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 #include "qemu/memalign.h" 27 28 #define MAX_IN_FLIGHT 16 29 #define MAX_IO_BYTES (1 << 20) /* 1 Mb */ 30 #define DEFAULT_MIRROR_BUF_SIZE (MAX_IN_FLIGHT * MAX_IO_BYTES) 31 32 /* The mirroring buffer is a list of granularity-sized chunks. 33 * Free chunks are organized in a list. 34 */ 35 typedef struct MirrorBuffer { 36 QSIMPLEQ_ENTRY(MirrorBuffer) next; 37 } MirrorBuffer; 38 39 typedef struct MirrorOp MirrorOp; 40 41 typedef struct MirrorBlockJob { 42 BlockJob common; 43 BlockBackend *target; 44 BlockDriverState *mirror_top_bs; 45 BlockDriverState *base; 46 BlockDriverState *base_overlay; 47 48 /* The name of the graph node to replace */ 49 char *replaces; 50 /* The BDS to replace */ 51 BlockDriverState *to_replace; 52 /* Used to block operations on the drive-mirror-replace target */ 53 Error *replace_blocker; 54 bool is_none_mode; 55 BlockMirrorBackingMode backing_mode; 56 /* Whether the target image requires explicit zero-initialization */ 57 bool zero_target; 58 MirrorCopyMode copy_mode; 59 BlockdevOnError on_source_error, on_target_error; 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 unsigned 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 MirrorOp *waiting_for_op; 111 112 QTAILQ_ENTRY(MirrorOp) next; 113 }; 114 115 typedef enum MirrorMethod { 116 MIRROR_METHOD_COPY, 117 MIRROR_METHOD_ZERO, 118 MIRROR_METHOD_DISCARD, 119 } MirrorMethod; 120 121 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read, 122 int error) 123 { 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 if (self) { 163 /* 164 * If the operation is already (indirectly) waiting for us, 165 * or will wait for us as soon as it wakes up, then just go 166 * on (instead of producing a deadlock in the former case). 167 */ 168 if (op->waiting_for_op) { 169 continue; 170 } 171 172 self->waiting_for_op = op; 173 } 174 175 qemu_co_queue_wait(&op->waiting_requests, NULL); 176 177 if (self) { 178 self->waiting_for_op = NULL; 179 } 180 181 break; 182 } 183 } 184 } 185 } 186 187 static void coroutine_fn mirror_iteration_done(MirrorOp *op, int ret) 188 { 189 MirrorBlockJob *s = op->s; 190 struct iovec *iov; 191 int64_t chunk_num; 192 int i, nb_chunks; 193 194 trace_mirror_iteration_done(s, op->offset, op->bytes, ret); 195 196 s->in_flight--; 197 s->bytes_in_flight -= op->bytes; 198 iov = op->qiov.iov; 199 for (i = 0; i < op->qiov.niov; i++) { 200 MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base; 201 QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next); 202 s->buf_free_count++; 203 } 204 205 chunk_num = op->offset / s->granularity; 206 nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity); 207 208 bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks); 209 QTAILQ_REMOVE(&s->ops_in_flight, op, next); 210 if (ret >= 0) { 211 if (s->cow_bitmap) { 212 bitmap_set(s->cow_bitmap, chunk_num, nb_chunks); 213 } 214 if (!s->initial_zeroing_ongoing) { 215 job_progress_update(&s->common.job, op->bytes); 216 } 217 } 218 qemu_iovec_destroy(&op->qiov); 219 220 qemu_co_queue_restart_all(&op->waiting_requests); 221 g_free(op); 222 } 223 224 static void coroutine_fn mirror_write_complete(MirrorOp *op, int ret) 225 { 226 MirrorBlockJob *s = op->s; 227 228 if (ret < 0) { 229 BlockErrorAction action; 230 231 bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes); 232 action = mirror_error_action(s, false, -ret); 233 if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) { 234 s->ret = ret; 235 } 236 } 237 238 mirror_iteration_done(op, ret); 239 } 240 241 static void coroutine_fn mirror_read_complete(MirrorOp *op, int ret) 242 { 243 MirrorBlockJob *s = op->s; 244 245 if (ret < 0) { 246 BlockErrorAction action; 247 248 bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes); 249 action = mirror_error_action(s, true, -ret); 250 if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) { 251 s->ret = ret; 252 } 253 254 mirror_iteration_done(op, ret); 255 return; 256 } 257 258 ret = blk_co_pwritev(s->target, op->offset, op->qiov.size, &op->qiov, 0); 259 mirror_write_complete(op, ret); 260 } 261 262 /* Clip bytes relative to offset to not exceed end-of-file */ 263 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s, 264 int64_t offset, 265 int64_t bytes) 266 { 267 return MIN(bytes, s->bdev_length - offset); 268 } 269 270 /* Round offset and/or bytes to target cluster if COW is needed, and 271 * return the offset of the adjusted tail against original. */ 272 static int mirror_cow_align(MirrorBlockJob *s, int64_t *offset, 273 uint64_t *bytes) 274 { 275 bool need_cow; 276 int ret = 0; 277 int64_t align_offset = *offset; 278 int64_t align_bytes = *bytes; 279 int max_bytes = s->granularity * s->max_iov; 280 281 need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap); 282 need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity, 283 s->cow_bitmap); 284 if (need_cow) { 285 bdrv_round_to_clusters(blk_bs(s->target), *offset, *bytes, 286 &align_offset, &align_bytes); 287 } 288 289 if (align_bytes > max_bytes) { 290 align_bytes = max_bytes; 291 if (need_cow) { 292 align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size); 293 } 294 } 295 /* Clipping may result in align_bytes unaligned to chunk boundary, but 296 * that doesn't matter because it's already the end of source image. */ 297 align_bytes = mirror_clip_bytes(s, align_offset, align_bytes); 298 299 ret = align_offset + align_bytes - (*offset + *bytes); 300 *offset = align_offset; 301 *bytes = align_bytes; 302 assert(ret >= 0); 303 return ret; 304 } 305 306 static inline void coroutine_fn 307 mirror_wait_for_any_operation(MirrorBlockJob *s, bool active) 308 { 309 MirrorOp *op; 310 311 QTAILQ_FOREACH(op, &s->ops_in_flight, next) { 312 /* Do not wait on pseudo ops, because it may in turn wait on 313 * some other operation to start, which may in fact be the 314 * caller of this function. Since there is only one pseudo op 315 * at any given time, we will always find some real operation 316 * to wait on. */ 317 if (!op->is_pseudo_op && op->is_in_flight && 318 op->is_active_write == active) 319 { 320 qemu_co_queue_wait(&op->waiting_requests, NULL); 321 return; 322 } 323 } 324 abort(); 325 } 326 327 static inline void coroutine_fn 328 mirror_wait_for_free_in_flight_slot(MirrorBlockJob *s) 329 { 330 /* Only non-active operations use up in-flight slots */ 331 mirror_wait_for_any_operation(s, false); 332 } 333 334 /* Perform a mirror copy operation. 335 * 336 * *op->bytes_handled is set to the number of bytes copied after and 337 * including offset, excluding any bytes copied prior to offset due 338 * to alignment. This will be op->bytes if no alignment is necessary, 339 * or (new_end - op->offset) if the tail is rounded up or down due to 340 * alignment or buffer limit. 