xref: /openbmc/qemu/block/mirror.c (revision 64a917d5)
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 "block/dirty-bitmap.h"
22 #include "sysemu/block-backend.h"
23 #include "qapi/error.h"
24 #include "qemu/ratelimit.h"
25 #include "qemu/bitmap.h"
26 #include "qemu/memalign.h"
27 
28 #define MAX_IN_FLIGHT 16
29 #define MAX_IO_BYTES (1 << 20) /* 1 Mb */
30 #define DEFAULT_MIRROR_BUF_SIZE (MAX_IN_FLIGHT * MAX_IO_BYTES)
31 
32 /* The mirroring buffer is a list of granularity-sized chunks.
33  * Free chunks are organized in a list.
34  */
35 typedef struct MirrorBuffer {
36     QSIMPLEQ_ENTRY(MirrorBuffer) next;
37 } MirrorBuffer;
38 
39 typedef struct MirrorOp MirrorOp;
40 
41 typedef struct MirrorBlockJob {
42     BlockJob common;
43     BlockBackend *target;
44     BlockDriverState *mirror_top_bs;
45     BlockDriverState *base;
46     BlockDriverState *base_overlay;
47 
48     /* The name of the graph node to replace */
49     char *replaces;
50     /* The BDS to replace */
51     BlockDriverState *to_replace;
52     /* Used to block operations on the drive-mirror-replace target */
53     Error *replace_blocker;
54     bool is_none_mode;
55     BlockMirrorBackingMode backing_mode;
56     /* Whether the target image requires explicit zero-initialization */
57     bool zero_target;
58     MirrorCopyMode copy_mode;
59     BlockdevOnError on_source_error, on_target_error;
60     /* Set when the target is synced (dirty bitmap is clean, nothing
61      * in flight) and the job is running in active mode */
62     bool actively_synced;
63     bool should_complete;
64     int64_t granularity;
65     size_t buf_size;
66     int64_t bdev_length;
67     unsigned long *cow_bitmap;
68     BdrvDirtyBitmap *dirty_bitmap;
69     BdrvDirtyBitmapIter *dbi;
70     uint8_t *buf;
71     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
72     int buf_free_count;
73 
74     uint64_t last_pause_ns;
75     unsigned long *in_flight_bitmap;
76     unsigned in_flight;
77     int64_t bytes_in_flight;
78     QTAILQ_HEAD(, MirrorOp) ops_in_flight;
79     int ret;
80     bool unmap;
81     int target_cluster_size;
82     int max_iov;
83     bool initial_zeroing_ongoing;
84     int in_active_write_counter;
85     int64_t active_write_bytes_in_flight;
86     bool prepared;
87     bool in_drain;
88 } MirrorBlockJob;
89 
90 typedef struct MirrorBDSOpaque {
91     MirrorBlockJob *job;
92     bool stop;
93     bool is_commit;
94 } MirrorBDSOpaque;
95 
96 struct MirrorOp {
97     MirrorBlockJob *s;
98     QEMUIOVector qiov;
99     int64_t offset;
100     uint64_t bytes;
101 
102     /* The pointee is set by mirror_co_read(), mirror_co_zero(), and
103      * mirror_co_discard() before yielding for the first time */
104     int64_t *bytes_handled;
105 
106     bool is_pseudo_op;
107     bool is_active_write;
108     bool is_in_flight;
109     CoQueue waiting_requests;
110     Coroutine *co;
111     MirrorOp *waiting_for_op;
112 
113     QTAILQ_ENTRY(MirrorOp) next;
114 };
115 
116 typedef enum MirrorMethod {
117     MIRROR_METHOD_COPY,
118     MIRROR_METHOD_ZERO,
119     MIRROR_METHOD_DISCARD,
120 } MirrorMethod;
121 
122 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
123                                             int error)
124 {
125     s->actively_synced = false;
126     if (read) {
127         return block_job_error_action(&s->common, s->on_source_error,
128                                       true, error);
129     } else {
130         return block_job_error_action(&s->common, s->on_target_error,
131                                       false, error);
132     }
133 }
134 
135 static void coroutine_fn mirror_wait_on_conflicts(MirrorOp *self,
136                                                   MirrorBlockJob *s,
137                                                   uint64_t offset,
138                                                   uint64_t bytes)
139 {
140     uint64_t self_start_chunk = offset / s->granularity;
141     uint64_t self_end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
142     uint64_t self_nb_chunks = self_end_chunk - self_start_chunk;
143 
144     while (find_next_bit(s->in_flight_bitmap, self_end_chunk,
145                          self_start_chunk) < self_end_chunk &&
146            s->ret >= 0)
147     {
148         MirrorOp *op;
149 
150         QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
151             uint64_t op_start_chunk = op->offset / s->granularity;
152             uint64_t op_nb_chunks = DIV_ROUND_UP(op->offset + op->bytes,
153                                                  s->granularity) -
154                                     op_start_chunk;
155 
156             if (op == self) {
157                 continue;
158             }
159 
160             if (ranges_overlap(self_start_chunk, self_nb_chunks,
161                                op_start_chunk, op_nb_chunks))
162             {
163                 if (self) {
164                     /*
165                      * If the operation is already (indirectly) waiting for us,
166                      * or will wait for us as soon as it wakes up, then just go
167                      * on (instead of producing a deadlock in the former case).
168                      */
169                     if (op->waiting_for_op) {
170                         continue;
171                     }
172 
173                     self->waiting_for_op = op;
174                 }
175 
176                 qemu_co_queue_wait(&op->waiting_requests, NULL);
177 
178                 if (self) {
179                     self->waiting_for_op = NULL;
180                 }
181 
182                 break;
183             }
184         }
185     }
186 }
187 
188 static void coroutine_fn mirror_iteration_done(MirrorOp *op, int ret)
189 {
190     MirrorBlockJob *s = op->s;
191     struct iovec *iov;
192     int64_t chunk_num;
193     int i, nb_chunks;
194 
195     trace_mirror_iteration_done(s, op->offset, op->bytes, ret);
196 
197     s->in_flight--;
198     s->bytes_in_flight -= op->bytes;
199     iov = op->qiov.iov;
200     for (i = 0; i < op->qiov.niov; i++) {
201         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
202         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
203         s->buf_free_count++;
204     }
205 
206     chunk_num = op->offset / s->granularity;
207     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
208 
209     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
210     QTAILQ_REMOVE(&s->ops_in_flight, op, next);
211     if (ret >= 0) {
212         if (s->cow_bitmap) {
213             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
214         }
215         if (!s->initial_zeroing_ongoing) {
216             job_progress_update(&s->common.job, op->bytes);
217         }
218     }
219     qemu_iovec_destroy(&op->qiov);
220 
221     qemu_co_queue_restart_all(&op->waiting_requests);
222     g_free(op);
223 }
224 
225 static void coroutine_fn mirror_write_complete(MirrorOp *op, int ret)
226 {
227     MirrorBlockJob *s = op->s;
228 
229     if (ret < 0) {
230         BlockErrorAction action;
231 
232         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
233         action = mirror_error_action(s, false, -ret);
234         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
235             s->ret = ret;
236         }
237     }
238 
239     mirror_iteration_done(op, ret);
240 }
241 
242 static void coroutine_fn mirror_read_complete(MirrorOp *op, int ret)
243 {
244     MirrorBlockJob *s = op->s;
245 
246     if (ret < 0) {
247         BlockErrorAction action;
248 
249         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
250         action = mirror_error_action(s, true, -ret);
251         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
252             s->ret = ret;
253         }
254 
255         mirror_iteration_done(op, ret);
256         return;
257     }
258 
259     ret = blk_co_pwritev(s->target, op->offset, op->qiov.size, &op->qiov, 0);
260     mirror_write_complete(op, ret);
261 }
262 
263 /* Clip bytes relative to offset to not exceed end-of-file */
264 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s,
265                                         int64_t offset,
266                                         int64_t bytes)
267 {
268     return MIN(bytes, s->bdev_length - offset);
269 }
270 
271 /* Round offset and/or bytes to target cluster if COW is needed, and
272  * return the offset of the adjusted tail against original. */
273 static int coroutine_fn mirror_cow_align(MirrorBlockJob *s, int64_t *offset,
274                                          uint64_t *bytes)
275 {
276     bool need_cow;
277     int ret = 0;
278     int64_t align_offset = *offset;
279     int64_t align_bytes = *bytes;
280     int max_bytes = s->granularity * s->max_iov;
281 
282     need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap);
283     need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity,
284                           s->cow_bitmap);
285     if (need_cow) {
286         bdrv_round_to_subclusters(blk_bs(s->target), *offset, *bytes,
287                                   &align_offset, &align_bytes);
288     }
289 
290     if (align_bytes > max_bytes) {
291         align_bytes = max_bytes;
292         if (need_cow) {
293             align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size);
294         }
295     }
296     /* Clipping may result in align_bytes unaligned to chunk boundary, but
297      * that doesn't matter because it's already the end of source image. */
298     align_bytes = mirror_clip_bytes(s, align_offset, align_bytes);
299 
300     ret = align_offset + align_bytes - (*offset + *bytes);
301     *offset = align_offset;
302     *bytes = align_bytes;
303     assert(ret >= 0);
304     return ret;
305 }
306 
307 static inline void coroutine_fn
308 mirror_wait_for_free_in_flight_slot(MirrorBlockJob *s)
309 {
310     MirrorOp *op;
311 
312     QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
313         /*
314          * Do not wait on pseudo ops, because it may in turn wait on
315          * some other operation to start, which may in fact be the
316          * caller of this function.  Since there is only one pseudo op
317          * at any given time, we will always find some real operation
318          * to wait on.
