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