xref: /openbmc/qemu/block/mirror.c (revision b6235a75)
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_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             WITH_GRAPH_RDLOCK_GUARD() {
580                 bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes,
581                                        &target_offset, &target_bytes);
582             }
583             if (target_offset == offset &&
584                 target_bytes == io_bytes) {
585                 mirror_method = ret & BDRV_BLOCK_ZERO ?
586                                     MIRROR_METHOD_ZERO :
587                                     MIRROR_METHOD_DISCARD;
588             }
589         }
590 
591         while (s->in_flight >= MAX_IN_FLIGHT) {
592             trace_mirror_yield_in_flight(s, offset, s->in_flight);
593             mirror_wait_for_free_in_flight_slot(s);
594         }
595 
596         if (s->ret < 0) {
597             ret = 0;
598             goto fail;
599         }
600 
601         io_bytes = mirror_clip_bytes(s, offset, io_bytes);
602         io_bytes = mirror_perform(s, offset, io_bytes, mirror_method);
603         if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) {
604             io_bytes_acct = 0;
605         } else {
606             io_bytes_acct = io_bytes;
607         }
608         assert(io_bytes);
609         offset += io_bytes;
610         nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity);
611         delay_ns = block_job_ratelimit_get_delay(&s->common, io_bytes_acct);
612     }
613 
614     ret = delay_ns;
615 fail:
616     QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next);
617     qemu_co_queue_restart_all(&pseudo_op->waiting_requests);
618     g_free(pseudo_op);
619 
620     return ret;
621 }
622 
623 static void mirror_free_init(MirrorBlockJob *s)
624 {
625     int granularity = s->granularity;
626     size_t buf_size = s->buf_size;
627     uint8_t *buf = s->buf;
628 
629     assert(s->buf_free_count == 0);
630     QSIMPLEQ_INIT(&s->buf_free);
631     while (buf_size != 0) {
632         MirrorBuffer *cur = (MirrorBuffer *)buf;
633         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
634         s->buf_free_count++;
635         buf_size -= granularity;
636         buf += granularity;
637     }
638 }
639 
640 /* This is also used for the .pause callback. There is no matching
641  * mirror_resume() because mirror_run() will begin iterating again
642  * when the job is resumed.
643  */
644 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s)
645 {
646     while (s->in_flight > 0) {
647         mirror_wait_for_free_in_flight_slot(s);
648     }
649 }
650 
651 /**
652  * mirror_exit_common: handle both abort() and prepare() cases.
653  * for .prepare, returns 0 on success and -errno on failure.
654  * for .abort cases, denoted by abort = true, MUST return 0.
655  */
656 static int mirror_exit_common(Job *job)
657 {
658     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
659     BlockJob *bjob = &s->common;
660     MirrorBDSOpaque *bs_opaque;
661     AioContext *replace_aio_context = NULL;
662     BlockDriverState *src;
663     BlockDriverState *target_bs;
664     BlockDriverState *mirror_top_bs;
665     Error *local_err = NULL;
666     bool abort = job->ret < 0;
667     int ret = 0;
668 
669     if (s->prepared) {
670         return 0;
671     }
672     s->prepared = true;
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     return ret;
797 }
798 
799 static int mirror_prepare(Job *job)
800 {
801     return mirror_exit_common(job);
802 }
803 
804 static void mirror_abort(Job *job)
805 {
806     int ret = mirror_exit_common(job);
807     assert(ret == 0);
808 }
809 
810 static void coroutine_fn mirror_throttle(MirrorBlockJob *s)
811 {
812     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
813 
814     if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) {
815         s->last_pause_ns = now;
816         job_sleep_ns(&s->common.job, 0);
817     } else {
818         job_pause_point(&s->common.job);
819     }
820 }
821 
822 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
823 {
824     int64_t offset;
825     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
826     BlockDriverState *target_bs = blk_bs(s->target);
827     int ret;
828     int64_t count;
829 
830     if (s->zero_target) {
831         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
832             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
833             return 0;
834         }
835 
836         s->initial_zeroing_ongoing = true;
837         for (offset = 0; offset < s->bdev_length; ) {
838             int bytes = MIN(s->bdev_length - offset,
839                             QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
840 
841             mirror_throttle(s);
842 
843             if (job_is_cancelled(&s->common.job)) {
844                 s->initial_zeroing_ongoing = false;
845                 return 0;
846             }
847 
848             if (s->in_flight >= MAX_IN_FLIGHT) {
849                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
850                                    s->in_flight);
851                 mirror_wait_for_free_in_flight_slot(s);
852                 continue;
853             }
854 
855             mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO);
856             offset += bytes;
857         }
858 
859         mirror_wait_for_all_io(s);
860         s->initial_zeroing_ongoing = false;
861     }
862 
863     /* First part, loop on the sectors and initialize the dirty bitmap.  */
864     for (offset = 0; offset < s->bdev_length; ) {
865         /* Just to make sure we are not exceeding int limit. */
866         int bytes = MIN(s->bdev_length - offset,
867                         QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
868 
869         mirror_throttle(s);
870 
871         if (job_is_cancelled(&s->common.job)) {
872             return 0;
873         }
874 
875         WITH_GRAPH_RDLOCK_GUARD() {
876             ret = bdrv_is_allocated_above(bs, s->base_overlay, true, offset,
877                                           bytes, &count);
878         }
879         if (ret < 0) {
880             return ret;
881         }
882 
883         assert(count);
884         if (ret > 0) {
885             bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
886         }
887         offset += count;
888     }
889     return 0;
890 }
891 
892 /* Called when going out of the streaming phase to flush the bulk of the
893  * data to the medium, or just before completing.