341 */ 342 static void coroutine_fn mirror_co_read(void *opaque) 343 { 344 MirrorOp *op = opaque; 345 MirrorBlockJob *s = op->s; 346 int nb_chunks; 347 uint64_t ret; 348 uint64_t max_bytes; 349 350 max_bytes = s->granularity * s->max_iov; 351 352 /* We can only handle as much as buf_size at a time. */ 353 op->bytes = MIN(s->buf_size, MIN(max_bytes, op->bytes)); 354 assert(op->bytes); 355 assert(op->bytes < BDRV_REQUEST_MAX_BYTES); 356 *op->bytes_handled = op->bytes; 357 358 if (s->cow_bitmap) { 359 *op->bytes_handled += mirror_cow_align(s, &op->offset, &op->bytes); 360 } 361 /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */ 362 assert(*op->bytes_handled <= UINT_MAX); 363 assert(op->bytes <= s->buf_size); 364 /* The offset is granularity-aligned because: 365 * 1) Caller passes in aligned values; 366 * 2) mirror_cow_align is used only when target cluster is larger. */ 367 assert(QEMU_IS_ALIGNED(op->offset, s->granularity)); 368 /* The range is sector-aligned, since bdrv_getlength() rounds up. */ 369 assert(QEMU_IS_ALIGNED(op->bytes, BDRV_SECTOR_SIZE)); 370 nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity); 371 372 while (s->buf_free_count < nb_chunks) { 373 trace_mirror_yield_in_flight(s, op->offset, s->in_flight); 374 mirror_wait_for_free_in_flight_slot(s); 375 } 376 377 /* Now make a QEMUIOVector taking enough granularity-sized chunks 378 * from s->buf_free. 379 */ 380 qemu_iovec_init(&op->qiov, nb_chunks); 381 while (nb_chunks-- > 0) { 382 MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free); 383 size_t remaining = op->bytes - op->qiov.size; 384 385 QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next); 386 s->buf_free_count--; 387 qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining)); 388 } 389 390 /* Copy the dirty cluster. */ 391 s->in_flight++; 392 s->bytes_in_flight += op->bytes; 393 op->is_in_flight = true; 394 trace_mirror_one_iteration(s, op->offset, op->bytes); 395 396 ret = bdrv_co_preadv(s->mirror_top_bs->backing, op->offset, op->bytes, 397 &op->qiov, 0); 398 mirror_read_complete(op, ret); 399 } 400 401 static void coroutine_fn mirror_co_zero(void *opaque) 402 { 403 MirrorOp *op = opaque; 404 int ret; 405 406 op->s->in_flight++; 407 op->s->bytes_in_flight += op->bytes; 408 *op->bytes_handled = op->bytes; 409 op->is_in_flight = true; 410 411 ret = blk_co_pwrite_zeroes(op->s->target, op->offset, op->bytes, 412 op->s->unmap ? BDRV_REQ_MAY_UNMAP : 0); 413 mirror_write_complete(op, ret); 414 } 415 416 static void coroutine_fn mirror_co_discard(void *opaque) 417 { 418 MirrorOp *op = opaque; 419 int ret; 420 421 op->s->in_flight++; 422 op->s->bytes_in_flight += op->bytes; 423 *op->bytes_handled = op->bytes; 424 op->is_in_flight = true; 425 426 ret = blk_co_pdiscard(op->s->target, op->offset, op->bytes); 427 mirror_write_complete(op, ret); 428 } 429 430 static unsigned mirror_perform(MirrorBlockJob *s, int64_t offset, 431 unsigned bytes, MirrorMethod mirror_method) 432 { 433 MirrorOp *op; 434 Coroutine *co; 435 int64_t bytes_handled = -1; 436 437 op = g_new(MirrorOp, 1); 438 *op = (MirrorOp){ 439 .s = s, 440 .offset = offset, 441 .bytes = bytes, 442 .bytes_handled = &bytes_handled, 443 }; 444 qemu_co_queue_init(&op->waiting_requests); 445 446 switch (mirror_method) { 447 case MIRROR_METHOD_COPY: 448 co = qemu_coroutine_create(mirror_co_read, op); 449 break; 450 case MIRROR_METHOD_ZERO: 451 co = qemu_coroutine_create(mirror_co_zero, op); 452 break; 453 case MIRROR_METHOD_DISCARD: 454 co = qemu_coroutine_create(mirror_co_discard, op); 455 break; 456 default: 457 abort(); 458 } 459 op->co = co; 460 461 QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next); 462 qemu_coroutine_enter(co); 463 /* At this point, ownership of op has been moved to the coroutine 464 * and the object may already be freed */ 465 466 /* Assert that this value has been set */ 467 assert(bytes_handled >= 0); 468 469 /* Same assertion as in mirror_co_read() (and for mirror_co_read() 470 * and mirror_co_discard(), bytes_handled == op->bytes, which 471 * is the @bytes parameter given to this function) */ 472 assert(bytes_handled <= UINT_MAX); 473 return bytes_handled; 474 } 475 476 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s) 477 { 478 BlockDriverState *source = s->mirror_top_bs->backing->bs; 479 MirrorOp *pseudo_op; 480 int64_t offset; 481 uint64_t delay_ns = 0, ret = 0; 482 /* At least the first dirty chunk is mirrored in one iteration. */ 483 int nb_chunks = 1; 484 bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target)); 485 int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES); 486 487 bdrv_dirty_bitmap_lock(s->dirty_bitmap); 488 offset = bdrv_dirty_iter_next(s->dbi); 489 if (offset < 0) { 490 bdrv_set_dirty_iter(s->dbi, 0); 491 offset = bdrv_dirty_iter_next(s->dbi); 492 trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap)); 493 assert(offset >= 0); 494 } 495 bdrv_dirty_bitmap_unlock(s->dirty_bitmap); 496 497 mirror_wait_on_conflicts(NULL, s, offset, 1); 498 499 job_pause_point(&s->common.job); 500 501 /* Find the number of consective dirty chunks following the first dirty 502 * one, and wait for in flight requests in them. */ 503 bdrv_dirty_bitmap_lock(s->dirty_bitmap); 504 while (nb_chunks * s->granularity < s->buf_size) { 505 int64_t next_dirty; 506 int64_t next_offset = offset + nb_chunks * s->granularity; 507 int64_t next_chunk = next_offset / s->granularity; 508 if (next_offset >= s->bdev_length || 509 !bdrv_dirty_bitmap_get_locked(s->dirty_bitmap, next_offset)) { 510 break; 511 } 512 if (test_bit(next_chunk, s->in_flight_bitmap)) { 513 break; 514 } 515 516 next_dirty = bdrv_dirty_iter_next(s->dbi); 517 if (next_dirty > next_offset || next_dirty < 0) { 518 /* The bitmap iterator's cache is stale, refresh it */ 519 bdrv_set_dirty_iter(s->dbi, next_offset); 520 next_dirty = bdrv_dirty_iter_next(s->dbi); 521 } 522 assert(next_dirty == next_offset); 523 nb_chunks++; 524 } 525 526 /* Clear dirty bits before querying the block status, because 527 * calling bdrv_block_status_above could yield - if some blocks are 528 * marked dirty in this window, we need to know. 529 */ 530 bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset, 531 nb_chunks * s->granularity); 532 bdrv_dirty_bitmap_unlock(s->dirty_bitmap); 533 534 /* Before claiming an area in the in-flight bitmap, we have to 535 * create a MirrorOp for it so that conflicting requests can wait 536 * for it. mirror_perform() will create the real MirrorOps later, 537 * for now we just create a pseudo operation that will wake up all 538 * conflicting requests once all real operations have been 539 * launched. */ 540 pseudo_op = g_new(MirrorOp, 1); 541 *pseudo_op = (MirrorOp){ 542 .offset = offset, 543 .bytes = nb_chunks * s->granularity, 544 .is_pseudo_op = true, 545 }; 546 qemu_co_queue_init(&pseudo_op->waiting_requests); 547 QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next); 548 549 bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks); 550 while (nb_chunks > 0 && offset < s->bdev_length) { 551 int ret; 552 int64_t io_bytes; 553 int64_t io_bytes_acct; 554 MirrorMethod mirror_method = MIRROR_METHOD_COPY; 555 556 assert(!