xref: /openbmc/qemu/block/mirror.c (revision f2b90109)
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 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_clusters(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 uint64_t 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     uint64_t delay_ns = 0, ret = 0;
480     /* At least the first dirty chunk is mirrored in one iteration. */
481     int nb_chunks = 1;
482     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
483     int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES);
484 
485     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
486     offset = bdrv_dirty_iter_next(s->dbi);
487     if (offset < 0) {
488         bdrv_set_dirty_iter(s->dbi, 0);
489         offset = bdrv_dirty_iter_next(s->dbi);
490         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
491         assert(offset >= 0);
492     }
493     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
494 
495     /*
496      * Wait for concurrent requests to @offset.  The next loop will limit the
497      * copied area based on in_flight_bitmap so we only copy an area that does
498      * not overlap with concurrent in-flight requests.  Still, we would like to
499      * copy something, so wait until there are at least no more requests to the
500      * very beginning of the area.
501      */
502     mirror_wait_on_conflicts(NULL, s, offset, 1);
503 
504     job_pause_point(&s->common.job);
505 
506     /* Find the number of consective dirty chunks following the first dirty
507      * one, and wait for in flight requests in them. */
508     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
509     while (nb_chunks * s->granularity < s->buf_size) {
510         int64_t next_dirty;
511         int64_t next_offset = offset + nb_chunks * s->granularity;
512         int64_t next_chunk = next_offset / s->granularity;
513         if (next_offset >= s->bdev_length ||
514             !bdrv_dirty_bitmap_get_locked(s->dirty_bitmap, next_offset)) {
515             break;
516         }
517         if (test_bit(next_chunk, s->in_flight_bitmap)) {
518             break;
519         }
520 
521         next_dirty = bdrv_dirty_iter_next(s->dbi);
522         if (next_dirty > next_offset || next_dirty < 0) {
523             /* The bitmap iterator's cache is stale, refresh it */
524             bdrv_set_dirty_iter(s->dbi, next_offset);
525             next_dirty = bdrv_dirty_iter_next(s->dbi);
526         }
527         assert(next_dirty == next_offset);
528         nb_chunks++;
529     }
530 
531     /* Clear dirty bits before querying the block status, because
532      * calling bdrv_block_status_above could yield - if some blocks are
533      * marked dirty in this window, we need to know.
534      */
535     bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset,
536                                    nb_chunks * s->granularity);
537     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
538 
539     /* Before claiming an area in the in-flight bitmap, we have to
540      * create a MirrorOp for it so that conflicting requests can wait
541      * for it.  mirror_perform() will create the real MirrorOps later,
542      * for now we just create a pseudo operation that will wake up all
543      * conflicting requests once all real operations have been
544      * launched. */
545     pseudo_op = g_new(MirrorOp, 1);
546     *pseudo_op = (MirrorOp){
547         .offset         = offset,
548         .bytes          = nb_chunks * s->granularity,
549         .is_pseudo_op   = true,
550     };
551     qemu_co_queue_init(&pseudo_op->waiting_requests);
552     QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next);
553 
554     bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks);
555     while (nb_chunks > 0 && offset < s->bdev_length) {
556         int ret;
557         int64_t io_bytes;
558         int64_t io_bytes_acct;
559         MirrorMethod mirror_method = MIRROR_METHOD_COPY;
560 
561         assert(!(offset % s->granularity));
562         WITH_GRAPH_RDLOCK_GUARD() {
563             ret = bdrv_block_status_above(source, NULL, offset,
564                                         nb_chunks * s->granularity,
565                                         &io_bytes, NULL, NULL);
566         }
567         if (ret < 0) {
568             io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes);
569         } else if (ret & BDRV_BLOCK_DATA) {
570             io_bytes = MIN(io_bytes, max_io_bytes);
571         }
572 
573         io_bytes -= io_bytes % s->granularity;
574         if (io_bytes < s->granularity) {
575             io_bytes = s->granularity;
576         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
577             int64_t target_offset;
578             int64_t target_bytes;
579             bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes,
580                                    &target_offset, &target_bytes);
581             if (target_offset == offset &&
582                 target_bytes == io_bytes) {
583                 mirror_method = ret & BDRV_BLOCK_ZERO ?
584                                     MIRROR_METHOD_ZERO :
585                                     MIRROR_METHOD_DISCARD;
586             }
587         }
588 
589         while (s->in_flight >= MAX_IN_FLIGHT) {
590             trace_mirror_yield_in_flight(s, offset, s->in_flight);
591             mirror_wait_for_free_in_flight_slot(s);
592         }
593 
594         if (s->ret < 0) {
595             ret = 0;
596             goto fail;
597         }
598 
599         io_bytes = mirror_clip_bytes(s, offset, io_bytes);
600         io_bytes = mirror_perform(s, offset, io_bytes, mirror_method);
601         if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) {
602             io_bytes_acct = 0;
603         } else {
604             io_bytes_acct = io_bytes;
605         }
606         assert(io_bytes);
607         offset += io_bytes;
608         nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity);
609         delay_ns = block_job_ratelimit_get_delay(&s->common, io_bytes_acct);
610     }
611 
612     ret = delay_ns;
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     return ret;
619 }
620 
621 static void mirror_free_init(MirrorBlockJob *s)
622 {
623     int granularity = s->granularity;
624     size_t buf_size = s->buf_size;
625     uint8_t *buf = s->buf;
626 
627     assert(s->buf_free_count == 0);
628     QSIMPLEQ_INIT(&s->buf_free);
629     while (buf_size != 0) {
630         MirrorBuffer *cur = (MirrorBuffer *)buf;
631         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
632         s->buf_free_count++;
633         buf_size -= granularity;
634         buf += granularity;
635     }
636 }
637 
638 /* This is also used for the .pause callback. There is no matching
639  * mirror_resume() because mirror_run() will begin iterating again
640  * when the job is resumed.
641  */
642 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s)
643 {
644     while (s->in_flight > 0) {
645         mirror_wait_for_free_in_flight_slot(s);
646     }
647 }
648 
649 /**
650  * mirror_exit_common: handle both abort() and prepare() cases.
651  * for .prepare, returns 0 on success and -errno on failure.
652  * for .abort cases, denoted by abort = true, MUST return 0.