894  */
895 static int coroutine_fn mirror_flush(MirrorBlockJob *s)
896 {
897     int ret = blk_co_flush(s->target);
898     if (ret < 0) {
899         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
900             s->ret = ret;
901         }
902     }
903     return ret;
904 }
905 
906 static int coroutine_fn mirror_run(Job *job, Error **errp)
907 {
908     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
909     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
910     MirrorBDSOpaque *mirror_top_opaque = s->mirror_top_bs->opaque;
911     BlockDriverState *target_bs = blk_bs(s->target);
912     bool need_drain = true;
913     BlockDeviceIoStatus iostatus;
914     int64_t length;
915     int64_t target_length;
916     BlockDriverInfo bdi;
917     char backing_filename[2]; /* we only need 2 characters because we are only
918                                  checking for a NULL string */
919     int ret = 0;
920 
921     if (job_is_cancelled(&s->common.job)) {
922         goto immediate_exit;
923     }
924 
925     bdrv_graph_co_rdlock();
926     s->bdev_length = bdrv_co_getlength(bs);
927     bdrv_graph_co_rdunlock();
928 
929     if (s->bdev_length < 0) {
930         ret = s->bdev_length;
931         goto immediate_exit;
932     }
933 
934     target_length = blk_co_getlength(s->target);
935     if (target_length < 0) {
936         ret = target_length;
937         goto immediate_exit;
938     }
939 
940     /* Active commit must resize the base image if its size differs from the
941      * active layer. */
942     if (s->base == blk_bs(s->target)) {
943         if (s->bdev_length > target_length) {
944             ret = blk_co_truncate(s->target, s->bdev_length, false,
945                                   PREALLOC_MODE_OFF, 0, NULL);
946             if (ret < 0) {
947                 goto immediate_exit;
948             }
949         }
950     } else if (s->bdev_length != target_length) {
951         error_setg(errp, "Source and target image have different sizes");
952         ret = -EINVAL;
953         goto immediate_exit;
954     }
955 
956     if (s->bdev_length == 0) {
957         /* Transition to the READY state and wait for complete. */
958         job_transition_to_ready(&s->common.job);
959         s->actively_synced = true;
960         while (!job_cancel_requested(&s->common.job) && !s->should_complete) {
961             job_yield(&s->common.job);
962         }
963         goto immediate_exit;
964     }
965 
966     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
967     s->in_flight_bitmap = bitmap_new(length);
968 
969     /* If we have no backing file yet in the destination, we cannot let
970      * the destination do COW.  Instead, we copy sectors around the
971      * dirty data if needed.  We need a bitmap to do that.
972      */
973     bdrv_get_backing_filename(target_bs, backing_filename,
974                               sizeof(backing_filename));
975     bdrv_graph_co_rdlock();
976     if (!bdrv_co_get_info(target_bs, &bdi) && bdi.cluster_size) {
977         s->target_cluster_size = bdi.cluster_size;
978     } else {
979         s->target_cluster_size = BDRV_SECTOR_SIZE;
980     }
981     bdrv_graph_co_rdunlock();
982     if (backing_filename[0] && !bdrv_backing_chain_next(target_bs) &&
983         s->granularity < s->target_cluster_size) {
984         s->buf_size = MAX(s->buf_size, s->target_cluster_size);
985         s->cow_bitmap = bitmap_new(length);
986     }
987     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
988 
989     s->buf = qemu_try_blockalign(bs, s->buf_size);
990     if (s->buf == NULL) {
991         ret = -ENOMEM;
992         goto immediate_exit;
993     }
994 
995     mirror_free_init(s);
996 
997     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
998     if (!s->is_none_mode) {
999         ret = mirror_dirty_init(s);
1000         if (ret < 0 || job_is_cancelled(&s->common.job)) {
1001             goto immediate_exit;
1002         }
1003     }
1004 
1005     /*
1006      * Only now the job is fully initialised and mirror_top_bs should start
1007      * accessing it.
1008      */
1009     mirror_top_opaque->job = s;
1010 
1011     assert(!s->dbi);
1012     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
1013     for (;;) {
1014         uint64_t delay_ns = 0;
1015         int64_t cnt, delta;
1016         bool should_complete;
1017 
1018         if (s->ret < 0) {
1019             ret = s->ret;
1020             goto immediate_exit;
1021         }
1022 
1023         job_pause_point(&s->common.job);
1024 
1025         if (job_is_cancelled(&s->common.job)) {
1026             ret = 0;
1027             goto immediate_exit;
1028         }
1029 
1030         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1031         /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
1032          * the number of bytes currently being processed; together those are
1033          * the current remaining operation length */
1034         job_progress_set_remaining(&s->common.job,
1035                                    s->bytes_in_flight + cnt +
1036                                    s->active_write_bytes_in_flight);
1037 
1038         /* Note that even when no rate limit is applied we need to yield
1039          * periodically with no pending I/O so that bdrv_drain_all() returns.
