xref: /openbmc/qemu/block/mirror.c (revision 3e3df0d8)
1 /*
2  * Image mirroring
3  *
4  * Copyright Red Hat, Inc. 2012
5  *
6  * Authors:
7  *  Paolo Bonzini  <pbonzini@redhat.com>
8  *
9  * This work is licensed under the terms of the GNU LGPL, version 2 or later.
10  * See the COPYING.LIB file in the top-level directory.
11  *
12  */
13 
14 #include "qemu/osdep.h"
15 #include "qemu/cutils.h"
16 #include "qemu/coroutine.h"
17 #include "qemu/range.h"
18 #include "trace.h"
19 #include "block/blockjob_int.h"
20 #include "block/block_int.h"
21 #include "block/dirty-bitmap.h"
22 #include "sysemu/block-backend.h"
23 #include "qapi/error.h"
24 #include "qemu/ratelimit.h"
25 #include "qemu/bitmap.h"
26 #include "qemu/memalign.h"
27 
28 #define MAX_IN_FLIGHT 16
29 #define MAX_IO_BYTES (1 << 20) /* 1 Mb */
30 #define DEFAULT_MIRROR_BUF_SIZE (MAX_IN_FLIGHT * MAX_IO_BYTES)
31 
32 /* The mirroring buffer is a list of granularity-sized chunks.
33  * Free chunks are organized in a list.
34  */
35 typedef struct MirrorBuffer {
36     QSIMPLEQ_ENTRY(MirrorBuffer) next;
37 } MirrorBuffer;
38 
39 typedef struct MirrorOp MirrorOp;
40 
41 typedef struct MirrorBlockJob {
42     BlockJob common;
43     BlockBackend *target;
44     BlockDriverState *mirror_top_bs;
45     BlockDriverState *base;
46     BlockDriverState *base_overlay;
47 
48     /* The name of the graph node to replace */
49     char *replaces;
50     /* The BDS to replace */
51     BlockDriverState *to_replace;
52     /* Used to block operations on the drive-mirror-replace target */
53     Error *replace_blocker;
54     bool is_none_mode;
55     BlockMirrorBackingMode backing_mode;
56     /* Whether the target image requires explicit zero-initialization */
57     bool zero_target;
58     MirrorCopyMode copy_mode;
59     BlockdevOnError on_source_error, on_target_error;
60     /* Set when the target is synced (dirty bitmap is clean, nothing
61      * in flight) and the job is running in active mode */
62     bool actively_synced;
63     bool should_complete;
64     int64_t granularity;
65     size_t buf_size;
66     int64_t bdev_length;
67     unsigned long *cow_bitmap;
68     BdrvDirtyBitmap *dirty_bitmap;
69     BdrvDirtyBitmapIter *dbi;
70     uint8_t *buf;
71     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
72     int buf_free_count;
73 
74     uint64_t last_pause_ns;
75     unsigned long *in_flight_bitmap;
76     unsigned in_flight;
77     int64_t bytes_in_flight;
78     QTAILQ_HEAD(, MirrorOp) ops_in_flight;
79     int ret;
80     bool unmap;
81     int target_cluster_size;
82     int max_iov;
83     bool initial_zeroing_ongoing;
84     int in_active_write_counter;
85     int64_t active_write_bytes_in_flight;
86     bool prepared;
87     bool in_drain;
88 } MirrorBlockJob;
89 
90 typedef struct MirrorBDSOpaque {
91     MirrorBlockJob *job;
92     bool stop;
93     bool is_commit;
94 } MirrorBDSOpaque;
95 
96 struct MirrorOp {
97     MirrorBlockJob *s;
98     QEMUIOVector qiov;
99     int64_t offset;
100     uint64_t bytes;
101 
102     /* The pointee is set by mirror_co_read(), mirror_co_zero(), and
103      * mirror_co_discard() before yielding for the first time */
104     int64_t *bytes_handled;
105 
106     bool is_pseudo_op;
107     bool is_active_write;
108     bool is_in_flight;
109     CoQueue waiting_requests;
110     Coroutine *co;
111     MirrorOp *waiting_for_op;
112 
113     QTAILQ_ENTRY(MirrorOp) next;
114 };
115 
116 typedef enum MirrorMethod {
117     MIRROR_METHOD_COPY,
118     MIRROR_METHOD_ZERO,
119     MIRROR_METHOD_DISCARD,
120 } MirrorMethod;
121 
122 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
123                                             int error)
124 {
125     s->actively_synced = false;
126     if (read) {
127         return block_job_error_action(&s->common, s->on_source_error,
128                                       true, error);
129     } else {
130         return block_job_error_action(&s->common, s->on_target_error,
131                                       false, error);
132     }
133 }
134 
135 static void coroutine_fn mirror_wait_on_conflicts(MirrorOp *self,
136                                                   MirrorBlockJob *s,
137                                                   uint64_t offset,
138                                                   uint64_t bytes)
139 {
140     uint64_t self_start_chunk = offset / s->granularity;
141     uint64_t self_end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
142     uint64_t self_nb_chunks = self_end_chunk - self_start_chunk;
143 
144     while (find_next_bit(s->in_flight_bitmap, self_end_chunk,
145                          self_start_chunk) < self_end_chunk &&
146            s->ret >= 0)
147     {
148         MirrorOp *op;
149 
150         QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
151             uint64_t op_start_chunk = op->offset / s->granularity;
152             uint64_t op_nb_chunks = DIV_ROUND_UP(op->offset + op->bytes,
153                                                  s->granularity) -
154                                     op_start_chunk;
155 
156             if (op == self) {
157                 continue;
158             }
159 
160             if (ranges_overlap(self_start_chunk, self_nb_chunks,
161                                op_start_chunk, op_nb_chunks))
162             {
163                 if (self) {
164                     /*
165                      * If the operation is already (indirectly) waiting for us,
166                      * or will wait for us as soon as it wakes up, then just go
167                      * on (instead of producing a deadlock in the former case).
