror_wait_for_free_in_flight_slot(s); 621 } 622 } 623 624 /** 625 * mirror_exit_common: handle both abort() and prepare() cases. 626 * for .prepare, returns 0 on success and -errno on failure. 627 * for .abort cases, denoted by abort = true, MUST return 0. 628 */ 629 static int mirror_exit_common(Job *job) 630 { 631 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 632 BlockJob *bjob = &s->common; 633 MirrorBDSOpaque *bs_opaque; 634 AioContext *replace_aio_context = NULL; 635 BlockDriverState *src; 636 BlockDriverState *target_bs; 637 BlockDriverState *mirror_top_bs; 638 Error *local_err = NULL; 639 bool abort = job->ret < 0; 640 int ret = 0; 641 642 if (s->prepared) { 643 return 0; 644 } 645 s->prepared = true; 646 647 mirror_top_bs = s->mirror_top_bs; 648 bs_opaque = mirror_top_bs->opaque; 649 src = mirror_top_bs->backing->bs; 650 target_bs = blk_bs(s->target); 651 652 if (bdrv_chain_contains(src, target_bs)) { 653 bdrv_unfreeze_backing_chain(mirror_top_bs, target_bs); 654 } 655 656 bdrv_release_dirty_bitmap(s->dirty_bitmap); 657 658 /* Make sure that the source BDS doesn't go away during bdrv_replace_node, 659 * before we can call bdrv_drained_end */ 660 bdrv_ref(src); 661 bdrv_ref(mirror_top_bs); 662 bdrv_ref(target_bs); 663 664 /* 665 * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before 666 * inserting target_bs at s->to_replace, where we might not be able to get 667 * these permissions. 668 */ 669 blk_unref(s->target); 670 s->target = NULL; 671 672 /* We don't access the source any more. Dropping any WRITE/RESIZE is 673 * required before it could become a backing file of target_bs. Not having 674 * these permissions any more means that we can't allow any new requests on 675 * mirror_top_bs from now on, so keep it drained. */ 676 bdrv_drained_begin(mirror_top_bs); 677 bs_opaque->stop = true; 678 bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing, 679 &error_abort); 680 if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) { 681 BlockDriverState *backing = s->is_none_mode ? src : s->base; 682 BlockDriverState *unfiltered_target = bdrv_skip_filters(target_bs); 683 684 if (bdrv_cow_bs(unfiltered_target) != backing) { 685 bdrv_set_backing_hd(unfiltered_target, backing, &local_err); 686 if (local_err) { 687 error_report_err(local_err); 688 local_err = NULL; 689 ret = -EPERM; 690 } 691 } 692 } else if (!abort && s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) { 693 assert(!bdrv_backing_chain_next(target_bs)); 694 ret = bdrv_open_backing_file(bdrv_skip_filters(target_bs), NULL, 695 "backing", &local_err); 696 if (ret < 0) { 697 error_report_err(local_err); 698 local_err = NULL; 699 } 700 } 701 702 if (s->to_replace) { 703 replace_aio_context = bdrv_get_aio_context(s->to_replace); 704 aio_context_acquire(replace_aio_context); 705 } 706 707 if (s->should_complete && !abort) { 708 BlockDriverState *to_replace = s->to_replace ?: src; 709 bool ro = bdrv_is_read_only(to_replace); 710 711 if (ro != bdrv_is_read_only(target_bs)) { 712 bdrv_reopen_set_read_only(target_bs, ro, NULL); 713 } 714 715 /* The mirror job has no requests in flight any more, but we need to 716 * drain potential other users of the BDS before changing the graph. */ 717 assert(s->in_drain); 718 bdrv_drained_begin(target_bs); 719 /* 720 * Cannot use check_to_replace_node() here, because that would 721 * check for an op blocker on @to_replace, and we have our own 722 * there. 723 */ 724 if (bdrv_recurse_can_replace(src, to_replace)) { 725 bdrv_replace_node(to_replace, target_bs, &local_err); 726 } else { 727 error_setg(&local_err, "Can no longer replace '%s' by '%s', " 728 "because it can no longer be guaranteed that doing so " 729 "would not lead to an abrupt change of visible data", 730 to_replace->node_name, target_bs->node_name); 731 } 732 bdrv_drained_end(target_bs); 733 if (local_err) { 734 error_report_err(local_err); 735 ret = -EPERM; 736 } 737 } 738 if (s->to_replace) { 739 bdrv_op_unblock_all(s->to_replace, s->replace_blocker); 740 error_free(s->replace_blocker); 741 bdrv_unref(s->to_replace); 742 } 743 if (replace_aio_context) { 744 aio_context_release(replace_aio_context); 745 } 746 g_free(s->replaces); 747 bdrv_unref(target_bs); 748 749 /* 750 * Remove the mirror filter driver from the graph. Before this, get rid of 751 * the blockers on the intermediate nodes so that the resulting state is 752 * valid. 