1040          * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
1041          * an error, or when the source is clean, whichever comes first. */
1042         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
1043         WITH_JOB_LOCK_GUARD() {
1044             iostatus = s->common.iostatus;
1045         }
1046         if (delta < BLOCK_JOB_SLICE_TIME &&
1047             iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
1048             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
1049                 (cnt == 0 && s->in_flight > 0)) {
1050                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
1051                 mirror_wait_for_free_in_flight_slot(s);
1052                 continue;
1053             } else if (cnt != 0) {
1054                 delay_ns = mirror_iteration(s);
1055             }
1056         }
1057 
1058         should_complete = false;
1059         if (s->in_flight == 0 && cnt == 0) {
1060             trace_mirror_before_flush(s);
1061             if (!job_is_ready(&s->common.job)) {
1062                 if (mirror_flush(s) < 0) {
1063                     /* Go check s->ret.  */
1064                     continue;
1065                 }
1066                 /* We're out of the streaming phase.  From now on, if the job
1067                  * is cancelled we will actually complete all pending I/O and
1068                  * report completion.  This way, block-job-cancel will leave
1069                  * the target in a consistent state.
1070                  */
1071                 job_transition_to_ready(&s->common.job);
1072                 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) {
1073                     s->actively_synced = true;
1074                 }
1075             }
1076 
1077             should_complete = s->should_complete ||
1078                 job_cancel_requested(&s->common.job);
1079             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1080         }
1081 
1082         if (cnt == 0 && should_complete) {
1083             /* The dirty bitmap is not updated while operations are pending.
1084              * If we're about to exit, wait for pending operations before
1085              * calling bdrv_get_dirty_count(bs), or we may exit while the
1086              * source has dirty data to copy!
1087              *
1088              * Note that I/O can be submitted by the guest while
1089              * mirror_populate runs, so pause it now.  Before deciding
1090              * whether to switch to target check one last time if I/O has
1091              * come in the meanwhile, and if not flush the data to disk.
1092              */
1093             trace_mirror_before_drain(s, cnt);
1094 
1095             s->in_drain = true;
1096             bdrv_drained_begin(bs);
1097 
1098             /* Must be zero because we are drained */
1099             assert(s->in_active_write_counter == 0);
1100 
1101             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1102             if (cnt > 0 || mirror_flush(s) < 0) {
1103                 bdrv_drained_end(bs);
1104                 s->in_drain = false;
1105                 continue;
1106             }
1107 
1108             /* The two disks are in sync.  Exit and report successful
1109              * completion.
1110              */
1111             assert(QLIST_EMPTY(&bs->tracked_requests));
1112             need_drain = false;
1113             break;
1114         }
1115 
1116         if (job_is_ready(&s->common.job) && !should_complete) {
1117             delay_ns = (s->in_flight == 0 &&
1118                         cnt == 0 ? BLOCK_JOB_SLICE_TIME : 0);
1119         }
1120         trace_mirror_before_sleep(s, cnt, job_is_ready(&s->common.job),
1121                                   delay_ns);
1122         job_sleep_ns(&s->common.job, delay_ns);
1123         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1124     }
1125 
1126 immediate_exit:
1127     if (s->in_flight > 0) {
1128         /* We get here only if something went wrong.  Either the job failed,
1129          * or it was cancelled prematurely so that we do not guarantee that
1130          * the target is a copy of the source.
1131          */
1132         assert(ret < 0 || job_is_cancelled(&s->common.job));
1133         assert(need_drain);
1134         mirror_wait_for_all_io(s);
1135     }
1136 
1137     assert(s->in_flight == 0);
1138     qemu_vfree(s->buf);
1139     g_free(s->cow_bitmap);
1140     g_free(s->in_flight_bitmap);
1141     bdrv_dirty_iter_free(s->dbi);
1142 
1143     if (need_drain) {
1144         s->in_drain = true;
1145         bdrv_drained_begin(bs);
1146     }
1147 
1148     return ret;
1149 }
1150 
1151 static void mirror_complete(Job *job, Error **errp)
1152 {
1153     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1154 
1155     if (!job_is_ready(job)) {
1156         error_setg(errp, "The active block job '%s' cannot be completed",
1157                    job->id);
1158         return;
1159     }
1160 
1161     /* block all operations on to_replace bs */
1162     if (s->replaces) {
1163         AioContext *replace_aio_context;
1164 
1165         s->to_replace = bdrv_find_node(s->replaces);
1166         if (!s->to_replace) {
1167             error_setg(errp, "Node name '%s' not found", s->replaces);
1168             return;
1169         }
1170 
1171         replace_aio_context = bdrv_get_aio_context(s->to_replace);
1172         aio_context_acquire(replace_aio_context);
1173 
1174         /* TODO Translate this into child freeze system. */
1175         error_setg(&s->replace_blocker,
1176                    "block device is in use by block-job-complete");
1177         bdrv_op_block_all(s->to_replace, s->replace_blocker);
1178         bdrv_ref(s->to_replace);
1179 
1180         aio_context_release(replace_aio_context);
1181     }
1182 
1183     s->should_complete = true;
1184 
1185     /* If the job is paused, it will be re-entered when it is resumed */
1186     WITH_JOB_LOCK_GUARD() {
1187         if (!job->paused) {
1188             job_enter_cond_locked(job, NULL);
1189         }
1190     }
1191 }
1192 
1193 static void coroutine_fn mirror_pause(Job *job)
1194 {
1195     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1196 
1197     mirror_wait_for_all_io(s);
1198 }
1199 
1200 static bool mirror_drained_poll(BlockJob *job)
1201 {
1202     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1203 
1204     /* If the job isn't paused nor cancelled, we can't be sure that it won't
1205      * issue more requests. We make an exception if we've reached this point
1206      * from one of our own drain sections, to avoid a deadlock waiting for
1207      * ourselves.
