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 /* Set when the target is synced (dirty bitmap is clean, nothing 60 * in flight) and the job is running in active mode */ 61 bool actively_synced; 62 bool should_complete; 63 int64_t granularity; 64 size_t buf_size; 65 int64_t bdev_length; 66 unsigned long *cow_bitmap; 67 BdrvDirtyBitmap *dirty_bitmap; 68 BdrvDirtyBitmapIter *dbi; 69 uint8_t *buf; 70 QSIMPLEQ_HEAD(, MirrorBuffer) buf_free; 71 int buf_free_count; 72 73 uint64_t last_pause_ns; 74 unsigned long *in_flight_bitmap; 75 int in_flight; 76 int64_t bytes_in_flight; 77 QTAILQ_HEAD(, MirrorOp) ops_in_flight; 78 int ret; 79 bool unmap; 80 int target_cluster_size; 81 int max_iov; 82 bool initial_zeroing_ongoing; 83 int in_active_write_counter; 84 bool prepared; 85 bool in_drain; 86 } MirrorBlockJob; 87 88 typedef struct MirrorBDSOpaque { 89 MirrorBlockJob *job; 90 bool stop; 91 bool is_commit; 92 } MirrorBDSOpaque; 93 94 struct MirrorOp { 95 MirrorBlockJob *s; 96 QEMUIOVector qiov; 97 int64_t offset; 98 uint64_t bytes; 99 100 /* The pointee is set by mirror_co_read(), mirror_co_zero(), and 101 * mirror_co_discard() before yielding for the first time */ 102 int64_t *bytes_handled; 103 104 bool is_pseudo_op; 105 bool is_active_write; 106 bool is_in_flight; 107 CoQueue waiting_requests; 108 Coroutine *co; 109 MirrorOp *waiting_for_op; 110 111 QTAILQ_ENTRY(MirrorOp) next; 112 }; 113 114 typedef enum MirrorMethod { 115 MIRROR_METHOD_COPY, 116 MIRROR_METHOD_ZERO, 117 MIRROR_METHOD_DISCARD, 118 } MirrorMethod; 119 120 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read, 121 int error) 122 { 123 s->actively_synced = false; 124 if (read) { 125 return block_job_error_action(&s->common, s->on_source_error, 126 true, error); 127 } else { 128 return block_job_error_action(&s->common, s->on_target_error, 129 false, error); 130 } 131 } 132 133 static void coroutine_fn mirror_wait_on_conflicts(MirrorOp *self, 134 MirrorBlockJob *s, 135 uint64_t offset, 136 uint64_t bytes) 137 { 138 uint64_t self_start_chunk = offset / s->granularity; 139 uint64_t self_end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity); 140 uint64_t self_nb_chunks = self_end_chunk - self_start_chunk; 141 142 while (find_next_bit(s->in_flight_bitmap, self_end_chunk, 143 self_start_chunk) < self_end_chunk && 144 s->ret >= 0) 145 { 146 MirrorOp *op; 147 148 QTAILQ_FOREACH(op, &s->ops_in_flight, next) { 149 uint64_t op_start_chunk = op->offset / s->granularity; 150 uint64_t op_nb_chunks = DIV_ROUND_UP(op->offset + op->bytes, 151 s->granularity) - 152 op_start_chunk; 153 154 if (op == self) { 155 continue; 156 } 157 158 if (ranges_overlap(self_start_chunk, self_nb_chunks, 159 op_start_chunk, op_nb_chunks)) 160 { 161 if (self) { 162 /* 163 * If the operation is already (indirectly) waiting for us, 164 * or will wait for us as soon as it wakes up, then just go 165 * on (instead of producing a deadlock in the former case). 166 */ 167 if (op->waiting_for_op) { 168 continue; 169 } 170 171 self->waiting_for_op = op; 172 } 173 174 qemu_co_queue_wait(&op->waiting_requests, NULL); 175 176 if (self) { 177 self->waiting_for_op = NULL; 178 } 179 180 break; 181 } 182 } 183 } 184 } 185 186 static void coroutine_fn mirror_iteration_done(MirrorOp *op, int ret) 187 { 188 MirrorBlockJob *s = op->s; 189 struct iovec *iov; 190 int64_t chunk_num; 191 int i, nb_chunks; 192 193 trace_mirror_iteration_done(s, op->offset, op->bytes, ret); 194 195 s->in_flight--; 196 s->bytes_in_flight -= op->bytes; 197 iov = op->qiov.iov; 198 for (i = 0; i < op->qiov.niov; i++) { 199 MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base; 200 QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next); 201 s->buf_free_count++; 202 } 203 204 chunk_num = op->offset / s->granularity; 205 nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity); 206 207 bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks); 208 QTAILQ_REMOVE(&s->ops_in_flight, op, next); 209 if (ret >= 0) { 210 if (s->cow_bitmap) { 211 bitmap_set(s->cow_bitmap, chunk_num, nb_chunks); 212 } 213 if (!s->initial_zeroing_ongoing) { 214 job_progress_update(&s->common.job, op->bytes); 215 } 216 } 217 qemu_iovec_destroy(&op->qiov); 218 219 qemu_co_queue_restart_all(&op->waiting_requests); 220 g_free(op); 221 } 222 223 static void coroutine_fn mirror_write_complete(MirrorOp *op, int ret) 224 { 225 MirrorBlockJob *s = op->s; 226 227 if (ret < 0) { 228 BlockErrorAction action; 229 230 bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes); 231 action = mirror_error_action(s, false, -ret); 232 if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) { 233 s->ret = ret; 234 } 235 } 236 237 mirror_iteration_done(op, ret); 238 } 239 240 static void coroutine_fn mirror_read_complete(MirrorOp *op, int ret) 241 { 242 MirrorBlockJob *s = op->s; 243 244 if (ret < 0) { 245 BlockErrorAction action; 246 247 bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes); 248 action = mirror_error_action(s, true, -ret); 249 if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) { 250 s->ret = ret; 251 } 252 253 mirror_iteration_done(op, ret); 254 return; 255 } 256 257 ret = blk_co_pwritev(s->target, op->offset, op->qiov.size, &op->qiov, 0); 258 mirror_write_complete(op, ret); 259 } 260 261 /* Clip bytes relative to offset to not exceed end-of-file */ 262 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s, 263 int64_t offset, 264 int64_t bytes) 265 { 266 return MIN(bytes, s->bdev_length - offset); 267 } 268 269 /* Round offset and/or bytes to target cluster if COW is needed, and 270 * return the offset of the adjusted tail against original. */ 271 static int mirror_cow_align(MirrorBlockJob *s, int64_t *offset, 272 uint64_t *bytes) 273 { 274 bool need_cow; 275 int ret = 0; 276 int64_t align_offset = *offset; 277 int64_t align_bytes = *bytes; 278 int max_bytes = s->granularity * s->max_iov; 279 280 need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap); 281 need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity, 282 s->cow_bitmap); 283 if (need_cow) { 284 bdrv_round_to_clusters(blk_bs(s->target), *offset, *bytes, 285 &align_offset, &align_bytes); 286 } 287 288 if (align_bytes > max_bytes) { 289 align_bytes = max_bytes; 290 if (need_cow) { 291 align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size); 292 } 293 } 294 /* Clipping may result in align_bytes unaligned to chunk boundary, but 295 * that doesn't matter because it's already the end of source image. */ 296 align_bytes = mirror_clip_bytes(s, align_offset, align_bytes); 297 298 ret = align_offset + align_bytes - (*offset + *bytes); 299 *offset = align_offset; 300 *bytes = align_bytes; 301 assert(ret >= 0); 302 return ret; 303 } 304 305 static inline void coroutine_fn 306 mirror_wait_for_any_operation(MirrorBlockJob *s, bool active) 307 { 308 MirrorOp *op; 309 310 QTAILQ_FOREACH(op, &s->ops_in_flight, next) { 311 /* Do not wait on pseudo ops, because it may in turn wait on 312 * some other operation to start, which may in fact be the 313 * caller of this function. Since there is only one pseudo op 314 * at any given time, we will always find some real operation 315 * to wait on. */ 316 if (!op->is_pseudo_op && op->is_in_flight && 317 op->is_active_write == active) 318 { 319 qemu_co_queue_wait(&op->waiting_requests, NULL); 320 return; 321 } 322 } 323 abort(); 324 } 325 326 static inline void coroutine_fn 327 mirror_wait_for_free_in_flight_slot(MirrorBlockJob *s) 328 { 329 /* Only non-active operations use up in-flight slots */ 330 mirror_wait_for_any_operation(s, false); 331 } 332 333 /* Perform a mirror copy operation. 