753 */ 754 block_job_remove_all_bdrv(bjob); 755 bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort); 756 757 /* We just changed the BDS the job BB refers to (with either or both of the 758 * bdrv_replace_node() calls), so switch the BB back so the cleanup does 759 * the right thing. We don't need any permissions any more now. */ 760 blk_remove_bs(bjob->blk); 761 blk_set_perm(bjob->blk, 0, BLK_PERM_ALL, &error_abort); 762 blk_insert_bs(bjob->blk, mirror_top_bs, &error_abort); 763 764 bs_opaque->job = NULL; 765 766 bdrv_drained_end(src); 767 bdrv_drained_end(mirror_top_bs); 768 s->in_drain = false; 769 bdrv_unref(mirror_top_bs); 770 bdrv_unref(src); 771 772 return ret; 773 } 774 775 static int mirror_prepare(Job *job) 776 { 777 return mirror_exit_common(job); 778 } 779 780 static void mirror_abort(Job *job) 781 { 782 int ret = mirror_exit_common(job); 783 assert(ret == 0); 784 } 785 786 static void coroutine_fn mirror_throttle(MirrorBlockJob *s) 787 { 788 int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); 789 790 if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) { 791 s->last_pause_ns = now; 792 job_sleep_ns(&s->common.job, 0); 793 } else { 794 job_pause_point(&s->common.job); 795 } 796 } 797 798 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s) 799 { 800 int64_t offset; 801 BlockDriverState *bs = s->mirror_top_bs->backing->bs; 802 BlockDriverState *target_bs = blk_bs(s->target); 803 int ret; 804 int64_t count; 805 806 if (s->zero_target) { 807 if (!bdrv_can_write_zeroes_with_unmap(target_bs)) { 808 bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length); 809 return 0; 810 } 811 812 s->initial_zeroing_ongoing = true; 813 for (offset = 0; offset < s->bdev_length; ) { 814 int bytes = MIN(s->bdev_length - offset, 815 QEMU_ALIGN_DOWN(INT_MAX, s->granularity)); 816 817 mirror_throttle(s); 818 819 if (job_is_cancelled(&s->common.job)) { 820 s->initial_zeroing_ongoing = false; 821 return 0; 822 } 823 824 if (s->in_flight >= MAX_IN_FLIGHT) { 825 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count, 826 s->in_flight); 827 mirror_wait_for_free_in_flight_slot(s); 828 continue; 829 } 830 831 mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO); 832 offset += bytes; 833 } 834 835 mirror_wait_for_all_io(s); 836 s->initial_zeroing_ongoing = false; 837 } 838 839 /* First part, loop on the sectors and initialize the dirty bitmap. */ 840 for (offset = 0; offset < s->bdev_length; ) { 841 /* Just to make sure we are not exceeding int limit. */ 842 int bytes = MIN(s->bdev_length - offset, 843 QEMU_ALIGN_DOWN(INT_MAX, s->granularity)); 844 845 mirror_throttle(s); 846 847 if (job_is_cancelled(&s->common.job)) { 848 return 0; 849 } 850 851 ret = bdrv_is_allocated_above(bs, s->base_overlay, true, offset, bytes, 852 &count); 853 if (ret < 0) { 854 return ret; 855 } 856 857 assert(count); 858 if (ret > 0) { 859 bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count); 860 } 861 offset += count; 862 } 863 return 0; 864 } 865 866 /* Called when going out of the streaming phase to flush the bulk of the 867 * data to the medium, or just before completing. 868 */ 869 static int mirror_flush(MirrorBlockJob *s) 870 { 871 int ret = blk_flush(s->target); 872 if (ret < 0) { 873 if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) { 874 s->ret = ret; 875 } 876 } 877 return ret; 878 } 879 880 static int coroutine_fn mirror_run(Job *job, Error **errp) 881 { 882 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 883 BlockDriverState *bs = s->mirror_top_bs->backing->bs; 884 BlockDriverState *target_bs = blk_bs(s->target); 885 bool need_drain = true; 886 int64_t length; 887 int64_t target_length; 888 BlockDriverInfo bdi; 889 char backing_filename[2]; /* we only need 2 characters because we are only 890 checking for a NULL string */ 891 int ret = 0; 892 893 if (job_is_cancelled(&s->common.job)) { 894 goto immediate_exit; 895 } 896 897 s->bdev_length = bdrv_getlength(bs); 898 if (s->bdev_length < 0) { 899 ret = s->bdev_length; 900 goto immediate_exit; 901 } 902 903 target_length = blk_getlength(s->target); 904 if (target_length < 0) { 905 ret = target_length; 906 goto immediate_exit; 907 } 908 909 /* Active commit must resize the base image if its size differs from the 910 * active layer. */ 911 if (s->base == blk_bs(s->target)) { 912 if (s->bdev_length > target_length) { 913 ret = blk_truncate(s->target, s->bdev_length, false, 914 PREALLOC_MODE_OFF, 0, NULL); 915 if (ret < 0) { 916 goto immediate_exit; 917 } 918 } 919 } else if (s->bdev_length != target_length) { 920 error_setg(errp, "Source and target image have different sizes"); 921 ret = -EINVAL; 922 goto immediate_exit; 923 } 924 925 if (s->bdev_length == 0) { 926 /* Transition to the READY state and wait for complete. */ 927 job_transition_to_ready(&s->common.job); 928 s->synced = true; 929 s->actively_synced = true; 930 while (!job_is_cancelled(&s->common.job) && !s->should_complete) { 931 job_yield(&s->common.job); 932 } 933 s->common.job.cancelled = false; 934 goto immediate_exit; 935 } 936 937 length = DIV_ROUND_UP(s->bdev_length, s->granularity); 938 s->in_flight_bitmap = bitmap_new(length); 939 940 /* If we have no backing file yet in the destination, we cannot let 941 * the destination do COW. Instead, we copy sectors around the 942 * dirty data if needed. We need a bitmap to do that. 943 */ 944 bdrv_get_backing_filename(target_bs, backing_filename, 945 sizeof(backing_filename)); 946 if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) { 947 s->target_cluster_size = bdi.cluster_size; 948 } else { 949 s->target_cluster_size = BDRV_SECTOR_SIZE; 950 } 951 if (backing_filename[0] && !bdrv_backing_chain_next(target_bs) && 952 s->granularity < s->target_cluster_size) { 953 s->buf_size = MAX(s->buf_size, s->target_cluster_size); 954 s->cow_bitmap = bitmap_new(length); 955 } 956 s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov); 957 958 s->buf = qemu_try_blockalign(bs, s->buf_size); 959 if (s->buf == NULL) { 960 ret = -ENOMEM; 961 goto immediate_exit; 962 } 963 964 mirror_free_init(s); 965 966 s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); 967 if (!s->is_none_mode) { 968 ret = mirror_dirty_init(s); 969 if (ret < 0 || job_is_cancelled(&s->common.job)) { 970 goto immediate_exit; 971 } 972 } 973 974 assert(!s->dbi); 975 s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap); 976 for (;;) { 977 uint64_t delay_ns = 0; 978 int64_t cnt, delta; 979 bool should_complete; 980 981 /* Do not start passive operations while there are active 982 * writes in progress */ 983 while (s->in_active_write_counter) { 984 mirror_wait_for_any_operation(s, true); 985 } 986 987 if (s->ret < 0) { 988 ret = s->ret; 989 goto immediate_exit; 990 } 991 992 job_pause_point(&s->common.job); 993 994 cnt = bdrv_get_dirty_count(s->dirty_bitmap); 995 /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is 996 * the number of bytes currently being processed; together those are 997 * the current remaining operation length */ 998 job_progress_set_remaining(&s->common.job, s->bytes_in_flight + cnt); 999 1000 /* Note that even when no rate limit is applied we need to yield 1001 * periodically with no pending I/O so that bdrv_drain_all() returns. 1002 * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is 1003 * an error, or when the source is clean, whichever comes first. */ 1004 delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns; 1005 if (delta < BLOCK_JOB_SLICE_TIME && 1006 s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) { 1007 if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 || 1008 (cnt == 0 && s->in_flight > 0)) { 1009 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight); 1010 mirror_wait_for_free_in_flight_slot(s); 1011 continue; 1012 } else if (cnt != 0) { 1013 delay_ns = mirror_iteration(s); 1014 } 1015 } 1016 1017 should_complete = false; 1018 if (s->in_flight == 0 && cnt == 0) { 1019 trace_mirror_before_flush(s); 1020 if (!s->synced) { 1021 if (mirror_flush(s) < 0) { 1022 /* Go check s->ret. */ 1023 continue; 1024 } 1025 /* We're out of the streaming phase. From now on, if the job 1026 * is cancelled we will actually complete all pending I/O and 1027 * report completion. This way, block-job-cancel will leave 1028 * the target in a consistent state. 1029 */ 1030 job_transition_to_ready(&s->common.job); 1031 s->synced = true; 1032 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) { 1033 s->actively_synced = true; 1034 } 1035 } 1036 1037 should_complete = s->should_complete || 1038 job_is_cancelled(&s->common.job); 1039 cnt = bdrv_get_dirty_count(s->dirty_bitmap); 1040 } 1041 1042 if (cnt == 0 && should_complete) { 1043 /* The dirty bitmap is not updated while operations are pending. 1044 * If we're about to exit, wait for pending operations before 1045 * calling bdrv_get_dirty_count(bs), or we may exit while the 1046 * source has dirty data to copy! 1047 * 1048 * Note that I/O can be submitted by the guest while 1049 * mirror_populate runs, so pause it now. Before deciding 1050 * whether to switch to target check one last time if I/O has 1051 * come in the meanwhile, and if not flush the data to disk. 1052 */ 1053 trace_mirror_before_drain(s, cnt); 1054 1055 s->in_drain = true; 1056 bdrv_drained_begin(bs); 1057 cnt = bdrv_get_dirty_count(s->dirty_bitmap); 1058 if (cnt > 0 || mirror_flush(s) < 0) { 1059 bdrv_drained_end(bs); 1060 s->in_drain = false; 1061 continue; 1062 } 1063 1064 /* The two disks are in sync. Exit and report successful 1065 * completion. 1066 */ 1067 assert(QLIST_EMPTY(&bs->tracked_requests)); 1068 s->common.job.cancelled = false; 1069 need_drain = false; 1070 break; 1071 } 1072 1073 ret = 0; 1074 1075 if (s->synced && !should_complete) { 1076 delay_ns = (s->in_flight == 0 && 1077 cnt == 0 ? BLOCK_JOB_SLICE_TIME : 0); 1078 } 1079 trace_mirror_before_sleep(s, cnt, s->synced, delay_ns); 1080 job_sleep_ns(&s->common.job, delay_ns); 1081 if (job_is_cancelled(&s->common.job) && 1082 (!s->synced || s->common.job.force_cancel)) 1083 { 1084 break; 1085 } 1086 s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); 1087 } 1088 1089 immediate_exit: 1090 if (s->in_flight > 0) { 1091 /* We get here only if something went wrong. Either the job failed, 1092 * or it was cancelled prematurely so that we do not guarantee that 1093 * the target is a copy of the source. 