1208      */
1209     WITH_JOB_LOCK_GUARD() {
1210         if (!s->common.job.paused && !job_is_cancelled_locked(&job->job)
1211             && !s->in_drain) {
1212             return true;
1213         }
1214     }
1215 
1216     return !!s->in_flight;
1217 }
1218 
1219 static bool mirror_cancel(Job *job, bool force)
1220 {
1221     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1222     BlockDriverState *target = blk_bs(s->target);
1223 
1224     /*
1225      * Before the job is READY, we treat any cancellation like a
1226      * force-cancellation.
1227      */
1228     force = force || !job_is_ready(job);
1229 
1230     if (force) {
1231         bdrv_cancel_in_flight(target);
1232     }
1233     return force;
1234 }
1235 
1236 static bool commit_active_cancel(Job *job, bool force)
1237 {
1238     /* Same as above in mirror_cancel() */
1239     return force || !job_is_ready(job);
1240 }
1241 
1242 static const BlockJobDriver mirror_job_driver = {
1243     .job_driver = {
1244         .instance_size          = sizeof(MirrorBlockJob),
1245         .job_type               = JOB_TYPE_MIRROR,
1246         .free                   = block_job_free,
1247         .user_resume            = block_job_user_resume,
1248         .run                    = mirror_run,
1249         .prepare                = mirror_prepare,
1250         .abort                  = mirror_abort,
1251         .pause                  = mirror_pause,
1252         .complete               = mirror_complete,
1253         .cancel                 = mirror_cancel,
1254     },
1255     .drained_poll           = mirror_drained_poll,
1256 };
1257 
1258 static const BlockJobDriver commit_active_job_driver = {
1259     .job_driver = {
1260         .instance_size          = sizeof(MirrorBlockJob),
1261         .job_type               = JOB_TYPE_COMMIT,
1262         .free                   = block_job_free,
1263         .user_resume            = block_job_user_resume,
1264         .run                    = mirror_run,
1265         .prepare                = mirror_prepare,
1266         .abort                  = mirror_abort,
1267         .pause                  = mirror_pause,
1268         .complete               = mirror_complete,
1269         .cancel                 = commit_active_cancel,
1270     },
1271     .drained_poll           = mirror_drained_poll,
1272 };
1273 
1274 static void coroutine_fn
1275 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method,
1276                      uint64_t offset, uint64_t bytes,
1277                      QEMUIOVector *qiov, int flags)
1278 {
1279     int ret;
1280     size_t qiov_offset = 0;
1281     int64_t bitmap_offset, bitmap_end;
1282 
1283     if (!QEMU_IS_ALIGNED(offset, job->granularity) &&
1284         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset))
1285     {
1286             /*
1287              * Dirty unaligned padding: ignore it.
1288              *
1289              * Reasoning:
1290              * 1. If we copy it, we can't reset corresponding bit in
1291              *    dirty_bitmap as there may be some "dirty" bytes still not
1292              *    copied.
1293              * 2. It's already dirty, so skipping it we don't diverge mirror
1294              *    progress.
1295              *
1296              * Note, that because of this, guest write may have no contribution
1297              * into mirror converge, but that's not bad, as we have background
1298              * process of mirroring. If under some bad circumstances (high guest
1299              * IO load) background process starve, we will not converge anyway,
1300              * even if each write will contribute, as guest is not guaranteed to
1301              * rewrite the whole disk.
1302              */
1303             qiov_offset = QEMU_ALIGN_UP(offset, job->granularity) - offset;
1304             if (bytes <= qiov_offset) {
1305                 /* nothing to do after shrink */
1306                 return;
1307             }
1308             offset += qiov_offset;
1309             bytes -= qiov_offset;
1310     }
1311 
1312     if (!QEMU_IS_ALIGNED(offset + bytes, job->granularity) &&
1313         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset + bytes - 1))
1314     {
1315         uint64_t tail = (offset + bytes) % job->granularity;
1316 
1317         if (bytes <= tail) {
1318             /* nothing to do after shrink */
1319             return;
1320         }
1321         bytes -= tail;
1322     }
1323 
1324     /*
1325      * Tails are either clean or shrunk, so for bitmap resetting
1326      * we safely align the range down.
1327      */
1328     bitmap_offset = QEMU_ALIGN_UP(offset, job->granularity);
1329     bitmap_end = QEMU_ALIGN_DOWN(offset + bytes, job->granularity);
1330     if (bitmap_offset < bitmap_end) {
1331         bdrv_reset_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1332                                 bitmap_end - bitmap_offset);
1333     }
1334 
1335     job_progress_increase_remaining(&job->common.job, bytes);
1336     job->active_write_bytes_in_flight += bytes;
1337 
1338     switch (method) {
1339     case MIRROR_METHOD_COPY:
1340         ret = blk_co_pwritev_part(job->target, offset, bytes,
1341                                   qiov, qiov_offset, flags);
1342         break;
1343 
1344     case MIRROR_METHOD_ZERO:
1345         assert(!qiov);
1346         ret = blk_co_pwrite_zeroes(job->target, offset, bytes, flags);
1347         break;
1348 
1349     case MIRROR_METHOD_DISCARD:
1350         assert(!qiov);
1351         ret = blk_co_pdiscard(job->target, offset, bytes);
1352         break;
1353 
1354     default:
1355         abort();
1356     }
1357 
1358     job->active_write_bytes_in_flight -= bytes;
1359     if (ret >= 0) {
1360         job_progress_update(&job->common.job, bytes);
1361     } else {
1362         BlockErrorAction action;
1363 
1364         /*
1365          * We failed, so we should mark dirty the whole area, aligned up.
1366          * Note that we don't care about shrunk tails if any: they were dirty
1367          * at function start, and they must be still dirty, as we've locked
1368          * the region for in-flight op.