334 * 335 * *op->bytes_handled is set to the number of bytes copied after and 336 * including offset, excluding any bytes copied prior to offset due 337 * to alignment. This will be op->bytes if no alignment is necessary, 338 * or (new_end - op->offset) if the tail is rounded up or down due to 339 * alignment or buffer limit. 340 */ 341 static void coroutine_fn mirror_co_read(void *opaque) 342 { 343 MirrorOp *op = opaque; 344 MirrorBlockJob *s = op->s; 345 int nb_chunks; 346 uint64_t ret; 347 uint64_t max_bytes; 348 349 max_bytes = s->granularity * s->max_iov; 350 351 /* We can only handle as much as buf_size at a time. */ 352 op->bytes = MIN(s->buf_size, MIN(max_bytes, op->bytes)); 353 assert(op->bytes); 354 assert(op->bytes < BDRV_REQUEST_MAX_BYTES); 355 *op->bytes_handled = op->bytes; 356 357 if (s->cow_bitmap) { 358 *op->bytes_handled += mirror_cow_align(s, &op->offset, &op->bytes); 359 } 360 /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */ 361 assert(*op->bytes_handled <= UINT_MAX); 362 assert(op->bytes <= s->buf_size); 363 /* The offset is granularity-aligned because: 364 * 1) Caller passes in aligned values; 365 * 2) mirror_cow_align is used only when target cluster is larger. */ 366 assert(QEMU_IS_ALIGNED(op->offset, s->granularity)); 367 /* The range is sector-aligned, since bdrv_getlength() rounds up. */ 368 assert(QEMU_IS_ALIGNED(op->bytes, BDRV_SECTOR_SIZE)); 369 nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity); 370 371 while (s->buf_free_count < nb_chunks) { 372 trace_mirror_yield_in_flight(s, op->offset, s->in_flight); 373 mirror_wait_for_free_in_flight_slot(s); 374 } 375 376 /* Now make a QEMUIOVector taking enough granularity-sized chunks 377 * from s->buf_free. 378 */ 379 qemu_iovec_init(&op->qiov, nb_chunks); 380 while (nb_chunks-- > 0) { 381 MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free); 382 size_t remaining = op->bytes - op->qiov.size; 383 384 QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next); 385 s->buf_free_count--; 386 qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining)); 387 } 388 389 /* Copy the dirty cluster. */ 390 s->in_flight++; 391 s->bytes_in_flight += op->bytes; 392 op->is_in_flight = true; 393 trace_mirror_one_iteration(s, op->offset, op->bytes); 394 395 ret = bdrv_co_preadv(s->mirror_top_bs->backing, op->offset, op->bytes, 396 &op->qiov, 0); 397 mirror_read_complete(op, ret); 398 } 399 400 static void coroutine_fn mirror_co_zero(void *opaque) 401 { 402 MirrorOp *op = opaque; 403 int ret; 404 405 op->s->in_flight++; 406 op->s->bytes_in_flight += op->bytes; 407 *op->bytes_handled = op->bytes; 408 op->is_in_flight = true; 409 410 ret = blk_co_pwrite_zeroes(op->s->target, op->offset, op->bytes, 411 op->s->unmap ? BDRV_REQ_MAY_UNMAP : 0); 412 mirror_write_complete(op, ret); 413 } 414 415 static void coroutine_fn mirror_co_discard(void *opaque) 416 { 417 MirrorOp *op = opaque; 418 int ret; 419 420 op->s->in_flight++; 421 op->s->bytes_in_flight += op->bytes; 422 *op->bytes_handled = op->bytes; 423 op->is_in_flight = true; 424 425 ret = blk_co_pdiscard(op->s->target, op->offset, op->bytes); 426 mirror_write_complete(op, ret); 427 } 428 429 static unsigned mirror_perform(MirrorBlockJob *s, int64_t offset, 430 unsigned bytes, MirrorMethod mirror_method) 431 { 432 MirrorOp *op; 433 Coroutine *co; 434 int64_t bytes_handled = -1; 435 436 op = g_new(MirrorOp, 1); 437 *op = (MirrorOp){ 438 .s = s, 439 .offset = offset, 440 .bytes = bytes, 441 .bytes_handled = &bytes_handled, 442 }; 443 qemu_co_queue_init(&op->waiting_requests); 444 445 switch (mirror_method) { 446 case MIRROR_METHOD_COPY: 447 co = qemu_coroutine_create(mirror_co_read, op); 448 break; 449 case MIRROR_METHOD_ZERO: 450 co = qemu_coroutine_create(mirror_co_zero, op); 451 break; 452 case MIRROR_METHOD_DISCARD: 453 co = qemu_coroutine_create(mirror_co_discard, op); 454 break; 455 default: 456 abort(); 457 } 458 op->co = co; 459 460 QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next); 461 qemu_coroutine_enter(co); 462 /* At this point, ownership of op has been moved to the coroutine 463 * and the object may already be freed */ 464 465 /* Assert that this value has been set */ 466 assert(bytes_handled >= 0); 467 468 /* Same assertion as in mirror_co_read() (and for mirror_co_read() 469 * and mirror_co_discard(), bytes_handled == op->bytes, which 470 * is the @bytes parameter given to this function) */ 471 assert(bytes_handled <= UINT_MAX); 472 return bytes_handled; 473 } 474 475 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s) 476 { 477 BlockDriverState *source = s->mirror_top_bs->backing->bs; 478 MirrorOp *pseudo_op; 479 int64_t offset; 480 uint64_t delay_ns = 0, ret = 0; 481 /* At least the first dirty chunk is mirrored in one iteration. */ 482 int nb_chunks = 1; 483 bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target)); 484 int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES); 485 486 bdrv_dirty_bitmap_lock(s->dirty_bitmap); 487 offset = bdrv_dirty_iter_next(s->dbi); 488 if (offset < 0) { 489 bdrv_set_dirty_iter(s->dbi, 0); 490 offset = bdrv_dirty_iter_next(s->dbi); 491 trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap)); 492 assert(offset >= 0); 493 } 494 bdrv_dirty_bitmap_unlock(s->dirty_bitmap); 495 496 mirror_wait_on_conflicts(NULL, s, offset, 1); 497 498 job_pause_point(&s->common.job); 499 500 /* Find the number of consective dirty chunks following the first dirty 501 * one, and wait for in flight requests in them. */ 502 bdrv_dirty_bitmap_lock(s->dirty_bitmap); 503 while (nb_chunks * s->granularity < s->buf_size) { 504 int64_t next_dirty; 505 int64_t next_offset = offset + nb_chunks * s->granularity; 506 int64_t next_chunk = next_offset / s->granularity; 507 if (next_offset >= s->bdev_length || 508 !bdrv_dirty_bitmap_get_locked(s->dirty_bitmap, next_offset)) { 509 break; 510 } 511 if (test_bit(next_chunk, s->in_flight_bitmap)) { 512 break; 513 } 514 515 next_dirty = bdrv_dirty_iter_next(s->dbi); 516 if (next_dirty > next_offset || next_dirty < 0) { 517 /* The bitmap iterator's cache is stale, refresh it */ 518 bdrv_set_dirty_iter(s->dbi, next_offset); 519 next_dirty = bdrv_dirty_iter_next(s->dbi); 520 } 521 assert(next_dirty == next_offset); 522 nb_chunks++; 523 } 524 525 /* Clear dirty bits before querying the block status, because 526 * calling bdrv_block_status_above could yield - if some blocks are 527 * marked dirty in this window, we need to know. 528 */ 529 bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset, 530 nb_chunks * s->granularity); 531 bdrv_dirty_bitmap_unlock(s->dirty_bitmap); 532 533 /* Before claiming an area in the in-flight bitmap, we have to 534 * create a MirrorOp for it so that conflicting requests can wait 535 * for it. mirror_perform() will create the real MirrorOps later, 536 * for now we just create a pseudo operation that will wake up all 537 * conflicting requests once all real operations have been 538 * launched. */ 539 pseudo_op = g_new(MirrorOp, 1); 540 *pseudo_op = (MirrorOp){ 541 .offset = offset, 542 .bytes = nb_chunks * s->granularity, 543 .is_pseudo_op = true, 544 }; 545 qemu_co_queue_init(&pseudo_op->waiting_requests); 546 QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next); 547 548 bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks); 549 while (nb_chunks > 0 && offset < s->bdev_length) { 550 int ret; 551 int64_t io_bytes; 552 int64_t io_bytes_acct; 553 MirrorMethod mirror_method = MIRROR_METHOD_COPY; 554 555 assert(!