653  */
654 static int mirror_exit_common(Job *job)
655 {
656     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
657     BlockJob *bjob = &s->common;
658     MirrorBDSOpaque *bs_opaque;
659     AioContext *replace_aio_context = NULL;
660     BlockDriverState *src;
661     BlockDriverState *target_bs;
662     BlockDriverState *mirror_top_bs;
663     Error *local_err = NULL;
664     bool abort = job->ret < 0;
665     int ret = 0;
666 
667     if (s->prepared) {
668         return 0;
669     }
670     s->prepared = true;
671 
672     mirror_top_bs = s->mirror_top_bs;
673     bs_opaque = mirror_top_bs->opaque;
674     src = mirror_top_bs->backing->bs;
675     target_bs = blk_bs(s->target);
676 
677     if (bdrv_chain_contains(src, target_bs)) {
678         bdrv_unfreeze_backing_chain(mirror_top_bs, target_bs);
679     }
680 
681     bdrv_release_dirty_bitmap(s->dirty_bitmap);
682 
683     /* Make sure that the source BDS doesn't go away during bdrv_replace_node,
684      * before we can call bdrv_drained_end */
685     bdrv_ref(src);
686     bdrv_ref(mirror_top_bs);
687     bdrv_ref(target_bs);
688 
689     /*
690      * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
691      * inserting target_bs at s->to_replace, where we might not be able to get
692      * these permissions.
693      */
694     blk_unref(s->target);
695     s->target = NULL;
696 
697     /* We don't access the source any more. Dropping any WRITE/RESIZE is
698      * required before it could become a backing file of target_bs. Not having
699      * these permissions any more means that we can't allow any new requests on
700      * mirror_top_bs from now on, so keep it drained. */
701     bdrv_drained_begin(mirror_top_bs);
702     bs_opaque->stop = true;
703     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
704                              &error_abort);
705     if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
706         BlockDriverState *backing = s->is_none_mode ? src : s->base;
707         BlockDriverState *unfiltered_target = bdrv_skip_filters(target_bs);
708 
709         if (bdrv_cow_bs(unfiltered_target) != backing) {
710             bdrv_set_backing_hd(unfiltered_target, backing, &local_err);
711             if (local_err) {
712                 error_report_err(local_err);
713                 local_err = NULL;
714                 ret = -EPERM;
715             }
716         }
717     } else if (!abort && s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
718         assert(!bdrv_backing_chain_next(target_bs));
719         ret = bdrv_open_backing_file(bdrv_skip_filters(target_bs), NULL,
720                                      "backing", &local_err);
721         if (ret < 0) {
722             error_report_err(local_err);
723             local_err = NULL;
724         }
725     }
726 
727     if (s->to_replace) {
728         replace_aio_context = bdrv_get_aio_context(s->to_replace);
729         aio_context_acquire(replace_aio_context);
730     }
731 
732     if (s->should_complete && !abort) {
733         BlockDriverState *to_replace = s->to_replace ?: src;
734         bool ro = bdrv_is_read_only(to_replace);
735 
736         if (ro != bdrv_is_read_only(target_bs)) {
737             bdrv_reopen_set_read_only(target_bs, ro, NULL);
738         }
739 
740         /* The mirror job has no requests in flight any more, but we need to
741          * drain potential other users of the BDS before changing the graph. */
742         assert(s->in_drain);
743         bdrv_drained_begin(target_bs);
744         /*
745          * Cannot use check_to_replace_node() here, because that would
746          * check for an op blocker on @to_replace, and we have our own
747          * there.
748          */
749         if (bdrv_recurse_can_replace(src, to_replace)) {
750             bdrv_replace_node(to_replace, target_bs, &local_err);
751         } else {
752             error_setg(&local_err, "Can no longer replace '%s' by '%s', "
753                        "because it can no longer be guaranteed that doing so "
754                        "would not lead to an abrupt change of visible data",
755                        to_replace->node_name, target_bs->node_name);
756         }
757         bdrv_drained_end(target_bs);
758         if (local_err) {
759             error_report_err(local_err);
760             ret = -EPERM;
761         }
762     }
763     if (s->to_replace) {
764         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
765         error_free(s->replace_blocker);
766         bdrv_unref(s->to_replace);
767     }
768     if (replace_aio_context) {
769         aio_context_release(replace_aio_context);
770     }
771     g_free(s->replaces);
772     bdrv_unref(target_bs);
773 
774     /*
775      * Remove the mirror filter driver from the graph. Before this, get rid of
776      * the blockers on the intermediate nodes so that the resulting state is
777      * valid.
778      */
779     block_job_remove_all_bdrv(bjob);
780     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
781 
782     bs_opaque->job = NULL;
783 
784     bdrv_drained_end(src);
785     bdrv_drained_end(mirror_top_bs);
786     s->in_drain = false;
787     bdrv_unref(mirror_top_bs);
788     bdrv_unref(src);
789 
790     return ret;
791 }
792 
793 static int mirror_prepare(Job *job)
794 {
795     return mirror_exit_common(job);
796 }
797 
798 static void mirror_abort(Job *job)
799 {
800     int ret = mirror_exit_common(job);
801     assert(ret == 0);
802 }
803 
804 static void coroutine_fn mirror_throttle(MirrorBlockJob *s)
805 {
806     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
807 
808     if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) {
809         s->last_pause_ns = now;
810         job_sleep_ns(&s->common.job, 0);
811     } else {
812         job_pause_point(&s->common.job);
813     }
814 }
815 
816 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
817 {
818     int64_t offset;
819     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
820     BlockDriverState *target_bs = blk_bs(s->target);
821     int ret;
822     int64_t count;
823 
824     if (s->zero_target) {
825         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
826             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
827             return 0;
828         }
829 
830         s->initial_zeroing_ongoing = true;
831         for (offset = 0; offset < s->bdev_length; ) {
832             int bytes = MIN(s->bdev_length - offset,
833                             QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
834 
835             mirror_throttle(s);
836 
837             if (job_is_cancelled(&s->common.job)) {
838                 s->initial_zeroing_ongoing = false;
839                 return 0;
840             }
841 
842             if (s->in_flight >= MAX_IN_FLIGHT) {
843                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
844                                    s->in_flight);
845                 mirror_wait_for_free_in_flight_slot(s);
846                 continue;
847             }
848 
849             mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO);
850             offset += bytes;
851         }
852 
853         mirror_wait_for_all_io(s);
854         s->initial_zeroing_ongoing = false;
855     }
856 
857     /* First part, loop on the sectors and initialize the dirty bitmap.  */
858     for (offset = 0; offset < s->bdev_length; ) {
859         /* Just to make sure we are not exceeding int limit. */
860         int bytes = MIN(s->bdev_length - offset,
861                         QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
862 
863         mirror_throttle(s);
864 
865         if (job_is_cancelled(&s->common.job)) {
866             return 0;
867         }
868 
869         WITH_GRAPH_RDLOCK_GUARD() {
870             ret = bdrv_is_allocated_above(bs, s->base_overlay, true, offset,
871                                           bytes, &count);
872         }
873         if (ret < 0) {
874             return ret;
875         }
876 
877         assert(count);
878         if (ret > 0) {
879             bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
880         }
881         offset += count;
882     }
883     return 0;
884 }
885 
886 /* Called when going out of the streaming phase to flush the bulk of the
887  * data to the medium, or just before completing.