319          * Also, do not wait on active operations, because they do not
320          * use up in-flight slots.
321          */
322         if (!op->is_pseudo_op && op->is_in_flight && !op->is_active_write) {
323             qemu_co_queue_wait(&op->waiting_requests, NULL);
324             return;
325         }
326     }
327     abort();
328 }
329 
330 /* Perform a mirror copy operation.
331  *
332  * *op->bytes_handled is set to the number of bytes copied after and
333  * including offset, excluding any bytes copied prior to offset due
334  * to alignment.  This will be op->bytes if no alignment is necessary,
335  * or (new_end - op->offset) if the tail is rounded up or down due to
336  * alignment or buffer limit.
337  */
338 static void coroutine_fn mirror_co_read(void *opaque)
339 {
340     MirrorOp *op = opaque;
341     MirrorBlockJob *s = op->s;
342     int nb_chunks;
343     uint64_t ret;
344     uint64_t max_bytes;
345 
346     max_bytes = s->granularity * s->max_iov;
347 
348     /* We can only handle as much as buf_size at a time. */
349     op->bytes = MIN(s->buf_size, MIN(max_bytes, op->bytes));
350     assert(op->bytes);
351     assert(op->bytes < BDRV_REQUEST_MAX_BYTES);
352     *op->bytes_handled = op->bytes;
353 
354     if (s->cow_bitmap) {
355         *op->bytes_handled += mirror_cow_align(s, &op->offset, &op->bytes);
356     }
357     /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */
358     assert(*op->bytes_handled <= UINT_MAX);
359     assert(op->bytes <= s->buf_size);
360     /* The offset is granularity-aligned because:
361      * 1) Caller passes in aligned values;
362      * 2) mirror_cow_align is used only when target cluster is larger. */
363     assert(QEMU_IS_ALIGNED(op->offset, s->granularity));
364     /* The range is sector-aligned, since bdrv_getlength() rounds up. */
365     assert(QEMU_IS_ALIGNED(op->bytes, BDRV_SECTOR_SIZE));
366     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
367 
368     while (s->buf_free_count < nb_chunks) {
369         trace_mirror_yield_in_flight(s, op->offset, s->in_flight);
370         mirror_wait_for_free_in_flight_slot(s);
371     }
372 
373     /* Now make a QEMUIOVector taking enough granularity-sized chunks
374      * from s->buf_free.
375      */
376     qemu_iovec_init(&op->qiov, nb_chunks);
377     while (nb_chunks-- > 0) {
378         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
379         size_t remaining = op->bytes - op->qiov.size;
380 
381         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
382         s->buf_free_count--;
383         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
384     }
385 
386     /* Copy the dirty cluster.  */
387     s->in_flight++;
388     s->bytes_in_flight += op->bytes;
389     op->is_in_flight = true;
390     trace_mirror_one_iteration(s, op->offset, op->bytes);
391 
392     WITH_GRAPH_RDLOCK_GUARD() {
393         ret = bdrv_co_preadv(s->mirror_top_bs->backing, op->offset, op->bytes,
394                              &op->qiov, 0);
395     }
396     mirror_read_complete(op, ret);
397 }
398 
399 static void coroutine_fn mirror_co_zero(void *opaque)
400 {
401     MirrorOp *op = opaque;
402     int ret;
403 
404     op->s->in_flight++;
405     op->s->bytes_in_flight += op->bytes;
406     *op->bytes_handled = op->bytes;
407     op->is_in_flight = true;
408 
409     ret = blk_co_pwrite_zeroes(op->s->target, op->offset, op->bytes,
410                                op->s->unmap ? BDRV_REQ_MAY_UNMAP : 0);
411     mirror_write_complete(op, ret);
412 }
413 
414 static void coroutine_fn mirror_co_discard(void *opaque)
415 {
416     MirrorOp *op = opaque;
417     int ret;
418 
419     op->s->in_flight++;
420     op->s->bytes_in_flight += op->bytes;
421     *op->bytes_handled = op->bytes;
422     op->is_in_flight = true;
423 
424     ret = blk_co_pdiscard(op->s->target, op->offset, op->bytes);
425     mirror_write_complete(op, ret);
426 }
427 
428 static unsigned mirror_perform(MirrorBlockJob *s, int64_t offset,
429                                unsigned bytes, MirrorMethod mirror_method)
430 {
431     MirrorOp *op;
432     Coroutine *co;
433     int64_t bytes_handled = -1;
434 
435     op = g_new(MirrorOp, 1);
436     *op = (MirrorOp){
437         .s              = s,
438         .offset         = offset,
439         .bytes          = bytes,
440         .bytes_handled  = &bytes_handled,
441     };
442     qemu_co_queue_init(&op->waiting_requests);
443 
444     switch (mirror_method) {
445     case MIRROR_METHOD_COPY:
446         co = qemu_coroutine_create(mirror_co_read, op);
447         break;
448     case MIRROR_METHOD_ZERO:
449         co = qemu_coroutine_create(mirror_co_zero, op);
450         break;
451     case MIRROR_METHOD_DISCARD:
452         co = qemu_coroutine_create(mirror_co_discard, op);
453         break;
454     default:
455         abort();
456     }
457     op->co = co;
458 
459     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
460     qemu_coroutine_enter(co);
461     /* At this point, ownership of op has been moved to the coroutine
462      * and the object may already be freed */
463 
464     /* Assert that this value has been set */
465     assert(bytes_handled >= 0);
466 
467     /* Same assertion as in mirror_co_read() (and for mirror_co_read()
468      * and mirror_co_discard(), bytes_handled == op->bytes, which
469      * is the @bytes parameter given to this function) */
470     assert(bytes_handled <= UINT_MAX);
471     return bytes_handled;
472 }
473 
474 static void coroutine_fn mirror_iteration(MirrorBlockJob *s)
475 {
476     BlockDriverState *source = s->mirror_top_bs->backing->bs;
477     MirrorOp *pseudo_op;
478     int64_t offset;
479     /* At least the first dirty chunk is mirrored in one iteration. */
480     int nb_chunks = 1;
481     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
482     int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES);
483 
484     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
485     offset = bdrv_dirty_iter_next(s->dbi);
486     if (offset < 0) {
487         bdrv_set_dirty_iter(s->dbi, 0);
488         offset = bdrv_dirty_iter_next(s->dbi);
489         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
490         assert(offset >= 0);
491     }
492     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
493 
494     /*
495      * Wait for concurrent requests to @offset.  The next loop will limit the
496      * copied area based on in_flight_bitmap so we only copy an area that does
497      * not overlap with concurrent in-flight requests.  Still, we would like to
498      * copy something, so wait until there are at least no more requests to the
499      * very beginning of the area.