(offset % s->granularity)); 556 ret = bdrv_block_status_above(source, NULL, offset, 557 nb_chunks * s->granularity, 558 &io_bytes, NULL, NULL); 559 if (ret < 0) { 560 io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes); 561 } else if (ret & BDRV_BLOCK_DATA) { 562 io_bytes = MIN(io_bytes, max_io_bytes); 563 } 564 565 io_bytes -= io_bytes % s->granularity; 566 if (io_bytes < s->granularity) { 567 io_bytes = s->granularity; 568 } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) { 569 int64_t target_offset; 570 int64_t target_bytes; 571 bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes, 572 &target_offset, &target_bytes); 573 if (target_offset == offset && 574 target_bytes == io_bytes) { 575 mirror_method = ret & BDRV_BLOCK_ZERO ? 576 MIRROR_METHOD_ZERO : 577 MIRROR_METHOD_DISCARD; 578 } 579 } 580 581 while (s->in_flight >= MAX_IN_FLIGHT) { 582 trace_mirror_yield_in_flight(s, offset, s->in_flight); 583 mirror_wait_for_free_in_flight_slot(s); 584 } 585 586 if (s->ret < 0) { 587 ret = 0; 588 goto fail; 589 } 590 591 io_bytes = mirror_clip_bytes(s, offset, io_bytes); 592 io_bytes = mirror_perform(s, offset, io_bytes, mirror_method); 593 if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) { 594 io_bytes_acct = 0; 595 } else { 596 io_bytes_acct = io_bytes; 597 } 598 assert(io_bytes); 599 offset += io_bytes; 600 nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity); 601 delay_ns = block_job_ratelimit_get_delay(&s->common, io_bytes_acct); 602 } 603 604 ret = delay_ns; 605 fail: 606 QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next); 607 qemu_co_queue_restart_all(&pseudo_op->waiting_requests); 608 g_free(pseudo_op); 609 610 return ret; 611 } 612 613 static void mirror_free_init(MirrorBlockJob *s) 614 { 615 int granularity = s->granularity; 616 size_t buf_size = s->buf_size; 617 uint8_t *buf = s->buf; 618 619 assert(s->buf_free_count == 0); 620 QSIMPLEQ_INIT(&s->buf_free); 621 while (buf_size != 0) { 622 MirrorBuffer *cur = (MirrorBuffer *)buf; 623 QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next); 624 s->buf_free_count++; 625 buf_size -= granularity; 626 buf += granularity; 627 } 628 } 629 630 /* This is also used for the .pause callback. There is no matching 631 * mirror_resume() because mirror_run() will begin iterating again 632 * when the job is resumed. 633 */ 634 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s) 635 { 636 while (s->in_flight > 0) { 637 mirror_wait_for_free_in_flight_slot(s); 638 } 639 } 640 641 /** 642 * mirror_exit_common: handle both abort() and prepare() cases. 643 * for .prepare, returns 0 on success and -errno on failure. 644 * for .abort cases, denoted by abort = true, MUST return 0. 645 */ 646 static int mirror_exit_common(Job *job) 647 { 648 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 649 BlockJob *bjob = &s->common; 650 MirrorBDSOpaque *bs_opaque; 651 AioContext *replace_aio_context = NULL; 652 BlockDriverState *src; 653 BlockDriverState *target_bs; 654 BlockDriverState *mirror_top_bs; 655 Error *local_err = NULL; 656 bool abort = job->ret < 0; 657 int ret = 0; 658 659 if (s->prepared) { 660 return 0; 661 } 662 s->prepared = true; 663 664 mirror_top_bs = s->mirror_top_bs; 665 bs_opaque = mirror_top_bs->opaque; 666 src = mirror_top_bs->backing->bs; 667 target_bs = blk_bs(s->target); 668 669 if (bdrv_chain_contains(src, target_bs)) { 670 bdrv_unfreeze_backing_chain(mirror_top_bs, target_bs); 671 } 672 673 bdrv_release_dirty_bitmap(s->dirty_bitmap); 674 675 /* Make sure that the source BDS doesn't go away during bdrv_replace_node, 676 * before we can call bdrv_drained_end */ 677 bdrv_ref(src); 678 bdrv_ref(mirror_top_bs); 679 bdrv_ref(target_bs); 680 681 /* 682 * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before 683 * inserting target_bs at s->to_replace, where we might not be able to get 684 * these permissions. 685 */ 686 blk_unref(s->target); 687 s->target = NULL; 688 689 /* We don't access the source any more. Dropping any WRITE/RESIZE is 690 * required before it could become a backing file of target_bs. Not having 691 * these permissions any more means that we can't allow any new requests on 692 * mirror_top_bs from now on, so keep it drained. */ 693 bdrv_drained_begin(mirror_top_bs); 694 bs_opaque->stop = true; 695 bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing, 696 &error_abort); 697 if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) { 698 BlockDriverState *backing = s->is_none_mode ? src : s->base; 699 BlockDriverState *unfiltered_target = bdrv_skip_filters(target_bs); 700 701 if (bdrv_cow_bs(unfiltered_target) != backing) { 702 bdrv_set_backing_hd(unfiltered_target, backing, &local_err); 703 if (local_err) { 704 error_report_err(local_err); 705 local_err = NULL; 706 ret = -EPERM; 707 } 708 } 709 } else if (!abort && s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) { 710 assert(!bdrv_backing_chain_next(target_bs)); 711 ret = bdrv_open_backing_file(bdrv_skip_filters(target_bs), NULL, 712 "backing", &local_err); 713 if (ret < 0) { 714 error_report_err(local_err); 715 local_err = NULL; 716 } 717 } 718 719 if (s->to_replace) { 720 replace_aio_context = bdrv_get_aio_context(s->to_replace); 721 aio_context_acquire(replace_aio_context); 722 } 723 724 if (s->should_complete && !abort) { 725 BlockDriverState *to_replace = s->to_replace ?: src; 726 bool ro = bdrv_is_read_only(to_replace); 727 728 if (ro != bdrv_is_read_only(target_bs)) { 729 bdrv_reopen_set_read_only(target_bs, ro, NULL); 730 } 731 732 /* The mirror job has no requests in flight any more, but we need to 733 * drain potential other users of the BDS before changing the graph. */ 734 assert(s->in_drain); 735 bdrv_drained_begin(target_bs); 736 /* 737 * Cannot use check_to_replace_node() here, because that would 738 * check for an op blocker on @to_replace, and we have our own 739 * there. 