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 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->synced = false; 125 s->actively_synced = false; 126 if (read) { 127 return block_job_error_action(&s->common, s->on_source_error, 128 true, error); 129 } else { 130 return block_job_error_action(&s->common, s->on_target_error, 131 false, error); 132 } 133 } 134 135 static void coroutine_fn mirror_wait_on_conflicts(MirrorOp *self, 136 MirrorBlockJob *s, 137 uint64_t offset, 138 uint64_t bytes) 139 { 140 uint64_t self_start_chunk = offset / s->granularity; 141 uint64_t self_end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity); 142 uint64_t self_nb_chunks = self_end_chunk - self_start_chunk; 143 144 while (find_next_bit(s->in_flight_bitmap, self_end_chunk, 145 self_start_chunk) < self_end_chunk && 146 s->ret >= 0) 147 { 148 MirrorOp *op; 149 150 QTAILQ_FOREACH(op, &s->ops_in_flight, next) { 151 uint64_t op_start_chunk = op->offset / s->granularity; 152 uint64_t op_nb_chunks = DIV_ROUND_UP(op->offset + op->bytes, 153 s->granularity) - 154 op_start_chunk; 155 156 if (op == self) { 157 continue; 158 } 159 160 if (ranges_overlap(self_start_chunk, self_nb_chunks, 161 op_start_chunk, op_nb_chunks)) 162 { 163 /* 164 * If the operation is already (indirectly) waiting for us, or 165 * will wait for us as soon as it wakes up, then just go on 166 * (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 qemu_co_queue_wait(&op->waiting_requests, NULL); 174 self->waiting_for_op = NULL; 175 break; 176 } 177 } 178 } 179 } 180 181 static void coroutine_fn mirror_iteration_done(MirrorOp *op, int ret) 182 { 183 MirrorBlockJob *s = op->s; 184 struct iovec *iov; 185 int64_t chunk_num; 186 int i, nb_chunks; 187 188 trace_mirror_iteration_done(s, op->offset, op->bytes, ret); 189 190 s->in_flight--; 191 s->bytes_in_flight -= op->bytes; 192 iov = op->qiov.iov; 193 for (i = 0; i < op->qiov.niov; i++) { 194 MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base; 195 QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next); 196 s->buf_free_count++; 197 } 198 199 chunk_num = op->offset / s->granularity; 200 nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity); 201 202 bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks); 203 QTAILQ_REMOVE(&s->ops_in_flight, op, next); 204 if (ret >= 0) { 205 if (s->cow_bitmap) { 206 bitmap_set(s->cow_bitmap, chunk_num, nb_chunks); 207 } 208 if (!s->initial_zeroing_ongoing) { 209 job_progress_update(&s->common.job, op->bytes); 210 } 211 } 212 qemu_iovec_destroy(&op->qiov); 213 214 qemu_co_queue_restart_all(&op->waiting_requests); 215 g_free(op); 216 } 217 218 static void coroutine_fn mirror_write_complete(MirrorOp *op, int ret) 219 { 220 MirrorBlockJob *s = op->s; 221 222 if (ret < 0) { 223 BlockErrorAction action; 224 225 bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes); 226 action = mirror_error_action(s, false, -ret); 227 if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) { 228 s->ret = ret; 229 } 230 } 231 232 mirror_iteration_done(op, ret); 233 } 234 235 static void coroutine_fn mirror_read_complete(MirrorOp *op, int ret) 236 { 237 MirrorBlockJob *s = op->s; 238 239 if (ret < 0) { 240 BlockErrorAction action; 241 242 bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes); 243 action = mirror_error_action(s, true, -ret); 244 if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) { 245 s->ret = ret; 246 } 247 248 mirror_iteration_done(op, ret); 249 return; 250 } 251 252 ret = blk_co_pwritev(s->target, op->offset, op->qiov.size, &op->qiov, 0); 253 mirror_write_complete(op, ret); 254 } 255 256 /* Clip bytes relative to offset to not exceed end-of-file */ 257 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s, 258 int64_t offset, 259 int64_t bytes) 260 { 261 return MIN(bytes, s->bdev_length - offset); 262 } 263 264 /* Round offset and/or bytes to target cluster if COW is needed, and 265 * return the offset of the adjusted tail against original. */ 266 static int mirror_cow_align(MirrorBlockJob *s, int64_t *offset, 267 uint64_t *bytes) 268 { 269 bool need_cow; 270 int ret = 0; 271 int64_t align_offset = *offset; 272 int64_t align_bytes = *bytes; 273 int max_bytes = s->granularity * s->max_iov; 274 275 need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap); 276 need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity, 277 s->cow_bitmap); 278 if (need_cow) { 279 bdrv_round_to_clusters(blk_bs(s->target), *offset, *bytes, 280 &align_offset, &align_bytes); 281 } 282 283 if (align_bytes > max_bytes) { 284 align_bytes = max_bytes; 285 if (need_cow) { 286 align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size); 287 } 288 } 289 /* Clipping may result in align_bytes unaligned to chunk boundary, but 290 * that doesn't matter because it's already the end of source image. */ 291 align_bytes = mirror_clip_bytes(s, align_offset, align_bytes); 292 293 ret = align_offset + align_bytes - (*offset + *bytes); 294 *offset = align_offset; 295 *bytes = align_bytes; 296 assert(ret >= 0); 297 return ret; 298 } 299 300 static inline void coroutine_fn 301 mirror_wait_for_any_operation(MirrorBlockJob *s, bool active) 302 { 303 MirrorOp *op; 304 305 QTAILQ_FOREACH(op, &s->ops_in_flight, next) { 306 /* Do not wait on pseudo ops, because it may in turn wait on 307 * some other operation to start, which may in fact be the 308 * caller of this function. Since there is only one pseudo op 309 * at any given time, we will always find some real operation 310 * to wait on. */ 311 if (!op->is_pseudo_op && op->is_in_flight && 312 op->is_active_write == active) 313 { 314 qemu_co_queue_wait(&op->waiting_requests, NULL); 315 return; 316 } 317 } 318 abort(); 319 } 320 321 static inline void coroutine_fn 322 mirror_wait_for_free_in_flight_slot(MirrorBlockJob *s) 323 { 324 /* Only non-active operations use up in-flight slots */ 325 mirror_wait_for_any_operation(s, false); 326 } 327 328 /* Perform a mirror copy operation. 329 * 330 * *op->bytes_handled is set to the number of bytes copied after and 331 * including offset, excluding any bytes copied prior to offset due 332 * to alignment. This will be op->bytes if no alignment is necessary, 333 * or (new_end - op->offset) if the tail is rounded up or down due to 334 * alignment or buffer limit. 