500      */
501     mirror_wait_on_conflicts(NULL, s, offset, 1);
502 
503     job_pause_point(&s->common.job);
504 
505     /* Find the number of consective dirty chunks following the first dirty
506      * one, and wait for in flight requests in them. */
507     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
508     while (nb_chunks * s->granularity < s->buf_size) {
509         int64_t next_dirty;
510         int64_t next_offset = offset + nb_chunks * s->granularity;
511         int64_t next_chunk = next_offset / s->granularity;
512         if (next_offset >= s->bdev_length ||
513             !bdrv_dirty_bitmap_get_locked(s->dirty_bitmap, next_offset)) {
514             break;
515         }
516         if (test_bit(next_chunk, s->in_flight_bitmap)) {
517             break;
518         }
519 
520         next_dirty = bdrv_dirty_iter_next(s->dbi);
521         if (next_dirty > next_offset || next_dirty < 0) {
522             /* The bitmap iterator's cache is stale, refresh it */
523             bdrv_set_dirty_iter(s->dbi, next_offset);
524             next_dirty = bdrv_dirty_iter_next(s->dbi);
525         }
526         assert(next_dirty == next_offset);
527         nb_chunks++;
528     }
529 
530     /* Clear dirty bits before querying the block status, because
531      * calling bdrv_block_status_above could yield - if some blocks are
532      * marked dirty in this window, we need to know.
533      */
534     bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset,
535                                    nb_chunks * s->granularity);
536     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
537 
538     /* Before claiming an area in the in-flight bitmap, we have to
539      * create a MirrorOp for it so that conflicting requests can wait
540      * for it.  mirror_perform() will create the real MirrorOps later,
541      * for now we just create a pseudo operation that will wake up all
542      * conflicting requests once all real operations have been
543      * launched. */
544     pseudo_op = g_new(MirrorOp, 1);
545     *pseudo_op = (MirrorOp){
546         .offset         = offset,
547         .bytes          = nb_chunks * s->granularity,
548         .is_pseudo_op   = true,
549     };
550     qemu_co_queue_init(&pseudo_op->waiting_requests);
551     QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next);
552 
553     bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks);
554     while (nb_chunks > 0 && offset < s->bdev_length) {
555         int ret;
556         int64_t io_bytes;
557         int64_t io_bytes_acct;
558         MirrorMethod mirror_method = MIRROR_METHOD_COPY;
559 
560         assert(!(offset % s->granularity));
561         WITH_GRAPH_RDLOCK_GUARD() {
562             ret = bdrv_block_status_above(source, NULL, offset,
563                                         nb_chunks * s->granularity,
564                                         &io_bytes, NULL, NULL);
565         }
566         if (ret < 0) {
567             io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes);
568         } else if (ret & BDRV_BLOCK_DATA) {
569             io_bytes = MIN(io_bytes, max_io_bytes);
570         }
571 
572         io_bytes -= io_bytes % s->granularity;
573         if (io_bytes < s->granularity) {
574             io_bytes = s->granularity;
575         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
576             int64_t target_offset;
577             int64_t target_bytes;
578             WITH_GRAPH_RDLOCK_GUARD() {
579                 bdrv_round_to_subclusters(blk_bs(s->target), offset, io_bytes,
580                                           &target_offset, &target_bytes);
581             }
582             if (target_offset == offset &&
583                 target_bytes == io_bytes) {
584                 mirror_method = ret & BDRV_BLOCK_ZERO ?
585                                     MIRROR_METHOD_ZERO :
586                                     MIRROR_METHOD_DISCARD;
587             }
588         }
589 
590         while (s->in_flight >= MAX_IN_FLIGHT) {
591             trace_mirror_yield_in_flight(s, offset, s->in_flight);
592             mirror_wait_for_free_in_flight_slot(s);
593         }
594 
595         if (s->ret < 0) {
596             ret = 0;
597             goto fail;
598         }
599 
600         io_bytes = mirror_clip_bytes(s, offset, io_bytes);
601         io_bytes = mirror_perform(s, offset, io_bytes, mirror_method);
602         if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) {
603             io_bytes_acct = 0;
604         } else {
605             io_bytes_acct = io_bytes;
606         }
607         assert(io_bytes);
608         offset += io_bytes;
609         nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity);
610         block_job_ratelimit_processed_bytes(&s->common, io_bytes_acct);
611     }
612 
613 fail:
614     QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next);
615     qemu_co_queue_restart_all(&pseudo_op->waiting_requests);
616     g_free(pseudo_op);
617 }
618 
619 static void mirror_free_init(MirrorBlockJob *s)
620 {
621     int granularity = s->granularity;
622     size_t buf_size = s->buf_size;
623     uint8_t *buf = s->buf;
624 
625     assert(s->buf_free_count == 0);
626     QSIMPLEQ_INIT(&s->buf_free);
627     while (buf_size != 0) {
628         MirrorBuffer *cur = (MirrorBuffer *)buf;
629         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
630         s->buf_free_count++;
631         buf_size -= granularity;
632         buf += granularity;
633     }
634 }
635 
636 /* This is also used for the .pause callback. There is no matching
637  * mirror_resume() because mirror_run() will begin iterating again
638  * when the job is resumed.
639  */
640 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s)
641 {
642     while (s->in_flight > 0) {
643         mirror_wait_for_free_in_flight_slot(s);
644     }
645 }
646 
647 /**
648  * mirror_exit_common: handle both abort() and prepare() cases.
649  * for .prepare, returns 0 on success and -errno on failure.
650  * for .abort cases, denoted by abort = true, MUST return 0.
651  */
652 static int mirror_exit_common(Job *job)
653 {
654     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
655     BlockJob *bjob = &s->common;
656     MirrorBDSOpaque *bs_opaque;
657     AioContext *replace_aio_context = NULL;
658     BlockDriverState *src;
659     BlockDriverState *target_bs;
660     BlockDriverState *mirror_top_bs;
661     Error *local_err = NULL;
662     bool abort = job->ret < 0;
663     int ret = 0;
664 
665     GLOBAL_STATE_CODE();
666 
667     if (s->prepared) {
668         return 0;
669     }
670     s->prepared = true;
671 
672     aio_context_acquire(qemu_get_aio_context());
673 
674     mirror_top_bs = s->mirror_top_bs;
675     bs_opaque = mirror_top_bs->opaque;
676     src = mirror_top_bs->backing->bs;
677     target_bs = blk_bs(s->target);
678 
679     if (bdrv_chain_contains(src, target_bs)) {
680         bdrv_unfreeze_backing_chain(mirror_top_bs, target_bs);
681     }
682 
683     bdrv_release_dirty_bitmap(s->dirty_bitmap);
684 
685     /* Make sure that the source BDS doesn't go away during bdrv_replace_node,
686      * before we can call bdrv_drained_end */
687     bdrv_ref(src);
688     bdrv_ref(mirror_top_bs);
689     bdrv_ref(target_bs);
690 
691     /*
692      * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
693      * inserting target_bs at s->to_replace, where we might not be able to get
694      * these permissions.
695      */
696     blk_unref(s->target);
697     s->target = NULL;
698 
699     /* We don't access the source any more. Dropping any WRITE/RESIZE is
700      * required before it could become a backing file of target_bs. Not having
701      * these permissions any more means that we can't allow any new requests on
702      * mirror_top_bs from now on, so keep it drained. */
703     bdrv_drained_begin(mirror_top_bs);
704     bs_opaque->stop = true;
705     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
706                              &error_abort);
707     if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
708         BlockDriverState *backing = s->is_none_mode ? src : s->base;
709         BlockDriverState *unfiltered_target = bdrv_skip_filters(target_bs);
710 
711         if (bdrv_cow_bs(unfiltered_target) != backing) {
712             bdrv_set_backing_hd(unfiltered_target, backing, &local_err);
713             if (local_err) {
714                 error_report_err(local_err);
715                 local_err = NULL;
716                 ret = -EPERM;
717             }
718         }
719     } else if (!abort && s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
720         assert(!bdrv_backing_chain_next(target_bs));
721         ret = bdrv_open_backing_file(bdrv_skip_filters(target_bs), NULL,
722                                      "backing", &local_err);
723         if (ret < 0) {
724             error_report_err(local_err);
725             local_err = NULL;
726         }
727     }
728 
729     if (s->to_replace) {
730         replace_aio_context = bdrv_get_aio_context(s->to_replace);
731         aio_context_acquire(replace_aio_context);
732     }
733 
734     if (s->should_complete && !abort) {
735         BlockDriverState *to_replace = s->to_replace ?: src;
736         bool ro = bdrv_is_read_only(to_replace);
737 
738         if (ro != bdrv_is_read_only(target_bs)) {
739             bdrv_reopen_set_read_only(target_bs, ro, NULL);
740         }
741 
742         /* The mirror job has no requests in flight any more, but we need to
743          * drain potential other users of the BDS before changing the graph. */
744         assert(s->in_drain);
745         bdrv_drained_begin(target_bs);
746         /*
747          * Cannot use check_to_replace_node() here, because that would
748          * check for an op blocker on @to_replace, and we have our own
749          * there.