1094 */ 1095 assert(ret < 0 || ((s->common.job.force_cancel || !s->synced) && 1096 job_is_cancelled(&s->common.job))); 1097 assert(need_drain); 1098 mirror_wait_for_all_io(s); 1099 } 1100 1101 assert(s->in_flight == 0); 1102 qemu_vfree(s->buf); 1103 g_free(s->cow_bitmap); 1104 g_free(s->in_flight_bitmap); 1105 bdrv_dirty_iter_free(s->dbi); 1106 1107 if (need_drain) { 1108 s->in_drain = true; 1109 bdrv_drained_begin(bs); 1110 } 1111 1112 return ret; 1113 } 1114 1115 static void mirror_complete(Job *job, Error **errp) 1116 { 1117 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 1118 1119 if (!s->synced) { 1120 error_setg(errp, "The active block job '%s' cannot be completed", 1121 job->id); 1122 return; 1123 } 1124 1125 /* block all operations on to_replace bs */ 1126 if (s->replaces) { 1127 AioContext *replace_aio_context; 1128 1129 s->to_replace = bdrv_find_node(s->replaces); 1130 if (!s->to_replace) { 1131 error_setg(errp, "Node name '%s' not found", s->replaces); 1132 return; 1133 } 1134 1135 replace_aio_context = bdrv_get_aio_context(s->to_replace); 1136 aio_context_acquire(replace_aio_context); 1137 1138 /* TODO Translate this into permission system. Current definition of 1139 * GRAPH_MOD would require to request it for the parents; they might 1140 * not even be BlockDriverStates, however, so a BdrvChild can't address 1141 * them. May need redefinition of GRAPH_MOD. */ 1142 error_setg(&s->replace_blocker, 1143 "block device is in use by block-job-complete"); 1144 bdrv_op_block_all(s->to_replace, s->replace_blocker); 1145 bdrv_ref(s->to_replace); 1146 1147 aio_context_release(replace_aio_context); 1148 } 1149 1150 s->should_complete = true; 1151 1152 /* If the job is paused, it will be re-entered when it is resumed */ 1153 if (!job->paused) { 1154 job_enter(job); 1155 } 1156 } 1157 1158 static void coroutine_fn mirror_pause(Job *job) 1159 { 1160 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 1161 1162 mirror_wait_for_all_io(s); 1163 } 1164 1165 static bool mirror_drained_poll(BlockJob *job) 1166 { 1167 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common); 1168 1169 /* If the job isn't paused nor cancelled, we can't be sure that it won't 1170 * issue more requests. We make an exception if we've reached this point 1171 * from one of our own drain sections, to avoid a deadlock waiting for 1172 * ourselves. 1173 */ 1174 if (!s->common.job.paused && !s->common.job.cancelled && !s->in_drain) { 1175 return true; 1176 } 1177 1178 return !!s->in_flight; 1179 } 1180 1181 static void mirror_cancel(Job *job) 1182 { 1183 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 1184 BlockDriverState *target = blk_bs(s->target); 1185 1186 bdrv_cancel_in_flight(target); 1187 } 1188 1189 static const BlockJobDriver mirror_job_driver = { 1190 .job_driver = { 1191 .instance_size = sizeof(MirrorBlockJob), 1192 .job_type = JOB_TYPE_MIRROR, 1193 .free = block_job_free, 1194 .user_resume = block_job_user_resume, 1195 .run = mirror_run, 1196 .prepare = mirror_prepare, 1197 .abort = mirror_abort, 1198 .pause = mirror_pause, 1199 .complete = mirror_complete, 1200 .cancel = mirror_cancel, 1201 }, 1202 .drained_poll = mirror_drained_poll, 1203 }; 1204 1205 static const BlockJobDriver commit_active_job_driver = { 1206 .job_driver = { 1207 .instance_size = sizeof(MirrorBlockJob), 1208 .job_type = JOB_TYPE_COMMIT, 1209 .free = block_job_free, 1210 .user_resume = block_job_user_resume, 1211 .run = mirror_run, 1212 .prepare = mirror_prepare, 1213 .abort = mirror_abort, 1214 .pause = mirror_pause, 1215 .complete = mirror_complete, 1216 }, 1217 .drained_poll = mirror_drained_poll, 1218 }; 1219 1220 static void coroutine_fn 1221 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method, 1222 uint64_t offset, uint64_t bytes, 1223 QEMUIOVector *qiov, int flags) 1224 { 1225 int ret; 1226 size_t qiov_offset = 0; 1227 int64_t bitmap_offset, bitmap_end; 1228 1229 if (!QEMU_IS_ALIGNED(offset, job->granularity) && 1230 bdrv_dirty_bitmap_get(job->dirty_bitmap, offset)) 1231 { 1232 /* 1233 * Dirty unaligned padding: ignore it. 1234 * 1235 * Reasoning: 1236 * 1. If we copy it, we can't reset corresponding bit in 1237 * dirty_bitmap as there may be some "dirty" bytes still not 1238 * copied. 1239 * 2. It's already dirty, so skipping it we don't diverge mirror 1240 * progress. 1241 * 1242 * Note, that because of this, guest write may have no contribution 1243 * into mirror converge, but that's not bad, as we have background 1244 * process of mirroring. If under some bad circumstances (high guest 1245 * IO load) background process starve, we will not converge anyway, 1246 * even if each write will contribute, as guest is not guaranteed to 1247 * rewrite the whole disk. 1248 */ 1249 qiov_offset = QEMU_ALIGN_UP(offset, job->granularity) - offset; 1250 if (bytes <= qiov_offset) { 1251 /* nothing to do after shrink */ 1252 return; 1253 } 1254 offset += qiov_offset; 1255 bytes -= qiov_offset; 1256 } 1257 1258 if (!QEMU_IS_ALIGNED(offset + bytes, job->granularity) && 1259 bdrv_dirty_bitmap_get(job->dirty_bitmap, offset + bytes - 1)) 1260 { 1261 uint64_t tail = (offset + bytes) % job->granularity; 1262 1263 if (bytes <= tail) { 1264 /* nothing to do after shrink */ 1265 