(offset % s->granularity)); 557 ret = bdrv_block_status_above(source, NULL, offset, 558 nb_chunks * s->granularity, 559 &io_bytes, NULL, NULL); 560 if (ret < 0) { 561 io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes); 562 } else if (ret & BDRV_BLOCK_DATA) { 563 io_bytes = MIN(io_bytes, max_io_bytes); 564 } 565 566 io_bytes -= io_bytes % s->granularity; 567 if (io_bytes < s->granularity) { 568 io_bytes = s->granularity; 569 } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) { 570 int64_t target_offset; 571 int64_t target_bytes; 572 bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes, 573 &target_offset, &target_bytes); 574 if (target_offset == offset && 575 target_bytes == io_bytes) { 576 mirror_method = ret & BDRV_BLOCK_ZERO ? 577 MIRROR_METHOD_ZERO : 578 MIRROR_METHOD_DISCARD; 579 } 580 } 581 582 while (s->in_flight >= MAX_IN_FLIGHT) { 583 trace_mirror_yield_in_flight(s, offset, s->in_flight); 584 mirror_wait_for_free_in_flight_slot(s); 585 } 586 587 if (s->ret < 0) { 588 ret = 0; 589 goto fail; 590 } 591 592 io_bytes = mirror_clip_bytes(s, offset, io_bytes); 593 io_bytes = mirror_perform(s, offset, io_bytes, mirror_method); 594 if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) { 595 io_bytes_acct = 0; 596 } else { 597 io_bytes_acct = io_bytes; 598 } 599 assert(io_bytes); 600 offset += io_bytes; 601 nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity); 602 delay_ns = block_job_ratelimit_get_delay(&s->common, io_bytes_acct); 603 } 604 605 ret = delay_ns; 606 fail: 607 QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next); 608 qemu_co_queue_restart_all(&pseudo_op->waiting_requests); 609 g_free(pseudo_op); 610 611 return ret; 612 } 613 614 static void mirror_free_init(MirrorBlockJob *s) 615 { 616 int granularity = s->granularity; 617 size_t buf_size = s->buf_size; 618 uint8_t *buf = s->buf; 619 620 assert(s->buf_free_count == 0); 621 QSIMPLEQ_INIT(&s->buf_free); 622 while (buf_size != 0) { 623 MirrorBuffer *cur = (MirrorBuffer *)buf; 624 QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next); 625 s->buf_free_count++; 626 buf_size -= granularity; 627 buf += granularity; 628 } 629 } 630 631 /* This is also used for the .pause callback. There is no matching 632 * mirror_resume() because mirror_run() will begin iterating again 633 * when the job is resumed. 634 */ 635 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s) 636 { 637 while (s->in_flight > 0) { 638 mirror_wait_for_free_in_flight_slot(s); 639 } 640 } 641 642 /** 643 * mirror_exit_common: handle both abort() and prepare() cases. 644 * for .prepare, returns 0 on success and -errno on failure. 645 * for .abort cases, denoted by abort = true, MUST return 0. 646 */ 647 static int mirror_exit_common(Job *job) 648 { 649 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 650 BlockJob *bjob = &s->common; 651 MirrorBDSOpaque *bs_opaque; 652 AioContext *replace_aio_context = NULL; 653 BlockDriverState *src; 654 BlockDriverState *target_bs; 655 BlockDriverState *mirror_top_bs; 656 Error *local_err = NULL; 657 bool abort = job->ret < 0; 658 int ret = 0; 659 660 if (s->prepared) { 661 return 0; 662 } 663 s->prepared = true; 664 665 mirror_top_bs = s->mirror_top_bs; 666 bs_opaque = mirror_top_bs->opaque; 667 src = mirror_top_bs->backing->bs; 668 target_bs = blk_bs(s->target); 669 670 if (bdrv_chain_contains(src, target_bs)) { 671 bdrv_unfreeze_backing_chain(mirror_top_bs, target_bs); 672 } 673 674 bdrv_release_dirty_bitmap(s->dirty_bitmap); 675 676 /* Make sure that the source BDS doesn't go away during bdrv_replace_node, 677 * before we can call bdrv_drained_end */ 678 bdrv_ref(src); 679 bdrv_ref(mirror_top_bs); 680 bdrv_ref(target_bs); 681 682 /* 683 * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before 684 * inserting target_bs at s->to_replace, where we might not be able to get 685 * these permissions. 686 */ 687 blk_unref(s->target); 688 s->target = NULL; 689 690 /* We don't access the source any more. Dropping any WRITE/RESIZE is 691 * required before it could become a backing file of target_bs. Not having 692 * these permissions any more means that we can't allow any new requests on 693 * mirror_top_bs from now on, so keep it drained. */ 694 bdrv_drained_begin(mirror_top_bs); 695 bs_opaque->stop = true; 696 bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing, 697 &error_abort); 698 if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) { 699 BlockDriverState *backing = s->is_none_mode ? src : s->base; 700 BlockDriverState *unfiltered_target = bdrv_skip_filters(target_bs); 701 702 if (bdrv_cow_bs(unfiltered_target) != backing) { 703 bdrv_set_backing_hd(unfiltered_target, backing, &local_err); 704 if (local_err) { 705 error_report_err(local_err); 706 local_err = NULL; 707 ret = -EPERM; 708 } 709 } 710 } else if (!abort && s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) { 711 assert(!bdrv_backing_chain_next(target_bs)); 712 ret = bdrv_open_backing_file(bdrv_skip_filters(target_bs), NULL, 713 "backing", &local_err); 714 if (ret < 0) { 715 error_report_err(local_err); 716 local_err = NULL; 717 } 718 } 719 720 if (s->to_replace) { 721 replace_aio_context = bdrv_get_aio_context(s->to_replace); 722 aio_context_acquire(replace_aio_context); 723 } 724 725 if (s->should_complete && !abort) { 726 BlockDriverState *to_replace = s->to_replace ?: src; 727 bool ro = bdrv_is_read_only(to_replace); 728 729 if (ro != bdrv_is_read_only(target_bs)) { 730 bdrv_reopen_set_read_only(target_bs, ro, NULL); 731 } 732 733 /* The mirror job has no requests in flight any more, but we need to 734 * drain potential other users of the BDS before changing the graph. */ 735 assert(s->in_drain); 736 bdrv_drained_begin(target_bs); 737 /* 738 * Cannot use check_to_replace_node() here, because that would 739 * check for an op blocker on @to_replace, and we have our own 740 * there. 741 */ 742 if (bdrv_recurse_can_replace(src, to_replace)) { 743 bdrv_replace_node(to_replace, target_bs, &local_err); 744 } else { 745 error_setg(&local_err, "Can no longer replace '%s' by '%s', " 746 "because it can no longer be guaranteed that doing so " 747 "would not lead to an abrupt change of visible data", 748 to_replace->node_name, target_bs->node_name); 749 } 750 bdrv_drained_end(target_bs); 751 if (local_err) { 752 error_report_err(local_err); 753 ret = -EPERM; 754 } 755 } 756 if (s->to_replace) { 757 bdrv_op_unblock_all(s->to_replace, s->replace_blocker); 758 error_free(s->replace_blocker); 759 bdrv_unref(s->to_replace); 760 } 761 if (replace_aio_context) { 762 aio_context_release(replace_aio_context); 763 } 764 g_free(s->replaces); 765 bdrv_unref(target_bs); 766 767 /* 768 * Remove the mirror filter driver from the graph. Before this, get rid of 769 * the blockers on the intermediate nodes so that the resulting state is 770 * valid. 