750          *
751          * TODO Pull out the writer lock from bdrv_replace_node() to here
752          */
753         bdrv_graph_rdlock_main_loop();
754         if (bdrv_recurse_can_replace(src, to_replace)) {
755             bdrv_replace_node(to_replace, target_bs, &local_err);
756         } else {
757             error_setg(&local_err, "Can no longer replace '%s' by '%s', "
758                        "because it can no longer be guaranteed that doing so "
759                        "would not lead to an abrupt change of visible data",
760                        to_replace->node_name, target_bs->node_name);
761         }
762         bdrv_graph_rdunlock_main_loop();
763         bdrv_drained_end(target_bs);
764         if (local_err) {
765             error_report_err(local_err);
766             ret = -EPERM;
767         }
768     }
769     if (s->to_replace) {
770         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
771         error_free(s->replace_blocker);
772         bdrv_unref(s->to_replace);
773     }
774     if (replace_aio_context) {
775         aio_context_release(replace_aio_context);
776     }
777     g_free(s->replaces);
778     bdrv_unref(target_bs);
779 
780     /*
781      * Remove the mirror filter driver from the graph. Before this, get rid of
782      * the blockers on the intermediate nodes so that the resulting state is
783      * valid.
784      */
785     block_job_remove_all_bdrv(bjob);
786     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
787 
788     bs_opaque->job = NULL;
789 
790     bdrv_drained_end(src);
791     bdrv_drained_end(mirror_top_bs);
792     s->in_drain = false;
793     bdrv_unref(mirror_top_bs);
794     bdrv_unref(src);
795 
796     aio_context_release(qemu_get_aio_context());
797 
798     return ret;
799 }
800 
801 static int mirror_prepare(Job *job)
802 {
803     return mirror_exit_common(job);
804 }
805 
806 static void mirror_abort(Job *job)
807 {
808     int ret = mirror_exit_common(job);
809     assert(ret == 0);
810 }
811 
812 static void coroutine_fn mirror_throttle(MirrorBlockJob *s)
813 {
814     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
815 
816     if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) {
817         s->last_pause_ns = now;
818         job_sleep_ns(&s->common.job, 0);
819     } else {
820         job_pause_point(&s->common.job);
821     }
822 }
823 
824 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
825 {
826     int64_t offset;
827     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
828     BlockDriverState *target_bs = blk_bs(s->target);
829     int ret;
830     int64_t count;
831 
832     if (s->zero_target) {
833         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
834             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
835             return 0;
836         }
837 
838         s->initial_zeroing_ongoing = true;
839         for (offset = 0; offset < s->bdev_length; ) {
840             int bytes = MIN(s->bdev_length - offset,
841                             QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
842 
843             mirror_throttle(s);
844 
845             if (job_is_cancelled(&s->common.job)) {
846                 s->initial_zeroing_ongoing = false;
847                 return 0;
848             }
849 
850             if (s->in_flight >= MAX_IN_FLIGHT) {
851                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
852                                    s->in_flight);
853                 mirror_wait_for_free_in_flight_slot(s);
854                 continue;
855             }
856 
857             mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO);
858             offset += bytes;
859         }
860 
861         mirror_wait_for_all_io(s);
862         s->initial_zeroing_ongoing = false;
863     }
864 
865     /* First part, loop on the sectors and initialize the dirty bitmap.  */
866     for (offset = 0; offset < s->bdev_length; ) {
867         /* Just to make sure we are not exceeding int limit. */
868         int bytes = MIN(s->bdev_length - offset,
869                         QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
870 
871         mirror_throttle(s);
872 
873         if (job_is_cancelled(&s->common.job)) {
874             return 0;
875         }
876 
877         WITH_GRAPH_RDLOCK_GUARD() {
878             ret = bdrv_is_allocated_above(bs, s->base_overlay, true, offset,
879                                           bytes, &count);
880         }
881         if (ret < 0) {
882             return ret;
883         }
884 
885         assert(count);
886         if (ret > 0) {
887             bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
888         }
889         offset += count;
890     }
891     return 0;
892 }
893 
894 /* Called when going out of the streaming phase to flush the bulk of the
895  * data to the medium, or just before completing.
896  */
897 static int coroutine_fn mirror_flush(MirrorBlockJob *s)
898 {
899     int ret = blk_co_flush(s->target);
900     if (ret < 0) {
901         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
902             s->ret = ret;
903         }
904     }
905     return ret;
906 }
907 
908 static int coroutine_fn mirror_run(Job *job, Error **errp)
909 {
910     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
911     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
912     MirrorBDSOpaque *mirror_top_opaque = s->mirror_top_bs->opaque;
913     BlockDriverState *target_bs = blk_bs(s->target);
914     bool need_drain = true;
915     BlockDeviceIoStatus iostatus;
916     int64_t length;
917     int64_t target_length;
918     BlockDriverInfo bdi;
919     char backing_filename[2]; /* we only need 2 characters because we are only
920                                  checking for a NULL string */
921     int ret = 0;
922 
923     if (job_is_cancelled(&s->common.job)) {
924         goto immediate_exit;
925     }
926 
927     bdrv_graph_co_rdlock();
928     s->bdev_length = bdrv_co_getlength(bs);
929     bdrv_graph_co_rdunlock();
930 
931     if (s->bdev_length < 0) {
932         ret = s->bdev_length;
933         goto immediate_exit;
934     }
935 
936     target_length = blk_co_getlength(s->target);
937     if (target_length < 0) {
938         ret = target_length;
939         goto immediate_exit;
940     }
941 
942     /* Active commit must resize the base image if its size differs from the
943      * active layer. */
944     if (s->base == blk_bs(s->target)) {
945         if (s->bdev_length > target_length) {
946             ret = blk_co_truncate(s->target, s->bdev_length, false,
947                                   PREALLOC_MODE_OFF, 0, NULL);
948             if (ret < 0) {
949                 goto immediate_exit;
950             }
951         }
952     } else if (s->bdev_length != target_length) {
953         error_setg(errp, "Source and target image have different sizes");
954         ret = -EINVAL;
955         goto immediate_exit;
956     }
957 
958     if (s->bdev_length == 0) {
959         /* Transition to the READY state and wait for complete. */
960         job_transition_to_ready(&s->common.job);
961         s->actively_synced = true;
962         while (!job_cancel_requested(&s->common.job) && !s->should_complete) {
963             job_yield(&s->common.job);
964         }
965         goto immediate_exit;
966     }
967 
968     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
969     s->in_flight_bitmap = bitmap_new(length);
970 
971     /* If we have no backing file yet in the destination, we cannot let
972      * the destination do COW.  Instead, we copy sectors around the
973      * dirty data if needed.  We need a bitmap to do that.
974      */
975     bdrv_get_backing_filename(target_bs, backing_filename,
976                               sizeof(backing_filename));
977     bdrv_graph_co_rdlock();
978     if (!bdrv_co_get_info(target_bs, &bdi) && bdi.cluster_size) {
979         s->target_cluster_size = bdi.cluster_size;
980     } else {
981         s->target_cluster_size = BDRV_SECTOR_SIZE;
982     }
983     bdrv_graph_co_rdunlock();
984     if (backing_filename[0] && !bdrv_backing_chain_next(target_bs) &&
985         s->granularity < s->target_cluster_size) {
986         s->buf_size = MAX(s->buf_size, s->target_cluster_size);
987         s->cow_bitmap = bitmap_new(length);
988     }
989     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
990 
991     s->buf = qemu_try_blockalign(bs, s->buf_size);
992     if (s->buf == NULL) {
993         ret = -ENOMEM;
994         goto immediate_exit;
995     }
996 
997     mirror_free_init(s);
998 
999     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1000     if (!s->is_none_mode) {
1001         ret = mirror_dirty_init(s);
1002         if (ret < 0 || job_is_cancelled(&s->common.job)) {
1003             goto immediate_exit;
1004         }
1005     }
1006 
1007     /*
1008      * Only now the job is fully initialised and mirror_top_bs should start
1009      * accessing it.
1010      */
1011     mirror_top_opaque->job = s;
1012 
1013     assert(!s->dbi);
1014     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
1015     for (;;) {
1016         int64_t cnt, delta;
1017         bool should_complete;
1018 
1019         if (s->ret < 0) {
1020             ret = s->ret;
1021             goto immediate_exit;
1022         }
1023 
1024         job_pause_point(&s->common.job);
1025 
1026         if (job_is_cancelled(&s->common.job)) {
1027             ret = 0;
1028             goto immediate_exit;
1029         }
1030 
1031         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1032         /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
1033          * the number of bytes currently being processed; together those are
1034          * the current remaining operation length */
1035         job_progress_set_remaining(&s->common.job,
1036                                    s->bytes_in_flight + cnt +
1037                                    s->active_write_bytes_in_flight);
1038 
1039         /* Note that even when no rate limit is applied we need to yield
1040          * periodically with no pending I/O so that bdrv_drain_all() returns.