return; 1266 } 1267 bytes -= tail; 1268 } 1269 1270 /* 1271 * Tails are either clean or shrunk, so for bitmap resetting 1272 * we safely align the range down. 1273 */ 1274 bitmap_offset = QEMU_ALIGN_UP(offset, job->granularity); 1275 bitmap_end = QEMU_ALIGN_DOWN(offset + bytes, job->granularity); 1276 if (bitmap_offset < bitmap_end) { 1277 bdrv_reset_dirty_bitmap(job->dirty_bitmap, bitmap_offset, 1278 bitmap_end - bitmap_offset); 1279 } 1280 1281 job_progress_increase_remaining(&job->common.job, bytes); 1282 1283 switch (method) { 1284 case MIRROR_METHOD_COPY: 1285 ret = blk_co_pwritev_part(job->target, offset, bytes, 1286 qiov, qiov_offset, flags); 1287 break; 1288 1289 case MIRROR_METHOD_ZERO: 1290 assert(!qiov); 1291 ret = blk_co_pwrite_zeroes(job->target, offset, bytes, flags); 1292 break; 1293 1294 case MIRROR_METHOD_DISCARD: 1295 assert(!qiov); 1296 ret = blk_co_pdiscard(job->target, offset, bytes); 1297 break; 1298 1299 default: 1300 abort(); 1301 } 1302 1303 if (ret >= 0) { 1304 job_progress_update(&job->common.job, bytes); 1305 } else { 1306 BlockErrorAction action; 1307 1308 /* 1309 * We failed, so we should mark dirty the whole area, aligned up. 1310 * Note that we don't care about shrunk tails if any: they were dirty 1311 * at function start, and they must be still dirty, as we've locked 1312 * the region for in-flight op. 1313 */ 1314 bitmap_offset = QEMU_ALIGN_DOWN(offset, job->granularity); 1315 bitmap_end = QEMU_ALIGN_UP(offset + bytes, job->granularity); 1316 bdrv_set_dirty_bitmap(job->dirty_bitmap, bitmap_offset, 1317 bitmap_end - bitmap_offset); 1318 job->actively_synced = false; 1319 1320 action = mirror_error_action(job, false, -ret); 1321 if (action == BLOCK_ERROR_ACTION_REPORT) { 1322 if (!job->ret) { 1323 job->ret = ret; 1324 } 1325 } 1326 } 1327 } 1328 1329 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s, 1330 uint64_t offset, 1331 uint64_t bytes) 1332 { 1333 MirrorOp *op; 1334 uint64_t start_chunk = offset / s->granularity; 1335 uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity); 1336 1337 op = g_new(MirrorOp, 1); 1338 *op = (MirrorOp){ 1339 .s = s, 1340 .offset = offset, 1341 .bytes = bytes, 1342 .is_active_write = true, 1343 .is_in_flight = true, 1344 }; 1345 qemu_co_queue_init(&op->waiting_requests); 1346 QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next); 1347 1348 s->in_active_write_counter++; 1349 1350 mirror_wait_on_conflicts(op, s, offset, bytes); 1351 1352 bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk); 1353 1354 return op; 1355 } 1356 1357 static void coroutine_fn active_write_settle(MirrorOp *op) 1358 { 1359 uint64_t start_chunk = op->offset / op->s->granularity; 1360 uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes, 1361 op->s->granularity); 1362 1363 if (!--op->s->in_active_write_counter && op->s->actively_synced) { 1364 BdrvChild *source = op->s->mirror_top_bs->backing; 1365 1366 if (QLIST_FIRST(&source->bs->parents) == source && 1367 QLIST_NEXT(source, next_parent) == NULL) 1368 { 1369 /* Assert that we are back in sync once all active write 1370 * operations are settled. 1371 * Note that we can only assert this if the mirror node 1372 * is the source node's only parent. */ 1373 assert(!bdrv_get_dirty_count(op->s->dirty_bitmap)); 1374 } 1375 } 1376 bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk); 1377 QTAILQ_REMOVE(&op->s->ops_in_flight, op, next); 1378 qemu_co_queue_restart_all(&op->waiting_requests); 1379 g_free(op); 1380 } 1381 1382 static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs, 1383 uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags) 1384 { 1385 return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags); 1386 } 1387 1388 static int coroutine_fn bdrv_mirror_top_do_write(BlockDriverState *bs, 1389 MirrorMethod method, uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, 1390 int flags) 1391 { 1392 MirrorOp *op = NULL; 1393 MirrorBDSOpaque *s = bs->opaque; 1394 int ret = 0; 1395 bool copy_to_target; 1396 1397 copy_to_target = s->job->ret >= 0 && 1398 s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING; 1399 1400 if (copy_to_target) { 1401 op = active_write_prepare(s->job, offset, bytes); 1402 } 1403 1404 switch (method) { 1405 case MIRROR_METHOD_COPY: 1406 ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags); 1407 break; 1408 1409 case MIRROR_METHOD_ZERO: 1410 ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags); 1411 break; 1412 1413 case MIRROR_METHOD_DISCARD: 1414 ret = bdrv_co_pdiscard(bs->backing, offset, bytes); 1415 break; 1416 1417 default: 1418 abort(); 1419 } 1420 1421 if (ret < 0) { 1422 goto out; 1423 } 1424 1425 if (copy_to_target) { 1426 do_sync_target_write(s->job, method, offset, bytes, qiov, flags); 1427 } 1428 1429 out: 1430 if (copy_to_target) { 1431 active_write_settle(op); 1432 } 1433 return ret; 1434 } 1435 1436 static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs, 1437 uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags) 1438 { 