888  */
889 static int mirror_flush(MirrorBlockJob *s)
890 {
891     int ret = blk_flush(s->target);
892     if (ret < 0) {
893         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
894             s->ret = ret;
895         }
896     }
897     return ret;
898 }
899 
900 static int coroutine_fn mirror_run(Job *job, Error **errp)
901 {
902     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
903     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
904     MirrorBDSOpaque *mirror_top_opaque = s->mirror_top_bs->opaque;
905     BlockDriverState *target_bs = blk_bs(s->target);
906     bool need_drain = true;
907     BlockDeviceIoStatus iostatus;
908     int64_t length;
909     int64_t target_length;
910     BlockDriverInfo bdi;
911     char backing_filename[2]; /* we only need 2 characters because we are only
912                                  checking for a NULL string */
913     int ret = 0;
914 
915     if (job_is_cancelled(&s->common.job)) {
916         goto immediate_exit;
917     }
918 
919     bdrv_graph_co_rdlock();
920     s->bdev_length = bdrv_co_getlength(bs);
921     bdrv_graph_co_rdunlock();
922 
923     if (s->bdev_length < 0) {
924         ret = s->bdev_length;
925         goto immediate_exit;
926     }
927 
928     target_length = blk_co_getlength(s->target);
929     if (target_length < 0) {
930         ret = target_length;
931         goto immediate_exit;
932     }
933 
934     /* Active commit must resize the base image if its size differs from the
935      * active layer. */
936     if (s->base == blk_bs(s->target)) {
937         if (s->bdev_length > target_length) {
938             ret = blk_co_truncate(s->target, s->bdev_length, false,
939                                   PREALLOC_MODE_OFF, 0, NULL);
940             if (ret < 0) {
941                 goto immediate_exit;
942             }
943         }
944     } else if (s->bdev_length != target_length) {
945         error_setg(errp, "Source and target image have different sizes");
946         ret = -EINVAL;
947         goto immediate_exit;
948     }
949 
950     if (s->bdev_length == 0) {
951         /* Transition to the READY state and wait for complete. */
952         job_transition_to_ready(&s->common.job);
953         s->actively_synced = true;
954         while (!job_cancel_requested(&s->common.job) && !s->should_complete) {
955             job_yield(&s->common.job);
956         }
957         goto immediate_exit;
958     }
959 
960     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
961     s->in_flight_bitmap = bitmap_new(length);
962 
963     /* If we have no backing file yet in the destination, we cannot let
964      * the destination do COW.  Instead, we copy sectors around the
965      * dirty data if needed.  We need a bitmap to do that.
966      */
967     bdrv_get_backing_filename(target_bs, backing_filename,
968                               sizeof(backing_filename));
969     if (!bdrv_co_get_info(target_bs, &bdi) && bdi.cluster_size) {
970         s->target_cluster_size = bdi.cluster_size;
971     } else {
972         s->target_cluster_size = BDRV_SECTOR_SIZE;
973     }
974     if (backing_filename[0] && !bdrv_backing_chain_next(target_bs) &&
975         s->granularity < s->target_cluster_size) {
976         s->buf_size = MAX(s->buf_size, s->target_cluster_size);
977         s->cow_bitmap = bitmap_new(length);
978     }
979     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
980 
981     s->buf = qemu_try_blockalign(bs, s->buf_size);
982     if (s->buf == NULL) {
983         ret = -ENOMEM;
984         goto immediate_exit;
985     }
986 
987     mirror_free_init(s);
988 
989     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
990     if (!s->is_none_mode) {
991         ret = mirror_dirty_init(s);
992         if (ret < 0 || job_is_cancelled(&s->common.job)) {
993             goto immediate_exit;
994         }
995     }
996 
997     /*
998      * Only now the job is fully initialised and mirror_top_bs should start
999      * accessing it.
1000      */
1001     mirror_top_opaque->job = s;
1002 
1003     assert(!s->dbi);
1004     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
1005     for (;;) {
1006         uint64_t delay_ns = 0;
1007         int64_t cnt, delta;
1008         bool should_complete;
1009 
1010         if (s->ret < 0) {
1011             ret = s->ret;
1012             goto immediate_exit;
1013         }
1014 
1015         job_pause_point(&s->common.job);
1016 
1017         if (job_is_cancelled(&s->common.job)) {
1018             ret = 0;
1019             goto immediate_exit;
1020         }
1021 
1022         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1023         /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
1024          * the number of bytes currently being processed; together those are
1025          * the current remaining operation length */
1026         job_progress_set_remaining(&s->common.job,
1027                                    s->bytes_in_flight + cnt +
1028                                    s->active_write_bytes_in_flight);
1029 
1030         /* Note that even when no rate limit is applied we need to yield
1031          * periodically with no pending I/O so that bdrv_drain_all() returns.
1032          * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
1033          * an error, or when the source is clean, whichever comes first. */
1034         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
1035         WITH_JOB_LOCK_GUARD() {
1036             iostatus = s->common.iostatus;
1037         }
1038         if (delta < BLOCK_JOB_SLICE_TIME &&
1039             iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
1040             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
1041                 (cnt == 0 && s->in_flight > 0)) {
1042                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
1043                 mirror_wait_for_free_in_flight_slot(s);
1044                 continue;
1045             } else if (cnt != 0) {
1046                 delay_ns = mirror_iteration(s);
1047             }
1048         }
1049 
1050         should_complete = false;
1051         if (s->in_flight == 0 && cnt == 0) {
1052             trace_mirror_before_flush(s);
1053             if (!job_is_ready(&s->common.job)) {
1054                 if (mirror_flush(s) < 0) {
1055                     /* Go check s->ret.  */
1056                     continue;
1057                 }
1058                 /* We're out of the streaming phase.  From now on, if the job
1059                  * is cancelled we will actually complete all pending I/O and
1060                  * report completion.  This way, block-job-cancel will leave
1061                  * the target in a consistent state.