1369          */
1370         bitmap_offset = QEMU_ALIGN_DOWN(offset, job->granularity);
1371         bitmap_end = QEMU_ALIGN_UP(offset + bytes, job->granularity);
1372         bdrv_set_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1373                               bitmap_end - bitmap_offset);
1374         job->actively_synced = false;
1375 
1376         action = mirror_error_action(job, false, -ret);
1377         if (action == BLOCK_ERROR_ACTION_REPORT) {
1378             if (!job->ret) {
1379                 job->ret = ret;
1380             }
1381         }
1382     }
1383 }
1384 
1385 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s,
1386                                                    uint64_t offset,
1387                                                    uint64_t bytes)
1388 {
1389     MirrorOp *op;
1390     uint64_t start_chunk = offset / s->granularity;
1391     uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
1392 
1393     op = g_new(MirrorOp, 1);
1394     *op = (MirrorOp){
1395         .s                  = s,
1396         .offset             = offset,
1397         .bytes              = bytes,
1398         .is_active_write    = true,
1399         .is_in_flight       = true,
1400         .co                 = qemu_coroutine_self(),
1401     };
1402     qemu_co_queue_init(&op->waiting_requests);
1403     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
1404 
1405     s->in_active_write_counter++;
1406 
1407     /*
1408      * Wait for concurrent requests affecting the area.  If there are already
1409      * running requests that are copying off now-to-be stale data in the area,
1410      * we must wait for them to finish before we begin writing fresh data to the
1411      * target so that the write operations appear in the correct order.
1412      * Note that background requests (see mirror_iteration()) in contrast only
1413      * wait for conflicting requests at the start of the dirty area, and then
1414      * (based on the in_flight_bitmap) truncate the area to copy so it will not
1415      * conflict with any requests beyond that.  For active writes, however, we
1416      * cannot truncate that area.  The request from our parent must be blocked
1417      * until the area is copied in full.  Therefore, we must wait for the whole
1418      * area to become free of concurrent requests.
1419      */
1420     mirror_wait_on_conflicts(op, s, offset, bytes);
1421 
1422     bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1423 
1424     return op;
1425 }
1426 
1427 static void coroutine_fn GRAPH_RDLOCK active_write_settle(MirrorOp *op)
1428 {
1429     uint64_t start_chunk = op->offset / op->s->granularity;
1430     uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes,
1431                                       op->s->granularity);
1432 
1433     if (!--op->s->in_active_write_counter && op->s->actively_synced) {
1434         BdrvChild *source = op->s->mirror_top_bs->backing;
1435 
1436         if (QLIST_FIRST(&source->bs->parents) == source &&
1437             QLIST_NEXT(source, next_parent) == NULL)
1438         {
1439             /* Assert that we are back in sync once all active write
1440              * operations are settled.
1441              * Note that we can only assert this if the mirror node
1442              * is the source node's only parent. */
1443             assert(!bdrv_get_dirty_count(op->s->dirty_bitmap));
1444         }
1445     }
1446     bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1447     QTAILQ_REMOVE(&op->s->ops_in_flight, op, next);
1448     qemu_co_queue_restart_all(&op->waiting_requests);
1449     g_free(op);
1450 }
1451 
1452 static int coroutine_fn GRAPH_RDLOCK
1453 bdrv_mirror_top_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
1454                        QEMUIOVector *qiov, BdrvRequestFlags flags)
1455 {
1456     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1457 }
1458 
1459 static int coroutine_fn GRAPH_RDLOCK
1460 bdrv_mirror_top_do_write(BlockDriverState *bs, MirrorMethod method,
1461                          uint64_t offset, uint64_t bytes, QEMUIOVector *qiov,
1462                          int flags)
1463 {
1464     MirrorOp *op = NULL;
1465     MirrorBDSOpaque *s = bs->opaque;
1466     int ret = 0;
1467     bool copy_to_target = false;
1468 
1469     if (s->job) {
1470         copy_to_target = s->job->ret >= 0 &&
1471                          !job_is_cancelled(&s->job->common.job) &&
1472                          s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1473     }
1474 
1475     if (copy_to_target) {
1476         op = active_write_prepare(s->job, offset, bytes);
1477     }
1478 
1479     switch (method) {
1480     case MIRROR_METHOD_COPY:
1481         ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1482         break;
1483 
1484     case MIRROR_METHOD_ZERO:
1485         ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1486         break;
1487 
1488     case MIRROR_METHOD_DISCARD:
1489         ret = bdrv_co_pdiscard(bs->backing, offset, bytes);
1490         break;
1491 
1492     default:
1493         abort();
1494     }
1495 
1496     if (ret < 0) {
1497         goto out;
1498     }
1499 
1500     if (copy_to_target) {
1501         do_sync_target_write(s->job, method, offset, bytes, qiov, flags);
1502     }
1503 
1504 out:
1505     if (copy_to_target) {
1506         active_write_settle(op);
1507     }
1508     return ret;
1509 }
1510 
1511 static int coroutine_fn GRAPH_RDLOCK
1512 bdrv_mirror_top_pwritev(BlockDriverState *bs, int64_t offset, int64_t bytes,
1513                         QEMUIOVector *qiov, BdrvRequestFlags flags)
1514 {
1515     MirrorBDSOpaque *s = bs->opaque;
1516     QEMUIOVector bounce_qiov;
1517     void *bounce_buf;
1518     int ret = 0;
1519     bool copy_to_target = false;
1520 
1521     if (s->job) {
1522         copy_to_target = s->job->ret >= 0 &&
1523                          !job_is_cancelled(&s->job->common.job) &&
1524                          s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1525     }
1526 
1527     if (copy_to_target) {
1528         /* The guest might concurrently modify the data to write; but
1529          * the data on source and destination must match, so we have
1530          * to use a bounce buffer if we are going to write to the
1531          * target now. */
1532         bounce_buf = qemu_blockalign(bs, bytes);
1533         iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes);
1534 
1535         qemu_iovec_init(&bounce_qiov, 1);