168                      */
169                     if (op->waiting_for_op) {
170                         continue;
171                     }
172 
173                     self->waiting_for_op = op;
174                 }
175 
176                 qemu_co_queue_wait(&op->waiting_requests, NULL);
177 
178                 if (self) {
179                     self->waiting_for_op = NULL;
180                 }
181 
182                 break;
183             }
184         }
185     }
186 }
187 
188 static void coroutine_fn mirror_iteration_done(MirrorOp *op, int ret)
189 {
190     MirrorBlockJob *s = op->s;
191     struct iovec *iov;
192     int64_t chunk_num;
193     int i, nb_chunks;
194 
195     trace_mirror_iteration_done(s, op->offset, op->bytes, ret);
196 
197     s->in_flight--;
198     s->bytes_in_flight -= op->bytes;
199     iov = op->qiov.iov;
200     for (i = 0; i < op->qiov.niov; i++) {
201         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
202         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
203         s->buf_free_count++;
204     }
205 
206     chunk_num = op->offset / s->granularity;
207     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
208 
209     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
210     QTAILQ_REMOVE(&s->ops_in_flight, op, next);
211     if (ret >= 0) {
212         if (s->cow_bitmap) {
213             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
214         }
215         if (!s->initial_zeroing_ongoing) {
216             job_progress_update(&s->common.job, op->bytes);
217         }
218     }
219     qemu_iovec_destroy(&op->qiov);
220 
221     qemu_co_queue_restart_all(&op->waiting_requests);
222     g_free(op);
223 }
224 
225 static void coroutine_fn mirror_write_complete(MirrorOp *op, int ret)
226 {
227     MirrorBlockJob *s = op->s;
228 
229     if (ret < 0) {
230         BlockErrorAction action;
231 
232         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
233         action = mirror_error_action(s, false, -ret);
234         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
235             s->ret = ret;
236         }
237     }
238 
239     mirror_iteration_done(op, ret);
240 }
241 
242 static void coroutine_fn mirror_read_complete(MirrorOp *op, int ret)
243 {
244     MirrorBlockJob *s = op->s;
245 
246     if (ret < 0) {
247         BlockErrorAction action;
248 
249         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
250         action = mirror_error_action(s, true, -ret);
251         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
252             s->ret = ret;
253         }
254 
255         mirror_iteration_done(op, ret);
256         return;
257     }
258 
259     ret = blk_co_pwritev(s->target, op->offset, op->qiov.size, &op->qiov, 0);
260     mirror_write_complete(op, ret);
261 }
262 
263 /* Clip bytes relative to offset to not exceed end-of-file */
264 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s,
265                                         int64_t offset,
266                                         int64_t bytes)
267 {
268     return MIN(bytes, s->bdev_length - offset);
269 }
270 
271 /* Round offset and/or bytes to target cluster if COW is needed, and
272  * return the offset of the adjusted tail against original. */
273 static int coroutine_fn mirror_cow_align(MirrorBlockJob *s, int64_t *offset,
274                                          uint64_t *bytes)
275 {
276     bool need_cow;
277     int ret = 0;
278     int64_t align_offset = *offset;
279     int64_t align_bytes = *bytes;
280     int max_bytes = s->granularity * s->max_iov;
281 
282     need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap);
283     need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity,
284                           s->cow_bitmap);
285     if (need_cow) {
286         bdrv_round_to_subclusters(blk_bs(s->target), *offset, *bytes,
287                                   &align_offset, &align_bytes);
288     }
289 
290     if (align_bytes > max_bytes) {
291         align_bytes = max_bytes;
292         if (need_cow) {
293             align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size);
294         }
295     }
296     /* Clipping may result in align_bytes unaligned to chunk boundary, but
297      * that doesn't matter because it's already the end of source image. */
298     align_bytes = mirror_clip_bytes(s, align_offset, align_bytes);
299 
300     ret = align_offset + align_bytes - (*offset + *bytes);
301     *offset = align_offset;
302     *bytes = align_bytes;
303     assert(ret >= 0);
304     return ret;
305 }
306 
307 static inline void coroutine_fn
308 mirror_wait_for_free_in_flight_slot(MirrorBlockJob *s)
309 {
310     MirrorOp *op;
311 
312     QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
313         /*
314          * Do not wait on pseudo ops, because it may in turn wait on
315          * some other operation to start, which may in fact be the
316          * caller of this function.  Since there is only one pseudo op
317          * at any given time, we will always find some real operation
318          * to wait on.
319          * Also, do not wait on active operations, because they do not
320          * use up in-flight slots.
321          */
322         if (!op->is_pseudo_op && op->is_in_flight && !op->is_active_write) {
323             qemu_co_queue_wait(&op->waiting_requests, NULL);
324             return;
325         }
326     }
327     abort();
328 }
329 
330 /* Perform a mirror copy operation.
331  *
332  * *op->bytes_handled is set to the number of bytes copied after and
333  * including offset, excluding any bytes copied prior to offset due
334  * to alignment.  This will be op->bytes if no alignment is necessary,
335  * or (new_end - op->offset) if the tail is rounded up or down due to
336  * alignment or buffer limit.
337  */
338 static void coroutine_fn mirror_co_read(void *opaque)
339 {
340     MirrorOp *op = opaque;
341     MirrorBlockJob *s = op->s;
342     int nb_chunks;
343     uint64_t ret;
344     uint64_t max_bytes;
345 
346     max_bytes = s->granularity * s->max_iov;
347 
348     /* We can only handle as much as buf_size at a time. */
349     op->bytes = MIN(s->buf_size, MIN(max_bytes, op->bytes));
350     assert(op->bytes);
351     assert(op->bytes < BDRV_REQUEST_MAX_BYTES);
352     *op->bytes_handled = op->bytes;
353 
354     if (s->cow_bitmap) {
355         *op->bytes_handled += mirror_cow_align(s, &op->offset, &op->bytes);
356     }
357     /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */
358     assert(*op->bytes_handled <= UINT_MAX);
359     assert(op->bytes <= s->buf_size);
360     /* The offset is granularity-aligned because:
361      * 1) Caller passes in aligned values;
362      * 2) mirror_cow_align is used only when target cluster is larger. */
363     assert(QEMU_IS_ALIGNED(op->offset, s->granularity));
364     /* The range is sector-aligned, since bdrv_getlength() rounds up. */
365     assert(QEMU_IS_ALIGNED(op->bytes, BDRV_SECTOR_SIZE));
366     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
367 
368     while (s->buf_free_count < nb_chunks) {
369         trace_mirror_yield_in_flight(s, op->offset, s->in_flight);
370         mirror_wait_for_free_in_flight_slot(s);
371     }
372 
373     /* Now make a QEMUIOVector taking enough granularity-sized chunks
374      * from s->buf_free.