740 */ 741 if (bdrv_recurse_can_replace(src, to_replace)) { 742 bdrv_replace_node(to_replace, target_bs, &local_err); 743 } else { 744 error_setg(&local_err, "Can no longer replace '%s' by '%s', " 745 "because it can no longer be guaranteed that doing so " 746 "would not lead to an abrupt change of visible data", 747 to_replace->node_name, target_bs->node_name); 748 } 749 bdrv_drained_end(target_bs); 750 if (local_err) { 751 error_report_err(local_err); 752 ret = -EPERM; 753 } 754 } 755 if (s->to_replace) { 756 bdrv_op_unblock_all(s->to_replace, s->replace_blocker); 757 error_free(s->replace_blocker); 758 bdrv_unref(s->to_replace); 759 } 760 if (replace_aio_context) { 761 aio_context_release(replace_aio_context); 762 } 763 g_free(s->replaces); 764 bdrv_unref(target_bs); 765 766 /* 767 * Remove the mirror filter driver from the graph. Before this, get rid of 768 * the blockers on the intermediate nodes so that the resulting state is 769 * valid. 770 */ 771 block_job_remove_all_bdrv(bjob); 772 bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort); 773 774 bs_opaque->job = NULL; 775 776 bdrv_drained_end(src); 777 bdrv_drained_end(mirror_top_bs); 778 s->in_drain = false; 779 bdrv_unref(mirror_top_bs); 780 bdrv_unref(src); 781 782 return ret; 783 } 784 785 static int mirror_prepare(Job *job) 786 { 787 return mirror_exit_common(job); 788 } 789 790 static void mirror_abort(Job *job) 791 { 792 int ret = mirror_exit_common(job); 793 assert(ret == 0); 794 } 795 796 static void coroutine_fn mirror_throttle(MirrorBlockJob *s) 797 { 798 int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); 799 800 if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) { 801 s->last_pause_ns = now; 802 job_sleep_ns(&s->common.job, 0); 803 } else { 804 job_pause_point(&s->common.job); 805 } 806 } 807 808 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s) 809 { 810 int64_t offset; 811 BlockDriverState *bs = s->mirror_top_bs->backing->bs; 812 BlockDriverState *target_bs = blk_bs(s->target); 813 int ret; 814 int64_t count; 815 816 if (s->zero_target) { 817 if (!bdrv_can_write_zeroes_with_unmap(target_bs)) { 818 bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length); 819 return 0; 820 } 821 822 s->initial_zeroing_ongoing = true; 823 for (offset = 0; offset < s->bdev_length; ) { 824 int bytes = MIN(s->bdev_length - offset, 825 QEMU_ALIGN_DOWN(INT_MAX, s->granularity)); 826 827 mirror_throttle(s); 828 829 if (job_is_cancelled(&s->common.job)) { 830 s->initial_zeroing_ongoing = false; 831 return 0; 832 } 833 834 if (s->in_flight >= MAX_IN_FLIGHT) { 835 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count, 836 s->in_flight); 837 mirror_wait_for_free_in_flight_slot(s); 838 continue; 839 } 840 841 mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO); 842 offset += bytes; 843 } 844 845 mirror_wait_for_all_io(s); 846 s->initial_zeroing_ongoing = false; 847 } 848 849 /* First part, loop on the sectors and initialize the dirty bitmap. */ 850 for (offset = 0; offset < s->bdev_length; ) { 851 /* Just to make sure we are not exceeding int limit. */ 852 int bytes = MIN(s->bdev_length - offset, 853 QEMU_ALIGN_DOWN(INT_MAX, s->granularity)); 854 855 mirror_throttle(s); 856 857 if (job_is_cancelled(&s->common.job)) { 858 return 0; 859 } 860 861 ret = bdrv_is_allocated_above(bs, s->base_overlay, true, offset, bytes, 862 &count); 863 if (ret < 0) { 864 return ret; 865 } 866 867 assert(count); 868 if (ret > 0) { 869 bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count); 870 } 871 offset += count; 872 } 873 return 0; 874 } 875 876 /* Called when going out of the streaming phase to flush the bulk of the 877 * data to the medium, or just before completing. 878 */ 879 static int mirror_flush(MirrorBlockJob *s) 880 { 881 int ret = blk_flush(s->target); 882 if (ret < 0) { 883 if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) { 884 s->ret = ret; 885 } 886 } 887 return ret; 888 } 889 890 static int coroutine_fn mirror_run(Job *job, Error **errp) 891 { 892 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 893 BlockDriverState *bs = s->mirror_top_bs->backing->bs; 894 BlockDriverState *target_bs = blk_bs(s->target); 895 bool need_drain = true; 896 int64_t length; 897 int64_t target_length; 898 BlockDriverInfo bdi; 899 char backing_filename[2]; /* we only need 2 characters because we are only 900 checking for a NULL string */ 901 int ret = 0; 902 903 if (job_is_cancelled(&s->common.job)) { 904 goto immediate_exit; 905 } 906 907 s->bdev_length = bdrv_getlength(bs); 908 if (s->bdev_length < 0) { 909 ret = s->bdev_length; 910 goto immediate_exit; 911 } 912 913 target_length = blk_getlength(s->target); 914 if (target_length < 0) { 915 ret = target_length; 916 goto immediate_exit; 917 } 918 919 /* Active commit must resize the base image if its size differs from the 920 * active layer. */ 921 if (s->base == blk_bs(s->target)) { 922 if (s->bdev_length > target_length) { 923 ret = blk_truncate(s->target, s->bdev_length, false, 924 PREALLOC_MODE_OFF, 0, NULL); 925 if (ret < 0) { 926 goto immediate_exit; 927 } 928 } 929 } else if (s->bdev_length != target_length) { 930 error_setg(errp, "Source and target image have different sizes"); 931 ret = -EINVAL; 932 goto immediate_exit; 933 } 934 935 if (s->bdev_length == 0) { 936 /* Transition to the READY state and wait for complete. */ 937 job_transition_to_ready(&s->common.job); 938 s->actively_synced = true; 939 while (!job_cancel_requested(&s->common.job) && !s->should_complete) { 940 job_yield(&s->common.job); 941 } 942 goto immediate_exit; 943 } 944 945 length = DIV_ROUND_UP(s->bdev_length, s->granularity); 946 s->in_flight_bitmap = bitmap_new(length); 947 948 /* If we have no backing file yet in the destination, we cannot let 949 * the destination do COW. Instead, we copy sectors around the 950 * dirty data if needed. We need a bitmap to do that. 951 */ 952 bdrv_get_backing_filename(target_bs, backing_filename, 953 sizeof(backing_filename)); 954 if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) { 955 s->target_cluster_size = bdi.cluster_size; 956 } else { 957 s->target_cluster_size = BDRV_SECTOR_SIZE; 958 } 959 if (backing_filename[0] && !bdrv_backing_chain_next(target_bs) && 960 s->granularity < s->target_cluster_size) { 961 s->buf_size = MAX(s->buf_size, s->target_cluster_size); 962 s->cow_bitmap = bitmap_new(length); 963 } 964 s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov); 965 966 s->buf = qemu_try_blockalign(bs, s->buf_size); 967 if (s->buf == NULL) { 968 ret = -ENOMEM; 969 goto immediate_exit; 970 } 971 972 mirror_free_init(s); 973 974 s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); 975 if (!s->is_none_mode) { 976 ret = mirror_dirty_init(s); 977 if (ret < 0 || job_is_cancelled(&s->common.job)) { 978 goto immediate_exit; 979 } 980 } 981 982 assert(!s->dbi); 983 s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap); 984 for (;;) { 985 uint64_t delay_ns = 0; 986 int64_t cnt, delta; 987 bool should_complete; 988 989 /* Do not start passive operations while there are active 990 * writes in progress */ 991 while (s->in_active_write_counter) { 992 mirror_wait_for_any_operation(s, true); 993 } 994 995 if (s->ret < 0) { 996 ret = s->ret; 997 goto immediate_exit; 998 } 999 1000 job_pause_point(&s->common.job); 1001 1002 if (job_is_cancelled(&s->common.job)) { 1003 ret = 0; 1004 goto immediate_exit; 1005 } 1006 1007 cnt = bdrv_get_dirty_count(s->dirty_bitmap); 1008 /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is 1009 * the number of bytes currently being processed; together those are 1010 * the current remaining operation length */ 1011 job_progress_set_remaining(&s->common.job, s->bytes_in_flight + cnt); 1012 1013 /* Note that even when no rate limit is applied we need to yield 1014 * periodically with no pending I/O so that bdrv_drain_all() returns. 