1062                  */
1063                 job_transition_to_ready(&s->common.job);
1064                 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) {
1065                     s->actively_synced = true;
1066                 }
1067             }
1068 
1069             should_complete = s->should_complete ||
1070                 job_cancel_requested(&s->common.job);
1071             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1072         }
1073 
1074         if (cnt == 0 && should_complete) {
1075             /* The dirty bitmap is not updated while operations are pending.
1076              * If we're about to exit, wait for pending operations before
1077              * calling bdrv_get_dirty_count(bs), or we may exit while the
1078              * source has dirty data to copy!
1079              *
1080              * Note that I/O can be submitted by the guest while
1081              * mirror_populate runs, so pause it now.  Before deciding
1082              * whether to switch to target check one last time if I/O has
1083              * come in the meanwhile, and if not flush the data to disk.
1084              */
1085             trace_mirror_before_drain(s, cnt);
1086 
1087             s->in_drain = true;
1088             bdrv_drained_begin(bs);
1089 
1090             /* Must be zero because we are drained */
1091             assert(s->in_active_write_counter == 0);
1092 
1093             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1094             if (cnt > 0 || mirror_flush(s) < 0) {
1095                 bdrv_drained_end(bs);
1096                 s->in_drain = false;
1097                 continue;
1098             }
1099 
1100             /* The two disks are in sync.  Exit and report successful
1101              * completion.
1102              */
1103             assert(QLIST_EMPTY(&bs->tracked_requests));
1104             need_drain = false;
1105             break;
1106         }
1107 
1108         if (job_is_ready(&s->common.job) && !should_complete) {
1109             delay_ns = (s->in_flight == 0 &&
1110                         cnt == 0 ? BLOCK_JOB_SLICE_TIME : 0);
1111         }
1112         trace_mirror_before_sleep(s, cnt, job_is_ready(&s->common.job),
1113                                   delay_ns);
1114         job_sleep_ns(&s->common.job, delay_ns);
1115         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1116     }
1117 
1118 immediate_exit:
1119     if (s->in_flight > 0) {
1120         /* We get here only if something went wrong.  Either the job failed,
1121          * or it was cancelled prematurely so that we do not guarantee that
1122          * the target is a copy of the source.
1123          */
1124         assert(ret < 0 || job_is_cancelled(&s->common.job));
1125         assert(need_drain);
1126         mirror_wait_for_all_io(s);
1127     }
1128 
1129     assert(s->in_flight == 0);
1130     qemu_vfree(s->buf);
1131     g_free(s->cow_bitmap);
1132     g_free(s->in_flight_bitmap);
1133     bdrv_dirty_iter_free(s->dbi);
1134 
1135     if (need_drain) {
1136         s->in_drain = true;
1137         bdrv_drained_begin(bs);
1138     }
1139 
1140     return ret;
1141 }
1142 
1143 static void mirror_complete(Job *job, Error **errp)
1144 {
1145     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1146 
1147     if (!job_is_ready(job)) {
1148         error_setg(errp, "The active block job '%s' cannot be completed",
1149                    job->id);
1150         return;
1151     }
1152 
1153     /* block all operations on to_replace bs */
1154     if (s->replaces) {
1155         AioContext *replace_aio_context;
1156 
1157         s->to_replace = bdrv_find_node(s->replaces);
1158         if (!s->to_replace) {
1159             error_setg(errp, "Node name '%s' not found", s->replaces);
1160             return;
1161         }
1162 
1163         replace_aio_context = bdrv_get_aio_context(s->to_replace);
1164         aio_context_acquire(replace_aio_context);
1165 
1166         /* TODO Translate this into child freeze system. */
1167         error_setg(&s->replace_blocker,
1168                    "block device is in use by block-job-complete");
1169         bdrv_op_block_all(s->to_replace, s->replace_blocker);
1170         bdrv_ref(s->to_replace);
1171 
1172         aio_context_release(replace_aio_context);
1173     }
1174 
1175     s->should_complete = true;
1176 
1177     /* If the job is paused, it will be re-entered when it is resumed */
1178     WITH_JOB_LOCK_GUARD() {
1179         if (!job->paused) {
1180             job_enter_cond_locked(job, NULL);
1181         }
1182     }
1183 }
1184 
1185 static void coroutine_fn mirror_pause(Job *job)
1186 {
1187     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1188 
1189     mirror_wait_for_all_io(s);
1190 }
1191 
1192 static bool mirror_drained_poll(BlockJob *job)
1193 {
1194     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1195 
1196     /* If the job isn't paused nor cancelled, we can't be sure that it won't
1197      * issue more requests. We make an exception if we've reached this point
1198      * from one of our own drain sections, to avoid a deadlock waiting for
1199      * ourselves.
1200      */
1201     WITH_JOB_LOCK_GUARD() {
1202         if (!s->common.job.paused && !job_is_cancelled_locked(&job->job)
1203             && !s->in_drain) {
1204             return true;
1205         }
1206     }
1207 
1208     return !!s->in_flight;
1209 }
1210 
1211 static bool mirror_cancel(Job *job, bool force)
1212 {
1213     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1214     BlockDriverState *target = blk_bs(s->target);
1215 
1216     /*
1217      * Before the job is READY, we treat any cancellation like a
1218      * force-cancellation.
1219      */
1220     force = force || !job_is_ready(job);
1221 
1222     if (force) {
1223         bdrv_cancel_in_flight(target);
1224     }
1225     return force;
1226 }
1227 
1228 static bool commit_active_cancel(Job *job, bool force)
1229 {
1230     /* Same as above in mirror_cancel() */
1231     return force || !job_is_ready(job);
1232 }
1233 
1234 static const BlockJobDriver mirror_job_driver = {
1235     .job_driver = {
1236         .instance_size          = sizeof(MirrorBlockJob),
1237         .job_type               = JOB_TYPE_MIRROR,
1238         .free                   = block_job_free,
1239         .user_resume            = block_job_user_resume,
1240         .run                    = mirror_run,
1241         .prepare                = mirror_prepare,
1242         .abort                  = mirror_abort,
1243         .pause                  = mirror_pause,
1244         .complete               = mirror_complete,
1245         .cancel                 = mirror_cancel,
1246     },
1247     .drained_poll           = mirror_drained_poll,
1248 };
1249 
1250 static const BlockJobDriver commit_active_job_driver = {
1251     .job_driver = {
1252         .instance_size          = sizeof(MirrorBlockJob),
1253         .job_type               = JOB_TYPE_COMMIT,
1254         .free                   = block_job_free,
1255         .user_resume            = block_job_user_resume,
1256         .run                    = mirror_run,
1257         .prepare                = mirror_prepare,
1258         .abort                  = mirror_abort,
1259         .pause                  = mirror_pause,
1260         .complete               = mirror_complete,
1261         .cancel                 = commit_active_cancel,
1262     },
1263     .drained_poll           = mirror_drained_poll,
1264 };
1265 
1266 static void coroutine_fn
1267 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method,
1268                      uint64_t offset, uint64_t bytes,
1269                      QEMUIOVector *qiov, int flags)
1270 {
1271     int ret;
1272     size_t qiov_offset = 0;
1273     int64_t bitmap_offset, bitmap_end;
1274 
1275     if (!QEMU_IS_ALIGNED(offset, job->granularity) &&
1276         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset))
1277     {
1278             /*
1279              * Dirty unaligned padding: ignore it.