375      */
376     qemu_iovec_init(&op->qiov, nb_chunks);
377     while (nb_chunks-- > 0) {
378         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
379         size_t remaining = op->bytes - op->qiov.size;
380 
381         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
382         s->buf_free_count--;
383         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
384     }
385 
386     /* Copy the dirty cluster.  */
387     s->in_flight++;
388     s->bytes_in_flight += op->bytes;
389     op->is_in_flight = true;
390     trace_mirror_one_iteration(s, op->offset, op->bytes);
391 
392     WITH_GRAPH_RDLOCK_GUARD() {
393         ret = bdrv_co_preadv(s->mirror_top_bs->backing, op->offset, op->bytes,
394                              &op->qiov, 0);
395     }
396     mirror_read_complete(op, ret);
397 }
398 
399 static void coroutine_fn mirror_co_zero(void *opaque)
400 {
401     MirrorOp *op = opaque;
402     int ret;
403 
404     op->s->in_flight++;
405     op->s->bytes_in_flight += op->bytes;
406     *op->bytes_handled = op->bytes;
407     op->is_in_flight = true;
408 
409     ret = blk_co_pwrite_zeroes(op->s->target, op->offset, op->bytes,
410                                op->s->unmap ? BDRV_REQ_MAY_UNMAP : 0);
411     mirror_write_complete(op, ret);
412 }
413 
414 static void coroutine_fn mirror_co_discard(void *opaque)
415 {
416     MirrorOp *op = opaque;
417     int ret;
418 
419     op->s->in_flight++;
420     op->s->bytes_in_flight += op->bytes;
421     *op->bytes_handled = op->bytes;
422     op->is_in_flight = true;
423 
424     ret = blk_co_pdiscard(op->s->target, op->offset, op->bytes);
425     mirror_write_complete(op, ret);
426 }
427 
428 static unsigned mirror_perform(MirrorBlockJob *s, int64_t offset,
429                                unsigned bytes, MirrorMethod mirror_method)
430 {
431     MirrorOp *op;
432     Coroutine *co;
433     int64_t bytes_handled = -1;
434 
435     op = g_new(MirrorOp, 1);
436     *op = (MirrorOp){
437         .s              = s,
438         .offset         = offset,
439         .bytes          = bytes,
440         .bytes_handled  = &bytes_handled,
441     };
442     qemu_co_queue_init(&op->waiting_requests);
443 
444     switch (mirror_method) {
445     case MIRROR_METHOD_COPY:
446         co = qemu_coroutine_create(mirror_co_read, op);
447         break;
448     case MIRROR_METHOD_ZERO:
449         co = qemu_coroutine_create(mirror_co_zero, op);
450         break;
451     case MIRROR_METHOD_DISCARD:
452         co = qemu_coroutine_create(mirror_co_discard, op);
453         break;
454     default:
455         abort();
456     }
457     op->co = co;
458 
459     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
460     qemu_coroutine_enter(co);
461     /* At this point, ownership of op has been moved to the coroutine
462      * and the object may already be freed */
463 
464     /* Assert that this value has been set */
465     assert(bytes_handled >= 0);
466 
467     /* Same assertion as in mirror_co_read() (and for mirror_co_read()
468      * and mirror_co_discard(), bytes_handled == op->bytes, which
469      * is the @bytes parameter given to this function) */
470     assert(bytes_handled <= UINT_MAX);
471     return bytes_handled;
472 }
473 
474 static void coroutine_fn mirror_iteration(MirrorBlockJob *s)
475 {
476     BlockDriverState *source = s->mirror_top_bs->backing->bs;
477     MirrorOp *pseudo_op;
478     int64_t offset;
479     /* At least the first dirty chunk is mirrored in one iteration. */
480     int nb_chunks = 1;
481     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
482     int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES);
483 
484     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
485     offset = bdrv_dirty_iter_next(s->dbi);
486     if (offset < 0) {
487         bdrv_set_dirty_iter(s->dbi, 0);
488         offset = bdrv_dirty_iter_next(s->dbi);
489         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
490         assert(offset >= 0);
491     }
492     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
493 
494     /*
495      * Wait for concurrent requests to @offset.  The next loop will limit the
496      * copied area based on in_flight_bitmap so we only copy an area that does
497      * not overlap with concurrent in-flight requests.  Still, we would like to
498      * copy something, so wait until there are at least no more requests to the
499      * very beginning of the area.
500      */
501     mirror_wait_on_conflicts(NULL, s, offset, 1);
502 
503     job_pause_point(&s->common.job);
504 
505     /* Find the number of consecutive dirty chunks following the first dirty
506      * one, and wait for in flight requests in them. */
507     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
508     while (nb_chunks * s->granularity < s->buf_size) {
509         int64_t next_dirty;
510         int64_t next_offset = offset + nb_chunks * s->granularity;
511         int64_t next_chunk = next_offset / s->granularity;
512         if (next_offset >= s->bdev_length ||
513             !bdrv_dirty_bitmap_get_locked(s->dirty_bitmap, next_offset)) {
514             break;
515         }
516         if (test_bit(next_chunk, s->in_flight_bitmap)) {
517             break;
518         }
519 
520         next_dirty = bdrv_dirty_iter_next(s->dbi);
521         if (next_dirty > next_offset || next_dirty < 0) {
522             /* The bitmap iterator's cache is stale, refresh it */
523             bdrv_set_dirty_iter(s->dbi, next_offset);
524             next_dirty = bdrv_dirty_iter_next(s->dbi);
525         }
526         assert(next_dirty == next_offset);
527         nb_chunks++;
528     }
529 
530     /* Clear dirty bits before querying the block status, because
531      * calling bdrv_block_status_above could yield - if some blocks are
532      * marked dirty in this window, we need to know.