771 */ 772 block_job_remove_all_bdrv(bjob); 773 bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort); 774 775 bs_opaque->job = NULL; 776 777 bdrv_drained_end(src); 778 bdrv_drained_end(mirror_top_bs); 779 s->in_drain = false; 780 bdrv_unref(mirror_top_bs); 781 bdrv_unref(src); 782 783 return ret; 784 } 785 786 static int mirror_prepare(Job *job) 787 { 788 return mirror_exit_common(job); 789 } 790 791 static void mirror_abort(Job *job) 792 { 793 int ret = mirror_exit_common(job); 794 assert(ret == 0); 795 } 796 797 static void coroutine_fn mirror_throttle(MirrorBlockJob *s) 798 { 799 int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); 800 801 if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) { 802 s->last_pause_ns = now; 803 job_sleep_ns(&s->common.job, 0); 804 } else { 805 job_pause_point(&s->common.job); 806 } 807 } 808 809 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s) 810 { 811 int64_t offset; 812 BlockDriverState *bs = s->mirror_top_bs->backing->bs; 813 BlockDriverState *target_bs = blk_bs(s->target); 814 int ret; 815 int64_t count; 816 817 if (s->zero_target) { 818 if (!bdrv_can_write_zeroes_with_unmap(target_bs)) { 819 bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length); 820 return 0; 821 } 822 823 s->initial_zeroing_ongoing = true; 824 for (offset = 0; offset < s->bdev_length; ) { 825 int bytes = MIN(s->bdev_length - offset, 826 QEMU_ALIGN_DOWN(INT_MAX, s->granularity)); 827 828 mirror_throttle(s); 829 830 if (job_is_cancelled(&s->common.job)) { 831 s->initial_zeroing_ongoing = false; 832 return 0; 833 } 834 835 if (s->in_flight >= MAX_IN_FLIGHT) { 836 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count, 837 s->in_flight); 838 mirror_wait_for_free_in_flight_slot(s); 839 continue; 840 } 841 842 mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO); 843 offset += bytes; 844 } 845 846 mirror_wait_for_all_io(s); 847 s->initial_zeroing_ongoing = false; 848 } 849 850 /* First part, loop on the sectors and initialize the dirty bitmap. */ 851 for (offset = 0; offset < s->bdev_length; ) { 852 /* Just to make sure we are not exceeding int limit. */ 853 int bytes = MIN(s->bdev_length - offset, 854 QEMU_ALIGN_DOWN(INT_MAX, s->granularity)); 855 856 mirror_throttle(s); 857 858 if (job_is_cancelled(&s->common.job)) { 859 return 0; 860 } 861 862 ret = bdrv_is_allocated_above(bs, s->base_overlay, true, offset, bytes, 863 &count); 864 if (ret < 0) { 865 return ret; 866 } 867 868 assert(count); 869 if (ret > 0) { 870 bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count); 871 } 872 offset += count; 873 } 874 return 0; 875 } 876 877 /* Called when going out of the streaming phase to flush the bulk of the 878 * data to the medium, or just before completing. 879 */ 880 static int mirror_flush(MirrorBlockJob *s) 881 { 882 int ret = blk_flush(s->target); 883 if (ret < 0) { 884 if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) { 885 s->ret = ret; 886 } 887 } 888 return ret; 889 } 890 891 static int coroutine_fn mirror_run(Job *job, Error **errp) 892 { 893 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 894 BlockDriverState *bs = s->mirror_top_bs->backing->bs; 895 BlockDriverState *target_bs = blk_bs(s->target); 896 bool need_drain = true; 897 BlockDeviceIoStatus iostatus; 898 int64_t length; 899 int64_t target_length; 900 BlockDriverInfo bdi; 901 char backing_filename[2]; /* we only need 2 characters because we are only 902 checking for a NULL string */ 903 int ret = 0; 904 905 if (job_is_cancelled(&s->common.job)) { 906 goto immediate_exit; 907 } 908 909 s->bdev_length = bdrv_getlength(bs); 910 if (s->bdev_length < 0) { 911 ret = s->bdev_length; 912 goto immediate_exit; 913 } 914 915 target_length = blk_getlength(s->target); 916 if (target_length < 0) { 917 ret = target_length; 918 goto immediate_exit; 919 } 920 921 /* Active commit must resize the base image if its size differs from the 922 * active layer. */ 923 if (s->base == blk_bs(s->target)) { 924 if (s->bdev_length > target_length) { 925 ret = blk_co_truncate(s->target, s->bdev_length, false, 926 PREALLOC_MODE_OFF, 0, NULL); 927 if (ret < 0) { 928 goto immediate_exit; 929 } 930 } 931 } else if (s->bdev_length != target_length) { 932 error_setg(errp, "Source and target image have different sizes"); 933 ret = -EINVAL; 934 goto immediate_exit; 935 } 936 937 if (s->bdev_length == 0) { 938 /* Transition to the READY state and wait for complete. */ 939 job_transition_to_ready(&s->common.job); 940 s->actively_synced = true; 941 while (!job_cancel_requested(&s->common.job) && !s->should_complete) { 942 job_yield(&s->common.job); 943 } 944 goto immediate_exit; 945 } 946 947 length = DIV_ROUND_UP(s->bdev_length, s->granularity); 948 s->in_flight_bitmap = bitmap_new(length); 949 950 /* If we have no backing file yet in the destination, we cannot let 951 * the destination do COW. Instead, we copy sectors around the 952 * dirty data if needed. We need a bitmap to do that. 953 */ 954 bdrv_get_backing_filename(target_bs, backing_filename, 955 sizeof(backing_filename)); 956 if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) { 957 s->target_cluster_size = bdi.cluster_size; 958 } else { 959 s->target_cluster_size = BDRV_SECTOR_SIZE; 960 } 961 if (backing_filename[0] && !bdrv_backing_chain_next(target_bs) && 962 s->granularity < s->target_cluster_size) { 963 s->buf_size = MAX(s->buf_size, s->target_cluster_size); 964 s->cow_bitmap = bitmap_new(length); 965 } 966 s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov); 967 968 s->buf = qemu_try_blockalign(bs, s->buf_size); 969 if (s->buf == NULL) { 970 ret = -ENOMEM; 971 goto immediate_exit; 972 } 973 974 mirror_free_init(s); 975 976 s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); 977 if (!s->is_none_mode) { 978 ret = mirror_dirty_init(s); 979 if (ret < 0 || job_is_cancelled(&s->common.job)) { 980 goto immediate_exit; 981 } 982 } 983 984 assert(!s->dbi); 985 s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap); 986 for (;;) { 987 uint64_t delay_ns = 0; 988 int64_t cnt, delta; 989 bool should_complete; 990 991 /* Do not start passive operations while there are active 992 * writes in progress */ 993 while (s->in_active_write_counter) { 994 mirror_wait_for_any_operation(s, true); 995 } 996 997 if (s->ret < 0) { 998 ret = s->ret; 999 goto immediate_exit; 1000 } 1001 1002 job_pause_point(&s->common.job); 1003 1004 if (job_is_cancelled(&s->common.job)) { 1005 ret = 0; 1006 goto immediate_exit; 1007 } 1008 1009 cnt = bdrv_get_dirty_count(s->dirty_bitmap); 1010 /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is 1011 * the number of bytes currently being processed; together those are 1012 * the current remaining operation length */ 1013 job_progress_set_remaining(&s->common.job, s->bytes_in_flight + cnt); 1014 1015 /* Note that even when no rate limit is applied we need to yield 1016 * periodically with no pending I/O so that bdrv_drain_all() returns. 