1041          * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
1042          * an error, or when the source is clean, whichever comes first. */
1043         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
1044         WITH_JOB_LOCK_GUARD() {
1045             iostatus = s->common.iostatus;
1046         }
1047         if (delta < BLOCK_JOB_SLICE_TIME &&
1048             iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
1049             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
1050                 (cnt == 0 && s->in_flight > 0)) {
1051                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
1052                 mirror_wait_for_free_in_flight_slot(s);
1053                 continue;
1054             } else if (cnt != 0) {
1055                 mirror_iteration(s);
1056             }
1057         }
1058 
1059         should_complete = false;
1060         if (s->in_flight == 0 && cnt == 0) {
1061             trace_mirror_before_flush(s);
1062             if (!job_is_ready(&s->common.job)) {
1063                 if (mirror_flush(s) < 0) {
1064                     /* Go check s->ret.  */
1065                     continue;
1066                 }
1067                 /* We're out of the streaming phase.  From now on, if the job
1068                  * is cancelled we will actually complete all pending I/O and
1069                  * report completion.  This way, block-job-cancel will leave
1070                  * the target in a consistent state.
1071                  */
1072                 job_transition_to_ready(&s->common.job);
1073                 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) {
1074                     s->actively_synced = true;
1075                 }
1076             }
1077 
1078             should_complete = s->should_complete ||
1079                 job_cancel_requested(&s->common.job);
1080             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1081         }
1082 
1083         if (cnt == 0 && should_complete) {
1084             /* The dirty bitmap is not updated while operations are pending.
1085              * If we're about to exit, wait for pending operations before
1086              * calling bdrv_get_dirty_count(bs), or we may exit while the
1087              * source has dirty data to copy!
1088              *
1089              * Note that I/O can be submitted by the guest while
1090              * mirror_populate runs, so pause it now.  Before deciding
1091              * whether to switch to target check one last time if I/O has
1092              * come in the meanwhile, and if not flush the data to disk.
1093              */
1094             trace_mirror_before_drain(s, cnt);
1095 
1096             s->in_drain = true;
1097             bdrv_drained_begin(bs);
1098 
1099             /* Must be zero because we are drained */
1100             assert(s->in_active_write_counter == 0);
1101 
1102             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1103             if (cnt > 0 || mirror_flush(s) < 0) {
1104                 bdrv_drained_end(bs);
1105                 s->in_drain = false;
1106                 continue;
1107             }
1108 
1109             /* The two disks are in sync.  Exit and report successful
1110              * completion.
1111              */
1112             assert(QLIST_EMPTY(&bs->tracked_requests));
1113             need_drain = false;
1114             break;
1115         }
1116 
1117         if (job_is_ready(&s->common.job) && !should_complete) {
1118             if (s->in_flight == 0 && cnt == 0) {
1119                 trace_mirror_before_sleep(s, cnt, job_is_ready(&s->common.job),
1120                                           BLOCK_JOB_SLICE_TIME);
1121                 job_sleep_ns(&s->common.job, BLOCK_JOB_SLICE_TIME);
1122             }
1123         } else {
1124             block_job_ratelimit_sleep(&s->common);
1125         }
1126         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1127     }
1128 
1129 immediate_exit:
1130     if (s->in_flight > 0) {
1131         /* We get here only if something went wrong.  Either the job failed,
1132          * or it was cancelled prematurely so that we do not guarantee that
1133          * the target is a copy of the source.
1134          */
1135         assert(ret < 0 || job_is_cancelled(&s->common.job));
1136         assert(need_drain);
1137         mirror_wait_for_all_io(s);
1138     }
1139 
1140     assert(s->in_flight == 0);
1141     qemu_vfree(s->buf);
1142     g_free(s->cow_bitmap);
1143     g_free(s->in_flight_bitmap);
1144     bdrv_dirty_iter_free(s->dbi);
1145 
1146     if (need_drain) {
1147         s->in_drain = true;
1148         bdrv_drained_begin(bs);
1149     }
1150 
1151     return ret;
1152 }
1153 
1154 static void mirror_complete(Job *job, Error **errp)
1155 {
1156     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1157 
1158     if (!job_is_ready(job)) {
1159         error_setg(errp, "The active block job '%s' cannot be completed",
1160                    job->id);
1161         return;
1162     }
1163 
1164     /* block all operations on to_replace bs */
1165     if (s->replaces) {
1166         AioContext *replace_aio_context;
1167 
1168         s->to_replace = bdrv_find_node(s->replaces);
1169         if (!s->to_replace) {
1170             error_setg(errp, "Node name '%s' not found", s->replaces);
1171             return;
1172         }
1173 
1174         replace_aio_context = bdrv_get_aio_context(s->to_replace);
1175         aio_context_acquire(replace_aio_context);
1176 
1177         /* TODO Translate this into child freeze system. */
1178         error_setg(&s->replace_blocker,
1179                    "block device is in use by block-job-complete");
1180         bdrv_op_block_all(s->to_replace, s->replace_blocker);
1181         bdrv_ref(s->to_replace);
1182 
1183         aio_context_release(replace_aio_context);
1184     }
1185 
1186     s->should_complete = true;
1187 
1188     /* If the job is paused, it will be re-entered when it is resumed */
1189     WITH_JOB_LOCK_GUARD() {
1190         if (!job->paused) {
1191             job_enter_cond_locked(job, NULL);
1192         }
1193     }
1194 }
1195 
1196 static void coroutine_fn mirror_pause(Job *job)
1197 {
1198     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1199 
1200     mirror_wait_for_all_io(s);
1201 }
1202 
1203 static bool mirror_drained_poll(BlockJob *job)
1204 {
1205     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1206 
1207     /* If the job isn't paused nor cancelled, we can't be sure that it won't
1208      * issue more requests. We make an exception if we've reached this point
1209      * from one of our own drain sections, to avoid a deadlock waiting for
1210      * ourselves.
1211      */
1212     WITH_JOB_LOCK_GUARD() {
1213         if (!s->common.job.paused && !job_is_cancelled_locked(&job->job)
1214             && !s->in_drain) {
1215             return true;
1216         }
1217     }
1218 
1219     return !!s->in_flight;
1220 }
1221 
1222 static bool mirror_cancel(Job *job, bool force)
1223 {
1224     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1225     BlockDriverState *target = blk_bs(s->target);
1226 
1227     /*
1228      * Before the job is READY, we treat any cancellation like a
1229      * force-cancellation.
1230      */
1231     force = force || !job_is_ready(job);
1232 
1233     if (force) {
1234         bdrv_cancel_in_flight(target);
1235     }
1236     return force;
1237 }
1238 
1239 static bool commit_active_cancel(Job *job, bool force)
1240 {
1241     /* Same as above in mirror_cancel() */
1242     return force || !job_is_ready(job);
1243 }
1244 
1245 static const BlockJobDriver mirror_job_driver = {
1246     .job_driver = {
1247         .instance_size          = sizeof(MirrorBlockJob),
1248         .job_type               = JOB_TYPE_MIRROR,
1249         .free                   = block_job_free,
1250         .user_resume            = block_job_user_resume,
1251         .run                    = mirror_run,
1252         .prepare                = mirror_prepare,
1253         .abort                  = mirror_abort,
1254         .pause                  = mirror_pause,
1255         .complete               = mirror_complete,
1256         .cancel                 = mirror_cancel,
1257     },
1258     .drained_poll           = mirror_drained_poll,
1259 };
1260 
1261 static const BlockJobDriver commit_active_job_driver = {
1262     .job_driver = {
1263         .instance_size          = sizeof(MirrorBlockJob),
1264         .job_type               = JOB_TYPE_COMMIT,
1265         .free                   = block_job_free,
1266         .user_resume            = block_job_user_resume,
1267         .run                    = mirror_run,
1268         .prepare                = mirror_prepare,
1269         .abort                  = mirror_abort,
1270         .pause                  = mirror_pause,
1271         .complete               = mirror_complete,
1272         .cancel                 = commit_active_cancel,
1273     },
1274     .drained_poll           = mirror_drained_poll,
1275 };
1276 
1277 static void coroutine_fn
1278 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method,
1279                      uint64_t offset, uint64_t bytes,
1280                      QEMUIOVector *qiov, int flags)
1281 {
1282     int ret;
1283     size_t qiov_offset = 0;
1284     int64_t bitmap_offset, bitmap_end;
1285 
1286     if (!QEMU_IS_ALIGNED(offset, job->granularity) &&
1287         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset))
1288     {
1289             /*
1290              * Dirty unaligned padding: ignore it.