335 */ 336 static void coroutine_fn mirror_co_read(void *opaque) 337 { 338 MirrorOp *op = opaque; 339 MirrorBlockJob *s = op->s; 340 int nb_chunks; 341 uint64_t ret; 342 uint64_t max_bytes; 343 344 max_bytes = s->granularity * s->max_iov; 345 346 /* We can only handle as much as buf_size at a time. */ 347 op->bytes = MIN(s->buf_size, MIN(max_bytes, op->bytes)); 348 assert(op->bytes); 349 assert(op->bytes < BDRV_REQUEST_MAX_BYTES); 350 *op->bytes_handled = op->bytes; 351 352 if (s->cow_bitmap) { 353 *op->bytes_handled += mirror_cow_align(s, &op->offset, &op->bytes); 354 } 355 /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */ 356 assert(*op->bytes_handled <= UINT_MAX); 357 assert(op->bytes <= s->buf_size); 358 /* The offset is granularity-aligned because: 359 * 1) Caller passes in aligned values; 360 * 2) mirror_cow_align is used only when target cluster is larger. */ 361 assert(QEMU_IS_ALIGNED(op->offset, s->granularity)); 362 /* The range is sector-aligned, since bdrv_getlength() rounds up. */ 363 assert(QEMU_IS_ALIGNED(op->bytes, BDRV_SECTOR_SIZE)); 364 nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity); 365 366 while (s->buf_free_count < nb_chunks) { 367 trace_mirror_yield_in_flight(s, op->offset, s->in_flight); 368 mirror_wait_for_free_in_flight_slot(s); 369 } 370 371 /* Now make a QEMUIOVector taking enough granularity-sized chunks 372 * from s->buf_free. 373 */ 374 qemu_iovec_init(&op->qiov, nb_chunks); 375 while (nb_chunks-- > 0) { 376 MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free); 377 size_t remaining = op->bytes - op->qiov.size; 378 379 QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next); 380 s->buf_free_count--; 381 qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining)); 382 } 383 384 /* Copy the dirty cluster. */ 385 s->in_flight++; 386 s->bytes_in_flight += op->bytes; 387 op->is_in_flight = true; 388 trace_mirror_one_iteration(s, op->offset, op->bytes); 389 390 ret = bdrv_co_preadv(s->mirror_top_bs->backing, op->offset, op->bytes, 391 &op->qiov, 0); 392 mirror_read_complete(op, ret); 393 } 394 395 static void coroutine_fn mirror_co_zero(void *opaque) 396 { 397 MirrorOp *op = opaque; 398 int ret; 399 400 op->s->in_flight++; 401 op->s->bytes_in_flight += op->bytes; 402 *op->bytes_handled = op->bytes; 403 op->is_in_flight = true; 404 405 ret = blk_co_pwrite_zeroes(op->s->target, op->offset, op->bytes, 406 op->s->unmap ? BDRV_REQ_MAY_UNMAP : 0); 407 mirror_write_complete(op, ret); 408 } 409 410 static void coroutine_fn mirror_co_discard(void *opaque) 411 { 412 MirrorOp *op = opaque; 413 int ret; 414 415 op->s->in_flight++; 416 op->s->bytes_in_flight += op->bytes; 417 *op->bytes_handled = op->bytes; 418 op->is_in_flight = true; 419 420 ret = blk_co_pdiscard(op->s->target, op->offset, op->bytes); 421 mirror_write_complete(op, ret); 422 } 423 424 static unsigned mirror_perform(MirrorBlockJob *s, int64_t offset, 425 unsigned bytes, MirrorMethod mirror_method) 426 { 427 MirrorOp *op; 428 Coroutine *co; 429 int64_t bytes_handled = -1; 430 431 op = g_new(MirrorOp, 1); 432 *op = (MirrorOp){ 433 .s = s, 434 .offset = offset, 435 .bytes = bytes, 436 .bytes_handled = &bytes_handled, 437 }; 438 qemu_co_queue_init(&op->waiting_requests); 439 440 switch (mirror_method) { 441 case MIRROR_METHOD_COPY: 442 co = qemu_coroutine_create(mirror_co_read, op); 443 break; 444 case MIRROR_METHOD_ZERO: 445 co = qemu_coroutine_create(mirror_co_zero, op); 446 break; 447 case MIRROR_METHOD_DISCARD: 448 co = qemu_coroutine_create(mirror_co_discard, op); 449 break; 450 default: 451 abort(); 452 } 453 op->co = co; 454 455 QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next); 456 qemu_coroutine_enter(co); 457 /* At this point, ownership of op has been moved to the coroutine 458 * and the object may already be freed */ 459 460 /* Assert that this value has been set */ 461 assert(bytes_handled >= 0); 462 463 /* Same assertion as in mirror_co_read() (and for mirror_co_read() 464 * and mirror_co_discard(), bytes_handled == op->bytes, which 465 * is the @bytes parameter given to this function) */ 466 assert(bytes_handled <= UINT_MAX); 467 return bytes_handled; 468 } 469 470 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s) 471 { 472 BlockDriverState *source = s->mirror_top_bs->backing->bs; 473 MirrorOp *pseudo_op; 474 int64_t offset; 475 uint64_t delay_ns = 0, ret = 0; 476 /* At least the first dirty chunk is mirrored in one iteration. */ 477 int nb_chunks = 1; 478 bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target)); 479 int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES); 480 481 bdrv_dirty_bitmap_lock(s->dirty_bitmap); 482 offset = bdrv_dirty_iter_next(s->dbi); 483 if (offset < 0) { 484 bdrv_set_dirty_iter(s->dbi, 0); 485 offset = bdrv_dirty_iter_next(s->dbi); 486 trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap)); 487 assert(offset >= 0); 488 } 489 bdrv_dirty_bitmap_unlock(s->dirty_bitmap); 490 491 mirror_wait_on_conflicts(NULL, s, offset, 1); 492 493 job_pause_point(&s->common.job); 494 495 /* Find the number of consective dirty chunks following the first dirty 496 * one, and wait for in flight requests in them. */ 497 bdrv_dirty_bitmap_lock(s->dirty_bitmap); 498 while (nb_chunks * s->granularity < s->buf_size) { 499 int64_t next_dirty; 500 int64_t next_offset = offset + nb_chunks * s->granularity; 501 int64_t next_chunk = next_offset / s->granularity; 502 if (next_offset >= s->bdev_length || 503 !bdrv_dirty_bitmap_get_locked(s->dirty_bitmap, next_offset)) { 504 break; 505 } 506 if (test_bit(next_chunk, s->in_flight_bitmap)) { 507 break; 508 } 509 510 next_dirty = bdrv_dirty_iter_next(s->dbi); 511 if (next_dirty > next_offset || next_dirty < 0) { 512 /* The bitmap iterator's cache is stale, refresh it */ 513 bdrv_set_dirty_iter(s->dbi, next_offset); 514 next_dirty = bdrv_dirty_iter_next(s->dbi); 515 } 516 assert(next_dirty == next_offset); 517 nb_chunks++; 518 } 519 520 /* Clear dirty bits before querying the block status, because 521 * calling bdrv_block_status_above could yield - if some blocks are 522 * marked dirty in this window, we need to know. 523 */ 524 bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset, 525 nb_chunks * s->granularity); 526 bdrv_dirty_bitmap_unlock(s->dirty_bitmap); 527 528 /* Before claiming an area in the in-flight bitmap, we have to 529 * create a MirrorOp for it so that conflicting requests can wait 530 * for it. mirror_perform() will create the real MirrorOps later, 531 * for now we just create a pseudo operation that will wake up all 532 * conflicting requests once all real operations have been 533 * launched. */ 534 pseudo_op = g_new(MirrorOp, 1); 535 *pseudo_op = (MirrorOp){ 536 .offset = offset, 537 .bytes = nb_chunks * s->granularity, 538 .is_pseudo_op = true, 539 }; 540 qemu_co_queue_init(&pseudo_op->waiting_requests); 541 QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next); 542 543 bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks); 544 while (nb_chunks > 0 && offset < s->bdev_length) { 545 int ret; 546 int64_t io_bytes; 547 int64_t io_bytes_acct; 548 MirrorMethod mirror_method = MIRROR_METHOD_COPY; 549 550 assert(!