1536         qemu_iovec_add(&bounce_qiov, bounce_buf, bytes);
1537         qiov = &bounce_qiov;
1538 
1539         flags &= ~BDRV_REQ_REGISTERED_BUF;
1540     }
1541 
1542     ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov,
1543                                    flags);
1544 
1545     if (copy_to_target) {
1546         qemu_iovec_destroy(&bounce_qiov);
1547         qemu_vfree(bounce_buf);
1548     }
1549 
1550     return ret;
1551 }
1552 
1553 static int coroutine_fn GRAPH_RDLOCK bdrv_mirror_top_flush(BlockDriverState *bs)
1554 {
1555     if (bs->backing == NULL) {
1556         /* we can be here after failed bdrv_append in mirror_start_job */
1557         return 0;
1558     }
1559     return bdrv_co_flush(bs->backing->bs);
1560 }
1561 
1562 static int coroutine_fn GRAPH_RDLOCK
1563 bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs, int64_t offset,
1564                               int64_t bytes, BdrvRequestFlags flags)
1565 {
1566     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL,
1567                                     flags);
1568 }
1569 
1570 static int coroutine_fn GRAPH_RDLOCK
1571 bdrv_mirror_top_pdiscard(BlockDriverState *bs, int64_t offset, int64_t bytes)
1572 {
1573     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes,
1574                                     NULL, 0);
1575 }
1576 
1577 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs)
1578 {
1579     if (bs->backing == NULL) {
1580         /* we can be here after failed bdrv_attach_child in
1581          * bdrv_set_backing_hd */
1582         return;
1583     }
1584     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1585             bs->backing->bs->filename);
1586 }
1587 
1588 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1589                                        BdrvChildRole role,
1590                                        BlockReopenQueue *reopen_queue,
1591                                        uint64_t perm, uint64_t shared,
1592                                        uint64_t *nperm, uint64_t *nshared)
1593 {
1594     MirrorBDSOpaque *s = bs->opaque;
1595 
1596     if (s->stop) {
1597         /*
1598          * If the job is to be stopped, we do not need to forward
1599          * anything to the real image.
1600          */
1601         *nperm = 0;
1602         *nshared = BLK_PERM_ALL;
1603         return;
1604     }
1605 
1606     bdrv_default_perms(bs, c, role, reopen_queue,
1607                        perm, shared, nperm, nshared);
1608 
1609     if (s->is_commit) {
1610         /*
1611          * For commit jobs, we cannot take CONSISTENT_READ, because
1612          * that permission is unshared for everything above the base
1613          * node (except for filters on the base node).
1614          * We also have to force-share the WRITE permission, or
1615          * otherwise we would block ourselves at the base node (if
1616          * writes are blocked for a node, they are also blocked for
1617          * its backing file).
1618          * (We could also share RESIZE, because it may be needed for
1619          * the target if its size is less than the top node's; but
1620          * bdrv_default_perms_for_cow() automatically shares RESIZE
1621          * for backing nodes if WRITE is shared, so there is no need
1622          * to do it here.)
1623          */
1624         *nperm &= ~BLK_PERM_CONSISTENT_READ;
1625         *nshared |= BLK_PERM_WRITE;
1626     }
1627 }
1628 
1629 /* Dummy node that provides consistent read to its users without requiring it
1630  * from its backing file and that allows writes on the backing file chain. */
1631 static BlockDriver bdrv_mirror_top = {
1632     .format_name                = "mirror_top",
1633     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1634     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1635     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1636     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1637     .bdrv_co_flush              = bdrv_mirror_top_flush,
1638     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1639     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1640 
1641     .is_filter                  = true,
1642     .filtered_child_is_backing  = true,
1643 };
1644 
1645 static BlockJob *mirror_start_job(
1646                              const char *job_id, BlockDriverState *bs,
1647                              int creation_flags, BlockDriverState *target,
1648                              const char *replaces, int64_t speed,
1649                              uint32_t granularity, int64_t buf_size,
1650                              BlockMirrorBackingMode backing_mode,
1651                              bool zero_target,
1652                              BlockdevOnError on_source_error,
1653                              BlockdevOnError on_target_error,
1654                              bool unmap,
1655                              BlockCompletionFunc *cb,
1656                              void *opaque,
1657                              const BlockJobDriver *driver,
1658                              bool is_none_mode, BlockDriverState *base,
1659                              bool auto_complete, const char *filter_node_name,
1660                              bool is_mirror, MirrorCopyMode copy_mode,
1661                              Error **errp)
1662 {
1663     MirrorBlockJob *s;
1664     MirrorBDSOpaque *bs_opaque;
1665     BlockDriverState *mirror_top_bs;
1666     bool target_is_backing;
1667     uint64_t target_perms, target_shared_perms;
1668     int ret;
1669 
1670     if (granularity == 0) {
1671         granularity = bdrv_get_default_bitmap_granularity(target);
1672     }
1673 
1674     assert(is_power_of_2(granularity));
1675 
1676     if (buf_size < 0) {
1677         error_setg(errp, "Invalid parameter 'buf-size'");
1678         return NULL;
1679     }
1680 
1681     if (buf_size == 0) {
1682         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1683     }
1684 
1685     if (bdrv_skip_filters(bs) == bdrv_skip_filters(target)) {
1686         error_setg(errp, "Can't mirror node into itself");
1687         return NULL;
1688     }
1689 
1690     target_is_backing = bdrv_chain_contains(bs, target);
1691 
1692     /* In the case of active commit, add dummy driver to provide consistent
1693      * reads on the top, while disabling it in the intermediate nodes, and make
1694      * the backing chain writable. */
1695     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1696                                          BDRV_O_RDWR, errp);