533      */
534     bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset,
535                                    nb_chunks * s->granularity);
536     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
537 
538     /* Before claiming an area in the in-flight bitmap, we have to
539      * create a MirrorOp for it so that conflicting requests can wait
540      * for it.  mirror_perform() will create the real MirrorOps later,
541      * for now we just create a pseudo operation that will wake up all
542      * conflicting requests once all real operations have been
543      * launched. */
544     pseudo_op = g_new(MirrorOp, 1);
545     *pseudo_op = (MirrorOp){
546         .offset         = offset,
547         .bytes          = nb_chunks * s->granularity,
548         .is_pseudo_op   = true,
549     };
550     qemu_co_queue_init(&pseudo_op->waiting_requests);
551     QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next);
552 
553     bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks);
554     while (nb_chunks > 0 && offset < s->bdev_length) {
555         int ret;
556         int64_t io_bytes;
557         int64_t io_bytes_acct;
558         MirrorMethod mirror_method = MIRROR_METHOD_COPY;
559 
560         assert(!(offset % s->granularity));
561         WITH_GRAPH_RDLOCK_GUARD() {
562             ret = bdrv_block_status_above(source, NULL, offset,
563                                         nb_chunks * s->granularity,
564                                         &io_bytes, NULL, NULL);
565         }
566         if (ret < 0) {
567             io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes);
568         } else if (ret & BDRV_BLOCK_DATA) {
569             io_bytes = MIN(io_bytes, max_io_bytes);
570         }
571 
572         io_bytes -= io_bytes % s->granularity;
573         if (io_bytes < s->granularity) {
574             io_bytes = s->granularity;
575         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
576             int64_t target_offset;
577             int64_t target_bytes;
578             WITH_GRAPH_RDLOCK_GUARD() {
579                 bdrv_round_to_subclusters(blk_bs(s->target), offset, io_bytes,
580                                           &target_offset, &target_bytes);
581             }
582             if (target_offset == offset &&
583                 target_bytes == io_bytes) {
584                 mirror_method = ret & BDRV_BLOCK_ZERO ?
585                                     MIRROR_METHOD_ZERO :
586                                     MIRROR_METHOD_DISCARD;
587             }
588         }
589 
590         while (s->in_flight >= MAX_IN_FLIGHT) {
591             trace_mirror_yield_in_flight(s, offset, s->in_flight);
592             mirror_wait_for_free_in_flight_slot(s);
593         }
594 
595         if (s->ret < 0) {
596             ret = 0;
597             goto fail;
598         }
599 
600         io_bytes = mirror_clip_bytes(s, offset, io_bytes);
601         io_bytes = mirror_perform(s, offset, io_bytes, mirror_method);
602         if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) {
603             io_bytes_acct = 0;
604         } else {
605             io_bytes_acct = io_bytes;
606         }
607         assert(io_bytes);
608         offset += io_bytes;
609         nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity);
610         block_job_ratelimit_processed_bytes(&s->common, io_bytes_acct);
611     }
612 
613 fail:
614     QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next);
615     qemu_co_queue_restart_all(&pseudo_op->waiting_requests);
616     g_free(pseudo_op);
617 }
618 
619 static void mirror_free_init(MirrorBlockJob *s)
620 {
621     int granularity = s->granularity;
622     size_t buf_size = s->buf_size;
623     uint8_t *buf = s->buf;
624 
625     assert(s->buf_free_count == 0);
626     QSIMPLEQ_INIT(&s->buf_free);
627     while (buf_size != 0) {
628         MirrorBuffer *cur = (MirrorBuffer *)buf;
629         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
630         s->buf_free_count++;
631         buf_size -= granularity;
632         buf += granularity;
633     }
634 }
635 
636 /* This is also used for the .pause callback. There is no matching
637  * mirror_resume() because mirror_run() will begin iterating again
638  * when the job is resumed.
639  */
640 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s)
641 {
642     while (s->in_flight > 0) {
643         mirror_wait_for_free_in_flight_slot(s);
644     }
645 }
646 
647 /**
648  * mirror_exit_common: handle both abort() and prepare() cases.
649  * for .prepare, returns 0 on success and -errno on failure.
650  * for .abort cases, denoted by abort = true, MUST return 0.
651  */
652 static int mirror_exit_common(Job *job)
653 {
654     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
655     BlockJob *bjob = &s->common;
656     MirrorBDSOpaque *bs_opaque;
657     AioContext *replace_aio_context = NULL;
658     BlockDriverState *src;
659     BlockDriverState *target_bs;
660     BlockDriverState *mirror_top_bs;
661     Error *local_err = NULL;
662     bool abort = job->ret < 0;
663     int ret = 0;
664 
665     GLOBAL_STATE_CODE();
666 
667     if (s->prepared) {
668         return 0;
669     }
670     s->prepared = true;
671 
672     aio_context_acquire(qemu_get_aio_context());
673 
674     mirror_top_bs = s->mirror_top_bs;
675     bs_opaque = mirror_top_bs->opaque;
676     src = mirror_top_bs->backing->bs;
677     target_bs = blk_bs(s->target);
678 
679     if (bdrv_chain_contains(src, target_bs)) {
680         bdrv_unfreeze_backing_chain(mirror_top_bs, target_bs);
681     }
682 
683     bdrv_release_dirty_bitmap(s->dirty_bitmap);
684 
685     /* Make sure that the source BDS doesn't go away during bdrv_replace_node,
686      * before we can call bdrv_drained_end */
687     bdrv_ref(src);
688     bdrv_ref(mirror_top_bs);
689     bdrv_ref(target_bs);
690 
691     /*
692      * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
693      * inserting target_bs at s->to_replace, where we might not be able to get
694      * these permissions.