1291              *
1292              * Reasoning:
1293              * 1. If we copy it, we can't reset corresponding bit in
1294              *    dirty_bitmap as there may be some "dirty" bytes still not
1295              *    copied.
1296              * 2. It's already dirty, so skipping it we don't diverge mirror
1297              *    progress.
1298              *
1299              * Note, that because of this, guest write may have no contribution
1300              * into mirror converge, but that's not bad, as we have background
1301              * process of mirroring. If under some bad circumstances (high guest
1302              * IO load) background process starve, we will not converge anyway,
1303              * even if each write will contribute, as guest is not guaranteed to
1304              * rewrite the whole disk.
1305              */
1306             qiov_offset = QEMU_ALIGN_UP(offset, job->granularity) - offset;
1307             if (bytes <= qiov_offset) {
1308                 /* nothing to do after shrink */
1309                 return;
1310             }
1311             offset += qiov_offset;
1312             bytes -= qiov_offset;
1313     }
1314 
1315     if (!QEMU_IS_ALIGNED(offset + bytes, job->granularity) &&
1316         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset + bytes - 1))
1317     {
1318         uint64_t tail = (offset + bytes) % job->granularity;
1319 
1320         if (bytes <= tail) {
1321             /* nothing to do after shrink */
1322             return;
1323         }
1324         bytes -= tail;
1325     }
1326 
1327     /*
1328      * Tails are either clean or shrunk, so for bitmap resetting
1329      * we safely align the range down.
1330      */
1331     bitmap_offset = QEMU_ALIGN_UP(offset, job->granularity);
1332     bitmap_end = QEMU_ALIGN_DOWN(offset + bytes, job->granularity);
1333     if (bitmap_offset < bitmap_end) {
1334         bdrv_reset_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1335                                 bitmap_end - bitmap_offset);
1336     }
1337 
1338     job_progress_increase_remaining(&job->common.job, bytes);
1339     job->active_write_bytes_in_flight += bytes;
1340 
1341     switch (method) {
1342     case MIRROR_METHOD_COPY:
1343         ret = blk_co_pwritev_part(job->target, offset, bytes,
1344                                   qiov, qiov_offset, flags);
1345         break;
1346 
1347     case MIRROR_METHOD_ZERO:
1348         assert(!qiov);
1349         ret = blk_co_pwrite_zeroes(job->target, offset, bytes, flags);
1350         break;
1351 
1352     case MIRROR_METHOD_DISCARD:
1353         assert(!qiov);
1354         ret = blk_co_pdiscard(job->target, offset, bytes);
1355         break;
1356 
1357     default:
1358         abort();
1359     }
1360 
1361     job->active_write_bytes_in_flight -= bytes;
1362     if (ret >= 0) {
1363         job_progress_update(&job->common.job, bytes);
1364     } else {
1365         BlockErrorAction action;
1366 
1367         /*
1368          * We failed, so we should mark dirty the whole area, aligned up.
1369          * Note that we don't care about shrunk tails if any: they were dirty
1370          * at function start, and they must be still dirty, as we've locked
1371          * the region for in-flight op.
1372          */
1373         bitmap_offset = QEMU_ALIGN_DOWN(offset, job->granularity);
1374         bitmap_end = QEMU_ALIGN_UP(offset + bytes, job->granularity);
1375         bdrv_set_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1376                               bitmap_end - bitmap_offset);
1377         job->actively_synced = false;
1378 
1379         action = mirror_error_action(job, false, -ret);
1380         if (action == BLOCK_ERROR_ACTION_REPORT) {
1381             if (!job->ret) {
1382                 job->ret = ret;
1383             }
1384         }
1385     }
1386 }
1387 
1388 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s,
1389                                                    uint64_t offset,
1390                                                    uint64_t bytes)
1391 {
1392     MirrorOp *op;
1393     uint64_t start_chunk = offset / s->granularity;
1394     uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
1395 
1396     op = g_new(MirrorOp, 1);
1397     *op = (MirrorOp){
1398         .s                  = s,
1399         .offset             = offset,
1400         .bytes              = bytes,
1401         .is_active_write    = true,
1402         .is_in_flight       = true,
1403         .co                 = qemu_coroutine_self(),
1404     };
1405     qemu_co_queue_init(&op->waiting_requests);
1406     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
1407 
1408     s->in_active_write_counter++;
1409 
1410     /*
1411      * Wait for concurrent requests affecting the area.  If there are already
1412      * running requests that are copying off now-to-be stale data in the area,
1413      * we must wait for them to finish before we begin writing fresh data to the
1414      * target so that the write operations appear in the correct order.
1415      * Note that background requests (see mirror_iteration()) in contrast only
1416      * wait for conflicting requests at the start of the dirty area, and then
1417      * (based on the in_flight_bitmap) truncate the area to copy so it will not
1418      * conflict with any requests beyond that.  For active writes, however, we
1419      * cannot truncate that area.  The request from our parent must be blocked
1420      * until the area is copied in full.  Therefore, we must wait for the whole
1421      * area to become free of concurrent requests.
1422      */
1423     mirror_wait_on_conflicts(op, s, offset, bytes);
1424 
1425     bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1426 
1427     return op;
1428 }
1429 
1430 static void coroutine_fn GRAPH_RDLOCK active_write_settle(MirrorOp *op)
1431 {
1432     uint64_t start_chunk = op->offset / op->s->granularity;
1433     uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes,
1434                                       op->s->granularity);
1435 
1436     if (!--op->s->in_active_write_counter && op->s->actively_synced) {
1437         BdrvChild *source = op->s->mirror_top_bs->backing;
1438 
1439         if (QLIST_FIRST(&source->bs->parents) == source &&
1440             QLIST_NEXT(source, next_parent) == NULL)
1441         {
1442             /* Assert that we are back in sync once all active write
1443              * operations are settled.