1697     if (mirror_top_bs == NULL) {
1698         return NULL;
1699     }
1700     if (!filter_node_name) {
1701         mirror_top_bs->implicit = true;
1702     }
1703 
1704     /* So that we can always drop this node */
1705     mirror_top_bs->never_freeze = true;
1706 
1707     mirror_top_bs->total_sectors = bs->total_sectors;
1708     mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
1709     mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
1710                                           BDRV_REQ_NO_FALLBACK;
1711     bs_opaque = g_new0(MirrorBDSOpaque, 1);
1712     mirror_top_bs->opaque = bs_opaque;
1713 
1714     bs_opaque->is_commit = target_is_backing;
1715 
1716     bdrv_drained_begin(bs);
1717     ret = bdrv_append(mirror_top_bs, bs, errp);
1718     bdrv_drained_end(bs);
1719 
1720     if (ret < 0) {
1721         bdrv_unref(mirror_top_bs);
1722         return NULL;
1723     }
1724 
1725     /* Make sure that the source is not resized while the job is running */
1726     s = block_job_create(job_id, driver, NULL, mirror_top_bs,
1727                          BLK_PERM_CONSISTENT_READ,
1728                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1729                          BLK_PERM_WRITE, speed,
1730                          creation_flags, cb, opaque, errp);
1731     if (!s) {
1732         goto fail;
1733     }
1734 
1735     /* The block job now has a reference to this node */
1736     bdrv_unref(mirror_top_bs);
1737 
1738     s->mirror_top_bs = mirror_top_bs;
1739 
1740     /* No resize for the target either; while the mirror is still running, a
1741      * consistent read isn't necessarily possible. We could possibly allow
1742      * writes and graph modifications, though it would likely defeat the
1743      * purpose of a mirror, so leave them blocked for now.
1744      *
1745      * In the case of active commit, things look a bit different, though,
1746      * because the target is an already populated backing file in active use.
1747      * We can allow anything except resize there.*/
1748 
1749     target_perms = BLK_PERM_WRITE;
1750     target_shared_perms = BLK_PERM_WRITE_UNCHANGED;
1751 
1752     if (target_is_backing) {
1753         int64_t bs_size, target_size;
1754         bs_size = bdrv_getlength(bs);
1755         if (bs_size < 0) {
1756             error_setg_errno(errp, -bs_size,
1757                              "Could not inquire top image size");
1758             goto fail;
1759         }
1760 
1761         target_size = bdrv_getlength(target);
1762         if (target_size < 0) {
1763             error_setg_errno(errp, -target_size,
1764                              "Could not inquire base image size");
1765             goto fail;
1766         }
1767 
1768         if (target_size < bs_size) {
1769             target_perms |= BLK_PERM_RESIZE;
1770         }
1771 
1772         target_shared_perms |= BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE;
1773     } else if (bdrv_chain_contains(bs, bdrv_skip_filters(target))) {
1774         /*
1775          * We may want to allow this in the future, but it would
1776          * require taking some extra care.
1777          */
1778         error_setg(errp, "Cannot mirror to a filter on top of a node in the "
1779                    "source's backing chain");
1780         goto fail;
1781     }
1782 
1783     s->target = blk_new(s->common.job.aio_context,
1784                         target_perms, target_shared_perms);
1785     ret = blk_insert_bs(s->target, target, errp);
1786     if (ret < 0) {
1787         goto fail;
1788     }
1789     if (is_mirror) {
1790         /* XXX: Mirror target could be a NBD server of target QEMU in the case
1791          * of non-shared block migration. To allow migration completion, we
1792          * have to allow "inactivate" of the target BB.  When that happens, we
1793          * know the job is drained, and the vcpus are stopped, so no write
1794          * operation will be performed. Block layer already has assertions to
1795          * ensure that. */
1796         blk_set_force_allow_inactivate(s->target);
1797     }
1798     blk_set_allow_aio_context_change(s->target, true);
1799     blk_set_disable_request_queuing(s->target, true);
1800 
1801     s->replaces = g_strdup(replaces);
1802     s->on_source_error = on_source_error;
1803     s->on_target_error = on_target_error;
1804     s->is_none_mode = is_none_mode;
1805     s->backing_mode = backing_mode;
1806     s->zero_target = zero_target;
1807     s->copy_mode = copy_mode;
1808     s->base = base;
1809     s->base_overlay = bdrv_find_overlay(bs, base);
1810     s->granularity = granularity;
1811     s->buf_size = ROUND_UP(buf_size, granularity);
1812     s->unmap = unmap;
1813     if (auto_complete) {
1814         s->should_complete = true;
1815     }
1816 
1817     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1818     if (!s->dirty_bitmap) {
1819         goto fail;
1820     }
1821     if (s->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING) {
1822         bdrv_disable_dirty_bitmap(s->dirty_bitmap);
1823     }
1824 
1825     ret = block_job_add_bdrv(&s->common, "source", bs, 0,
1826                              BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE |
1827                              BLK_PERM_CONSISTENT_READ,
1828                              errp);
1829     if (ret < 0) {
1830         goto fail;
1831     }
1832 
1833     /* Required permissions are already taken with blk_new() */
1834     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1835                        &error_abort);
1836 
1837     /* In commit_active_start() all intermediate nodes disappear, so
1838      * any jobs in them must be blocked */
1839     if (target_is_backing) {
1840         BlockDriverState *iter, *filtered_target;
1841         uint64_t iter_shared_perms;
1842 
1843         /*
1844          * The topmost node with
1845          * bdrv_skip_filters(filtered_target) == bdrv_skip_filters(target)
1846          */
1847         filtered_target = bdrv_cow_bs(bdrv_find_overlay(bs, target));
1848 
1849         assert(bdrv_skip_filters(filtered_target) ==
1850                bdrv_skip_filters(target));
1851 
1852         /*
1853          * XXX BLK_PERM_WRITE needs to be allowed so we don't block
1854          * ourselves at s->base (if writes are blocked for a node, they are
1855          * also blocked for its backing file). The other options would be a
1856          * second filter driver above s->base (== target).