1439 MirrorBDSOpaque *s = bs->opaque; 1440 QEMUIOVector bounce_qiov; 1441 void *bounce_buf; 1442 int ret = 0; 1443 bool copy_to_target; 1444 1445 copy_to_target = s->job->ret >= 0 && 1446 s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING; 1447 1448 if (copy_to_target) { 1449 /* The guest might concurrently modify the data to write; but 1450 * the data on source and destination must match, so we have 1451 * to use a bounce buffer if we are going to write to the 1452 * target now. */ 1453 bounce_buf = qemu_blockalign(bs, bytes); 1454 iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes); 1455 1456 qemu_iovec_init(&bounce_qiov, 1); 1457 qemu_iovec_add(&bounce_qiov, bounce_buf, bytes); 1458 qiov = &bounce_qiov; 1459 } 1460 1461 ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov, 1462 flags); 1463 1464 if (copy_to_target) { 1465 qemu_iovec_destroy(&bounce_qiov); 1466 qemu_vfree(bounce_buf); 1467 } 1468 1469 return ret; 1470 } 1471 1472 static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs) 1473 { 1474 if (bs->backing == NULL) { 1475 /* we can be here after failed bdrv_append in mirror_start_job */ 1476 return 0; 1477 } 1478 return bdrv_co_flush(bs->backing->bs); 1479 } 1480 1481 static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs, 1482 int64_t offset, int bytes, BdrvRequestFlags flags) 1483 { 1484 return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL, 1485 flags); 1486 } 1487 1488 static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs, 1489 int64_t offset, int bytes) 1490 { 1491 return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes, 1492 NULL, 0); 1493 } 1494 1495 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs) 1496 { 1497 if (bs->backing == NULL) { 1498 /* we can be here after failed bdrv_attach_child in 1499 * bdrv_set_backing_hd */ 1500 return; 1501 } 1502 pstrcpy(bs->exact_filename, sizeof(bs->exact_filename), 1503 bs->backing->bs->filename); 1504 } 1505 1506 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c, 1507 BdrvChildRole role, 1508 BlockReopenQueue *reopen_queue, 1509 uint64_t perm, uint64_t shared, 1510 uint64_t *nperm, uint64_t *nshared) 1511 { 1512 MirrorBDSOpaque *s = bs->opaque; 1513 1514 if (s->stop) { 1515 /* 1516 * If the job is to be stopped, we do not need to forward 1517 * anything to the real image. 1518 */ 1519 *nperm = 0; 1520 *nshared = BLK_PERM_ALL; 1521 return; 1522 } 1523 1524 bdrv_default_perms(bs, c, role, reopen_queue, 1525 perm, shared, nperm, nshared); 1526 1527 if (s->is_commit) { 1528 /* 1529 * For commit jobs, we cannot take CONSISTENT_READ, because 1530 * that permission is unshared for everything above the base 1531 * node (except for filters on the base node). 1532 * We also have to force-share the WRITE permission, or 1533 * otherwise we would block ourselves at the base node (if 1534 * writes are blocked for a node, they are also blocked for 1535 * its backing file). 1536 * (We could also share RESIZE, because it may be needed for 1537 * the target if its size is less than the top node's; but 1538 * bdrv_default_perms_for_cow() automatically shares RESIZE 1539 * for backing nodes if WRITE is shared, so there is no need 1540 * to do it here.) 1541 */ 1542 *nperm &= ~BLK_PERM_CONSISTENT_READ; 1543 *nshared |= BLK_PERM_WRITE; 1544 } 1545 } 1546 1547 /* Dummy node that provides consistent read to its users without requiring it 1548 * from its backing file and that allows writes on the backing file chain. */ 1549 static BlockDriver bdrv_mirror_top = { 1550 .format_name = "mirror_top", 1551 .bdrv_co_preadv = bdrv_mirror_top_preadv, 1552 .bdrv_co_pwritev = bdrv_mirror_top_pwritev, 1553 .bdrv_co_pwrite_zeroes = bdrv_mirror_top_pwrite_zeroes, 1554 .bdrv_co_pdiscard = bdrv_mirror_top_pdiscard, 1555 .bdrv_co_flush = bdrv_mirror_top_flush, 1556 .bdrv_refresh_filename = bdrv_mirror_top_refresh_filename, 1557 .bdrv_child_perm = bdrv_mirror_top_child_perm, 1558 1559 .is_filter = true, 1560 }; 1561 1562 static BlockJob *mirror_start_job( 1563 const char *job_id, BlockDriverState *bs, 1564 int creation_flags, BlockDriverState *target, 1565 const char *replaces, int64_t speed, 1566 uint32_t granularity, int64_t buf_size, 1567 BlockMirrorBackingMode backing_mode, 1568 bool zero_target, 1569 BlockdevOnError on_source_error, 1570 BlockdevOnError on_target_error, 1571 bool unmap, 1572 BlockCompletionFunc *cb, 1573 void *opaque, 1574 const BlockJobDriver *driver, 1575 bool is_none_mode, BlockDriverState *base, 1576 bool auto_complete, const char *filter_node_name, 1577 bool is_mirror, MirrorCopyMode copy_mode, 1578 Error **errp) 1579 { 1580 MirrorBlockJob *s; 1581 MirrorBDSOpaque *bs_opaque; 1582 BlockDriverState *mirror_top_bs; 1583 bool target_is_backing; 1584 uint64_t target_perms, target_shared_perms; 1585 int ret; 1586 1587 if (granularity == 0) { 1588 granularity = bdrv_get_default_bitmap_granularity(target); 1589 } 1590 1591 assert(is_power_of_2(granularity)); 1592 1593 if (buf_size < 0) { 1594 error_setg(errp, "Invalid