1280              *
1281              * Reasoning:
1282              * 1. If we copy it, we can't reset corresponding bit in
1283              *    dirty_bitmap as there may be some "dirty" bytes still not
1284              *    copied.
1285              * 2. It's already dirty, so skipping it we don't diverge mirror
1286              *    progress.
1287              *
1288              * Note, that because of this, guest write may have no contribution
1289              * into mirror converge, but that's not bad, as we have background
1290              * process of mirroring. If under some bad circumstances (high guest
1291              * IO load) background process starve, we will not converge anyway,
1292              * even if each write will contribute, as guest is not guaranteed to
1293              * rewrite the whole disk.
1294              */
1295             qiov_offset = QEMU_ALIGN_UP(offset, job->granularity) - offset;
1296             if (bytes <= qiov_offset) {
1297                 /* nothing to do after shrink */
1298                 return;
1299             }
1300             offset += qiov_offset;
1301             bytes -= qiov_offset;
1302     }
1303 
1304     if (!QEMU_IS_ALIGNED(offset + bytes, job->granularity) &&
1305         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset + bytes - 1))
1306     {
1307         uint64_t tail = (offset + bytes) % job->granularity;
1308 
1309         if (bytes <= tail) {
1310             /* nothing to do after shrink */
1311             return;
1312         }
1313         bytes -= tail;
1314     }
1315 
1316     /*
1317      * Tails are either clean or shrunk, so for bitmap resetting
1318      * we safely align the range down.
1319      */
1320     bitmap_offset = QEMU_ALIGN_UP(offset, job->granularity);
1321     bitmap_end = QEMU_ALIGN_DOWN(offset + bytes, job->granularity);
1322     if (bitmap_offset < bitmap_end) {
1323         bdrv_reset_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1324                                 bitmap_end - bitmap_offset);
1325     }
1326 
1327     job_progress_increase_remaining(&job->common.job, bytes);
1328     job->active_write_bytes_in_flight += bytes;
1329 
1330     switch (method) {
1331     case MIRROR_METHOD_COPY:
1332         ret = blk_co_pwritev_part(job->target, offset, bytes,
1333                                   qiov, qiov_offset, flags);
1334         break;
1335 
1336     case MIRROR_METHOD_ZERO:
1337         assert(!qiov);
1338         ret = blk_co_pwrite_zeroes(job->target, offset, bytes, flags);
1339         break;
1340 
1341     case MIRROR_METHOD_DISCARD:
1342         assert(!qiov);
1343         ret = blk_co_pdiscard(job->target, offset, bytes);
1344         break;
1345 
1346     default:
1347         abort();
1348     }
1349 
1350     job->active_write_bytes_in_flight -= bytes;
1351     if (ret >= 0) {
1352         job_progress_update(&job->common.job, bytes);
1353     } else {
1354         BlockErrorAction action;
1355 
1356         /*
1357          * We failed, so we should mark dirty the whole area, aligned up.
1358          * Note that we don't care about shrunk tails if any: they were dirty
1359          * at function start, and they must be still dirty, as we've locked
1360          * the region for in-flight op.
1361          */
1362         bitmap_offset = QEMU_ALIGN_DOWN(offset, job->granularity);
1363         bitmap_end = QEMU_ALIGN_UP(offset + bytes, job->granularity);
1364         bdrv_set_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1365                               bitmap_end - bitmap_offset);
1366         job->actively_synced = false;
1367 
1368         action = mirror_error_action(job, false, -ret);
1369         if (action == BLOCK_ERROR_ACTION_REPORT) {
1370             if (!job->ret) {
1371                 job->ret = ret;
1372             }
1373         }
1374     }
1375 }
1376 
1377 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s,
1378                                                    uint64_t offset,
1379                                                    uint64_t bytes)
1380 {
1381     MirrorOp *op;
1382     uint64_t start_chunk = offset / s->granularity;
1383     uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
1384 
1385     op = g_new(MirrorOp, 1);
1386     *op = (MirrorOp){
1387         .s                  = s,
1388         .offset             = offset,
1389         .bytes              = bytes,
1390         .is_active_write    = true,
1391         .is_in_flight       = true,
1392         .co                 = qemu_coroutine_self(),
1393     };
1394     qemu_co_queue_init(&op->waiting_requests);
1395     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
1396 
1397     s->in_active_write_counter++;
1398 
1399     /*
1400      * Wait for concurrent requests affecting the area.  If there are already
1401      * running requests that are copying off now-to-be stale data in the area,
1402      * we must wait for them to finish before we begin writing fresh data to the
1403      * target so that the write operations appear in the correct order.
1404      * Note that background requests (see mirror_iteration()) in contrast only
1405      * wait for conflicting requests at the start of the dirty area, and then
1406      * (based on the in_flight_bitmap) truncate the area to copy so it will not
1407      * conflict with any requests beyond that.  For active writes, however, we
1408      * cannot truncate that area.  The request from our parent must be blocked
1409      * until the area is copied in full.  Therefore, we must wait for the whole
1410      * area to become free of concurrent requests.
1411      */
1412     mirror_wait_on_conflicts(op, s, offset, bytes);
1413 
1414     bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1415 
1416     return op;
1417 }
1418 
1419 static void coroutine_fn active_write_settle(MirrorOp *op)
1420 {
1421     uint64_t start_chunk = op->offset / op->s->granularity;
1422     uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes,
1423                                       op->s->granularity);
1424 
1425     if (!--op->s->in_active_write_counter && op->s->actively_synced) {
1426         BdrvChild *source = op->s->mirror_top_bs->backing;
1427 
1428         if (QLIST_FIRST(&source->bs->parents) == source &&
1429             QLIST_NEXT(source, next_parent) == NULL)
1430         {
1431             /* Assert that we are back in sync once all active write
1432              * operations are settled.
