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