(offset % s->granularity)); 551 ret = bdrv_block_status_above(source, NULL, offset, 552 nb_chunks * s->granularity, 553 &io_bytes, NULL, NULL); 554 if (ret < 0) { 555 io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes); 556 } else if (ret & BDRV_BLOCK_DATA) { 557 io_bytes = MIN(io_bytes, max_io_bytes); 558 } 559 560 io_bytes -= io_bytes % s->granularity; 561 if (io_bytes < s->granularity) { 562 io_bytes = s->granularity; 563 } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) { 564 int64_t target_offset; 565 int64_t target_bytes; 566 bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes, 567 &target_offset, &target_bytes); 568 if (target_offset == offset && 569 target_bytes == io_bytes) { 570 mirror_method = ret & BDRV_BLOCK_ZERO ? 571 MIRROR_METHOD_ZERO : 572 MIRROR_METHOD_DISCARD; 573 } 574 } 575 576 while (s->in_flight >= MAX_IN_FLIGHT) { 577 trace_mirror_yield_in_flight(s, offset, s->in_flight); 578 mirror_wait_for_free_in_flight_slot(s); 579 } 580 581 if (s->ret < 0) { 582 ret = 0; 583 goto fail; 584 } 585 586 io_bytes = mirror_clip_bytes(s, offset, io_bytes); 587 io_bytes = mirror_perform(s, offset, io_bytes, mirror_method); 588 if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) { 589 io_bytes_acct = 0; 590 } else { 591 io_bytes_acct = io_bytes; 592 } 593 assert(io_bytes); 594 offset += io_bytes; 595 nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity); 596 delay_ns = block_job_ratelimit_get_delay(&s->common, io_bytes_acct); 597 } 598 599 ret = delay_ns; 600 fail: 601 QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next); 602 qemu_co_queue_restart_all(&pseudo_op->waiting_requests); 603 g_free(pseudo_op); 604 605 return ret; 606 } 607 608 static void mirror_free_init(MirrorBlockJob *s) 609 { 610 int granularity = s->granularity; 611 size_t buf_size = s->buf_size; 612 uint8_t *buf = s->buf; 613 614 assert(s->buf_free_count == 0); 615 QSIMPLEQ_INIT(&s->buf_free); 616 while (buf_size != 0) { 617 MirrorBuffer *cur = (MirrorBuffer *)buf; 618 QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next); 619 s->buf_free_count++; 620 buf_size -= granularity; 621 buf += granularity; 622 } 623 } 624 625 /* This is also used for the .pause callback. There is no matching 626 * mirror_resume() because mirror_run() will begin iterating again 627 * when the job is resumed. 628 */ 629 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s) 630 { 631 while (s->in_flight > 0) { 632 mirror_wait_for_free_in_flight_slot(s); 633 } 634 } 635 636 /** 637 * mirror_exit_common: handle both abort() and prepare() cases. 638 * for .prepare, returns 0 on success and -errno on failure. 639 * for .abort cases, denoted by abort = true, MUST return 0. 640 */ 641 static int mirror_exit_common(Job *job) 642 { 643 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 644 BlockJob *bjob = &s->common; 645 MirrorBDSOpaque *bs_opaque; 646 AioContext *replace_aio_context = NULL; 647 BlockDriverState *src; 648 BlockDriverState *target_bs; 649 BlockDriverState *mirror_top_bs; 650 Error *local_err = NULL; 651 bool abort = job->ret < 0; 652 int ret = 0; 653 654 if (s->prepared) { 655 return 0; 656 } 657 s->prepared = true; 658 659 mirror_top_bs = s->mirror_top_bs; 660 bs_opaque = mirror_top_bs->opaque; 661 src = mirror_top_bs->backing->bs; 662 target_bs = blk_bs(s->target); 663 664 if (bdrv_chain_contains(src, target_bs)) { 665 bdrv_unfreeze_backing_chain(mirror_top_bs, target_bs); 666 } 667 668 bdrv_release_dirty_bitmap(s->dirty_bitmap); 669 670 /* Make sure that the source BDS doesn't go away during bdrv_replace_node, 671 * before we can call bdrv_drained_end */ 672 bdrv_ref(src); 673 bdrv_ref(mirror_top_bs); 674 bdrv_ref(target_bs); 675 676 /* 677 * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before 678 * inserting target_bs at s->to_replace, where we might not be able to get 679 * these permissions. 680 */ 681 blk_unref(s->target); 682 s->target = NULL; 683 684 /* We don't access the source any more. Dropping any WRITE/RESIZE is 685 * required before it could become a backing file of target_bs. Not having 686 * these permissions any more means that we can't allow any new requests on 687 * mirror_top_bs from now on, so keep it drained. */ 688 bdrv_drained_begin(mirror_top_bs); 689 bs_opaque->stop = true; 690 bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing, 691 &error_abort); 692 if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) { 693 BlockDriverState *backing = s->is_none_mode ? src : s->base; 694 BlockDriverState *unfiltered_target = bdrv_skip_filters(target_bs); 695 696 if (bdrv_cow_bs(unfiltered_target) != backing) { 697 bdrv_set_backing_hd(unfiltered_target, backing, &local_err); 698 if (local_err) { 699 error_report_err(local_err); 700 local_err = NULL; 701 ret = -EPERM; 702 } 703 } 704 } else if (!abort && s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) { 705 assert(!bdrv_backing_chain_next(target_bs)); 706 ret = bdrv_open_backing_file(bdrv_skip_filters(target_bs), NULL, 707 "backing", &local_err); 708 if (ret < 0) { 709 error_report_err(local_err); 710 local_err = NULL; 711 } 712 } 713 714 if (s->to_replace) { 715 replace_aio_context = bdrv_get_aio_context(s->to_replace); 716 aio_context_acquire(replace_aio_context); 717 } 718 719 if (s->should_complete && !abort) { 720 BlockDriverState *to_replace = s->to_replace ?: src; 721 bool ro = bdrv_is_read_only(to_replace); 722 723 if (ro != bdrv_is_read_only(target_bs)) { 724 bdrv_reopen_set_read_only(target_bs, ro, NULL); 725 } 726 727 /* The mirror job has no requests in flight any more, but we need to 728 * drain potential other users of the BDS before changing the graph. */ 729 assert(s->in_drain); 730 bdrv_drained_begin(target_bs); 731 /* 732 * Cannot use check_to_replace_node() here, because that would 733 * check for an op blocker on @to_replace, and we have our own 734 * there. 735 */ 736 if (bdrv_recurse_can_replace(src, to_replace)) { 737 bdrv_replace_node(to_replace, target_bs, &local_err); 738 } else { 739 error_setg(&local_err, "Can no longer replace '%s' by '%s', " 740 "because it can no longer be guaranteed that doing so " 741 "would not lead to an abrupt change of visible data", 742 to_replace->node_name, target_bs->node_name); 743 } 744 bdrv_drained_end(target_bs); 745 if (local_err) { 746 error_report_err(local_err); 747 ret = -EPERM; 748 } 749 } 750 if (s->to_replace) { 751 bdrv_op_unblock_all(s->to_replace, s->replace_blocker); 752 error_free(s->replace_blocker); 753 bdrv_unref(s->to_replace); 754 } 755 if (replace_aio_context) { 756 aio_context_release(replace_aio_context); 757 } 758 g_free(s->replaces); 759 bdrv_unref(target_bs); 760 761 /* 762 * Remove the mirror filter driver from the graph. Before this, get rid of 763 * the blockers on the intermediate nodes so that the resulting state is 764 * valid. 