1017 * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is 1018 * an error, or when the source is clean, whichever comes first. */ 1019 delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns; 1020 WITH_JOB_LOCK_GUARD() { 1021 iostatus = s->common.iostatus; 1022 } 1023 if (delta < BLOCK_JOB_SLICE_TIME && 1024 iostatus == BLOCK_DEVICE_IO_STATUS_OK) { 1025 if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 || 1026 (cnt == 0 && s->in_flight > 0)) { 1027 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight); 1028 mirror_wait_for_free_in_flight_slot(s); 1029 continue; 1030 } else if (cnt != 0) { 1031 delay_ns = mirror_iteration(s); 1032 } 1033 } 1034 1035 should_complete = false; 1036 if (s->in_flight == 0 && cnt == 0) { 1037 trace_mirror_before_flush(s); 1038 if (!job_is_ready(&s->common.job)) { 1039 if (mirror_flush(s) < 0) { 1040 /* Go check s->ret. */ 1041 continue; 1042 } 1043 /* We're out of the streaming phase. From now on, if the job 1044 * is cancelled we will actually complete all pending I/O and 1045 * report completion. This way, block-job-cancel will leave 1046 * the target in a consistent state. 1047 */ 1048 job_transition_to_ready(&s->common.job); 1049 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) { 1050 s->actively_synced = true; 1051 } 1052 } 1053 1054 should_complete = s->should_complete || 1055 job_cancel_requested(&s->common.job); 1056 cnt = bdrv_get_dirty_count(s->dirty_bitmap); 1057 } 1058 1059 if (cnt == 0 && should_complete) { 1060 /* The dirty bitmap is not updated while operations are pending. 1061 * If we're about to exit, wait for pending operations before 1062 * calling bdrv_get_dirty_count(bs), or we may exit while the 1063 * source has dirty data to copy! 1064 * 1065 * Note that I/O can be submitted by the guest while 1066 * mirror_populate runs, so pause it now. Before deciding 1067 * whether to switch to target check one last time if I/O has 1068 * come in the meanwhile, and if not flush the data to disk. 1069 */ 1070 trace_mirror_before_drain(s, cnt); 1071 1072 s->in_drain = true; 1073 bdrv_drained_begin(bs); 1074 cnt = bdrv_get_dirty_count(s->dirty_bitmap); 1075 if (cnt > 0 || mirror_flush(s) < 0) { 1076 bdrv_drained_end(bs); 1077 s->in_drain = false; 1078 continue; 1079 } 1080 1081 /* The two disks are in sync. Exit and report successful 1082 * completion. 1083 */ 1084 assert(QLIST_EMPTY(&bs->tracked_requests)); 1085 need_drain = false; 1086 break; 1087 } 1088 1089 if (job_is_ready(&s->common.job) && !should_complete) { 1090 delay_ns = (s->in_flight == 0 && 1091 cnt == 0 ? BLOCK_JOB_SLICE_TIME : 0); 1092 } 1093 trace_mirror_before_sleep(s, cnt, job_is_ready(&s->common.job), 1094 delay_ns); 1095 job_sleep_ns(&s->common.job, delay_ns); 1096 s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); 1097 } 1098 1099 immediate_exit: 1100 if (s->in_flight > 0) { 1101 /* We get here only if something went wrong. Either the job failed, 1102 * or it was cancelled prematurely so that we do not guarantee that 1103 * the target is a copy of the source. 1104 */ 1105 assert(ret < 0 || job_is_cancelled(&s->common.job)); 1106 assert(need_drain); 1107 mirror_wait_for_all_io(s); 1108 } 1109 1110 assert(s->in_flight == 0); 1111 qemu_vfree(s->buf); 1112 g_free(s->cow_bitmap); 1113 g_free(s->in_flight_bitmap); 1114 bdrv_dirty_iter_free(s->dbi); 1115 1116 if (need_drain) { 1117 s->in_drain = true; 1118 bdrv_drained_begin(bs); 1119 } 1120 1121 return ret; 1122 } 1123 1124 static void mirror_complete(Job *job, Error **errp) 1125 { 1126 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 1127 1128 if (!job_is_ready(job)) { 1129 error_setg(errp, "The active block job '%s' cannot be completed", 1130 job->id); 1131 return; 1132 } 1133 1134 /* block all operations on to_replace bs */ 1135 if (s->replaces) { 1136 AioContext *replace_aio_context; 1137 1138 s->to_replace = bdrv_find_node(s->replaces); 1139 if (!s->to_replace) { 1140 error_setg(errp, "Node name '%s' not found", s->replaces); 1141 return; 1142 } 1143 1144 replace_aio_context = bdrv_get_aio_context(s->to_replace); 1145 aio_context_acquire(replace_aio_context); 1146 1147 /* TODO Translate this into child freeze system. */ 1148 error_setg(&s->replace_blocker, 1149 "block device is in use by block-job-complete"); 1150 bdrv_op_block_all(s->to_replace, s->replace_blocker); 1151 bdrv_ref(s->to_replace); 1152 1153 aio_context_release(replace_aio_context); 1154 } 1155 1156 s->should_complete = true; 1157 1158 /* If the job is paused, it will be re-entered when it is resumed */ 1159 WITH_JOB_LOCK_GUARD() { 1160 if (!job->paused) { 1161 job_enter_cond_locked(job, NULL); 1162 } 1163 } 1164 } 1165 1166 static void coroutine_fn mirror_pause(Job *job) 1167 { 1168 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 1169 1170 mirror_wait_for_all_io(s); 1171 } 1172 1173 static bool mirror_drained_poll(BlockJob *job) 1174 { 1175 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common); 1176 1177 /* If the job isn't paused nor cancelled, we can't be sure that it won't 1178 * issue more requests. We make an exception if we've reached this point 1179 * from one of our own drain sections, to avoid a deadlock waiting for 1180 * ourselves. 1181 */ 1182 WITH_JOB_LOCK_GUARD() { 1183 if (!s->common.job.paused && !job_is_cancelled_locked(&job->job) 1184 && !s->in_drain) { 1185 return true; 1186 } 1187 } 1188 1189 return !!s->in_flight; 1190 } 1191 1192 static bool mirror_cancel(Job *job, bool force) 1193 { 1194 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 1195 BlockDriverState *target = blk_bs(s->target); 1196 1197 /* 1198 * Before the job is READY, we treat any cancellation like a 1199 * force-cancellation. 1200 */ 1201 force = force || !job_is_ready(job); 1202 1203 if (force) { 1204 bdrv_cancel_in_flight(target); 1205 } 1206 return force; 1207 } 1208 1209 static bool commit_active_cancel(Job *job, bool force) 1210 { 1211 /* Same as above in mirror_cancel() */ 1212 return force || !job_is_ready(job); 1213 } 1214 1215 static const BlockJobDriver mirror_job_driver = { 1216 .job_driver = { 1217 .instance_size = sizeof(MirrorBlockJob), 1218 .job_type = JOB_TYPE_MIRROR, 1219 .free = block_job_free, 1220 .user_resume = block_job_user_resume, 1221 .run = mirror_run, 1222 .prepare = mirror_prepare, 1223 .abort = mirror_abort, 1224 .pause = mirror_pause, 1225 .complete = mirror_complete, 1226 .cancel = mirror_cancel, 1227 }, 1228 .drained_poll = mirror_drained_poll, 1229 }; 1230 1231 static const BlockJobDriver commit_active_job_driver = { 1232 .job_driver = { 1233 .instance_size = sizeof(MirrorBlockJob), 1234 .job_type = JOB_TYPE_COMMIT, 1235 .free = block_job_free, 1236 .user_resume = block_job_user_resume, 1237 .run = mirror_run, 1238 .prepare = mirror_prepare, 1239 .abort = mirror_abort, 1240 .pause = mirror_pause, 1241 .complete = mirror_complete, 1242 .cancel = commit_active_cancel, 1243 }, 1244 .drained_poll = mirror_drained_poll, 1245 }; 1246 1247 static void coroutine_fn 1248 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method, 1249 uint64_t offset, uint64_t bytes, 1250 QEMUIOVector *qiov, int flags) 1251 { 1252 int ret; 1253 size_t qiov_offset = 0; 1254 int64_t bitmap_offset, bitmap_end; 1255 1256 if (!QEMU_IS_ALIGNED(offset, job->granularity) && 1257 bdrv_dirty_bitmap_get(job->dirty_bitmap, offset)) 1258 { 1259 /* 1260 * Dirty unaligned padding: ignore it. 1261 * 1262 * Reasoning: 1263 * 1. If we copy it, we can't reset corresponding bit in 1264 * dirty_bitmap as there may be some "dirty" bytes still not 1265 * copied. 