1015 * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is 1016 * an error, or when the source is clean, whichever comes first. */ 1017 delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns; 1018 if (delta < BLOCK_JOB_SLICE_TIME && 1019 s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) { 1020 if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 || 1021 (cnt == 0 && s->in_flight > 0)) { 1022 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight); 1023 mirror_wait_for_free_in_flight_slot(s); 1024 continue; 1025 } else if (cnt != 0) { 1026 delay_ns = mirror_iteration(s); 1027 } 1028 } 1029 1030 should_complete = false; 1031 if (s->in_flight == 0 && cnt == 0) { 1032 trace_mirror_before_flush(s); 1033 if (!job_is_ready(&s->common.job)) { 1034 if (mirror_flush(s) < 0) { 1035 /* Go check s->ret. */ 1036 continue; 1037 } 1038 /* We're out of the streaming phase. From now on, if the job 1039 * is cancelled we will actually complete all pending I/O and 1040 * report completion. This way, block-job-cancel will leave 1041 * the target in a consistent state. 1042 */ 1043 job_transition_to_ready(&s->common.job); 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_cancel_requested(&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 need_drain = false; 1081 break; 1082 } 1083 1084 if (job_is_ready(&s->common.job) && !should_complete) { 1085 delay_ns = (s->in_flight == 0 && 1086 cnt == 0 ? BLOCK_JOB_SLICE_TIME : 0); 1087 } 1088 trace_mirror_before_sleep(s, cnt, job_is_ready(&s->common.job), 1089 delay_ns); 1090 job_sleep_ns(&s->common.job, delay_ns); 1091 s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); 1092 } 1093 1094 immediate_exit: 1095 if (s->in_flight > 0) { 1096 /* We get here only if something went wrong. Either the job failed, 1097 * or it was cancelled prematurely so that we do not guarantee that 1098 * the target is a copy of the source. 1099 */ 1100 assert(ret < 0 || job_is_cancelled(&s->common.job)); 1101 assert(need_drain); 1102 mirror_wait_for_all_io(s); 1103 } 1104 1105 assert(s->in_flight == 0); 1106 qemu_vfree(s->buf); 1107 g_free(s->cow_bitmap); 1108 g_free(s->in_flight_bitmap); 1109 bdrv_dirty_iter_free(s->dbi); 1110 1111 if (need_drain) { 1112 s->in_drain = true; 1113 bdrv_drained_begin(bs); 1114 } 1115 1116 return ret; 1117 } 1118 1119 static void mirror_complete(Job *job, Error **errp) 1120 { 1121 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 1122 1123 if (!job_is_ready(job)) { 1124 error_setg(errp, "The active block job '%s' cannot be completed", 1125 job->id); 1126 return; 1127 } 1128 1129 /* block all operations on to_replace bs */ 1130 if (s->replaces) { 1131 AioContext *replace_aio_context; 1132 1133 s->to_replace = bdrv_find_node(s->replaces); 1134 if (!s->to_replace) { 1135 error_setg(errp, "Node name '%s' not found", s->replaces); 1136 return; 1137 } 1138 1139 replace_aio_context = bdrv_get_aio_context(s->to_replace); 1140 aio_context_acquire(replace_aio_context); 1141 1142 /* TODO Translate this into permission system. Current definition of 1143 * GRAPH_MOD would require to request it for the parents; they might 1144 * not even be BlockDriverStates, however, so a BdrvChild can't address 1145 * them. May need redefinition of GRAPH_MOD. */ 1146 error_setg(&s->replace_blocker, 1147 "block device is in use by block-job-complete"); 1148 bdrv_op_block_all(s->to_replace, s->replace_blocker); 1149 bdrv_ref(s->to_replace); 1150 1151 aio_context_release(replace_aio_context); 1152 } 1153 1154 s->should_complete = true; 1155 1156 /* If the job is paused, it will be re-entered when it is resumed */ 1157 if (!job->paused) { 1158 job_enter(job); 1159 } 1160 } 1161 1162 static void coroutine_fn mirror_pause(Job *job) 1163 { 1164 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 1165 1166 mirror_wait_for_all_io(s); 1167 } 1168 1169 static bool mirror_drained_poll(BlockJob *job) 1170 { 1171 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common); 1172 1173 /* If the job isn't paused nor cancelled, we can't be sure that it won't 1174 * issue more requests. We make an exception if we've reached this point 1175 * from one of our own drain sections, to avoid a deadlock waiting for 1176 * ourselves. 1177 */ 1178 if (!s->common.job.paused && !job_is_cancelled(&job->job) && !s->in_drain) { 1179 return true; 1180 } 1181 1182 return !!s->in_flight; 1183 } 1184 1185 static bool mirror_cancel(Job *job, bool force) 1186 { 1187 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job); 1188 BlockDriverState *target = blk_bs(s->target); 1189 1190 /* 1191 * Before the job is READY, we treat any cancellation like a 1192 * force-cancellation. 1193 */ 1194 force = force || !job_is_ready(job); 1195 1196 if (force) { 1197 bdrv_cancel_in_flight(target); 1198 } 1199 return force; 1200 } 1201 1202 static bool commit_active_cancel(Job *job, bool force) 1203 { 1204 /* Same as above in mirror_cancel() */ 1205 return force || !job_is_ready(job); 1206 } 1207 1208 static const BlockJobDriver mirror_job_driver = { 1209 .job_driver = { 1210 .instance_size = sizeof(MirrorBlockJob), 1211 .job_type = JOB_TYPE_MIRROR, 1212 .free = block_job_free, 1213 .user_resume = block_job_user_resume, 1214 .run = mirror_run, 1215 .prepare = mirror_prepare, 1216 .abort = mirror_abort, 1217 .pause = mirror_pause, 1218 .complete = mirror_complete, 1219 .cancel = mirror_cancel, 1220 }, 1221 .drained_poll = mirror_drained_poll, 1222 }; 1223 1224 static const BlockJobDriver commit_active_job_driver = { 1225 .job_driver = { 1226 .instance_size = sizeof(MirrorBlockJob), 1227 .job_type = JOB_TYPE_COMMIT, 1228 .free = block_job_free, 1229 .user_resume = block_job_user_resume, 1230 .run = mirror_run, 1231 .prepare = mirror_prepare, 1232 .abort = mirror_abort, 1233 .pause = mirror_pause, 1234 .complete = mirror_complete, 1235 .cancel = commit_active_cancel, 1236 }, 1237 .drained_poll = mirror_drained_poll, 1238 }; 1239 1240 static void coroutine_fn 1241 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method, 1242 uint64_t offset, uint64_t bytes, 1243 QEMUIOVector *qiov, int flags) 1244 { 1245 int ret; 1246 size_t qiov_offset = 0; 1247 int64_t bitmap_offset, bitmap_end; 1248 1249 if (!QEMU_IS_ALIGNED(offset, job->granularity) && 1250 bdrv_dirty_bitmap_get(job->dirty_bitmap, offset)) 1251 { 1252 /* 1253 * Dirty unaligned padding: ignore it. 1254 * 1255 * Reasoning: 1256 * 1. If we copy it, we can't reset corresponding bit in 1257 * dirty_bitmap as there may be some "dirty" bytes still not 1258 * copied. 