1857          */
1858         iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE;
1859 
1860         for (iter = bdrv_filter_or_cow_bs(bs); iter != target;
1861              iter = bdrv_filter_or_cow_bs(iter))
1862         {
1863             if (iter == filtered_target) {
1864                 /*
1865                  * From here on, all nodes are filters on the base.
1866                  * This allows us to share BLK_PERM_CONSISTENT_READ.
1867                  */
1868                 iter_shared_perms |= BLK_PERM_CONSISTENT_READ;
1869             }
1870 
1871             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1872                                      iter_shared_perms, errp);
1873             if (ret < 0) {
1874                 goto fail;
1875             }
1876         }
1877 
1878         if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) {
1879             goto fail;
1880         }
1881     }
1882 
1883     QTAILQ_INIT(&s->ops_in_flight);
1884 
1885     trace_mirror_start(bs, s, opaque);
1886     job_start(&s->common.job);
1887 
1888     return &s->common;
1889 
1890 fail:
1891     if (s) {
1892         /* Make sure this BDS does not go away until we have completed the graph
1893          * changes below */
1894         bdrv_ref(mirror_top_bs);
1895 
1896         g_free(s->replaces);
1897         blk_unref(s->target);
1898         bs_opaque->job = NULL;
1899         if (s->dirty_bitmap) {
1900             bdrv_release_dirty_bitmap(s->dirty_bitmap);
1901         }
1902         job_early_fail(&s->common.job);
1903     }
1904 
1905     bs_opaque->stop = true;
1906     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
1907                              &error_abort);
1908     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
1909 
1910     bdrv_unref(mirror_top_bs);
1911 
1912     return NULL;
1913 }
1914 
1915 void mirror_start(const char *job_id, BlockDriverState *bs,
1916                   BlockDriverState *target, const char *replaces,
1917                   int creation_flags, int64_t speed,
1918                   uint32_t granularity, int64_t buf_size,
1919                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1920                   bool zero_target,
1921                   BlockdevOnError on_source_error,
1922                   BlockdevOnError on_target_error,
1923                   bool unmap, const char *filter_node_name,
1924                   MirrorCopyMode copy_mode, Error **errp)
1925 {
1926     bool is_none_mode;
1927     BlockDriverState *base;
1928 
1929     GLOBAL_STATE_CODE();
1930 
1931     if ((mode == MIRROR_SYNC_MODE_INCREMENTAL) ||
1932         (mode == MIRROR_SYNC_MODE_BITMAP)) {
1933         error_setg(errp, "Sync mode '%s' not supported",
1934                    MirrorSyncMode_str(mode));
1935         return;
1936     }
1937     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1938     base = mode == MIRROR_SYNC_MODE_TOP ? bdrv_backing_chain_next(bs) : NULL;
1939     mirror_start_job(job_id, bs, creation_flags, target, replaces,
1940                      speed, granularity, buf_size, backing_mode, zero_target,
1941                      on_source_error, on_target_error, unmap, NULL, NULL,
1942                      &mirror_job_driver, is_none_mode, base, false,
1943                      filter_node_name, true, copy_mode, errp);
1944 }
1945 
1946 BlockJob *commit_active_start(const char *job_id, BlockDriverState *bs,
1947                               BlockDriverState *base, int creation_flags,
1948                               int64_t speed, BlockdevOnError on_error,
1949                               const char *filter_node_name,
1950                               BlockCompletionFunc *cb, void *opaque,
1951                               bool auto_complete, Error **errp)
1952 {
1953     bool base_read_only;
1954     BlockJob *job;
1955 
1956     GLOBAL_STATE_CODE();
1957 
1958     base_read_only = bdrv_is_read_only(base);
1959 
1960     if (base_read_only) {
1961         if (bdrv_reopen_set_read_only(base, false, errp) < 0) {
1962             return NULL;
1963         }
1964     }
1965 
1966     job = mirror_start_job(
1967                      job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1968                      MIRROR_LEAVE_BACKING_CHAIN, false,
1969                      on_error, on_error, true, cb, opaque,
1970                      &commit_active_job_driver, false, base, auto_complete,
1971                      filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND,
1972                      errp);
1973     if (!job) {
1974         goto error_restore_flags;
1975     }
1976 
1977     return job;
1978 
1979 error_restore_flags:
1980     /* ignore error and errp for bdrv_reopen, because we want to propagate
1981      * the original error */
1982     if (base_read_only) {
1983         bdrv_reopen_set_read_only(base, true, NULL);
1984     }
1985     return NULL;
1986 }
1987