1444              * Note that we can only assert this if the mirror node
1445              * is the source node's only parent. */
1446             assert(!bdrv_get_dirty_count(op->s->dirty_bitmap));
1447         }
1448     }
1449     bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1450     QTAILQ_REMOVE(&op->s->ops_in_flight, op, next);
1451     qemu_co_queue_restart_all(&op->waiting_requests);
1452     g_free(op);
1453 }
1454 
1455 static int coroutine_fn GRAPH_RDLOCK
1456 bdrv_mirror_top_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
1457                        QEMUIOVector *qiov, BdrvRequestFlags flags)
1458 {
1459     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1460 }
1461 
1462 static int coroutine_fn GRAPH_RDLOCK
1463 bdrv_mirror_top_do_write(BlockDriverState *bs, MirrorMethod method,
1464                          uint64_t offset, uint64_t bytes, QEMUIOVector *qiov,
1465                          int flags)
1466 {
1467     MirrorOp *op = NULL;
1468     MirrorBDSOpaque *s = bs->opaque;
1469     int ret = 0;
1470     bool copy_to_target = false;
1471 
1472     if (s->job) {
1473         copy_to_target = s->job->ret >= 0 &&
1474                          !job_is_cancelled(&s->job->common.job) &&
1475                          s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1476     }
1477 
1478     if (copy_to_target) {
1479         op = active_write_prepare(s->job, offset, bytes);
1480     }
1481 
1482     switch (method) {
1483     case MIRROR_METHOD_COPY:
1484         ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1485         break;
1486 
1487     case MIRROR_METHOD_ZERO:
1488         ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1489         break;
1490 
1491     case MIRROR_METHOD_DISCARD:
1492         ret = bdrv_co_pdiscard(bs->backing, offset, bytes);
1493         break;
1494 
1495     default:
1496         abort();
1497     }
1498 
1499     if (ret < 0) {
1500         goto out;
1501     }
1502 
1503     if (copy_to_target) {
1504         do_sync_target_write(s->job, method, offset, bytes, qiov, flags);
1505     }
1506 
1507 out:
1508     if (copy_to_target) {
1509         active_write_settle(op);
1510     }
1511     return ret;
1512 }
1513 
1514 static int coroutine_fn GRAPH_RDLOCK
1515 bdrv_mirror_top_pwritev(BlockDriverState *bs, int64_t offset, int64_t bytes,
1516                         QEMUIOVector *qiov, BdrvRequestFlags flags)
1517 {
1518     MirrorBDSOpaque *s = bs->opaque;
1519     QEMUIOVector bounce_qiov;
1520     void *bounce_buf;
1521     int ret = 0;
1522     bool copy_to_target = false;
1523 
1524     if (s->job) {
1525         copy_to_target = s->job->ret >= 0 &&
1526                          !job_is_cancelled(&s->job->common.job) &&
1527                          s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1528     }
1529 
1530     if (copy_to_target) {
1531         /* The guest might concurrently modify the data to write; but
1532          * the data on source and destination must match, so we have
1533          * to use a bounce buffer if we are going to write to the
1534          * target now. */
1535         bounce_buf = qemu_blockalign(bs, bytes);
1536         iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes);
1537 
1538         qemu_iovec_init(&bounce_qiov, 1);
1539         qemu_iovec_add(&bounce_qiov, bounce_buf, bytes);
1540         qiov = &bounce_qiov;
1541 
1542         flags &= ~BDRV_REQ_REGISTERED_BUF;
1543     }
1544 
1545     ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov,
1546                                    flags);
1547 
1548     if (copy_to_target) {
1549         qemu_iovec_destroy(&bounce_qiov);
1550         qemu_vfree(bounce_buf);
1551     }
1552 
1553     return ret;
1554 }
1555 
1556 static int coroutine_fn GRAPH_RDLOCK bdrv_mirror_top_flush(BlockDriverState *bs)
1557 {
1558     if (bs->backing == NULL) {
1559         /* we can be here after failed bdrv_append in mirror_start_job */
1560         return 0;
1561     }
1562     return bdrv_co_flush(bs->backing->bs);
1563 }
1564 
1565 static int coroutine_fn GRAPH_RDLOCK
1566 bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs, int64_t offset,
1567                               int64_t bytes, BdrvRequestFlags flags)
1568 {
1569     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL,
1570                                     flags);
1571 }
1572 
1573 static int coroutine_fn GRAPH_RDLOCK
1574 bdrv_mirror_top_pdiscard(BlockDriverState *bs, int64_t offset, int64_t bytes)
1575 {
1576     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes,
1577                                     NULL, 0);
1578 }
1579 
1580 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs)
1581 {
1582     if (bs->backing == NULL) {
1583         /* we can be here after failed bdrv_attach_child in
1584          * bdrv_set_backing_hd */
1585         return;
1586     }
1587     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1588             bs->backing->bs->filename);
1589 }
1590 
1591 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1592                                        BdrvChildRole role,
1593                                        BlockReopenQueue *reopen_queue,
1594                                        uint64_t perm, uint64_t shared,
1595                                        uint64_t *nperm, uint64_t *nshared)
1596 {
1597     MirrorBDSOpaque *s = bs->opaque;
1598 
1599     if (s->stop) {
1600         /*
1601          * If the job is to be stopped, we do not need to forward
1602          * anything to the real image.
1603          */
1604         *nperm = 0;
1605         *nshared = BLK_PERM_ALL;
1606         return;
1607     }
1608 
1609     bdrv_default_perms(bs, c, role, reopen_queue,
1610                        perm, shared, nperm, nshared);
1611 
1612     if (s->is_commit) {
1613         /*
1614          * For commit jobs, we cannot take CONSISTENT_READ, because
1615          * that permission is unshared for everything above the base
1616          * node (except for filters on the base node).
1617          * We also have to force-share the WRITE permission, or
1618          * otherwise we would block ourselves at the base node (if
1619          * writes are blocked for a node, they are also blocked for
1620          * its backing file).
1621          * (We could also share RESIZE, because it may be needed for
1622          * the target if its size is less than the top node's; but
1623          * bdrv_default_perms_for_cow() automatically shares RESIZE
1624          * for backing nodes if WRITE is shared, so there is no need
1625          * to do it here.)
1626          */
1627         *nperm &= ~BLK_PERM_CONSISTENT_READ;
1628         *nshared |= BLK_PERM_WRITE;
1629     }
1630 }
1631 
1632 /* Dummy node that provides consistent read to its users without requiring it
1633  * from its backing file and that allows writes on the backing file chain. */
1634 static BlockDriver bdrv_mirror_top = {
1635     .format_name                = "mirror_top",
1636     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1637     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1638     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1639     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1640     .bdrv_co_flush              = bdrv_mirror_top_flush,
1641     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1642     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1643 
1644     .is_filter                  = true,
1645     .filtered_child_is_backing  = true,
1646 };
1647 
1648 static BlockJob *mirror_start_job(
1649                              const char *job_id, BlockDriverState *bs,
1650                              int creation_flags, BlockDriverState *target,
1651                              const char *replaces, int64_t speed,
1652                              uint32_t granularity, int64_t buf_size,
1653                              BlockMirrorBackingMode backing_mode,
1654                              bool zero_target,
1655                              BlockdevOnError on_source_error,
1656                              BlockdevOnError on_target_error,
1657                              bool unmap,
1658                              BlockCompletionFunc *cb,
1659                              void *opaque,
1660                              const BlockJobDriver *driver,
1661                              bool is_none_mode, BlockDriverState *base,
1662                              bool auto_complete, const char *filter_node_name,
1663                              bool is_mirror, MirrorCopyMode copy_mode,
1664                              Error **errp)
1665 {
1666     MirrorBlockJob *s;
1667     MirrorBDSOpaque *bs_opaque;
1668     BlockDriverState *mirror_top_bs;
1669     bool target_is_backing;
1670     uint64_t target_perms, target_shared_perms;
1671     int ret;
1672 
1673     if (granularity == 0) {
1674         granularity = bdrv_get_default_bitmap_granularity(target);
1675     }
1676 
1677     assert(is_power_of_2(granularity));
1678 
1679     if (buf_size < 0) {
1680         error_setg(errp, "Invalid parameter 'buf-size'");
1681         return NULL;
1682     }
1683 
1684     if (buf_size == 0) {
1685         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1686     }
1687 
1688     if (bdrv_skip_filters(bs) == bdrv_skip_filters(target)) {
1689         error_setg(errp, "Can't mirror node into itself");
1690         return NULL;
1691     }
1692 
1693     target_is_backing = bdrv_chain_contains(bs, target);
1694 
1695     /* In the case of active commit, add dummy driver to provide consistent
1696      * reads on the top, while disabling it in the intermediate nodes, and make
1697      * the backing chain writable. */
1698     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1699                                          BDRV_O_RDWR, errp);
1700     if (mirror_top_bs == NULL) {
1701         return NULL;
1702     }
1703     if (!filter_node_name) {
1704         mirror_top_bs->implicit = true;
1705     }
1706 
1707     /* So that we can always drop this node */
1708     mirror_top_bs->never_freeze = true;
1709 
1710     mirror_top_bs->total_sectors = bs->total_sectors;
1711     mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
1712     mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
1713                                           BDRV_REQ_NO_FALLBACK;
1714     bs_opaque = g_new0(MirrorBDSOpaque, 1);
1715     mirror_top_bs->opaque = bs_opaque;
1716 
1717     bs_opaque->is_commit = target_is_backing;
1718 
1719     bdrv_drained_begin(bs);
1720     ret = bdrv_append(mirror_top_bs, bs, errp);
1721     bdrv_drained_end(bs);
1722 
1723     if (ret < 0) {
1724         bdrv_unref(mirror_top_bs);
1725         return NULL;
1726     }
1727 
1728     /* Make sure that the source is not resized while the job is running */
1729     s = block_job_create(job_id, driver, NULL, mirror_top_bs,
1730                          BLK_PERM_CONSISTENT_READ,
1731                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1732                          BLK_PERM_WRITE, speed,
1733                          creation_flags, cb, opaque, errp);
1734     if (!s) {
1735         goto fail;
1736     }
1737 
1738     /* The block job now has a reference to this node */
1739     bdrv_unref(mirror_top_bs);
1740 
1741     s->mirror_top_bs = mirror_top_bs;
1742 
1743     /* No resize for the target either; while the mirror is still running, a
1744      * consistent read isn't necessarily possible. We could possibly allow
1745      * writes and graph modifications, though it would likely defeat the
1746      * purpose of a mirror, so leave them blocked for now.
1747      *
1748      * In the case of active commit, things look a bit different, though,
1749      * because the target is an already populated backing file in active use.