765 */ 766 block_job_remove_all_bdrv(bjob); 767 bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort); 768 769 /* We just changed the BDS the job BB refers to (with either or both of the 770 * bdrv_replace_node() calls), so switch the BB back so the cleanup does 771 * the right thing. We don't need any permissions any more now. */ 772 blk_remove_bs(bjob->blk); 773 blk_set_perm(bjob->blk, 0, BLK_PERM_ALL, &error_abort); 774 blk_insert_bs(bjob->blk, mirror_top_bs, &error_abort); 775 776 bs_opaque->job = NULL; 777 778 bdrv_drained_end(src); 779 bdrv_drained_end(mirror_top_bs); 780 s->in_drain = false; 781 bdrv_unref(mirror_top_bs); 782 bdrv_unref(src); 783 784 return ret; 785 } 786 787 static int mirror_prepare(Job *job) 788 { 789 return mirror_exit_common(job); 790 } 791 792 static void mirror_abort(Job *job) 793 { 794 int ret = mirror_exit_common(job); 795 assert(ret == 0); 796 } 797 798 static void coroutine_fn mirror_throttle(MirrorBlockJob *s) 799 { 800 int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); 801 802 if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) { 803 s->last_pause_ns = now; 804 job_sleep_ns(&s->common.job, 0); 805 } else { 806 job_pause_point(&s->common.job); 807 } 808 } 809 810 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s) 811 { 812 int64_t offset; 813 BlockDriverState *bs = s->mirror_top_bs->backing->bs; 814 BlockDriverState *target_bs = blk_bs(s->target); 815 int ret; 816 int64_t count; 817 818 if (s->zero_target) { 819 if (!bdrv_can_write_zeroes_with_unmap(target_bs)) { 820 bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length); 821 return 0; 822 } 823 824 s->initial_zeroing_ongoing = true; 825 for (offset = 0; offset < s->bdev_length; ) { 826 int bytes = MIN(s->bdev_length - offset, 827 QEMU_ALIGN_DOWN(INT_MAX, s->granularity)); 828 829 mirror_throttle(s); 830 831 if (job_is_cancelled(&s->common.job)) { 832 s->initial_zeroing_ongoing = false; 833 return 0; 834 } 835 836 if (s->in_flight >= MAX_IN_FLIGHT) { 837 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count, 838 s->in_flight); 839 mirror_wait_for_free_in_flight_slot(s); 840 continue; 841 } 842 843 mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO); 844 offset += bytes; 845 } 846 847 mirror_wait_for_all_io(s); 848 s->initial_zeroing_ongoing = false; 849 } 850 851 /* First part, loop on the sectors and initialize the dirty bitmap. */ 852 for (offset = 0; offset < s->bdev_length; ) { 853 /* Just to make sure we are not exceeding int limit. */ 854 int bytes = MIN(s->bdev_length - offset, 855 QEMU_ALIGN_DOWN(INT_MAX, s->granularity)); 856 857 mirror_throttle(s); 858 859 if (job_is_cancelled(&s->common.job)) { 860 return 0; 861 } 862 863 ret = bdrv_is_allocated_above(bs, s->base_overlay, true, offset, bytes, 864 &count); 865 if (ret < 0) { 866 return ret; 867 } 868 869 assert(count); 870 if (ret > 0) { 871 bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count); 872 } 873 offset += count; 874 } 875 return 0; 876 } 877 878 /* Called when going out of the streaming phase to flush the bulk of the 879 * data to the medium, or just before completing. 880 */ 881 static int mirror_flush(MirrorBlockJob *s) 882 { 883 int ret = blk_flush(s->target); 884 if (ret < 0) { 885 if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) { 886 s->ret = ret; 887 } 888 } 889 return ret; 890 } 891 892 static int coroutine_fn mirror_run(Job *job, Error **errp) 893 { 894 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 895 BlockDriverState *bs = s->mirror_top_bs->backing->bs; 896 BlockDriverState *target_bs = blk_bs(s->target); 897 bool need_drain = true; 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_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->synced = true; 941 s->actively_synced = true; 942 while (!job_is_cancelled(&s->common.job) && !s->should_complete) { 943 job_yield(&s->common.job); 944 } 945 s->common.job.cancelled = false; 946 goto immediate_exit; 947 } 948 949 length = DIV_ROUND_UP(s->bdev_length, s->granularity); 950 s->in_flight_bitmap = bitmap_new(length); 951 952 /* If we have no backing file yet in the destination, we cannot let 953 * the destination do COW. Instead, we copy sectors around the 954 * dirty data if needed. We need a bitmap to do that. 955 */ 956 bdrv_get_backing_filename(target_bs, backing_filename, 957 sizeof(backing_filename)); 958 if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) { 959 s->target_cluster_size = bdi.cluster_size; 960 } else { 961 s->target_cluster_size = BDRV_SECTOR_SIZE; 962 } 963 if (backing_filename[0] && !bdrv_backing_chain_next(target_bs) && 964 s->granularity < s->target_cluster_size) { 965 s->buf_size = MAX(s->buf_size, s->target_cluster_size); 966 s->cow_bitmap = bitmap_new(length); 967 } 968 s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov); 969 970 s->buf = qemu_try_blockalign(bs, s->buf_size); 971 if (s->buf == NULL) { 972 ret = -ENOMEM; 973 goto immediate_exit; 974 } 975 976 mirror_free_init(s); 977 978 s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); 979 if (!s->is_none_mode) { 980 ret = mirror_dirty_init(s); 981 if (ret < 0 || job_is_cancelled(&s->common.job)) { 982 goto immediate_exit; 983 } 984 } 985 986 assert(!s->dbi); 987 s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap); 988 for (;;) { 989 uint64_t delay_ns = 0; 990 int64_t cnt, delta; 991 bool should_complete; 992 993 /* Do not start passive operations while there are active 994 * writes in progress */ 995 while (s->in_active_write_counter) { 996 mirror_wait_for_any_operation(s, true); 997 } 998 999 if (s->ret < 0) { 1000 ret = s->ret; 1001 goto immediate_exit; 1002 } 1003 1004 job_pause_point(&s->common.job); 1005 1006 cnt = bdrv_get_dirty_count(s->dirty_bitmap); 1007 /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is 1008 * the number of bytes currently being processed; together those are 1009 * the current remaining operation length */ 1010 job_progress_set_remaining(&s->common.job, s->bytes_in_flight + cnt); 1011 1012 /* Note that even when no rate limit is applied we need to yield 1013 * periodically with no pending I/O so that bdrv_drain_all() returns. 1014 * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is 1015 * an error, or when the source is clean, whichever comes first. */ 1016 delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns; 1017 if (delta < BLOCK_JOB_SLICE_TIME && 1018 s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) { 1019 if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 || 1020 (cnt == 0 && s->in_flight > 0)) { 1021 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight); 1022 mirror_wait_for_free_in_flight_slot(s); 1023 continue; 1024 } else if (cnt != 0) { 1025 delay_ns = mirror_iteration(s); 1026 } 1027 } 1028 1029 should_complete = false; 1030 if (s->in_flight == 0 && cnt == 0) { 1031 trace_mirror_before_flush(s); 1032 if (!s->synced) { 1033 if (mirror_flush(s) < 0) { 1034 /* Go check s->ret. */ 1035 continue; 1036 } 1037 /* We're out of the streaming phase. From now on, if the job 1038 * is cancelled we will actually complete all pending I/O and 1039 * report completion. This way, block-job-cancel will leave 1040 * the target in a consistent state. 