695      */
696     blk_unref(s->target);
697     s->target = NULL;
698 
699     /* We don't access the source any more. Dropping any WRITE/RESIZE is
700      * required before it could become a backing file of target_bs. Not having
701      * these permissions any more means that we can't allow any new requests on
702      * mirror_top_bs from now on, so keep it drained. */
703     bdrv_drained_begin(mirror_top_bs);
704     bs_opaque->stop = true;
705 
706     bdrv_graph_rdlock_main_loop();
707     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
708                              &error_abort);
709     bdrv_graph_rdunlock_main_loop();
710 
711     if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
712         BlockDriverState *backing = s->is_none_mode ? src : s->base;
713         BlockDriverState *unfiltered_target = bdrv_skip_filters(target_bs);
714 
715         if (bdrv_cow_bs(unfiltered_target) != backing) {
716             bdrv_set_backing_hd(unfiltered_target, backing, &local_err);
717             if (local_err) {
718                 error_report_err(local_err);
719                 local_err = NULL;
720                 ret = -EPERM;
721             }
722         }
723     } else if (!abort && s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
724         assert(!bdrv_backing_chain_next(target_bs));
725         ret = bdrv_open_backing_file(bdrv_skip_filters(target_bs), NULL,
726                                      "backing", &local_err);
727         if (ret < 0) {
728             error_report_err(local_err);
729             local_err = NULL;
730         }
731     }
732 
733     if (s->to_replace) {
734         replace_aio_context = bdrv_get_aio_context(s->to_replace);
735         aio_context_acquire(replace_aio_context);
736     }
737 
738     if (s->should_complete && !abort) {
739         BlockDriverState *to_replace = s->to_replace ?: src;
740         bool ro = bdrv_is_read_only(to_replace);
741 
742         if (ro != bdrv_is_read_only(target_bs)) {
743             bdrv_reopen_set_read_only(target_bs, ro, NULL);
744         }
745 
746         /* The mirror job has no requests in flight any more, but we need to
747          * drain potential other users of the BDS before changing the graph. */
748         assert(s->in_drain);
749         bdrv_drained_begin(target_bs);
750         /*
751          * Cannot use check_to_replace_node() here, because that would
752          * check for an op blocker on @to_replace, and we have our own
753          * there.
754          *
755          * TODO Pull out the writer lock from bdrv_replace_node() to here
756          */
757         bdrv_graph_rdlock_main_loop();
758         if (bdrv_recurse_can_replace(src, to_replace)) {
759             bdrv_replace_node(to_replace, target_bs, &local_err);
760         } else {
761             error_setg(&local_err, "Can no longer replace '%s' by '%s', "
762                        "because it can no longer be guaranteed that doing so "
763                        "would not lead to an abrupt change of visible data",
764                        to_replace->node_name, target_bs->node_name);
765         }
766         bdrv_graph_rdunlock_main_loop();
767         bdrv_drained_end(target_bs);
768         if (local_err) {
769             error_report_err(local_err);
770             ret = -EPERM;
771         }
772     }
773     if (s->to_replace) {
774         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
775         error_free(s->replace_blocker);
776         bdrv_unref(s->to_replace);
777     }
778     if (replace_aio_context) {
779         aio_context_release(replace_aio_context);
780     }
781     g_free(s->replaces);
782     bdrv_unref(target_bs);
783 
784     /*
785      * Remove the mirror filter driver from the graph. Before this, get rid of
786      * the blockers on the intermediate nodes so that the resulting state is
787      * valid.
788      */
789     block_job_remove_all_bdrv(bjob);
790     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
791 
792     bs_opaque->job = NULL;
793 
794     bdrv_drained_end(src);
795     bdrv_drained_end(mirror_top_bs);
796     s->in_drain = false;
797     bdrv_unref(mirror_top_bs);
798     bdrv_unref(src);
799 
800     aio_context_release(qemu_get_aio_context());
801 
802     return ret;
803 }
804 
805 static int mirror_prepare(Job *job)
806 {
807     return mirror_exit_common(job);
808 }
809 
810 static void mirror_abort(Job *job)
811 {
812     int ret = mirror_exit_common(job);
813     assert(ret == 0);
814 }
815 
816 static void coroutine_fn mirror_throttle(MirrorBlockJob *s)
817 {
818     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
819 
820     if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) {
821         s->last_pause_ns = now;
822         job_sleep_ns(&s->common.job, 0);
823     } else {
824         job_pause_point(&s->common.job);
825     }
826 }
827 
828 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
829 {
830     int64_t offset;
831     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
832     BlockDriverState *target_bs = blk_bs(s->target);
833     int ret;
834     int64_t count;
835 
836     if (s->zero_target) {
837         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
838             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
839             return 0;
840         }
841 
842         s->initial_zeroing_ongoing = true;
843         for (offset = 0; offset < s->bdev_length; ) {
844             int bytes = MIN(s->bdev_length - offset,
845                             QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
846 
847             mirror_throttle(s);
848 
849             if (job_is_cancelled(&s->common.job)) {
850                 s->initial_zeroing_ongoing = false;
851                 return 0;
852             }
853 
854             if (s->in_flight >= MAX_IN_FLIGHT) {
855                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
856                                    s->in_flight);
857                 mirror_wait_for_free_in_flight_slot(s);
858                 continue;
859             }
860 
861             mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO);
862             offset += bytes;
863         }
864 
865         mirror_wait_for_all_io(s);
866         s->initial_zeroing_ongoing = false;
867     }
868 
869     /* First part, loop on the sectors and initialize the dirty bitmap.  */
870     for (offset = 0; offset < s->bdev_length; ) {
871         /* Just to make sure we are not exceeding int limit. */
872         int bytes = MIN(s->bdev_length - offset,
873                         QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
874 
875         mirror_throttle(s);
876 
877         if (job_is_cancelled(&s->common.job)) {
878             return 0;
879         }
880 
881         WITH_GRAPH_RDLOCK_GUARD() {
882             ret = bdrv_is_allocated_above(bs, s->base_overlay, true, offset,
883                                           bytes, &count);
884         }
885         if (ret < 0) {
886             return ret;
887         }
888 
889         assert(count);
890         if (ret > 0) {
891             bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
892         }
893         offset += count;
894     }
895     return 0;
896 }
897 
898 /* Called when going out of the streaming phase to flush the bulk of the
899  * data to the medium, or just before completing.
