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