1041 */ 1042 job_transition_to_ready(&s->common.job); 1043 s->synced = true; 1044 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) { 1045 s->actively_synced = true; 1046 } 1047 } 1048 1049 should_complete = s->should_complete || 1050 job_is_cancelled(&s->common.job); 1051 cnt = bdrv_get_dirty_count(s->dirty_bitmap); 1052 } 1053 1054 if (cnt == 0 && should_complete) { 1055 /* The dirty bitmap is not updated while operations are pending. 1056 * If we're about to exit, wait for pending operations before 1057 * calling bdrv_get_dirty_count(bs), or we may exit while the 1058 * source has dirty data to copy! 1059 * 1060 * Note that I/O can be submitted by the guest while 1061 * mirror_populate runs, so pause it now. Before deciding 1062 * whether to switch to target check one last time if I/O has 1063 * come in the meanwhile, and if not flush the data to disk. 1064 */ 1065 trace_mirror_before_drain(s, cnt); 1066 1067 s->in_drain = true; 1068 bdrv_drained_begin(bs); 1069 cnt = bdrv_get_dirty_count(s->dirty_bitmap); 1070 if (cnt > 0 || mirror_flush(s) < 0) { 1071 bdrv_drained_end(bs); 1072 s->in_drain = false; 1073 continue; 1074 } 1075 1076 /* The two disks are in sync. Exit and report successful 1077 * completion. 1078 */ 1079 assert(QLIST_EMPTY(&bs->tracked_requests)); 1080 s->common.job.cancelled = false; 1081 need_drain = false; 1082 break; 1083 } 1084 1085 ret = 0; 1086 1087 if (s->synced && !should_complete) { 1088 delay_ns = (s->in_flight == 0 && 1089 cnt == 0 ? BLOCK_JOB_SLICE_TIME : 0); 1090 } 1091 trace_mirror_before_sleep(s, cnt, s->synced, delay_ns); 1092 job_sleep_ns(&s->common.job, delay_ns); 1093 if (job_is_cancelled(&s->common.job) && 1094 (!s->synced || s->common.job.force_cancel)) 1095 { 1096 break; 1097 } 1098 s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); 1099 } 1100 1101 immediate_exit: 1102 if (s->in_flight > 0) { 1103 /* We get here only if something went wrong. Either the job failed, 1104 * or it was cancelled prematurely so that we do not guarantee that 1105 * the target is a copy of the source. 1106 */ 1107 assert(ret < 0 || ((s->common.job.force_cancel || !s->synced) && 1108 job_is_cancelled(&s->common.job))); 1109 assert(need_drain); 1110 mirror_wait_for_all_io(s); 1111 } 1112 1113 assert(s->in_flight == 0); 1114 qemu_vfree(s->buf); 1115 g_free(s->cow_bitmap); 1116 g_free(s->in_flight_bitmap); 1117 bdrv_dirty_iter_free(s->dbi); 1118 1119 if (need_drain) { 1120 s->in_drain = true; 1121 bdrv_drained_begin(bs); 1122 } 1123 1124 return ret; 1125 } 1126 1127 static void mirror_complete(Job *job, Error **errp) 1128 { 1129 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 1130 1131 if (!s->synced) { 1132 error_setg(errp, "The active block job '%s' cannot be completed", 1133 job->id); 1134 return; 1135 } 1136 1137 /* block all operations on to_replace bs */ 1138 if (s->replaces) { 1139 AioContext *replace_aio_context; 1140 1141 s->to_replace = bdrv_find_node(s->replaces); 1142 if (!s->to_replace) { 1143 error_setg(errp, "Node name '%s' not found", s->replaces); 1144 return; 1145 } 1146 1147 replace_aio_context = bdrv_get_aio_context(s->to_replace); 1148 aio_context_acquire(replace_aio_context); 1149 1150 /* TODO Translate this into permission system. Current definition of 1151 * GRAPH_MOD would require to request it for the parents; they might 1152 * not even be BlockDriverStates, however, so a BdrvChild can't address 1153 * them. May need redefinition of GRAPH_MOD. */ 1154 error_setg(&s->replace_blocker, 1155 "block device is in use by block-job-complete"); 1156 bdrv_op_block_all(s->to_replace, s->replace_blocker); 1157 bdrv_ref(s->to_replace); 1158 1159 aio_context_release(replace_aio_context); 1160 } 1161 1162 s->should_complete = true; 1163 1164 /* If the job is paused, it will be re-entered when it is resumed */ 1165 if (!job->paused) { 1166 job_enter(job); 1167 } 1168 } 1169 1170 static void coroutine_fn mirror_pause(Job *job) 1171 { 1172 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 1173 1174 mirror_wait_for_all_io(s); 1175 } 1176 1177 static bool mirror_drained_poll(BlockJob *job) 1178 { 1179 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common); 1180 1181 /* If the job isn't paused nor cancelled, we can't be sure that it won't 1182 * issue more requests. We make an exception if we've reached this point 1183 * from one of our own drain sections, to avoid a deadlock waiting for 1184 * ourselves. 1185 */ 1186 if (!s->common.job.paused && !s->common.job.cancelled && !s->in_drain) { 1187 return true; 1188 } 1189 1190 return !!s->in_flight; 1191 } 1192 1193 static void mirror_cancel(Job *job, bool force) 1194 { 1195 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 1196 BlockDriverState *target = blk_bs(s->target); 1197 1198 if (force || !job_is_ready(job)) { 1199 bdrv_cancel_in_flight(target); 1200 } 1201 } 1202 1203 static const BlockJobDriver mirror_job_driver = { 1204 .job_driver = { 1205 .instance_size = sizeof(MirrorBlockJob), 1206 .job_type = JOB_TYPE_MIRROR, 1207 .free = block_job_free, 1208 .user_resume = block_job_user_resume, 1209 .run = mirror_run, 1210 .prepare = mirror_prepare, 1211 .abort = mirror_abort, 1212 .pause = mirror_pause, 1213 .complete = mirror_complete, 1214 .cancel = mirror_cancel, 1215 }, 1216 .drained_poll = mirror_drained_poll, 1217 }; 1218 1219 static const BlockJobDriver commit_active_job_driver = { 1220 .job_driver = { 1221 .instance_size = sizeof(MirrorBlockJob), 1222 .job_type = JOB_TYPE_COMMIT, 1223 .free = block_job_free, 1224 .user_resume = block_job_user_resume, 1225 .run = mirror_run, 1226 .prepare = mirror_prepare, 1227 .abort = mirror_abort, 1228 .pause = mirror_pause, 1229 .complete = mirror_complete, 1230 }, 1231 .drained_poll = mirror_drained_poll, 1232 }; 1233 1234 static void coroutine_fn 1235 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method, 1236 uint64_t offset, uint64_t bytes, 1237 QEMUIOVector *qiov, int flags) 1238 { 1239 int ret; 1240 size_t qiov_offset = 0; 1241 int64_t bitmap_offset, bitmap_end; 1242 1243 if (!QEMU_IS_ALIGNED(offset, job->granularity) && 1244 bdrv_dirty_bitmap_get(job->dirty_bitmap, offset)) 1245 { 1246 /* 1247 * Dirty unaligned padding: ignore it. 1248 * 1249 * Reasoning: 1250 * 1. If we copy it, we can't reset corresponding bit in 1251 * dirty_bitmap as there may be some "dirty" bytes still not 1252 * copied. 1253 * 2. It's already dirty, so skipping it we don't diverge mirror 1254 * progress. 