1266 * 2. It's already dirty, so skipping it we don't diverge mirror 1267 * progress. 1268 * 1269 * Note, that because of this, guest write may have no contribution 1270 * into mirror converge, but that's not bad, as we have background 1271 * process of mirroring. If under some bad circumstances (high guest 1272 * IO load) background process starve, we will not converge anyway, 1273 * even if each write will contribute, as guest is not guaranteed to 1274 * rewrite the whole disk. 1275 */ 1276 qiov_offset = QEMU_ALIGN_UP(offset, job->granularity) - offset; 1277 if (bytes <= qiov_offset) { 1278 /* nothing to do after shrink */ 1279 return; 1280 } 1281 offset += qiov_offset; 1282 bytes -= qiov_offset; 1283 } 1284 1285 if (!QEMU_IS_ALIGNED(offset + bytes, job->granularity) && 1286 bdrv_dirty_bitmap_get(job->dirty_bitmap, offset + bytes - 1)) 1287 { 1288 uint64_t tail = (offset + bytes) % job->granularity; 1289 1290 if (bytes <= tail) { 1291 /* nothing to do after shrink */ 1292 return; 1293 } 1294 bytes -= tail; 1295 } 1296 1297 /* 1298 * Tails are either clean or shrunk, so for bitmap resetting 1299 * we safely align the range down. 1300 */ 1301 bitmap_offset = QEMU_ALIGN_UP(offset, job->granularity); 1302 bitmap_end = QEMU_ALIGN_DOWN(offset + bytes, job->granularity); 1303 if (bitmap_offset < bitmap_end) { 1304 bdrv_reset_dirty_bitmap(job->dirty_bitmap, bitmap_offset, 1305 bitmap_end - bitmap_offset); 1306 } 1307 1308 job_progress_increase_remaining(&job->common.job, bytes); 1309 1310 switch (method) { 1311 case MIRROR_METHOD_COPY: 1312 ret = blk_co_pwritev_part(job->target, offset, bytes, 1313 qiov, qiov_offset, flags); 1314 break; 1315 1316 case MIRROR_METHOD_ZERO: 1317 assert(!qiov); 1318 ret = blk_co_pwrite_zeroes(job->target, offset, bytes, flags); 1319 break; 1320 1321 case MIRROR_METHOD_DISCARD: 1322 assert(!qiov); 1323 ret = blk_co_pdiscard(job->target, offset, bytes); 1324 break; 1325 1326 default: 1327 abort(); 1328 } 1329 1330 if (ret >= 0) { 1331 job_progress_update(&job->common.job, bytes); 1332 } else { 1333 BlockErrorAction action; 1334 1335 /* 1336 * We failed, so we should mark dirty the whole area, aligned up. 1337 * Note that we don't care about shrunk tails if any: they were dirty 1338 * at function start, and they must be still dirty, as we've locked 1339 * the region for in-flight op. 1340 */ 1341 bitmap_offset = QEMU_ALIGN_DOWN(offset, job->granularity); 1342 bitmap_end = QEMU_ALIGN_UP(offset + bytes, job->granularity); 1343 bdrv_set_dirty_bitmap(job->dirty_bitmap, bitmap_offset, 1344 bitmap_end - bitmap_offset); 1345 job->actively_synced = false; 1346 1347 action = mirror_error_action(job, false, -ret); 1348 if (action == BLOCK_ERROR_ACTION_REPORT) { 1349 if (!job->ret) { 1350 job->ret = ret; 1351 } 1352 } 1353 } 1354 } 1355 1356 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s, 1357 uint64_t offset, 1358 uint64_t bytes) 1359 { 1360 MirrorOp *op; 1361 uint64_t start_chunk = offset / s->granularity; 1362 uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity); 1363 1364 op = g_new(MirrorOp, 1); 1365 *op = (MirrorOp){ 1366 .s = s, 1367 .offset = offset, 1368 .bytes = bytes, 1369 .is_active_write = true, 1370 .is_in_flight = true, 1371 .co = qemu_coroutine_self(), 1372 }; 1373 qemu_co_queue_init(&op->waiting_requests); 1374 QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next); 1375 1376 s->in_active_write_counter++; 1377 1378 mirror_wait_on_conflicts(op, s, offset, bytes); 1379 1380 bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk); 1381 1382 return op; 1383 } 1384 1385 static void coroutine_fn active_write_settle(MirrorOp *op) 1386 { 1387 uint64_t start_chunk = op->offset / op->s->granularity; 1388 uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes, 1389 op->s->granularity); 1390 1391 if (!--op->s->in_active_write_counter && op->s->actively_synced) { 1392 BdrvChild *source = op->s->mirror_top_bs->backing; 1393 1394 if (QLIST_FIRST(&source->bs->parents) == source && 1395 QLIST_NEXT(source, next_parent) == NULL) 1396 { 1397 /* Assert that we are back in sync once all active write 1398 * operations are settled. 1399 * Note that we can only assert this if the mirror node 1400 * is the source node's only parent. */ 1401 assert(!bdrv_get_dirty_count(op->s->dirty_bitmap)); 1402 } 1403 } 1404 bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk); 1405 QTAILQ_REMOVE(&op->s->ops_in_flight, op, next); 1406 qemu_co_queue_restart_all(&op->waiting_requests); 1407 g_free(op); 1408 } 1409 1410 static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs, 1411 int64_t offset, int64_t bytes, QEMUIOVector *qiov, BdrvRequestFlags flags) 1412 { 1413 return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags); 1414 } 1415 1416 static int coroutine_fn bdrv_mirror_top_do_write(BlockDriverState *bs, 1417 MirrorMethod method, uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, 1418 int flags) 1419 { 1420 MirrorOp *op = NULL; 1421 MirrorBDSOpaque *s = bs->opaque; 1422 int ret = 0; 1423 bool copy_to_target; 1424 1425 copy_to_target = s->job->ret >= 0 && 1426 !job_is_cancelled(&s->job->common.job) && 1427 s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING; 1428 1429 if (copy_to_target) { 1430 op = active_write_prepare(s->job, offset, bytes); 1431 } 1432 1433 switch (method) { 1434 case MIRROR_METHOD_COPY: 1435 ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags); 1436 break; 1437 1438 case MIRROR_METHOD_ZERO: 1439 ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags); 1440 break; 1441 1442 case MIRROR_METHOD_DISCARD: 1443 ret = bdrv_co_pdiscard(bs->backing, offset, bytes); 1444 break; 1445 1446 default: 1447 abort(); 1448 } 1449 1450 if (ret < 0) { 1451 goto out; 1452 } 1453 1454 if (copy_to_target) { 1455 do_sync_target_write(s->job, method, offset, bytes, qiov, flags); 1456 } 1457 1458 out: 1459 if (copy_to_target) { 1460 active_write_settle(op); 1461 } 1462 return ret; 1463 } 1464 1465 static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs, 1466 int64_t offset, int64_t bytes, QEMUIOVector *qiov, BdrvRequestFlags flags) 1467 { 1468 MirrorBDSOpaque *s = bs->opaque; 1469 QEMUIOVector bounce_qiov; 1470 void *bounce_buf; 1471 int ret = 0; 1472 bool copy_to_target; 1473 1474 copy_to_target = s->job->ret >= 0 && 1475 !job_is_cancelled(&s->job->common.job) && 1476 s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING; 1477 1478 if (copy_to_target) { 1479 /* The guest might concurrently modify the data to write; but 1480 * the data on source and destination must match, so we have 1481 * to use a bounce buffer if we are going to write to the 1482 * target now. */ 1483 bounce_buf = qemu_blockalign(bs, bytes); 1484 iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes); 1485 1486 qemu_iovec_init(&bounce_qiov, 1); 1487 qemu_iovec_add(&bounce_qiov, bounce_buf, bytes); 1488 qiov = &bounce_qiov; 1489 1490 flags &= ~BDRV_REQ_REGISTERED_BUF; 1491 } 1492 1493 ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov, 1494 flags); 1495 1496 if (copy_to_target) { 