1259 * 2. It's already dirty, so skipping it we don't diverge mirror 1260 * progress. 1261 * 1262 * Note, that because of this, guest write may have no contribution 1263 * into mirror converge, but that's not bad, as we have background 1264 * process of mirroring. If under some bad circumstances (high guest 1265 * IO load) background process starve, we will not converge anyway, 1266 * even if each write will contribute, as guest is not guaranteed to 1267 * rewrite the whole disk. 1268 */ 1269 qiov_offset = QEMU_ALIGN_UP(offset, job->granularity) - offset; 1270 if (bytes <= qiov_offset) { 1271 /* nothing to do after shrink */ 1272 return; 1273 } 1274 offset += qiov_offset; 1275 bytes -= qiov_offset; 1276 } 1277 1278 if (!QEMU_IS_ALIGNED(offset + bytes, job->granularity) && 1279 bdrv_dirty_bitmap_get(job->dirty_bitmap, offset + bytes - 1)) 1280 { 1281 uint64_t tail = (offset + bytes) % job->granularity; 1282 1283 if (bytes <= tail) { 1284 /* nothing to do after shrink */ 1285 return; 1286 } 1287 bytes -= tail; 1288 } 1289 1290 /* 1291 * Tails are either clean or shrunk, so for bitmap resetting 1292 * we safely align the range down. 1293 */ 1294 bitmap_offset = QEMU_ALIGN_UP(offset, job->granularity); 1295 bitmap_end = QEMU_ALIGN_DOWN(offset + bytes, job->granularity); 1296 if (bitmap_offset < bitmap_end) { 1297 bdrv_reset_dirty_bitmap(job->dirty_bitmap, bitmap_offset, 1298 bitmap_end - bitmap_offset); 1299 } 1300 1301 job_progress_increase_remaining(&job->common.job, bytes); 1302 1303 switch (method) { 1304 case MIRROR_METHOD_COPY: 1305 ret = blk_co_pwritev_part(job->target, offset, bytes, 1306 qiov, qiov_offset, flags); 1307 break; 1308 1309 case MIRROR_METHOD_ZERO: 1310 assert(!qiov); 1311 ret = blk_co_pwrite_zeroes(job->target, offset, bytes, flags); 1312 break; 1313 1314 case MIRROR_METHOD_DISCARD: 1315 assert(!qiov); 1316 ret = blk_co_pdiscard(job->target, offset, bytes); 1317 break; 1318 1319 default: 1320 abort(); 1321 } 1322 1323 if (ret >= 0) { 1324 job_progress_update(&job->common.job, bytes); 1325 } else { 1326 BlockErrorAction action; 1327 1328 /* 1329 * We failed, so we should mark dirty the whole area, aligned up. 1330 * Note that we don't care about shrunk tails if any: they were dirty 1331 * at function start, and they must be still dirty, as we've locked 1332 * the region for in-flight op. 1333 */ 1334 bitmap_offset = QEMU_ALIGN_DOWN(offset, job->granularity); 1335 bitmap_end = QEMU_ALIGN_UP(offset + bytes, job->granularity); 1336 bdrv_set_dirty_bitmap(job->dirty_bitmap, bitmap_offset, 1337 bitmap_end - bitmap_offset); 1338 job->actively_synced = false; 1339 1340 action = mirror_error_action(job, false, -ret); 1341 if (action == BLOCK_ERROR_ACTION_REPORT) { 1342 if (!job->ret) { 1343 job->ret = ret; 1344 } 1345 } 1346 } 1347 } 1348 1349 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s, 1350 uint64_t offset, 1351 uint64_t bytes) 1352 { 1353 MirrorOp *op; 1354 uint64_t start_chunk = offset / s->granularity; 1355 uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity); 1356 1357 op = g_new(MirrorOp, 1); 1358 *op = (MirrorOp){ 1359 .s = s, 1360 .offset = offset, 1361 .bytes = bytes, 1362 .is_active_write = true, 1363 .is_in_flight = true, 1364 .co = qemu_coroutine_self(), 1365 }; 1366 qemu_co_queue_init(&op->waiting_requests); 1367 QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next); 1368 1369 s->in_active_write_counter++; 1370 1371 mirror_wait_on_conflicts(op, s, offset, bytes); 1372 1373 bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk); 1374 1375 return op; 1376 } 1377 1378 static void coroutine_fn active_write_settle(MirrorOp *op) 1379 { 1380 uint64_t start_chunk = op->offset / op->s->granularity; 1381 uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes, 1382 op->s->granularity); 1383 1384 if (!--op->s->in_active_write_counter && op->s->actively_synced) { 1385 BdrvChild *source = op->s->mirror_top_bs->backing; 1386 1387 if (QLIST_FIRST(&source->bs->parents) == source && 1388 QLIST_NEXT(source, next_parent) == NULL) 1389 { 1390 /* Assert that we are back in sync once all active write 1391 * operations are settled. 1392 * Note that we can only assert this if the mirror node 1393 * is the source node's only parent. */ 1394 assert(!bdrv_get_dirty_count(op->s->dirty_bitmap)); 1395 } 1396 } 1397 bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk); 1398 QTAILQ_REMOVE(&op->s->ops_in_flight, op, next); 1399 qemu_co_queue_restart_all(&op->waiting_requests); 1400 g_free(op); 1401 } 1402 1403 static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs, 1404 int64_t offset, int64_t bytes, QEMUIOVector *qiov, BdrvRequestFlags flags) 1405 { 1406 return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags); 1407 } 1408 1409 static int coroutine_fn bdrv_mirror_top_do_write(BlockDriverState *bs, 1410 MirrorMethod method, uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, 1411 int flags) 1412 { 1413 MirrorOp *op = NULL; 1414 MirrorBDSOpaque *s = bs->opaque; 1415 int ret = 0; 1416 bool copy_to_target; 1417 1418 copy_to_target = s->job->ret >= 0 && 1419 !job_is_cancelled(&s->job->common.job) && 1420 s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING; 1421 1422 if (copy_to_target) { 1423 op = active_write_prepare(s->job, offset, bytes); 1424 } 1425 1426 switch (method) { 1427 case MIRROR_METHOD_COPY: 1428 ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags); 1429 break; 1430 1431 case MIRROR_METHOD_ZERO: 1432 ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags); 1433 break; 1434 1435 case MIRROR_METHOD_DISCARD: 1436 ret = bdrv_co_pdiscard(bs->backing, offset, bytes); 1437 break; 1438 1439 default: 1440 abort(); 1441 } 1442 1443 if (ret < 0) { 1444 goto out; 1445 } 1446 1447 if (copy_to_target) { 1448 do_sync_target_write(s->job, method, offset, bytes, qiov, flags); 1449 } 1450 1451 out: 1452 if (copy_to_target) { 1453 active_write_settle(op); 1454 } 1455 return ret; 1456 } 1457 1458 static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs, 1459 int64_t offset, int64_t bytes, QEMUIOVector *qiov, BdrvRequestFlags flags) 1460 { 1461 MirrorBDSOpaque *s = bs->opaque; 1462 QEMUIOVector bounce_qiov; 1463 void *bounce_buf; 1464 int ret = 0; 1465 bool copy_to_target; 1466 1467 copy_to_target = s->job->ret >= 0 && 1468 !job_is_cancelled(&s->job->common.job) && 1469 s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING; 1470 1471 if (copy_to_target) { 1472 /* The guest might concurrently modify the data to write; but 1473 * the data on source and destination must match, so we have 1474 * to use a bounce buffer if we are going to write to the 1475 * target now. */ 1476 bounce_buf = qemu_blockalign(bs, bytes); 1477 iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes); 1478 1479 qemu_iovec_init(&bounce_qiov, 1); 1480 qemu_iovec_add(&bounce_qiov, bounce_buf, bytes); 1481 qiov = &bounce_qiov; 1482 } 1483 1484 ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov, 1485 flags); 1486 1487 if (copy_to_target) { 1488 