900  */
901 static int coroutine_fn mirror_flush(MirrorBlockJob *s)
902 {
903     int ret = blk_co_flush(s->target);
904     if (ret < 0) {
905         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
906             s->ret = ret;
907         }
908     }
909     return ret;
910 }
911 
912 static int coroutine_fn mirror_run(Job *job, Error **errp)
913 {
914     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
915     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
916     MirrorBDSOpaque *mirror_top_opaque = s->mirror_top_bs->opaque;
917     BlockDriverState *target_bs = blk_bs(s->target);
918     bool need_drain = true;
919     BlockDeviceIoStatus iostatus;
920     int64_t length;
921     int64_t target_length;
922     BlockDriverInfo bdi;
923     char backing_filename[2]; /* we only need 2 characters because we are only
924                                  checking for a NULL string */
925     int ret = 0;
926 
927     if (job_is_cancelled(&s->common.job)) {
928         goto immediate_exit;
929     }
930 
931     bdrv_graph_co_rdlock();
932     s->bdev_length = bdrv_co_getlength(bs);
933     bdrv_graph_co_rdunlock();
934 
935     if (s->bdev_length < 0) {
936         ret = s->bdev_length;
937         goto immediate_exit;
938     }
939 
940     target_length = blk_co_getlength(s->target);
941     if (target_length < 0) {
942         ret = target_length;
943         goto immediate_exit;
944     }
945 
946     /* Active commit must resize the base image if its size differs from the
947      * active layer. */
948     if (s->base == blk_bs(s->target)) {
949         if (s->bdev_length > target_length) {
950             ret = blk_co_truncate(s->target, s->bdev_length, false,
951                                   PREALLOC_MODE_OFF, 0, NULL);
952             if (ret < 0) {
953                 goto immediate_exit;
954             }
955         }
956     } else if (s->bdev_length != target_length) {
957         error_setg(errp, "Source and target image have different sizes");
958         ret = -EINVAL;
959         goto immediate_exit;
960     }
961 
962     if (s->bdev_length == 0) {
963         /* Transition to the READY state and wait for complete. */
964         job_transition_to_ready(&s->common.job);
965         s->actively_synced = true;
966         while (!job_cancel_requested(&s->common.job) && !s->should_complete) {
967             job_yield(&s->common.job);
968         }
969         goto immediate_exit;
970     }
971 
972     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
973     s->in_flight_bitmap = bitmap_new(length);
974 
975     /* If we have no backing file yet in the destination, we cannot let
976      * the destination do COW.  Instead, we copy sectors around the
977      * dirty data if needed.  We need a bitmap to do that.
978      */
979     bdrv_get_backing_filename(target_bs, backing_filename,
980                               sizeof(backing_filename));
981     bdrv_graph_co_rdlock();
982     if (!bdrv_co_get_info(target_bs, &bdi) && bdi.cluster_size) {
983         s->target_cluster_size = bdi.cluster_size;
984     } else {
985         s->target_cluster_size = BDRV_SECTOR_SIZE;
986     }
987     bdrv_graph_co_rdunlock();
988     if (backing_filename[0] && !bdrv_backing_chain_next(target_bs) &&
989         s->granularity < s->target_cluster_size) {
990         s->buf_size = MAX(s->buf_size, s->target_cluster_size);
991         s->cow_bitmap = bitmap_new(length);
992     }
993     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
994 
995     s->buf = qemu_try_blockalign(bs, s->buf_size);
996     if (s->buf == NULL) {
997         ret = -ENOMEM;
998         goto immediate_exit;
999     }
1000 
1001     mirror_free_init(s);
1002 
1003     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1004     if (!s->is_none_mode) {
1005         ret = mirror_dirty_init(s);
1006         if (ret < 0 || job_is_cancelled(&s->common.job)) {
1007             goto immediate_exit;
1008         }
1009     }
1010 
1011     /*
1012      * Only now the job is fully initialised and mirror_top_bs should start
1013      * accessing it.
1014      */
1015     mirror_top_opaque->job = s;
1016 
1017     assert(!s->dbi);
1018     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
1019     for (;;) {
1020         int64_t cnt, delta;
1021         bool should_complete;
1022 
1023         if (s->ret < 0) {
1024             ret = s->ret;
1025             goto immediate_exit;
1026         }
1027 
1028         job_pause_point(&s->common.job);
1029 
1030         if (job_is_cancelled(&s->common.job)) {
1031             ret = 0;
1032             goto immediate_exit;
1033         }
1034 
1035         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1036         /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
1037          * the number of bytes currently being processed; together those are
1038          * the current remaining operation length */
1039         job_progress_set_remaining(&s->common.job,
1040                                    s->bytes_in_flight + cnt +
1041                                    s->active_write_bytes_in_flight);
1042 
1043         /* Note that even when no rate limit is applied we need to yield
1044          * periodically with no pending I/O so that bdrv_drain_all() returns.
1045          * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
1046          * an error, or when the source is clean, whichever comes first. */
1047         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
1048         WITH_JOB_LOCK_GUARD() {
1049             iostatus = s->common.iostatus;
1050         }
1051         if (delta < BLOCK_JOB_SLICE_TIME &&
1052             iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
1053             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
1054                 (cnt == 0 && s->in_flight > 0)) {
1055                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
1056                 mirror_wait_for_free_in_flight_slot(s);
1057                 continue;
1058             } else if (cnt != 0) {
1059                 mirror_iteration(s);
1060             }
1061         }
1062 
1063         should_complete = false;
1064         if (s->in_flight == 0 && cnt == 0) {
1065             trace_mirror_before_flush(s);
1066             if (!job_is_ready(&s->common.job)) {
1067                 if (mirror_flush(s) < 0) {
1068                     /* Go check s->ret.  */
1069                     continue;
1070                 }
1071                 /* We're out of the streaming phase.  From now on, if the job
1072                  * is cancelled we will actually complete all pending I/O and
1073                  * report completion.  This way, block-job-cancel will leave
1074                  * the target in a consistent state.
1075                  */
1076                 job_transition_to_ready(&s->common.job);
1077                 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) {
1078                     s->actively_synced = true;
1079                 }
1080             }
1081 
1082             should_complete = s->should_complete ||
1083                 job_cancel_requested(&s->common.job);
1084             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1085         }
1086 
1087         if (cnt == 0 && should_complete) {
1088             /* The dirty bitmap is not updated while operations are pending.
1089              * If we're about to exit, wait for pending operations before
1090              * calling bdrv_get_dirty_count(bs), or we may exit while the
1091              * source has dirty data to copy!