1255 * 1256 * Note, that because of this, guest write may have no contribution 1257 * into mirror converge, but that's not bad, as we have background 1258 * process of mirroring. If under some bad circumstances (high guest 1259 * IO load) background process starve, we will not converge anyway, 1260 * even if each write will contribute, as guest is not guaranteed to 1261 * rewrite the whole disk. 1262 */ 1263 qiov_offset = QEMU_ALIGN_UP(offset, job->granularity) - offset; 1264 if (bytes <= qiov_offset) { 1265 /* nothing to do after shrink */ 1266 return; 1267 } 1268 offset += qiov_offset; 1269 bytes -= qiov_offset; 1270 } 1271 1272 if (!QEMU_IS_ALIGNED(offset + bytes, job->granularity) && 1273 bdrv_dirty_bitmap_get(job->dirty_bitmap, offset + bytes - 1)) 1274 { 1275 uint64_t tail = (offset + bytes) % job->granularity; 1276 1277 if (bytes <= tail) { 1278 /* nothing to do after shrink */ 1279 return; 1280 } 1281 bytes -= tail; 1282 } 1283 1284 /* 1285 * Tails are either clean or shrunk, so for bitmap resetting 1286 * we safely align the range down. 1287 */ 1288 bitmap_offset = QEMU_ALIGN_UP(offset, job->granularity); 1289 bitmap_end = QEMU_ALIGN_DOWN(offset + bytes, job->granularity); 1290 if (bitmap_offset < bitmap_end) { 1291 bdrv_reset_dirty_bitmap(job->dirty_bitmap, bitmap_offset, 1292 bitmap_end - bitmap_offset); 1293 } 1294 1295 job_progress_increase_remaining(&job->common.job, bytes); 1296 1297 switch (method) { 1298 case MIRROR_METHOD_COPY: 1299 ret = blk_co_pwritev_part(job->target, offset, bytes, 1300 qiov, qiov_offset, flags); 1301 break; 1302 1303 case MIRROR_METHOD_ZERO: 1304 assert(!qiov); 1305 ret = blk_co_pwrite_zeroes(job->target, offset, bytes, flags); 1306 break; 1307 1308 case MIRROR_METHOD_DISCARD: 1309 assert(!qiov); 1310 ret = blk_co_pdiscard(job->target, offset, bytes); 1311 break; 1312 1313 default: 1314 abort(); 1315 } 1316 1317 if (ret >= 0) { 1318 job_progress_update(&job->common.job, bytes); 1319 } else { 1320 BlockErrorAction action; 1321 1322 /* 1323 * We failed, so we should mark dirty the whole area, aligned up. 1324 * Note that we don't care about shrunk tails if any: they were dirty 1325 * at function start, and they must be still dirty, as we've locked 1326 * the region for in-flight op. 1327 */ 1328 bitmap_offset = QEMU_ALIGN_DOWN(offset, job->granularity); 1329 bitmap_end = QEMU_ALIGN_UP(offset + bytes, job->granularity); 1330 bdrv_set_dirty_bitmap(job->dirty_bitmap, bitmap_offset, 1331 bitmap_end - bitmap_offset); 1332 job->actively_synced = false; 1333 1334 action = mirror_error_action(job, false, -ret); 1335 if (action == BLOCK_ERROR_ACTION_REPORT) { 1336 if (!job->ret) { 1337 job->ret = ret; 1338 } 1339 } 1340 } 1341 } 1342 1343 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s, 1344 uint64_t offset, 1345 uint64_t bytes) 1346 { 1347 MirrorOp *op; 1348 uint64_t start_chunk = offset / s->granularity; 1349 uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity); 1350 1351 op = g_new(MirrorOp, 1); 1352 *op = (MirrorOp){ 1353 .s = s, 1354 .offset = offset, 1355 .bytes = bytes, 1356 .is_active_write = true, 1357 .is_in_flight = true, 1358 .co = qemu_coroutine_self(), 1359 }; 1360 qemu_co_queue_init(&op->waiting_requests); 1361 QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next); 1362 1363 s->in_active_write_counter++; 1364 1365 mirror_wait_on_conflicts(op, s, offset, bytes); 1366 1367 bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk); 1368 1369 return op; 1370 } 1371 1372 static void coroutine_fn active_write_settle(MirrorOp *op) 1373 { 1374 uint64_t start_chunk = op->offset / op->s->granularity; 1375 uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes, 1376 op->s->granularity); 1377 1378 if (!--op->s->in_active_write_counter && op->s->actively_synced) { 1379 BdrvChild *source = op->s->mirror_top_bs->backing; 1380 1381 if (QLIST_FIRST(&source->bs->parents) == source && 1382 QLIST_NEXT(source, next_parent) == NULL) 1383 { 1384 /* Assert that we are back in sync once all active write 1385 * operations are settled. 1386 * Note that we can only assert this if the mirror node 1387 * is the source node's only parent. */ 1388 assert(!bdrv_get_dirty_count(op->s->dirty_bitmap)); 1389 } 1390 } 1391 bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk); 1392 QTAILQ_REMOVE(&op->s->ops_in_flight, op, next); 1393 qemu_co_queue_restart_all(&op->waiting_requests); 1394 g_free(op); 1395 } 1396 1397 static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs, 1398 uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags) 1399 { 1400 return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags); 1401 } 1402 1403 static int coroutine_fn bdrv_mirror_top_do_write(BlockDriverState *bs, 1404 MirrorMethod method, uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, 1405 int flags) 1406 { 1407 MirrorOp *op = NULL; 1408 MirrorBDSOpaque *s = bs->opaque; 1409 int ret = 0; 1410 bool copy_to_target; 1411 1412 copy_to_target = s->job->ret >= 0 && 1413 s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING; 1414 1415 if (copy_to_target) { 1416 op = active_write_prepare(s->job, offset, bytes); 1417 } 1418 1419 switch (method) { 1420 case MIRROR_METHOD_COPY: 1421 ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags); 1422 break; 1423 1424 case MIRROR_METHOD_ZERO: 1425 ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags); 1426 break; 1427 1428 case MIRROR_METHOD_DISCARD: 1429 ret = bdrv_co_pdiscard(bs->backing, offset, bytes); 1430 break; 1431 1432 default: 1433 abort(); 1434 } 1435 1436 if (ret < 0) { 1437 goto out; 1438 } 1439 1440 if (copy_to_target) { 1441 do_sync_target_write(s->job, method, offset, bytes, qiov, flags); 1442 } 1443 1444 out: 1445 if (copy_to_target) { 1446 active_write_settle(op); 1447 } 1448 return ret; 1449 } 1450 1451 static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs, 1452 uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags) 1453 { 1454 MirrorBDSOpaque *s = bs->opaque; 1455 QEMUIOVector bounce_qiov; 1456 void *bounce_buf; 1457 int ret = 0; 1458 bool copy_to_target; 1459 1460 copy_to_target = s->job->ret >= 0 && 1461 s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING; 1462 1463 if (copy_to_target) { 1464 /* The guest might concurrently modify the data to write; but 1465 * the data on source and destination must match, so we have 1466 * to use a bounce buffer if we are going to write to the 1467 * target now. */ 1468 bounce_buf = qemu_blockalign(bs, bytes); 1469 iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes); 1470 1471 qemu_iovec_init(&bounce_qiov, 1); 1472 qemu_iovec_add(&bounce_qiov, bounce_buf, bytes); 1473 qiov = &bounce_qiov; 1474 } 1475 1476 ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov, 1477 flags); 1478 1479 if (copy_to_target) { 1480 qemu_iovec_destroy(&bounce_qiov); 1481 qemu_vfree(bounce_buf); 1482 } 1483 1484 return ret; 1485 } 1486 1487 static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs) 1488 { 1489 if (bs->backing == NULL) { 1490 /* we can be here after failed bdrv_append in mirror_start_job */ 1491 return 0; 1492 } 1493 return bdrv_co_flush(bs->backing->bs); 1494 } 1495 1496 static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs, 