qemu_iovec_destroy(&bounce_qiov); 1489 qemu_vfree(bounce_buf); 1490 } 1491 1492 return ret; 1493 } 1494 1495 static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs) 1496 { 1497 if (bs->backing == NULL) { 1498 /* we can be here after failed bdrv_append in mirror_start_job */ 1499 return 0; 1500 } 1501 return bdrv_co_flush(bs->backing->bs); 1502 } 1503 1504 static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs, 1505 int64_t offset, int64_t bytes, BdrvRequestFlags flags) 1506 { 1507 return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL, 1508 flags); 1509 } 1510 1511 static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs, 1512 int64_t offset, int64_t bytes) 1513 { 1514 return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes, 1515 NULL, 0); 1516 } 1517 1518 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs) 1519 { 1520 if (bs->backing == NULL) { 1521 /* we can be here after failed bdrv_attach_child in 1522 * bdrv_set_backing_hd */ 1523 return; 1524 } 1525 pstrcpy(bs->exact_filename, sizeof(bs->exact_filename), 1526 bs->backing->bs->filename); 1527 } 1528 1529 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c, 1530 BdrvChildRole role, 1531 BlockReopenQueue *reopen_queue, 1532 uint64_t perm, uint64_t shared, 1533 uint64_t *nperm, uint64_t *nshared) 1534 { 1535 MirrorBDSOpaque *s = bs->opaque; 1536 1537 if (s->stop) { 1538 /* 1539 * If the job is to be stopped, we do not need to forward 1540 * anything to the real image. 1541 */ 1542 *nperm = 0; 1543 *nshared = BLK_PERM_ALL; 1544 return; 1545 } 1546 1547 bdrv_default_perms(bs, c, role, reopen_queue, 1548 perm, shared, nperm, nshared); 1549 1550 if (s->is_commit) { 1551 /* 1552 * For commit jobs, we cannot take CONSISTENT_READ, because 1553 * that permission is unshared for everything above the base 1554 * node (except for filters on the base node). 1555 * We also have to force-share the WRITE permission, or 1556 * otherwise we would block ourselves at the base node (if 1557 * writes are blocked for a node, they are also blocked for 1558 * its backing file). 1559 * (We could also share RESIZE, because it may be needed for 1560 * the target if its size is less than the top node's; but 1561 * bdrv_default_perms_for_cow() automatically shares RESIZE 1562 * for backing nodes if WRITE is shared, so there is no need 1563 * to do it here.) 1564 */ 1565 *nperm &= ~BLK_PERM_CONSISTENT_READ; 1566 *nshared |= BLK_PERM_WRITE; 1567 } 1568 } 1569 1570 /* Dummy node that provides consistent read to its users without requiring it 1571 * from its backing file and that allows writes on the backing file chain. */ 1572 static BlockDriver bdrv_mirror_top = { 1573 .format_name = "mirror_top", 1574 .bdrv_co_preadv = bdrv_mirror_top_preadv, 1575 .bdrv_co_pwritev = bdrv_mirror_top_pwritev, 1576 .bdrv_co_pwrite_zeroes = bdrv_mirror_top_pwrite_zeroes, 1577 .bdrv_co_pdiscard = bdrv_mirror_top_pdiscard, 1578 .bdrv_co_flush = bdrv_mirror_top_flush, 1579 .bdrv_refresh_filename = bdrv_mirror_top_refresh_filename, 1580 .bdrv_child_perm = bdrv_mirror_top_child_perm, 1581 1582 .is_filter = true, 1583 }; 1584 1585 static BlockJob *mirror_start_job( 1586 const char *job_id, BlockDriverState *bs, 1587 int creation_flags, BlockDriverState *target, 1588 const char *replaces, int64_t speed, 1589 uint32_t granularity, int64_t buf_size, 1590 BlockMirrorBackingMode backing_mode, 1591 bool zero_target, 1592 BlockdevOnError on_source_error, 1593 BlockdevOnError on_target_error, 1594 bool unmap, 1595 BlockCompletionFunc *cb, 1596 void *opaque, 1597 const BlockJobDriver *driver, 1598 bool is_none_mode, BlockDriverState *base, 1599 bool auto_complete, const char *filter_node_name, 1600 bool is_mirror, MirrorCopyMode copy_mode, 1601 Error **errp) 1602 { 1603 MirrorBlockJob *s; 1604 MirrorBDSOpaque *bs_opaque; 1605 BlockDriverState *mirror_top_bs; 1606 bool target_is_backing; 1607 uint64_t target_perms, target_shared_perms; 1608 int ret; 1609 1610 if (granularity == 0) { 1611 granularity = bdrv_get_default_bitmap_granularity(target); 1612 } 1613 1614 assert(is_power_of_2(granularity)); 1615 1616 if (buf_size < 0) { 1617 error_setg(errp, "Invalid parameter 'buf-size'"); 1618 return NULL; 1619 } 1620 1621 if (buf_size == 0) { 1622 buf_size = DEFAULT_MIRROR_BUF_SIZE; 1623 } 1624 1625 if (bdrv_skip_filters(bs) == bdrv_skip_filters(target)) { 1626 error_setg(errp, "Can't mirror node into itself"); 1627 return NULL; 1628 } 1629 1630 target_is_backing = bdrv_chain_contains(bs, target); 1631 1632 /* In the case of active commit, add dummy driver to provide consistent 1633 * reads on the top, while disabling it in the intermediate nodes, and make 1634 * the backing chain writable. */ 1635 mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name, 1636 BDRV_O_RDWR, errp); 1637 if (mirror_top_bs == NULL) { 1638 return NULL; 1639 } 1640 if (!filter_node_name) { 1641 mirror_top_bs->implicit = true; 1642 } 1643 1644 /* So that we can always drop this node */ 1645 mirror_top_bs->never_freeze = true; 1646 1647 mirror_top_bs->total_sectors = bs->total_sectors; 1648 mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED; 1649 mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED | 1650 BDRV_REQ_NO_FALLBACK; 1651 bs_opaque = g_new0(MirrorBDSOpaque, 1); 1652 mirror_top_bs->opaque = bs_opaque; 1653 1654 bs_opaque->is_commit = target_is_backing; 1655 1656 bdrv_drained_begin(bs); 1657 ret = bdrv_append(mirror_top_bs, bs, errp); 1658 bdrv_drained_end(bs); 1659 1660 if (ret < 0) { 1661 bdrv_unref(mirror_top_bs); 1662 return NULL; 1663 } 1664 1665 /* Make sure that the source is not resized while the job is running */ 1666 s = block_job_create(job_id, driver, NULL, mirror_top_bs, 1667 BLK_PERM_CONSISTENT_READ, 1668 BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED | 1669 BLK_PERM_WRITE | BLK_PERM_GRAPH_MOD, speed, 1670 creation_flags, cb, opaque, errp); 1671 if (!s) { 1672 goto fail; 1673 } 1674 bs_opaque->job = s; 1675 1676 /* The block job now has a reference to this node */ 1677 bdrv_unref(mirror_top_bs); 1678 1679 s->mirror_top_bs = mirror_top_bs; 1680 1681 /* No resize for the target either; while the mirror is still running, a 1682 * consistent read isn't necessarily possible. We could possibly allow 1683 * writes and graph modifications, though it would likely defeat the 1684 * purpose of a mirror, so leave them blocked for now. 1685 * 1686 * In the case of active commit, things look a bit different, though, 1687 * because the target is an already populated backing file in active use. 1688 * We can allow anything except resize there.