1433              * Note that we can only assert this if the mirror node
1434              * is the source node's only parent. */
1435             assert(!bdrv_get_dirty_count(op->s->dirty_bitmap));
1436         }
1437     }
1438     bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1439     QTAILQ_REMOVE(&op->s->ops_in_flight, op, next);
1440     qemu_co_queue_restart_all(&op->waiting_requests);
1441     g_free(op);
1442 }
1443 
1444 static int coroutine_fn GRAPH_RDLOCK
1445 bdrv_mirror_top_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
1446                        QEMUIOVector *qiov, BdrvRequestFlags flags)
1447 {
1448     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1449 }
1450 
1451 static int coroutine_fn GRAPH_RDLOCK
1452 bdrv_mirror_top_do_write(BlockDriverState *bs, MirrorMethod method,
1453                          uint64_t offset, uint64_t bytes, QEMUIOVector *qiov,
1454                          int flags)
1455 {
1456     MirrorOp *op = NULL;
1457     MirrorBDSOpaque *s = bs->opaque;
1458     int ret = 0;
1459     bool copy_to_target = false;
1460 
1461     if (s->job) {
1462         copy_to_target = s->job->ret >= 0 &&
1463                          !job_is_cancelled(&s->job->common.job) &&
1464                          s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1465     }
1466 
1467     if (copy_to_target) {
1468         op = active_write_prepare(s->job, offset, bytes);
1469     }
1470 
1471     switch (method) {
1472     case MIRROR_METHOD_COPY:
1473         ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1474         break;
1475 
1476     case MIRROR_METHOD_ZERO:
1477         ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1478         break;
1479 
1480     case MIRROR_METHOD_DISCARD:
1481         ret = bdrv_co_pdiscard(bs->backing, offset, bytes);
1482         break;
1483 
1484     default:
1485         abort();
1486     }
1487 
1488     if (ret < 0) {
1489         goto out;
1490     }
1491 
1492     if (copy_to_target) {
1493         do_sync_target_write(s->job, method, offset, bytes, qiov, flags);
1494     }
1495 
1496 out:
1497     if (copy_to_target) {
1498         active_write_settle(op);
1499     }
1500     return ret;
1501 }
1502 
1503 static int coroutine_fn GRAPH_RDLOCK
1504 bdrv_mirror_top_pwritev(BlockDriverState *bs, int64_t offset, int64_t bytes,
1505                         QEMUIOVector *qiov, BdrvRequestFlags flags)
1506 {
1507     MirrorBDSOpaque *s = bs->opaque;
1508     QEMUIOVector bounce_qiov;
1509     void *bounce_buf;
1510     int ret = 0;
1511     bool copy_to_target = false;
1512 
1513     if (s->job) {
1514         copy_to_target = s->job->ret >= 0 &&
1515                          !job_is_cancelled(&s->job->common.job) &&
1516                          s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1517     }
1518 
1519     if (copy_to_target) {
1520         /* The guest might concurrently modify the data to write; but
1521          * the data on source and destination must match, so we have
1522          * to use a bounce buffer if we are going to write to the
1523          * target now. */
1524         bounce_buf = qemu_blockalign(bs, bytes);
1525         iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes);
1526 
1527         qemu_iovec_init(&bounce_qiov, 1);
1528         qemu_iovec_add(&bounce_qiov, bounce_buf, bytes);
1529         qiov = &bounce_qiov;
1530 
1531         flags &= ~BDRV_REQ_REGISTERED_BUF;
1532     }
1533 
1534     ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov,
1535                                    flags);
1536 
1537     if (copy_to_target) {
1538         qemu_iovec_destroy(&bounce_qiov);
1539         qemu_vfree(bounce_buf);
1540     }
1541 
1542     return ret;
1543 }
1544 
1545 static int coroutine_fn GRAPH_RDLOCK bdrv_mirror_top_flush(BlockDriverState *bs)
1546 {
1547     if (bs->backing == NULL) {
1548         /* we can be here after failed bdrv_append in mirror_start_job */
1549         return 0;
1550     }
1551     return bdrv_co_flush(bs->backing->bs);
1552 }
1553 
1554 static int coroutine_fn GRAPH_RDLOCK
1555 bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs, int64_t offset,
1556                               int64_t bytes, BdrvRequestFlags flags)
1557 {
1558     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL,
1559                                     flags);
1560 }
1561 
1562 static int coroutine_fn GRAPH_RDLOCK
1563 bdrv_mirror_top_pdiscard(BlockDriverState *bs, int64_t offset, int64_t bytes)
1564 {
1565     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes,
1566                                     NULL, 0);
1567 }
1568 
1569 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs)
1570 {
1571     if (bs->backing == NULL) {
1572         /* we can be here after failed bdrv_attach_child in
1573          * bdrv_set_backing_hd */
1574         return;
1575     }
1576     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1577             bs->backing->bs->filename);
1578 }
1579 
1580 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1581                                        BdrvChildRole role,
1582                                        BlockReopenQueue *reopen_queue,
1583                                        uint64_t perm, uint64_t shared,
1584                                        uint64_t *nperm, uint64_t *nshared)
1585 {
1586     MirrorBDSOpaque *s = bs->opaque;
1587 
1588     if (s->stop) {
1589         /*
1590          * If the job is to be stopped, we do not need to forward
1591          * anything to the real image.
1592          */
1593         *nperm = 0;
1594         *nshared = BLK_PERM_ALL;
1595         return;
1596     }
1597 
1598     bdrv_default_perms(bs, c, role, reopen_queue,
1599                        perm, shared, nperm, nshared);
1600 
1601     if (s->is_commit) {
1602         /*
1603          * For commit jobs, we cannot take CONSISTENT_READ, because
1604          * that permission is unshared for everything above the base
1605          * node (except for filters on the base node).
1606          * We also have to force-share the WRITE permission, or
1607          * otherwise we would block ourselves at the base node (if
1608          * writes are blocked for a node, they are also blocked for
1609          * its backing file).
1610          * (We could also share RESIZE, because it may be needed for
1611          * the target if its size is less than the top node's; but
1612          * bdrv_default_perms_for_cow() automatically shares RESIZE
1613          * for backing nodes if WRITE is shared, so there is no need
1614          * to do it here.)