1497 int64_t offset, int bytes, BdrvRequestFlags flags) 1498 { 1499 return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL, 1500 flags); 1501 } 1502 1503 static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs, 1504 int64_t offset, int bytes) 1505 { 1506 return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes, 1507 NULL, 0); 1508 } 1509 1510 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs) 1511 { 1512 if (bs->backing == NULL) { 1513 /* we can be here after failed bdrv_attach_child in 1514 * bdrv_set_backing_hd */ 1515 return; 1516 } 1517 pstrcpy(bs->exact_filename, sizeof(bs->exact_filename), 1518 bs->backing->bs->filename); 1519 } 1520 1521 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c, 1522 BdrvChildRole role, 1523 BlockReopenQueue *reopen_queue, 1524 uint64_t perm, uint64_t shared, 1525 uint64_t *nperm, uint64_t *nshared) 1526 { 1527 MirrorBDSOpaque *s = bs->opaque; 1528 1529 if (s->stop) { 1530 /* 1531 * If the job is to be stopped, we do not need to forward 1532 * anything to the real image. 1533 */ 1534 *nperm = 0; 1535 *nshared = BLK_PERM_ALL; 1536 return; 1537 } 1538 1539 bdrv_default_perms(bs, c, role, reopen_queue, 1540 perm, shared, nperm, nshared); 1541 1542 if (s->is_commit) { 1543 /* 1544 * For commit jobs, we cannot take CONSISTENT_READ, because 1545 * that permission is unshared for everything above the base 1546 * node (except for filters on the base node). 1547 * We also have to force-share the WRITE permission, or 1548 * otherwise we would block ourselves at the base node (if 1549 * writes are blocked for a node, they are also blocked for 1550 * its backing file). 1551 * (We could also share RESIZE, because it may be needed for 1552 * the target if its size is less than the top node's; but 1553 * bdrv_default_perms_for_cow() automatically shares RESIZE 1554 * for backing nodes if WRITE is shared, so there is no need 1555 * to do it here.) 1556 */ 1557 *nperm &= ~BLK_PERM_CONSISTENT_READ; 1558 *nshared |= BLK_PERM_WRITE; 1559 } 1560 } 1561 1562 /* Dummy node that provides consistent read to its users without requiring it 1563 * from its backing file and that allows writes on the backing file chain. */ 1564 static BlockDriver bdrv_mirror_top = { 1565 .format_name = "mirror_top", 1566 .bdrv_co_preadv = bdrv_mirror_top_preadv, 1567 .bdrv_co_pwritev = bdrv_mirror_top_pwritev, 1568 .bdrv_co_pwrite_zeroes = bdrv_mirror_top_pwrite_zeroes, 1569 .bdrv_co_pdiscard = bdrv_mirror_top_pdiscard, 1570 .bdrv_co_flush = bdrv_mirror_top_flush, 1571 .bdrv_refresh_filename = bdrv_mirror_top_refresh_filename, 1572 .bdrv_child_perm = bdrv_mirror_top_child_perm, 1573 1574 .is_filter = true, 1575 }; 1576 1577 static BlockJob *mirror_start_job( 1578 const char *job_id, BlockDriverState *bs, 1579 int creation_flags, BlockDriverState *target, 1580 const char *replaces, int64_t speed, 1581 uint32_t granularity, int64_t buf_size, 1582 BlockMirrorBackingMode backing_mode, 1583 bool zero_target, 1584 BlockdevOnError on_source_error, 1585 BlockdevOnError on_target_error, 1586 bool unmap, 1587 BlockCompletionFunc *cb, 1588 void *opaque, 1589 const BlockJobDriver *driver, 1590 bool is_none_mode, BlockDriverState *base, 1591 bool auto_complete, const char *filter_node_name, 1592 bool is_mirror, MirrorCopyMode copy_mode, 1593 Error **errp) 1594 { 1595 MirrorBlockJob *s; 1596 MirrorBDSOpaque *bs_opaque; 1597 BlockDriverState *mirror_top_bs; 1598 bool target_is_backing; 1599 uint64_t target_perms, target_shared_perms; 1600 int ret; 1601 1602 if (granularity == 0) { 1603 granularity = bdrv_get_default_bitmap_granularity(target); 1604 } 1605 1606 assert(is_power_of_2(granularity)); 1607 1608 if (buf_size < 0) { 1609 error_setg(errp, "Invalid parameter 'buf-size'"); 1610 return NULL; 1611 } 1612 1613 if (buf_size == 0) { 1614 buf_size = DEFAULT_MIRROR_BUF_SIZE; 1615 } 1616 1617 if (bdrv_skip_filters(bs) == bdrv_skip_filters(target)) { 1618 error_setg(errp, "Can't mirror node into itself"); 1619 return NULL; 1620 } 1621 1622 target_is_backing = bdrv_chain_contains(bs, target); 1623 1624 /* In the case of active commit, add dummy driver to provide consistent 1625 * reads on the top, while disabling it in the intermediate nodes, and make 1626 * the backing chain writable. */ 1627 mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name, 1628 BDRV_O_RDWR, errp); 1629 if (mirror_top_bs == NULL) { 1630 return NULL; 1631 } 1632 if (!filter_node_name) { 1633 mirror_top_bs->implicit = true; 1634 } 1635 1636 /* So that we can always drop this node */ 1637 mirror_top_bs->never_freeze = true; 1638 1639 mirror_top_bs->total_sectors = bs->total_sectors; 1640 mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED; 1641 mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED | 1642 BDRV_REQ_NO_FALLBACK; 1643 bs_opaque = g_new0(MirrorBDSOpaque, 1); 1644 mirror_top_bs->opaque = bs_opaque; 1645 1646 bs_opaque->is_commit = target_is_backing; 1647 1648 bdrv_drained_begin(bs); 1649 ret = bdrv_append(mirror_top_bs, bs, errp); 1650 bdrv_drained_end(bs); 1651 1652 if (ret < 0) { 1653 bdrv_unref(mirror_top_bs); 1654 return NULL; 1655 } 1656 1657 /* Make sure that the source is not resized while the job is running */ 1658 s = block_job_create(job_id, driver, NULL, mirror_top_bs, 1659 BLK_PERM_CONSISTENT_READ, 1660 BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED | 1661 BLK_PERM_WRITE | BLK_PERM_GRAPH_MOD, speed, 1662 creation_flags, cb, opaque, errp); 1663 if (!s) { 1664 goto fail; 1665 } 1666 bs_opaque->job = s; 1667 1668 /* The block job now has a reference to this node */ 1669 bdrv_unref(mirror_top_bs); 1670 1671 s->mirror_top_bs = mirror_top_bs; 1672 1673 /* No resize for the target either; while the mirror is still running, a 1674 * consistent read isn't necessarily possible. We could possibly allow 1675 * writes and graph modifications, though it would likely defeat the 1676 * purpose of a mirror, so leave them blocked for now. 1677 * 1678 * In the case of active commit, things look a bit different, though, 1679 * because the target is an already populated backing file in active use. 1680 * We can allow anything except resize there.