1092              *
1093              * Note that I/O can be submitted by the guest while
1094              * mirror_populate runs, so pause it now.  Before deciding
1095              * whether to switch to target check one last time if I/O has
1096              * come in the meanwhile, and if not flush the data to disk.
1097              */
1098             trace_mirror_before_drain(s, cnt);
1099 
1100             s->in_drain = true;
1101             bdrv_drained_begin(bs);
1102 
1103             /* Must be zero because we are drained */
1104             assert(s->in_active_write_counter == 0);
1105 
1106             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1107             if (cnt > 0 || mirror_flush(s) < 0) {
1108                 bdrv_drained_end(bs);
1109                 s->in_drain = false;
1110                 continue;
1111             }
1112 
1113             /* The two disks are in sync.  Exit and report successful
1114              * completion.
1115              */
1116             assert(QLIST_EMPTY(&bs->tracked_requests));
1117             need_drain = false;
1118             break;
1119         }
1120 
1121         if (job_is_ready(&s->common.job) && !should_complete) {
1122             if (s->in_flight == 0 && cnt == 0) {
1123                 trace_mirror_before_sleep(s, cnt, job_is_ready(&s->common.job),
1124                                           BLOCK_JOB_SLICE_TIME);
1125                 job_sleep_ns(&s->common.job, BLOCK_JOB_SLICE_TIME);
1126             }
1127         } else {
1128             block_job_ratelimit_sleep(&s->common);
1129         }
1130         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1131     }
1132 
1133 immediate_exit:
1134     if (s->in_flight > 0) {
1135         /* We get here only if something went wrong.  Either the job failed,
1136          * or it was cancelled prematurely so that we do not guarantee that
1137          * the target is a copy of the source.
1138          */
1139         assert(ret < 0 || job_is_cancelled(&s->common.job));
1140         assert(need_drain);
1141         mirror_wait_for_all_io(s);
1142     }
1143 
1144     assert(s->in_flight == 0);
1145     qemu_vfree(s->buf);
1146     g_free(s->cow_bitmap);
1147     g_free(s->in_flight_bitmap);
1148     bdrv_dirty_iter_free(s->dbi);
1149 
1150     if (need_drain) {
1151         s->in_drain = true;
1152         bdrv_drained_begin(bs);
1153     }
1154 
1155     return ret;
1156 }
1157 
1158 static void mirror_complete(Job *job, Error **errp)
1159 {
1160     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1161 
1162     if (!job_is_ready(job)) {
1163         error_setg(errp, "The active block job '%s' cannot be completed",
1164                    job->id);
1165         return;
1166     }
1167 
1168     /* block all operations on to_replace bs */
1169     if (s->replaces) {
1170         AioContext *replace_aio_context;
1171 
1172         s->to_replace = bdrv_find_node(s->replaces);
1173         if (!s->to_replace) {
1174             error_setg(errp, "Node name '%s' not found", s->replaces);
1175             return;
1176         }
1177 
1178         replace_aio_context = bdrv_get_aio_context(s->to_replace);
1179         aio_context_acquire(replace_aio_context);
1180 
1181         /* TODO Translate this into child freeze system. */
1182         error_setg(&s->replace_blocker,
1183                    "block device is in use by block-job-complete");
1184         bdrv_op_block_all(s->to_replace, s->replace_blocker);
1185         bdrv_ref(s->to_replace);
1186 
1187         aio_context_release(replace_aio_context);
1188     }
1189 
1190     s->should_complete = true;
1191 
1192     /* If the job is paused, it will be re-entered when it is resumed */
1193     WITH_JOB_LOCK_GUARD() {
1194         if (!job->paused) {
1195             job_enter_cond_locked(job, NULL);
1196         }
1197     }
1198 }
1199 
1200 static void coroutine_fn mirror_pause(Job *job)
1201 {
1202     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1203 
1204     mirror_wait_for_all_io(s);
1205 }
1206 
1207 static bool mirror_drained_poll(BlockJob *job)
1208 {
1209     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1210 
1211     /* If the job isn't paused nor cancelled, we can't be sure that it won't
1212      * issue more requests. We make an exception if we've reached this point
1213      * from one of our own drain sections, to avoid a deadlock waiting for
1214      * ourselves.
1215      */
1216     WITH_JOB_LOCK_GUARD() {
1217         if (!s->common.job.paused && !job_is_cancelled_locked(&job->job)
1218             && !s->in_drain) {
1219             return true;
1220         }
1221     }
1222 
1223     return !!s->in_flight;
1224 }
1225 
1226 static bool mirror_cancel(Job *job, bool force)
1227 {
1228     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1229     BlockDriverState *target = blk_bs(s->target);
1230 
1231     /*
1232      * Before the job is READY, we treat any cancellation like a
1233      * force-cancellation.
1234      */
1235     force = force || !job_is_ready(job);
1236 
1237     if (force) {
1238         bdrv_cancel_in_flight(target);
1239     }
1240     return force;
1241 }
1242 
1243 static bool commit_active_cancel(Job *job, bool force)
1244 {
1245     /* Same as above in mirror_cancel() */
1246     return force || !job_is_ready(job);
1247 }
1248 
1249 static const BlockJobDriver mirror_job_driver = {
1250     .job_driver = {
1251         .instance_size          = sizeof(MirrorBlockJob),
1252         .job_type               = JOB_TYPE_MIRROR,
1253         .free                   = block_job_free,
1254         .user_resume            = block_job_user_resume,
1255         .run                    = mirror_run,
1256         .prepare                = mirror_prepare,
1257         .abort                  = mirror_abort,
1258         .pause                  = mirror_pause,
1259         .complete               = mirror_complete,
1260         .cancel                 = mirror_cancel,
1261     },
1262     .drained_poll           = mirror_drained_poll,
1263 };
1264 
1265 static const BlockJobDriver commit_active_job_driver = {
1266     .job_driver = {
1267         .instance_size          = sizeof(MirrorBlockJob),
1268         .job_type               = JOB_TYPE_COMMIT,
1269         .free                   = block_job_free,
1270         .user_resume            = block_job_user_resume,
1271         .run                    = mirror_run,
1272         .prepare                = mirror_prepare,
1273         .abort                  = mirror_abort,
1274         .pause                  = mirror_pause,
1275         .complete               = mirror_complete,
1276         .cancel                 = commit_active_cancel,
1277     },
1278     .drained_poll           = mirror_drained_poll,
1279 };
1280 
1281 static void coroutine_fn
1282 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method,
1283                      uint64_t offset, uint64_t bytes,
1284                      QEMUIOVector *qiov, int flags)
1285 {
1286     int ret;
1287     size_t qiov_offset = 0;
1288     int64_t bitmap_offset, bitmap_end;
1289 
1290     if (!QEMU_IS_ALIGNED(offset, job->granularity) &&
1291         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset))
1292     {
1293             /*
1294              * Dirty unaligned padding: ignore it.