1615          */
1616         *nperm &= ~BLK_PERM_CONSISTENT_READ;
1617         *nshared |= BLK_PERM_WRITE;
1618     }
1619 }
1620 
1621 /* Dummy node that provides consistent read to its users without requiring it
1622  * from its backing file and that allows writes on the backing file chain. */
1623 static BlockDriver bdrv_mirror_top = {
1624     .format_name                = "mirror_top",
1625     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1626     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1627     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1628     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1629     .bdrv_co_flush              = bdrv_mirror_top_flush,
1630     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1631     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1632 
1633     .is_filter                  = true,
1634     .filtered_child_is_backing  = true,
1635 };
1636 
1637 static BlockJob *mirror_start_job(
1638                              const char *job_id, BlockDriverState *bs,
1639                              int creation_flags, BlockDriverState *target,
1640                              const char *replaces, int64_t speed,
1641                              uint32_t granularity, int64_t buf_size,
1642                              BlockMirrorBackingMode backing_mode,
1643                              bool zero_target,
1644                              BlockdevOnError on_source_error,
1645                              BlockdevOnError on_target_error,
1646                              bool unmap,
1647                              BlockCompletionFunc *cb,
1648                              void *opaque,
1649                              const BlockJobDriver *driver,
1650                              bool is_none_mode, BlockDriverState *base,
1651                              bool auto_complete, const char *filter_node_name,
1652                              bool is_mirror, MirrorCopyMode copy_mode,
1653                              Error **errp)
1654 {
1655     MirrorBlockJob *s;
1656     MirrorBDSOpaque *bs_opaque;
1657     BlockDriverState *mirror_top_bs;
1658     bool target_is_backing;
1659     uint64_t target_perms, target_shared_perms;
1660     int ret;
1661 
1662     if (granularity == 0) {
1663         granularity = bdrv_get_default_bitmap_granularity(target);
1664     }
1665 
1666     assert(is_power_of_2(granularity));
1667 
1668     if (buf_size < 0) {
1669         error_setg(errp, "Invalid parameter 'buf-size'");
1670         return NULL;
1671     }
1672 
1673     if (buf_size == 0) {
1674         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1675     }
1676 
1677     if (bdrv_skip_filters(bs) == bdrv_skip_filters(target)) {
1678         error_setg(errp, "Can't mirror node into itself");
1679         return NULL;
1680     }
1681 
1682     target_is_backing = bdrv_chain_contains(bs, target);
1683 
1684     /* In the case of active commit, add dummy driver to provide consistent
1685      * reads on the top, while disabling it in the intermediate nodes, and make
1686      * the backing chain writable. */
1687     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1688                                          BDRV_O_RDWR, errp);
1689     if (mirror_top_bs == NULL) {
1690         return NULL;
1691     }
1692     if (!filter_node_name) {
1693         mirror_top_bs->implicit = true;
1694     }
1695 
1696     /* So that we can always drop this node */
1697     mirror_top_bs->never_freeze = true;
1698 
1699     mirror_top_bs->total_sectors = bs->total_sectors;
1700     mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
1701     mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
1702                                           BDRV_REQ_NO_FALLBACK;
1703     bs_opaque = g_new0(MirrorBDSOpaque, 1);
1704     mirror_top_bs->opaque = bs_opaque;
1705 
1706     bs_opaque->is_commit = target_is_backing;
1707 
1708     bdrv_drained_begin(bs);
1709     ret = bdrv_append(mirror_top_bs, bs, errp);
1710     bdrv_drained_end(bs);
1711 
1712     if (ret < 0) {
1713         bdrv_unref(mirror_top_bs);
1714         return NULL;
1715     }
1716 
1717     /* Make sure that the source is not resized while the job is running */
1718     s = block_job_create(job_id, driver, NULL, mirror_top_bs,
1719                          BLK_PERM_CONSISTENT_READ,
1720                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1721                          BLK_PERM_WRITE, speed,
1722                          creation_flags, cb, opaque, errp);
1723     if (!s) {
1724         goto fail;
1725     }
1726 
1727     /* The block job now has a reference to this node */
1728     bdrv_unref(mirror_top_bs);
1729 
1730     s->mirror_top_bs = mirror_top_bs;
1731 
1732     /* No resize for the target either; while the mirror is still running, a
1733      * consistent read isn't necessarily possible. We could possibly allow
1734      * writes and graph modifications, though it would likely defeat the
1735      * purpose of a mirror, so leave them blocked for now.
1736      *
1737      * In the case of active commit, things look a bit different, though,
1738      * because the target is an already populated backing file in active use.
1739      * We can allow anything except resize there.*/
1740 
1741     target_perms = BLK_PERM_WRITE;
1742     target_shared_perms = BLK_PERM_WRITE_UNCHANGED;
1743 
1744     if (target_is_backing) {
1745         int64_t bs_size, target_size;
1746         bs_size = bdrv_getlength(bs);
1747         if (bs_size < 0) {
1748             error_setg_errno(errp, -bs_size,
1749                              "Could not inquire top image size");
1750             goto fail;
1751         }
1752 
1753         target_size = bdrv_getlength(target);
1754         if (target_size < 0) {
1755             error_setg_errno(errp, -target_size,
1756                              "Could not inquire base image size");
1757             goto fail;
1758         }
1759 
1760         if (target_size < bs_size) {
1761             target_perms |= BLK_PERM_RESIZE;
1762         }
1763 
1764         target_shared_perms |= BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE;
1765     } else if (bdrv_chain_contains(bs, bdrv_skip_filters(target))) {
1766         /*
1767          * We may want to allow this in the future, but it would
1768          * require taking some extra care.