parameter 'buf-size'"); 1595 return NULL; 1596 } 1597 1598 if (buf_size == 0) { 1599 buf_size = DEFAULT_MIRROR_BUF_SIZE; 1600 } 1601 1602 if (bdrv_skip_filters(bs) == bdrv_skip_filters(target)) { 1603 error_setg(errp, "Can't mirror node into itself"); 1604 return NULL; 1605 } 1606 1607 target_is_backing = bdrv_chain_contains(bs, target); 1608 1609 /* In the case of active commit, add dummy driver to provide consistent 1610 * reads on the top, while disabling it in the intermediate nodes, and make 1611 * the backing chain writable. */ 1612 mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name, 1613 BDRV_O_RDWR, errp); 1614 if (mirror_top_bs == NULL) { 1615 return NULL; 1616 } 1617 if (!filter_node_name) { 1618 mirror_top_bs->implicit = true; 1619 } 1620 1621 /* So that we can always drop this node */ 1622 mirror_top_bs->never_freeze = true; 1623 1624 mirror_top_bs->total_sectors = bs->total_sectors; 1625 mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED; 1626 mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED | 1627 BDRV_REQ_NO_FALLBACK; 1628 bs_opaque = g_new0(MirrorBDSOpaque, 1); 1629 mirror_top_bs->opaque = bs_opaque; 1630 1631 bs_opaque->is_commit = target_is_backing; 1632 1633 bdrv_drained_begin(bs); 1634 ret = bdrv_append(mirror_top_bs, bs, errp); 1635 bdrv_drained_end(bs); 1636 1637 if (ret < 0) { 1638 bdrv_unref(mirror_top_bs); 1639 return NULL; 1640 } 1641 1642 /* Make sure that the source is not resized while the job is running */ 1643 s = block_job_create(job_id, driver, NULL, mirror_top_bs, 1644 BLK_PERM_CONSISTENT_READ, 1645 BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED | 1646 BLK_PERM_WRITE | BLK_PERM_GRAPH_MOD, speed, 1647 creation_flags, cb, opaque, errp); 1648 if (!s) { 1649 goto fail; 1650 } 1651 bs_opaque->job = s; 1652 1653 /* The block job now has a reference to this node */ 1654 bdrv_unref(mirror_top_bs); 1655 1656 s->mirror_top_bs = mirror_top_bs; 1657 1658 /* No resize for the target either; while the mirror is still running, a 1659 * consistent read isn't necessarily possible. We could possibly allow 1660 * writes and graph modifications, though it would likely defeat the 1661 * purpose of a mirror, so leave them blocked for now. 1662 * 1663 * In the case of active commit, things look a bit different, though, 1664 * because the target is an already populated backing file in active use. 1665 * We can allow anything except resize there.*/ 1666 1667 target_perms = BLK_PERM_WRITE; 1668 target_shared_perms = BLK_PERM_WRITE_UNCHANGED; 1669 1670 if (target_is_backing) { 1671 int64_t bs_size, target_size; 1672 bs_size = bdrv_getlength(bs); 1673 if (bs_size < 0) { 1674 error_setg_errno(errp, -bs_size, 1675 "Could not inquire top image size"); 1676 goto fail; 1677 } 1678 1679 target_size = bdrv_getlength(target); 1680 if (target_size < 0) { 1681 error_setg_errno(errp, -target_size, 1682 "Could not inquire base image size"); 1683 goto fail; 1684 } 1685 1686 if (target_size < bs_size) { 1687 target_perms |= BLK_PERM_RESIZE; 1688 } 1689 1690 target_shared_perms |= BLK_PERM_CONSISTENT_READ 1691 | BLK_PERM_WRITE 1692 | BLK_PERM_GRAPH_MOD; 1693 } else if (bdrv_chain_contains(bs, bdrv_skip_filters(target))) { 1694 /* 1695 * We may want to allow this in the future, but it would 1696 * require taking some extra care. 1697 */ 1698 error_setg(errp, "Cannot mirror to a filter on top of a node in the " 1699 "source's backing chain"); 1700 goto fail; 1701 } 1702 1703 if (backing_mode != MIRROR_LEAVE_BACKING_CHAIN) { 1704 target_perms |= BLK_PERM_GRAPH_MOD; 1705 } 1706 1707 s->target = blk_new(s->common.job.aio_context, 1708 target_perms, target_shared_perms); 1709 ret = blk_insert_bs(s->target, target, errp); 1710 if (ret < 0) { 1711 goto fail; 1712 } 1713 if (is_mirror) { 1714 /* XXX: Mirror target could be a NBD server of target QEMU in the case 1715 * of non-shared block migration. To allow migration completion, we 1716 * have to allow "inactivate" of the target BB. When that happens, we 1717 * know the job is drained, and the vcpus are stopped, so no write 1718 * operation will be performed. Block layer already has assertions to 1719 * ensure that. */ 1720 blk_set_force_allow_inactivate(s->target); 1721 } 1722 blk_set_allow_aio_context_change(s->target, true); 1723 blk_set_disable_request_queuing(s->target, true); 1724 1725 s->replaces = g_strdup(replaces); 1726 s->on_source_error = on_source_error; 1727 s->on_target_error = on_target_error; 1728 s->is_none_mode = is_none_mode; 1729 s->backing_mode = backing_mode; 1730 s->zero_target = zero_target; 1731 s->copy_mode = copy_mode; 1732 s->base = base; 1733 s->base_overlay = bdrv_find_overlay(bs, base); 1734 s->granularity = granularity; 1735 s->buf_size = ROUND_UP(buf_size, granularity); 1736 s->unmap = unmap; 1737 if (auto_complete) { 1738 s->should_complete = true; 1739 } 1740 1741 s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp); 1742 if (!s->dirty_bitmap) { 1743 goto fail; 1744 } 1745 if (s->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING) { 1746 bdrv_disable_dirty_bitmap(s->dirty_bitmap); 1747 } 1748 1749 ret = block_job_add_bdrv(&s->common, "source", bs, 0, 1750 BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE | 1751 BLK_PERM_CONSISTENT_READ, 1752 errp); 1753 if (ret < 0) { 1754 goto fail; 1755 } 1756 1757 /* Required permissions are already taken with blk_new() */ 1758 block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL, 1759 &error_abort); 1760 1761 /* In commit_active_start() all intermediate nodes disappear, so 1762 * any jobs in them must be blocked */ 1763 if (target_is_backing) { 1764 BlockDriverState *iter, *filtered_target; 1765 uint64_t iter_shared_perms; 1766 1767 /* 1768 * The topmost node with 1769 * bdrv_skip_filters(filtered_target) == bdrv_skip_filters(target) 1770 */ 1771 filtered_target = bdrv_cow_bs(bdrv_find_overlay(bs, target)); 1772 1773 assert(bdrv_skip_filters(filtered_target) == 1774 bdrv_skip_filters(target)); 1775 1776 /* 1777 * XXX BLK_PERM_WRITE needs to be allowed so we don't block 1778 * ourselves at s->base (if writes are blocked for a node, they are 1779 * also blocked for its backing file). The other options would be a 1780 * second filter driver above s->base (== target). 1781 */ 1782 iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE; 1783 1784 for (iter = bdrv_filter_or_cow_bs(bs); iter != target; 1785 iter = bdrv_filter_or_cow_bs(iter)) 1786 { 1787 if (iter == filtered_target) { 1788 /* 1789 * From here on, all nodes are filters on the base. 1790 * This allows us to share BLK_PERM_CONSISTENT_READ. 1791 */ 1792 iter_shared_perms |= BLK_PERM_CONSISTENT_READ; 1793 } 1794 1795 ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0, 1796 iter_shared_perms, errp); 1797 if (ret < 0) { 1798 goto fail; 1799 } 1800 } 1801 1802 if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) { 1803 goto fail; 1804 } 1805 } 1806 1807 QTAILQ_INIT(&s->ops_in_flight); 1808 1809 trace_mirror_start(bs, s, opaque); 1810 job_start(&s->common.job); 1811 1812 return &s->common; 1813 1814 fail: 1815 if (s) { 1816 /* Make sure this BDS does not go away until we have completed the graph 1817 * changes below */ 1818 bdrv_ref(mirror_top_bs); 1819 1820 g_free(s->replaces); 1821 blk_unref(s->target); 1822 bs_opaque->job = NULL; 1823 if (s->dirty_bitmap) { 1824 bdrv_release_dirty_bitmap(s->dirty_bitmap); 1825 } 1826 job_early_fail(&s->common.job); 1827 } 1828 1829 bs_opaque->stop = true; 1830 bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing, 1831 &error_abort); 1832 bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort); 1833 1834 bdrv_unref(mirror_top_bs); 1835 1836 return NULL; 1837 } 1838 1839 void mirror_start(const char *job_id, BlockDriverState *bs, 1840 BlockDriverState *target, const char *replaces, 1841 int creation_flags, int64_t speed, 1842 uint32_t granularity, int64_t buf_size, 1843 MirrorSyncMode mode, BlockMirrorBackingMode backing_mode, 1844 bool zero_target, 1845 BlockdevOnError on_source_error, 1846 BlockdevOnError on_target_error, 1847 bool unmap, const char *filter_node_name, 1848 MirrorCopyMode copy_mode, Error **errp) 1849 { 1850 bool is_none_mode; 1851 BlockDriverState *base; 1852 1853 if ((mode == MIRROR_SYNC_MODE_INCREMENTAL) || 1854 (mode == MIRROR_SYNC_MODE_BITMAP)) { 1855 error_setg(errp, "Sync mode '%s' not supported", 1856 MirrorSyncMode_str(mode)); 1857 return; 1858 } 1859 is_none_mode = mode == MIRROR_SYNC_MODE_NONE; 1860 base = mode == MIRROR_SYNC_MODE_TOP ? bdrv_backing_chain_next(bs) : NULL; 1861 mirror_start_job(job_id, bs, creation_flags, target, replaces, 1862 speed, granularity, buf_size, backing_mode, zero_target, 1863 on_source_error, on_target_error, unmap, NULL, NULL, 1864 &mirror_job_driver, is_none_mode, base, false, 1865 filter_node_name, true, copy_mode, errp); 1866 } 1867 1868 BlockJob *commit_active_start(const char *job_id, BlockDriverState *bs, 1869 BlockDriverState *base, int creation_flags, 1870 int64_t speed, BlockdevOnError on_error, 1871 const char *filter_node_name, 1872 BlockCompletionFunc *cb, void *opaque, 1873 bool auto_complete, Error **errp) 1874 { 1875 bool base_read_only; 1876 BlockJob *job; 1877 1878 base_read_only = bdrv_is_read_only(base); 1879 1880 if (base_read_only) { 1881 if (bdrv_reopen_set_read_only(base, false, errp) < 0) { 1882 return NULL; 1883 } 1884 } 1885 1886 job = mirror_start_job( 1887 job_id, bs, creation_flags, base, NULL, speed, 0, 0, 1888 MIRROR_LEAVE_BACKING_CHAIN, false, 1889 on_error, on_error, true, cb, opaque, 1890 &commit_active_job_driver, false, base, auto_complete, 1891 filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND, 1892 errp); 1893 if (!job) { 1894 goto error_restore_flags; 1895 } 1896 1897 return job; 1898 1899 error_restore_flags: 1900 /* ignore error and errp for bdrv_reopen, because we want to propagate 1901 * the original error */ 1902 if (base_read_only) { 1903 bdrv_reopen_set_read_only(base, true, NULL); 1904 } 1905 return NULL; 1906 } 1907