*/ 1681 1682 target_perms = BLK_PERM_WRITE; 1683 target_shared_perms = BLK_PERM_WRITE_UNCHANGED; 1684 1685 if (target_is_backing) { 1686 int64_t bs_size, target_size; 1687 bs_size = bdrv_getlength(bs); 1688 if (bs_size < 0) { 1689 error_setg_errno(errp, -bs_size, 1690 "Could not inquire top image size"); 1691 goto fail; 1692 } 1693 1694 target_size = bdrv_getlength(target); 1695 if (target_size < 0) { 1696 error_setg_errno(errp, -target_size, 1697 "Could not inquire base image size"); 1698 goto fail; 1699 } 1700 1701 if (target_size < bs_size) { 1702 target_perms |= BLK_PERM_RESIZE; 1703 } 1704 1705 target_shared_perms |= BLK_PERM_CONSISTENT_READ 1706 | BLK_PERM_WRITE 1707 | BLK_PERM_GRAPH_MOD; 1708 } else if (bdrv_chain_contains(bs, bdrv_skip_filters(target))) { 1709 /* 1710 * We may want to allow this in the future, but it would 1711 * require taking some extra care. 1712 */ 1713 error_setg(errp, "Cannot mirror to a filter on top of a node in the " 1714 "source's backing chain"); 1715 goto fail; 1716 } 1717 1718 if (backing_mode != MIRROR_LEAVE_BACKING_CHAIN) { 1719 target_perms |= BLK_PERM_GRAPH_MOD; 1720 } 1721 1722 s->target = blk_new(s->common.job.aio_context, 1723 target_perms, target_shared_perms); 1724 ret = blk_insert_bs(s->target, target, errp); 1725 if (ret < 0) { 1726 goto fail; 1727 } 1728 if (is_mirror) { 1729 /* XXX: Mirror target could be a NBD server of target QEMU in the case 1730 * of non-shared block migration. To allow migration completion, we 1731 * have to allow "inactivate" of the target BB. When that happens, we 1732 * know the job is drained, and the vcpus are stopped, so no write 1733 * operation will be performed. Block layer already has assertions to 1734 * ensure that. */ 1735 blk_set_force_allow_inactivate(s->target); 1736 } 1737 blk_set_allow_aio_context_change(s->target, true); 1738 blk_set_disable_request_queuing(s->target, true); 1739 1740 s->replaces = g_strdup(replaces); 1741 s->on_source_error = on_source_error; 1742 s->on_target_error = on_target_error; 1743 s->is_none_mode = is_none_mode; 1744 s->backing_mode = backing_mode; 1745 s->zero_target = zero_target; 1746 s->copy_mode = copy_mode; 1747 s->base = base; 1748 s->base_overlay = bdrv_find_overlay(bs, base); 1749 s->granularity = granularity; 1750 s->buf_size = ROUND_UP(buf_size, granularity); 1751 s->unmap = unmap; 1752 if (auto_complete) { 1753 s->should_complete = true; 1754 } 1755 1756 s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp); 1757 if (!s->dirty_bitmap) { 1758 goto fail; 1759 } 1760 if (s->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING) { 1761 bdrv_disable_dirty_bitmap(s->dirty_bitmap); 1762 } 1763 1764 ret = block_job_add_bdrv(&s->common, "source", bs, 0, 1765 BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE | 1766 BLK_PERM_CONSISTENT_READ, 1767 errp); 1768 if (ret < 0) { 1769 goto fail; 1770 } 1771 1772 /* Required permissions are already taken with blk_new() */ 1773 block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL, 1774 &error_abort); 1775 1776 /* In commit_active_start() all intermediate nodes disappear, so 1777 * any jobs in them must be blocked */ 1778 if (target_is_backing) { 1779 BlockDriverState *iter, *filtered_target; 1780 uint64_t iter_shared_perms; 1781 1782 /* 1783 * The topmost node with 1784 * bdrv_skip_filters(filtered_target) == bdrv_skip_filters(target) 1785 */ 1786 filtered_target = bdrv_cow_bs(bdrv_find_overlay(bs, target)); 1787 1788 assert(bdrv_skip_filters(filtered_target) == 1789 bdrv_skip_filters(target)); 1790 1791 /* 1792 * XXX BLK_PERM_WRITE needs to be allowed so we don't block 1793 * ourselves at s->base (if writes are blocked for a node, they are 1794 * also blocked for its backing file). The other options would be a 1795 * second filter driver above s->base (== target). 1796 */ 1797 iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE; 1798 1799 for (iter = bdrv_filter_or_cow_bs(bs); iter != target; 1800 iter = bdrv_filter_or_cow_bs(iter)) 1801 { 1802 if (iter == filtered_target) { 1803 /* 1804 * From here on, all nodes are filters on the base. 1805 * This allows us to share BLK_PERM_CONSISTENT_READ. 1806 */ 1807 iter_shared_perms |= BLK_PERM_CONSISTENT_READ; 1808 } 1809 1810 ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0, 1811 iter_shared_perms, errp); 1812 if (ret < 0) { 1813 goto fail; 1814 } 1815 } 1816 1817 if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) { 1818 goto fail; 1819 } 1820 } 1821 1822 QTAILQ_INIT(&s->ops_in_flight); 1823 1824 trace_mirror_start(bs, s, opaque); 1825 job_start(&s->common.job); 1826 1827 return &s->common; 1828 1829 fail: 1830 if (s) { 1831 /* Make sure this BDS does not go away until we have completed the graph 1832 * changes below */ 1833 bdrv_ref(mirror_top_bs); 1834 1835 g_free(s->replaces); 1836 blk_unref(s->target); 1837 bs_opaque->job = NULL; 1838 if (s->dirty_bitmap) { 1839 bdrv_release_dirty_bitmap(s->dirty_bitmap); 1840 } 1841 job_early_fail(&s->common.job); 1842 } 1843 1844 bs_opaque->stop = true; 1845 bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing, 1846 &error_abort); 1847 bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort); 1848 1849 bdrv_unref(mirror_top_bs); 1850 1851 return NULL; 1852 } 1853 1854 void mirror_start(const char *job_id, BlockDriverState *bs, 1855 BlockDriverState *target, const char *replaces, 1856 int creation_flags, int64_t speed, 1857 uint32_t granularity, int64_t buf_size, 1858 MirrorSyncMode mode, BlockMirrorBackingMode backing_mode, 1859 bool zero_target, 1860 BlockdevOnError on_source_error, 1861 BlockdevOnError on_target_error, 1862 bool unmap, const char *filter_node_name, 1863 MirrorCopyMode copy_mode, Error **errp) 1864 { 1865 bool is_none_mode; 1866 BlockDriverState *base; 1867 1868 if ((mode == MIRROR_SYNC_MODE_INCREMENTAL) || 1869 (mode == MIRROR_SYNC_MODE_BITMAP)) { 1870 error_setg(errp, "Sync mode '%s' not supported", 1871 MirrorSyncMode_str(mode)); 1872 return; 1873 } 1874 is_none_mode = mode == MIRROR_SYNC_MODE_NONE; 1875 base = mode == MIRROR_SYNC_MODE_TOP ? bdrv_backing_chain_next(bs) : NULL; 1876 mirror_start_job(job_id, bs, creation_flags, target, replaces, 1877 speed, granularity, buf_size, backing_mode, zero_target, 1878 on_source_error, on_target_error, unmap, NULL, NULL, 1879 &mirror_job_driver, is_none_mode, base, false, 1880 filter_node_name, true, copy_mode, errp); 1881 } 1882 1883 BlockJob *commit_active_start(const char *job_id, BlockDriverState *bs, 1884 BlockDriverState *base, int creation_flags, 1885 int64_t speed, BlockdevOnError on_error, 1886 const char *filter_node_name, 1887 BlockCompletionFunc *cb, void *opaque, 1888 bool auto_complete, Error **errp) 1889 { 1890 bool base_read_only; 1891 BlockJob *job; 1892 1893 base_read_only = bdrv_is_read_only(base); 1894 1895 if (base_read_only) { 1896 if (bdrv_reopen_set_read_only(base, false, errp) < 0) { 1897 return NULL; 1898 } 1899 } 1900 1901 job = mirror_start_job( 1902 job_id, bs, creation_flags, base, NULL, speed, 0, 0, 1903 MIRROR_LEAVE_BACKING_CHAIN, false, 1904 on_error, on_error, true, cb, opaque, 1905 &commit_active_job_driver, false, base, auto_complete, 1906 filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND, 1907 errp); 1908 if (!job) { 1909 goto error_restore_flags; 1910 } 1911 1912 return job; 1913 1914 error_restore_flags: 1915 /* ignore error and errp for bdrv_reopen, because we want to propagate 1916 * the original error */ 1917 if (base_read_only) { 1918 bdrv_reopen_set_read_only(base, true, NULL); 1919 } 1920 return NULL; 1921 } 1922