1497 qemu_iovec_destroy(&bounce_qiov); 1498 qemu_vfree(bounce_buf); 1499 } 1500 1501 return ret; 1502 } 1503 1504 static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs) 1505 { 1506 if (bs->backing == NULL) { 1507 /* we can be here after failed bdrv_append in mirror_start_job */ 1508 return 0; 1509 } 1510 return bdrv_co_flush(bs->backing->bs); 1511 } 1512 1513 static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs, 1514 int64_t offset, int64_t bytes, BdrvRequestFlags flags) 1515 { 1516 return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL, 1517 flags); 1518 } 1519 1520 static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs, 1521 int64_t offset, int64_t bytes) 1522 { 1523 return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes, 1524 NULL, 0); 1525 } 1526 1527 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs) 1528 { 1529 if (bs->backing == NULL) { 1530 /* we can be here after failed bdrv_attach_child in 1531 * bdrv_set_backing_hd */ 1532 return; 1533 } 1534 pstrcpy(bs->exact_filename, sizeof(bs->exact_filename), 1535 bs->backing->bs->filename); 1536 } 1537 1538 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c, 1539 BdrvChildRole role, 1540 BlockReopenQueue *reopen_queue, 1541 uint64_t perm, uint64_t shared, 1542 uint64_t *nperm, uint64_t *nshared) 1543 { 1544 MirrorBDSOpaque *s = bs->opaque; 1545 1546 if (s->stop) { 1547 /* 1548 * If the job is to be stopped, we do not need to forward 1549 * anything to the real image. 1550 */ 1551 *nperm = 0; 1552 *nshared = BLK_PERM_ALL; 1553 return; 1554 } 1555 1556 bdrv_default_perms(bs, c, role, reopen_queue, 1557 perm, shared, nperm, nshared); 1558 1559 if (s->is_commit) { 1560 /* 1561 * For commit jobs, we cannot take CONSISTENT_READ, because 1562 * that permission is unshared for everything above the base 1563 * node (except for filters on the base node). 1564 * We also have to force-share the WRITE permission, or 1565 * otherwise we would block ourselves at the base node (if 1566 * writes are blocked for a node, they are also blocked for 1567 * its backing file). 1568 * (We could also share RESIZE, because it may be needed for 1569 * the target if its size is less than the top node's; but 1570 * bdrv_default_perms_for_cow() automatically shares RESIZE 1571 * for backing nodes if WRITE is shared, so there is no need 1572 * to do it here.) 1573 */ 1574 *nperm &= ~BLK_PERM_CONSISTENT_READ; 1575 *nshared |= BLK_PERM_WRITE; 1576 } 1577 } 1578 1579 /* Dummy node that provides consistent read to its users without requiring it 1580 * from its backing file and that allows writes on the backing file chain. */ 1581 static BlockDriver bdrv_mirror_top = { 1582 .format_name = "mirror_top", 1583 .bdrv_co_preadv = bdrv_mirror_top_preadv, 1584 .bdrv_co_pwritev = bdrv_mirror_top_pwritev, 1585 .bdrv_co_pwrite_zeroes = bdrv_mirror_top_pwrite_zeroes, 1586 .bdrv_co_pdiscard = bdrv_mirror_top_pdiscard, 1587 .bdrv_co_flush = bdrv_mirror_top_flush, 1588 .bdrv_refresh_filename = bdrv_mirror_top_refresh_filename, 1589 .bdrv_child_perm = bdrv_mirror_top_child_perm, 1590 1591 .is_filter = true, 1592 .filtered_child_is_backing = true, 1593 }; 1594 1595 static BlockJob *mirror_start_job( 1596 const char *job_id, BlockDriverState *bs, 1597 int creation_flags, BlockDriverState *target, 1598 const char *replaces, int64_t speed, 1599 uint32_t granularity, int64_t buf_size, 1600 BlockMirrorBackingMode backing_mode, 1601 bool zero_target, 1602 BlockdevOnError on_source_error, 1603 BlockdevOnError on_target_error, 1604 bool unmap, 1605 BlockCompletionFunc *cb, 1606 void *opaque, 1607 const BlockJobDriver *driver, 1608 bool is_none_mode, BlockDriverState *base, 1609 bool auto_complete, const char *filter_node_name, 1610 bool is_mirror, MirrorCopyMode copy_mode, 1611 Error **errp) 1612 { 1613 MirrorBlockJob *s; 1614 MirrorBDSOpaque *bs_opaque; 1615 BlockDriverState *mirror_top_bs; 1616 bool target_is_backing; 1617 uint64_t target_perms, target_shared_perms; 1618 int ret; 1619 1620 if (granularity == 0) { 1621 granularity = bdrv_get_default_bitmap_granularity(target); 1622 } 1623 1624 assert(is_power_of_2(granularity)); 1625 1626 if (buf_size < 0) { 1627 error_setg(errp, "Invalid parameter 'buf-size'"); 1628 return NULL; 1629 } 1630 1631 if (buf_size == 0) { 1632 buf_size = DEFAULT_MIRROR_BUF_SIZE; 1633 } 1634 1635 if (bdrv_skip_filters(bs) == bdrv_skip_filters(target)) { 1636 error_setg(errp, "Can't mirror node into itself"); 1637 return NULL; 1638 } 1639 1640 target_is_backing = bdrv_chain_contains(bs, target); 1641 1642 /* In the case of active commit, add dummy driver to provide consistent 1643 * reads on the top, while disabling it in the intermediate nodes, and make 1644 * the backing chain writable. */ 1645 mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name, 1646 BDRV_O_RDWR, errp); 1647 if (mirror_top_bs == NULL) { 1648 return NULL; 1649 } 1650 if (!filter_node_name) { 1651 mirror_top_bs->implicit = true; 1652 } 1653 1654 /* So that we can always drop this node */ 1655 mirror_top_bs->never_freeze = true; 1656 1657 mirror_top_bs->total_sectors = bs->total_sectors; 1658 mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED; 1659 mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED | 1660 BDRV_REQ_NO_FALLBACK; 1661 bs_opaque = g_new0(MirrorBDSOpaque, 1); 1662 mirror_top_bs->opaque = bs_opaque; 1663 1664 bs_opaque->is_commit = target_is_backing; 1665 1666 bdrv_drained_begin(bs); 1667 ret = bdrv_append(mirror_top_bs, bs, errp); 1668 bdrv_drained_end(bs); 1669 1670 if (ret < 0) { 1671 bdrv_unref(mirror_top_bs); 1672 return NULL; 1673 } 1674 1675 /* Make sure that the source is not resized while the job is running */ 1676 s = block_job_create(job_id, driver, NULL, mirror_top_bs, 1677 BLK_PERM_CONSISTENT_READ, 1678 BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED | 1679 BLK_PERM_WRITE, speed, 1680 creation_flags, cb, opaque, errp); 1681 if (!s) { 1682 goto fail; 1683 } 1684 bs_opaque->job = s; 1685 1686 /* The block job now has a reference to this node */ 1687 bdrv_unref(mirror_top_bs); 1688 1689 s->mirror_top_bs = mirror_top_bs; 1690 1691 /* No resize for the target either; while the mirror is still running, a 1692 * consistent read isn't necessarily possible. We could possibly allow 1693 * writes and graph modifications, though it would likely defeat the 1694 * purpose of a mirror, so leave them blocked for now. 1695 * 1696 * In the case of active commit, things look a bit different, though, 1697 * because the target is an already populated backing file in active use. 1698 * We can allow anything except resize there.