1769          */
1770         error_setg(errp, "Cannot mirror to a filter on top of a node in the "
1771                    "source's backing chain");
1772         goto fail;
1773     }
1774 
1775     s->target = blk_new(s->common.job.aio_context,
1776                         target_perms, target_shared_perms);
1777     ret = blk_insert_bs(s->target, target, errp);
1778     if (ret < 0) {
1779         goto fail;
1780     }
1781     if (is_mirror) {
1782         /* XXX: Mirror target could be a NBD server of target QEMU in the case
1783          * of non-shared block migration. To allow migration completion, we
1784          * have to allow "inactivate" of the target BB.  When that happens, we
1785          * know the job is drained, and the vcpus are stopped, so no write
1786          * operation will be performed. Block layer already has assertions to
1787          * ensure that. */
1788         blk_set_force_allow_inactivate(s->target);
1789     }
1790     blk_set_allow_aio_context_change(s->target, true);
1791     blk_set_disable_request_queuing(s->target, true);
1792 
1793     s->replaces = g_strdup(replaces);
1794     s->on_source_error = on_source_error;
1795     s->on_target_error = on_target_error;
1796     s->is_none_mode = is_none_mode;
1797     s->backing_mode = backing_mode;
1798     s->zero_target = zero_target;
1799     s->copy_mode = copy_mode;
1800     s->base = base;
1801     s->base_overlay = bdrv_find_overlay(bs, base);
1802     s->granularity = granularity;
1803     s->buf_size = ROUND_UP(buf_size, granularity);
1804     s->unmap = unmap;
1805     if (auto_complete) {
1806         s->should_complete = true;
1807     }
1808 
1809     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1810     if (!s->dirty_bitmap) {
1811         goto fail;
1812     }
1813     if (s->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING) {
1814         bdrv_disable_dirty_bitmap(s->dirty_bitmap);
1815     }
1816 
1817     ret = block_job_add_bdrv(&s->common, "source", bs, 0,
1818                              BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE |
1819                              BLK_PERM_CONSISTENT_READ,
1820                              errp);
1821     if (ret < 0) {
1822         goto fail;
1823     }
1824 
1825     /* Required permissions are already taken with blk_new() */
1826     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1827                        &error_abort);
1828 
1829     /* In commit_active_start() all intermediate nodes disappear, so
1830      * any jobs in them must be blocked */
1831     if (target_is_backing) {
1832         BlockDriverState *iter, *filtered_target;
1833         uint64_t iter_shared_perms;
1834 
1835         /*
1836          * The topmost node with
1837          * bdrv_skip_filters(filtered_target) == bdrv_skip_filters(target)
1838          */
1839         filtered_target = bdrv_cow_bs(bdrv_find_overlay(bs, target));
1840 
1841         assert(bdrv_skip_filters(filtered_target) ==
1842                bdrv_skip_filters(target));
1843 
1844         /*
1845          * XXX BLK_PERM_WRITE needs to be allowed so we don't block
1846          * ourselves at s->base (if writes are blocked for a node, they are
1847          * also blocked for its backing file). The other options would be a
1848          * second filter driver above s->base (== target).
1849          */
1850         iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE;
1851 
1852         for (iter = bdrv_filter_or_cow_bs(bs); iter != target;
1853              iter = bdrv_filter_or_cow_bs(iter))
1854         {
1855             if (iter == filtered_target) {
1856                 /*
1857                  * From here on, all nodes are filters on the base.
1858                  * This allows us to share BLK_PERM_CONSISTENT_READ.
1859                  */
1860                 iter_shared_perms |= BLK_PERM_CONSISTENT_READ;
1861             }
1862 
1863             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1864                                      iter_shared_perms, errp);
1865             if (ret < 0) {
1866                 goto fail;
1867             }
1868         }
1869 
1870         if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) {
1871             goto fail;
1872         }
1873     }
1874 
1875     QTAILQ_INIT(&s->ops_in_flight);
1876 
1877     trace_mirror_start(bs, s, opaque);
1878     job_start(&s->common.job);
1879 
1880     return &s->common;
1881 
1882 fail:
1883     if (s) {
1884         /* Make sure this BDS does not go away until we have completed the graph
1885          * changes below */
1886         bdrv_ref(mirror_top_bs);
1887 
1888         g_free(s->replaces);
1889         blk_unref(s->target);
1890         bs_opaque->job = NULL;
1891         if (s->dirty_bitmap) {
1892             bdrv_release_dirty_bitmap(s->dirty_bitmap);
1893         }
1894         job_early_fail(&s->common.job);
1895     }
1896 
1897     bs_opaque->stop = true;
1898     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
1899                              &error_abort);
1900     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
1901 
1902     bdrv_unref(mirror_top_bs);
1903 
1904     return NULL;
1905 }
1906 
1907 void mirror_start(const char *job_id, BlockDriverState *bs,
1908                   BlockDriverState *target, const char *replaces,
1909                   int creation_flags, int64_t speed,
1910                   uint32_t granularity, int64_t buf_size,
1911                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1912                   bool zero_target,
1913                   BlockdevOnError on_source_error,
1914                   BlockdevOnError on_target_error,
1915                   bool unmap, const char *filter_node_name,
1916                   MirrorCopyMode copy_mode, Error **errp)
1917 {
1918     bool is_none_mode;
1919     BlockDriverState *base;
1920 
1921     GLOBAL_STATE_CODE();
1922 
1923     if ((mode == MIRROR_SYNC_MODE_INCREMENTAL) ||
1924         (mode == MIRROR_SYNC_MODE_BITMAP)) {
1925         error_setg(errp, "Sync mode '%s' not supported",
1926                    MirrorSyncMode_str(mode));
1927         return;
1928     }
1929     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1930     base = mode == MIRROR_SYNC_MODE_TOP ? bdrv_backing_chain_next(bs) : NULL;
1931     mirror_start_job(job_id, bs, creation_flags, target, replaces,
1932                      speed, granularity, buf_size, backing_mode, zero_target,
1933                      on_source_error, on_target_error, unmap, NULL, NULL,
1934                      &mirror_job_driver, is_none_mode, base, false,
1935                      filter_node_name, true, copy_mode, errp);
1936 }
1937 
1938 BlockJob *commit_active_start(const char *job_id, BlockDriverState *bs,
1939                               BlockDriverState *base, int creation_flags,
1940                               int64_t speed, BlockdevOnError on_error,
1941                               const char *filter_node_name,
1942                               BlockCompletionFunc *cb, void *opaque,
1943                               bool auto_complete, Error **errp)
1944 {
1945     bool base_read_only;
1946     BlockJob *job;
1947 
1948     GLOBAL_STATE_CODE();
1949 
1950     base_read_only = bdrv_is_read_only(base);
1951 
1952     if (base_read_only) {
1953         if (bdrv_reopen_set_read_only(base, false, errp) < 0) {
1954             return NULL;
1955         }
1956     }
1957 
1958     job = mirror_start_job(
1959                      job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1960                      MIRROR_LEAVE_BACKING_CHAIN, false,
1961                      on_error, on_error, true, cb, opaque,
1962                      &commit_active_job_driver, false, base, auto_complete,
1963                      filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND,
1964                      errp);
1965     if (!job) {
1966         goto error_restore_flags;
1967     }
1968 
1969     return job;
1970 
1971 error_restore_flags:
1972     /* ignore error and errp for bdrv_reopen, because we want to propagate
1973      * the original error */
1974     if (base_read_only) {
1975         bdrv_reopen_set_read_only(base, true, NULL);
1976     }
1977     return NULL;
1978 }
1979