*/ 1689 1690 target_perms = BLK_PERM_WRITE; 1691 target_shared_perms = BLK_PERM_WRITE_UNCHANGED; 1692 1693 if (target_is_backing) { 1694 int64_t bs_size, target_size; 1695 bs_size = bdrv_getlength(bs); 1696 if (bs_size < 0) { 1697 error_setg_errno(errp, -bs_size, 1698 "Could not inquire top image size"); 1699 goto fail; 1700 } 1701 1702 target_size = bdrv_getlength(target); 1703 if (target_size < 0) { 1704 error_setg_errno(errp, -target_size, 1705 "Could not inquire base image size"); 1706 goto fail; 1707 } 1708 1709 if (target_size < bs_size) { 1710 target_perms |= BLK_PERM_RESIZE; 1711 } 1712 1713 target_shared_perms |= BLK_PERM_CONSISTENT_READ 1714 | BLK_PERM_WRITE 1715 | BLK_PERM_GRAPH_MOD; 1716 } else if (bdrv_chain_contains(bs, bdrv_skip_filters(target))) { 1717 /* 1718 * We may want to allow this in the future, but it would 1719 * require taking some extra care. 1720 */ 1721 error_setg(errp, "Cannot mirror to a filter on top of a node in the " 1722 "source's backing chain"); 1723 goto fail; 1724 } 1725 1726 if (backing_mode != MIRROR_LEAVE_BACKING_CHAIN) { 1727 target_perms |= BLK_PERM_GRAPH_MOD; 1728 } 1729 1730 s->target = blk_new(s->common.job.aio_context, 1731 target_perms, target_shared_perms); 1732 ret = blk_insert_bs(s->target, target, errp); 1733 if (ret < 0) { 1734 goto fail; 1735 } 1736 if (is_mirror) { 1737 /* XXX: Mirror target could be a NBD server of target QEMU in the case 1738 * of non-shared block migration. To allow migration completion, we 1739 * have to allow "inactivate" of the target BB. When that happens, we 1740 * know the job is drained, and the vcpus are stopped, so no write 1741 * operation will be performed. Block layer already has assertions to 1742 * ensure that. */ 1743 blk_set_force_allow_inactivate(s->target); 1744 } 1745 blk_set_allow_aio_context_change(s->target, true); 1746 blk_set_disable_request_queuing(s->target, true); 1747 1748 s->replaces = g_strdup(replaces); 1749 s->on_source_error = on_source_error; 1750 s->on_target_error = on_target_error; 1751 s->is_none_mode = is_none_mode; 1752 s->backing_mode = backing_mode; 1753 s->zero_target = zero_target; 1754 s->copy_mode = copy_mode; 1755 s->base = base; 1756 s->base_overlay = bdrv_find_overlay(bs, base); 1757 s->granularity = granularity; 1758 s->buf_size = ROUND_UP(buf_size, granularity); 1759 s->unmap = unmap; 1760 if (auto_complete) { 1761 s->should_complete = true; 1762 } 1763 1764 s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp); 1765 if (!s->dirty_bitmap) { 1766 goto fail; 1767 } 1768 if (s->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING) { 1769 bdrv_disable_dirty_bitmap(s->dirty_bitmap); 1770 } 1771 1772 ret = block_job_add_bdrv(&s->common, "source", bs, 0, 1773 BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE | 1774 BLK_PERM_CONSISTENT_READ, 1775 errp); 1776 if (ret < 0) { 1777 goto fail; 1778 } 1779 1780 /* Required permissions are already taken with blk_new() */ 1781 block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL, 1782 &error_abort); 1783 1784 /* In commit_active_start() all intermediate nodes disappear, so 1785 * any jobs in them must be blocked */ 1786 if (target_is_backing) { 1787 BlockDriverState *iter, *filtered_target; 1788 uint64_t iter_shared_perms; 1789 1790 /* 1791 * The topmost node with 1792 * bdrv_skip_filters(filtered_target) == bdrv_skip_filters(target) 1793 */ 1794 filtered_target = bdrv_cow_bs(bdrv_find_overlay(bs, target)); 1795 1796 assert(bdrv_skip_filters(filtered_target) == 1797 bdrv_skip_filters(target)); 1798 1799 /* 1800 * XXX BLK_PERM_WRITE needs to be allowed so we don't block 1801 * ourselves at s->base (if writes are blocked for a node, they are 1802 * also blocked for its backing file). The other options would be a 1803 * second filter driver above s->base (== target). 1804 */ 1805 iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE; 1806 1807 for (iter = bdrv_filter_or_cow_bs(bs); iter != target; 1808 iter = bdrv_filter_or_cow_bs(iter)) 1809 { 1810 if (iter == filtered_target) { 1811 /* 1812 * From here on, all nodes are filters on the base. 1813 * This allows us to share BLK_PERM_CONSISTENT_READ. 1814 */ 1815 iter_shared_perms |= BLK_PERM_CONSISTENT_READ; 1816 } 1817 1818 ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0, 1819 iter_shared_perms, errp); 1820 if (ret < 0) { 1821 goto fail; 1822 } 1823 } 1824 1825 if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) { 1826 goto fail; 1827 } 1828 } 1829 1830 QTAILQ_INIT(&s->ops_in_flight); 1831 1832 trace_mirror_start(bs, s, opaque); 1833 job_start(&s->common.job); 1834 1835 return &s->common; 1836 1837 fail: 1838 if (s) { 1839 /* Make sure this BDS does not go away until we have completed the graph 1840 * changes below */ 1841 bdrv_ref(mirror_top_bs); 1842 1843 g_free(s->replaces); 1844 blk_unref(s->target); 1845 bs_opaque->job = NULL; 1846 if (s->dirty_bitmap) { 1847 bdrv_release_dirty_bitmap(s->dirty_bitmap); 1848 } 1849 job_early_fail(&s->common.job); 1850 } 1851 1852 bs_opaque->stop = true; 1853 bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing, 1854 &error_abort); 1855 bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort); 1856 1857 bdrv_unref(mirror_top_bs); 1858 1859 return NULL; 1860 } 1861 1862 void mirror_start(const char *job_id, BlockDriverState *bs, 1863 BlockDriverState *target, const char *replaces, 1864 int creation_flags, int64_t speed, 1865 uint32_t granularity, int64_t buf_size, 1866 MirrorSyncMode mode, BlockMirrorBackingMode backing_mode, 1867 bool zero_target, 1868 BlockdevOnError on_source_error, 1869 BlockdevOnError on_target_error, 1870 bool unmap, const char *filter_node_name, 1871 MirrorCopyMode copy_mode, Error **errp) 1872 { 1873 bool is_none_mode; 1874 BlockDriverState *base; 1875 1876 if ((mode == MIRROR_SYNC_MODE_INCREMENTAL) || 1877 (mode == MIRROR_SYNC_MODE_BITMAP)) { 1878 error_setg(errp, "Sync mode '%s' not supported", 1879 MirrorSyncMode_str(mode)); 1880 return; 1881 } 1882 is_none_mode = mode == MIRROR_SYNC_MODE_NONE; 1883 base = mode == MIRROR_SYNC_MODE_TOP ? bdrv_backing_chain_next(bs) : NULL; 1884 mirror_start_job(job_id, bs, creation_flags, target, replaces, 1885 speed, granularity, buf_size, backing_mode, zero_target, 1886 on_source_error, on_target_error, unmap, NULL, NULL, 1887 &mirror_job_driver, is_none_mode, base, false, 1888 filter_node_name, true, copy_mode, errp); 1889 } 1890 1891 BlockJob *commit_active_start(const char *job_id, BlockDriverState *bs, 1892 BlockDriverState *base, int creation_flags, 1893 int64_t speed, BlockdevOnError on_error, 1894 const char *filter_node_name, 1895 BlockCompletionFunc *cb, void *opaque, 1896 bool auto_complete, Error **errp) 1897 { 1898 bool base_read_only; 1899 BlockJob *job; 1900 1901 base_read_only = bdrv_is_read_only(base); 1902 1903 if (base_read_only) { 1904 if (bdrv_reopen_set_read_only(base, false, errp) < 0) { 1905 return NULL; 1906 } 1907 } 1908 1909 job = mirror_start_job( 1910 job_id, bs, creation_flags, base, NULL, speed, 0, 0, 1911 MIRROR_LEAVE_BACKING_CHAIN, false, 1912 on_error, on_error, true, cb, opaque, 1913 &commit_active_job_driver, false, base, auto_complete, 1914 filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND, 1915 errp); 1916 if (!job) { 1917 goto error_restore_flags; 1918 } 1919 1920 return job; 1921 1922 error_restore_flags: 1923 /* ignore error and errp for bdrv_reopen, because we want to propagate 1924 * the original error */ 1925 if (base_read_only) { 1926 bdrv_reopen_set_read_only(base, true, NULL); 1927 } 1928 return NULL; 1929 } 1930