*/ 1699 1700 target_perms = BLK_PERM_WRITE; 1701 target_shared_perms = BLK_PERM_WRITE_UNCHANGED; 1702 1703 if (target_is_backing) { 1704 int64_t bs_size, target_size; 1705 bs_size = bdrv_getlength(bs); 1706 if (bs_size < 0) { 1707 error_setg_errno(errp, -bs_size, 1708 "Could not inquire top image size"); 1709 goto fail; 1710 } 1711 1712 target_size = bdrv_getlength(target); 1713 if (target_size < 0) { 1714 error_setg_errno(errp, -target_size, 1715 "Could not inquire base image size"); 1716 goto fail; 1717 } 1718 1719 if (target_size < bs_size) { 1720 target_perms |= BLK_PERM_RESIZE; 1721 } 1722 1723 target_shared_perms |= BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE; 1724 } else if (bdrv_chain_contains(bs, bdrv_skip_filters(target))) { 1725 /* 1726 * We may want to allow this in the future, but it would 1727 * require taking some extra care. 1728 */ 1729 error_setg(errp, "Cannot mirror to a filter on top of a node in the " 1730 "source's backing chain"); 1731 goto fail; 1732 } 1733 1734 s->target = blk_new(s->common.job.aio_context, 1735 target_perms, target_shared_perms); 1736 ret = blk_insert_bs(s->target, target, errp); 1737 if (ret < 0) { 1738 goto fail; 1739 } 1740 if (is_mirror) { 1741 /* XXX: Mirror target could be a NBD server of target QEMU in the case 1742 * of non-shared block migration. To allow migration completion, we 1743 * have to allow "inactivate" of the target BB. When that happens, we 1744 * know the job is drained, and the vcpus are stopped, so no write 1745 * operation will be performed. Block layer already has assertions to 1746 * ensure that. */ 1747 blk_set_force_allow_inactivate(s->target); 1748 } 1749 blk_set_allow_aio_context_change(s->target, true); 1750 blk_set_disable_request_queuing(s->target, true); 1751 1752 s->replaces = g_strdup(replaces); 1753 s->on_source_error = on_source_error; 1754 s->on_target_error = on_target_error; 1755 s->is_none_mode = is_none_mode; 1756 s->backing_mode = backing_mode; 1757 s->zero_target = zero_target; 1758 s->copy_mode = copy_mode; 1759 s->base = base; 1760 s->base_overlay = bdrv_find_overlay(bs, base); 1761 s->granularity = granularity; 1762 s->buf_size = ROUND_UP(buf_size, granularity); 1763 s->unmap = unmap; 1764 if (auto_complete) { 1765 s->should_complete = true; 1766 } 1767 1768 s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp); 1769 if (!s->dirty_bitmap) { 1770 goto fail; 1771 } 1772 if (s->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING) { 1773 bdrv_disable_dirty_bitmap(s->dirty_bitmap); 1774 } 1775 1776 ret = block_job_add_bdrv(&s->common, "source", bs, 0, 1777 BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE | 1778 BLK_PERM_CONSISTENT_READ, 1779 errp); 1780 if (ret < 0) { 1781 goto fail; 1782 } 1783 1784 /* Required permissions are already taken with blk_new() */ 1785 block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL, 1786 &error_abort); 1787 1788 /* In commit_active_start() all intermediate nodes disappear, so 1789 * any jobs in them must be blocked */ 1790 if (target_is_backing) { 1791 BlockDriverState *iter, *filtered_target; 1792 uint64_t iter_shared_perms; 1793 1794 /* 1795 * The topmost node with 1796 * bdrv_skip_filters(filtered_target) == bdrv_skip_filters(target) 1797 */ 1798 filtered_target = bdrv_cow_bs(bdrv_find_overlay(bs, target)); 1799 1800 assert(bdrv_skip_filters(filtered_target) == 1801 bdrv_skip_filters(target)); 1802 1803 /* 1804 * XXX BLK_PERM_WRITE needs to be allowed so we don't block 1805 * ourselves at s->base (if writes are blocked for a node, they are 1806 * also blocked for its backing file). The other options would be a 1807 * second filter driver above s->base (== target). 1808 */ 1809 iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE; 1810 1811 for (iter = bdrv_filter_or_cow_bs(bs); iter != target; 1812 iter = bdrv_filter_or_cow_bs(iter)) 1813 { 1814 if (iter == filtered_target) { 1815 /* 1816 * From here on, all nodes are filters on the base. 1817 * This allows us to share BLK_PERM_CONSISTENT_READ. 1818 */ 1819 iter_shared_perms |= BLK_PERM_CONSISTENT_READ; 1820 } 1821 1822 ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0, 1823 iter_shared_perms, errp); 1824 if (ret < 0) { 1825 goto fail; 1826 } 1827 } 1828 1829 if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) { 1830 goto fail; 1831 } 1832 } 1833 1834 QTAILQ_INIT(&s->ops_in_flight); 1835 1836 trace_mirror_start(bs, s, opaque); 1837 job_start(&s->common.job); 1838 1839 return &s->common; 1840 1841 fail: 1842 if (s) { 1843 /* Make sure this BDS does not go away until we have completed the graph 1844 * changes below */ 1845 bdrv_ref(mirror_top_bs); 1846 1847 g_free(s->replaces); 1848 blk_unref(s->target); 1849 bs_opaque->job = NULL; 1850 if (s->dirty_bitmap) { 1851 bdrv_release_dirty_bitmap(s->dirty_bitmap); 1852 } 1853 job_early_fail(&s->common.job); 1854 } 1855 1856 bs_opaque->stop = true; 1857 bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing, 1858 &error_abort); 1859 bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort); 1860 1861 bdrv_unref(mirror_top_bs); 1862 1863 return NULL; 1864 } 1865 1866 void mirror_start(const char *job_id, BlockDriverState *bs, 1867 BlockDriverState *target, const char *replaces, 1868 int creation_flags, int64_t speed, 1869 uint32_t granularity, int64_t buf_size, 1870 MirrorSyncMode mode, BlockMirrorBackingMode backing_mode, 1871 bool zero_target, 1872 BlockdevOnError on_source_error, 1873 BlockdevOnError on_target_error, 1874 bool unmap, const char *filter_node_name, 1875 MirrorCopyMode copy_mode, Error **errp) 1876 { 1877 bool is_none_mode; 1878 BlockDriverState *base; 1879 1880 GLOBAL_STATE_CODE(); 1881 1882 if ((mode == MIRROR_SYNC_MODE_INCREMENTAL) || 1883 (mode == MIRROR_SYNC_MODE_BITMAP)) { 1884 error_setg(errp, "Sync mode '%s' not supported", 1885 MirrorSyncMode_str(mode)); 1886 return; 1887 } 1888 is_none_mode = mode == MIRROR_SYNC_MODE_NONE; 1889 base = mode == MIRROR_SYNC_MODE_TOP ? bdrv_backing_chain_next(bs) : NULL; 1890 mirror_start_job(job_id, bs, creation_flags, target, replaces, 1891 speed, granularity, buf_size, backing_mode, zero_target, 1892 on_source_error, on_target_error, unmap, NULL, NULL, 1893 &mirror_job_driver, is_none_mode, base, false, 1894 filter_node_name, true, copy_mode, errp); 1895 } 1896 1897 BlockJob *commit_active_start(const char *job_id, BlockDriverState *bs, 1898 BlockDriverState *base, int creation_flags, 1899 int64_t speed, BlockdevOnError on_error, 1900 const char *filter_node_name, 1901 BlockCompletionFunc *cb, void *opaque, 1902 bool auto_complete, Error **errp) 1903 { 1904 bool base_read_only; 1905 BlockJob *job; 1906 1907 GLOBAL_STATE_CODE(); 1908 1909 base_read_only = bdrv_is_read_only(base); 1910 1911 if (base_read_only) { 1912 if (bdrv_reopen_set_read_only(base, false, errp) < 0) { 1913 return NULL; 1914 } 1915 } 1916 1917 job = mirror_start_job( 1918 job_id, bs, creation_flags, base, NULL, speed, 0, 0, 1919 MIRROR_LEAVE_BACKING_CHAIN, false, 1920 on_error, on_error, true, cb, opaque, 1921 &commit_active_job_driver, false, base, auto_complete, 1922 filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND, 1923 errp); 1924 if (!job) { 1925 goto error_restore_flags; 1926 } 1927 1928 return job; 1929 1930 error_restore_flags: 1931 /* ignore error and errp for bdrv_reopen, because we want to propagate 1932 * the original error */ 1933 if (base_read_only) { 1934 bdrv_reopen_set_read_only(base, true, NULL); 1935 } 1936 return NULL; 1937 } 1938