1750      * We can allow anything except resize there.*/
1751 
1752     target_perms = BLK_PERM_WRITE;
1753     target_shared_perms = BLK_PERM_WRITE_UNCHANGED;
1754 
1755     if (target_is_backing) {
1756         int64_t bs_size, target_size;
1757         bs_size = bdrv_getlength(bs);
1758         if (bs_size < 0) {
1759             error_setg_errno(errp, -bs_size,
1760                              "Could not inquire top image size");
1761             goto fail;
1762         }
1763 
1764         target_size = bdrv_getlength(target);
1765         if (target_size < 0) {
1766             error_setg_errno(errp, -target_size,
1767                              "Could not inquire base image size");
1768             goto fail;
1769         }
1770 
1771         if (target_size < bs_size) {
1772             target_perms |= BLK_PERM_RESIZE;
1773         }
1774 
1775         target_shared_perms |= BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE;
1776     } else if (bdrv_chain_contains(bs, bdrv_skip_filters(target))) {
1777         /*
1778          * We may want to allow this in the future, but it would
1779          * require taking some extra care.
1780          */
1781         error_setg(errp, "Cannot mirror to a filter on top of a node in the "
1782                    "source's backing chain");
1783         goto fail;
1784     }
1785 
1786     s->target = blk_new(s->common.job.aio_context,
1787                         target_perms, target_shared_perms);
1788     ret = blk_insert_bs(s->target, target, errp);
1789     if (ret < 0) {
1790         goto fail;
1791     }
1792     if (is_mirror) {
1793         /* XXX: Mirror target could be a NBD server of target QEMU in the case
1794          * of non-shared block migration. To allow migration completion, we
1795          * have to allow "inactivate" of the target BB.  When that happens, we
1796          * know the job is drained, and the vcpus are stopped, so no write
1797          * operation will be performed. Block layer already has assertions to
1798          * ensure that. */
1799         blk_set_force_allow_inactivate(s->target);
1800     }
1801     blk_set_allow_aio_context_change(s->target, true);
1802     blk_set_disable_request_queuing(s->target, true);
1803 
1804     s->replaces = g_strdup(replaces);
1805     s->on_source_error = on_source_error;
1806     s->on_target_error = on_target_error;
1807     s->is_none_mode = is_none_mode;
1808     s->backing_mode = backing_mode;
1809     s->zero_target = zero_target;
1810     s->copy_mode = copy_mode;
1811     s->base = base;
1812     s->base_overlay = bdrv_find_overlay(bs, base);
1813     s->granularity = granularity;
1814     s->buf_size = ROUND_UP(buf_size, granularity);
1815     s->unmap = unmap;
1816     if (auto_complete) {
1817         s->should_complete = true;
1818     }
1819 
1820     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1821     if (!s->dirty_bitmap) {
1822         goto fail;
1823     }
1824     if (s->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING) {
1825         bdrv_disable_dirty_bitmap(s->dirty_bitmap);
1826     }
1827 
1828     ret = block_job_add_bdrv(&s->common, "source", bs, 0,
1829                              BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE |
1830                              BLK_PERM_CONSISTENT_READ,
1831                              errp);
1832     if (ret < 0) {
1833         goto fail;
1834     }
1835 
1836     /* Required permissions are already taken with blk_new() */
1837     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1838                        &error_abort);
1839 
1840     /* In commit_active_start() all intermediate nodes disappear, so
1841      * any jobs in them must be blocked */
1842     if (target_is_backing) {
1843         BlockDriverState *iter, *filtered_target;
1844         uint64_t iter_shared_perms;
1845 
1846         /*
1847          * The topmost node with
1848          * bdrv_skip_filters(filtered_target) == bdrv_skip_filters(target)
1849          */
1850         filtered_target = bdrv_cow_bs(bdrv_find_overlay(bs, target));
1851 
1852         assert(bdrv_skip_filters(filtered_target) ==
1853                bdrv_skip_filters(target));
1854 
1855         /*
1856          * XXX BLK_PERM_WRITE needs to be allowed so we don't block
1857          * ourselves at s->base (if writes are blocked for a node, they are
1858          * also blocked for its backing file). The other options would be a
1859          * second filter driver above s->base (== target).
1860          */
1861         iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE;
1862 
1863         for (iter = bdrv_filter_or_cow_bs(bs); iter != target;
1864              iter = bdrv_filter_or_cow_bs(iter))
1865         {
1866             if (iter == filtered_target) {
1867                 /*
1868                  * From here on, all nodes are filters on the base.
1869                  * This allows us to share BLK_PERM_CONSISTENT_READ.
1870                  */
1871                 iter_shared_perms |= BLK_PERM_CONSISTENT_READ;
1872             }
1873 
1874             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1875                                      iter_shared_perms, errp);
1876             if (ret < 0) {
1877                 goto fail;
1878             }
1879         }
1880 
1881         if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) {
1882             goto fail;
1883         }
1884     }
1885 
1886     QTAILQ_INIT(&s->ops_in_flight);
1887 
1888     trace_mirror_start(bs, s, opaque);
1889     job_start(&s->common.job);
1890 
1891     return &s->common;
1892 
1893 fail:
1894     if (s) {
1895         /* Make sure this BDS does not go away until we have completed the graph
1896          * changes below */
1897         bdrv_ref(mirror_top_bs);
1898 
1899         g_free(s->replaces);
1900         blk_unref(s->target);
1901         bs_opaque->job = NULL;
1902         if (s->dirty_bitmap) {
1903             bdrv_release_dirty_bitmap(s->dirty_bitmap);
1904         }
1905         job_early_fail(&s->common.job);
1906     }
1907 
1908     bs_opaque->stop = true;
1909     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
1910                              &error_abort);
1911     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
1912 
1913     bdrv_unref(mirror_top_bs);
1914 
1915     return NULL;
1916 }
1917 
1918 void mirror_start(const char *job_id, BlockDriverState *bs,
1919                   BlockDriverState *target, const char *replaces,
1920                   int creation_flags, int64_t speed,
1921                   uint32_t granularity, int64_t buf_size,
1922                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1923                   bool zero_target,
1924                   BlockdevOnError on_source_error,
1925                   BlockdevOnError on_target_error,
1926                   bool unmap, const char *filter_node_name,
1927                   MirrorCopyMode copy_mode, Error **errp)
1928 {
1929     bool is_none_mode;
1930     BlockDriverState *base;
1931 
1932     GLOBAL_STATE_CODE();
1933 
1934     if ((mode == MIRROR_SYNC_MODE_INCREMENTAL) ||
1935         (mode == MIRROR_SYNC_MODE_BITMAP)) {
1936         error_setg(errp, "Sync mode '%s' not supported",
1937                    MirrorSyncMode_str(mode));
1938         return;
1939     }
1940     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1941     base = mode == MIRROR_SYNC_MODE_TOP ? bdrv_backing_chain_next(bs) : NULL;
1942     mirror_start_job(job_id, bs, creation_flags, target, replaces,
1943                      speed, granularity, buf_size, backing_mode, zero_target,
1944                      on_source_error, on_target_error, unmap, NULL, NULL,
1945                      &mirror_job_driver, is_none_mode, base, false,
1946                      filter_node_name, true, copy_mode, errp);
1947 }
1948 
1949 BlockJob *commit_active_start(const char *job_id, BlockDriverState *bs,
1950                               BlockDriverState *base, int creation_flags,
1951                               int64_t speed, BlockdevOnError on_error,
1952                               const char *filter_node_name,
1953                               BlockCompletionFunc *cb, void *opaque,
1954                               bool auto_complete, Error **errp)
1955 {
1956     bool base_read_only;
1957     BlockJob *job;
1958 
1959     GLOBAL_STATE_CODE();
1960 
1961     base_read_only = bdrv_is_read_only(base);
1962 
1963     if (base_read_only) {
1964         if (bdrv_reopen_set_read_only(base, false, errp) < 0) {
1965             return NULL;
1966         }
1967     }
1968 
1969     job = mirror_start_job(
1970                      job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1971                      MIRROR_LEAVE_BACKING_CHAIN, false,
1972                      on_error, on_error, true, cb, opaque,
1973                      &commit_active_job_driver, false, base, auto_complete,
1974                      filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND,
1975                      errp);
1976     if (!job) {
1977         goto error_restore_flags;
1978     }
1979 
1980     return job;
1981 
1982 error_restore_flags:
1983     /* ignore error and errp for bdrv_reopen, because we want to propagate
1984      * the original error */
1985     if (base_read_only) {
1986         bdrv_reopen_set_read_only(base, true, NULL);
1987     }
1988     return NULL;
1989 }
1990