1295              *
1296              * Reasoning:
1297              * 1. If we copy it, we can't reset corresponding bit in
1298              *    dirty_bitmap as there may be some "dirty" bytes still not
1299              *    copied.
1300              * 2. It's already dirty, so skipping it we don't diverge mirror
1301              *    progress.
1302              *
1303              * Note, that because of this, guest write may have no contribution
1304              * into mirror converge, but that's not bad, as we have background
1305              * process of mirroring. If under some bad circumstances (high guest
1306              * IO load) background process starve, we will not converge anyway,
1307              * even if each write will contribute, as guest is not guaranteed to
1308              * rewrite the whole disk.
1309              */
1310             qiov_offset = QEMU_ALIGN_UP(offset, job->granularity) - offset;
1311             if (bytes <= qiov_offset) {
1312                 /* nothing to do after shrink */
1313                 return;
1314             }
1315             offset += qiov_offset;
1316             bytes -= qiov_offset;
1317     }
1318 
1319     if (!QEMU_IS_ALIGNED(offset + bytes, job->granularity) &&
1320         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset + bytes - 1))
1321     {
1322         uint64_t tail = (offset + bytes) % job->granularity;
1323 
1324         if (bytes <= tail) {
1325             /* nothing to do after shrink */
1326             return;
1327         }
1328         bytes -= tail;
1329     }
1330 
1331     /*
1332      * Tails are either clean or shrunk, so for bitmap resetting
1333      * we safely align the range down.
1334      */
1335     bitmap_offset = QEMU_ALIGN_UP(offset, job->granularity);
1336     bitmap_end = QEMU_ALIGN_DOWN(offset + bytes, job->granularity);
1337     if (bitmap_offset < bitmap_end) {
1338         bdrv_reset_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1339                                 bitmap_end - bitmap_offset);
1340     }
1341 
1342     job_progress_increase_remaining(&job->common.job, bytes);
1343     job->active_write_bytes_in_flight += bytes;
1344 
1345     switch (method) {
1346     case MIRROR_METHOD_COPY:
1347         ret = blk_co_pwritev_part(job->target, offset, bytes,
1348                                   qiov, qiov_offset, flags);
1349         break;
1350 
1351     case MIRROR_METHOD_ZERO:
1352         assert(!qiov);
1353         ret = blk_co_pwrite_zeroes(job->target, offset, bytes, flags);
1354         break;
1355 
1356     case MIRROR_METHOD_DISCARD:
1357         assert(!qiov);
1358         ret = blk_co_pdiscard(job->target, offset, bytes);
1359         break;
1360 
1361     default:
1362         abort();
1363     }
1364 
1365     job->active_write_bytes_in_flight -= bytes;
1366     if (ret >= 0) {
1367         job_progress_update(&job->common.job, bytes);
1368     } else {
1369         BlockErrorAction action;
1370 
1371         /*
1372          * We failed, so we should mark dirty the whole area, aligned up.
1373          * Note that we don't care about shrunk tails if any: they were dirty
1374          * at function start, and they must be still dirty, as we've locked
1375          * the region for in-flight op.
1376          */
1377         bitmap_offset = QEMU_ALIGN_DOWN(offset, job->granularity);
1378         bitmap_end = QEMU_ALIGN_UP(offset + bytes, job->granularity);
1379         bdrv_set_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1380                               bitmap_end - bitmap_offset);
1381         job->actively_synced = false;
1382 
1383         action = mirror_error_action(job, false, -ret);
1384         if (action == BLOCK_ERROR_ACTION_REPORT) {
1385             if (!job->ret) {
1386                 job->ret = ret;
1387             }
1388         }
1389     }
1390 }
1391 
1392 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s,
1393                                                    uint64_t offset,
1394                                                    uint64_t bytes)
1395 {
1396     MirrorOp *op;
1397     uint64_t start_chunk = offset / s->granularity;
1398     uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
1399 
1400     op = g_new(MirrorOp, 1);
1401     *op = (MirrorOp){
1402         .s                  = s,
1403         .offset             = offset,
1404         .bytes              = bytes,
1405         .is_active_write    = true,
1406         .is_in_flight       = true,
1407         .co                 = qemu_coroutine_self(),
1408     };
1409     qemu_co_queue_init(&op->waiting_requests);
1410     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
1411 
1412     s->in_active_write_counter++;
1413 
1414     /*
1415      * Wait for concurrent requests affecting the area.  If there are already
1416      * running requests that are copying off now-to-be stale data in the area,
1417      * we must wait for them to finish before we begin writing fresh data to the
1418      * target so that the write operations appear in the correct order.
1419      * Note that background requests (see mirror_iteration()) in contrast only
1420      * wait for conflicting requests at the start of the dirty area, and then
1421      * (based on the in_flight_bitmap) truncate the area to copy so it will not
1422      * conflict with any requests beyond that.  For active writes, however, we
1423      * cannot truncate that area.  The request from our parent must be blocked
1424      * until the area is copied in full.  Therefore, we must wait for the whole
1425      * area to become free of concurrent requests.
1426      */
1427     mirror_wait_on_conflicts(op, s, offset, bytes);
1428 
1429     bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1430 
1431     return op;
1432 }
1433 
1434 static void coroutine_fn GRAPH_RDLOCK active_write_settle(MirrorOp *op)
1435 {
1436     uint64_t start_chunk = op->offset / op->s->granularity;
1437     uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes,
1438                                       op->s->granularity);
1439 
1440     if (!--op->s->in_active_write_counter && op->s->actively_synced) {
1441         BdrvChild *source = op->s->mirror_top_bs->backing;
1442 
1443         if (QLIST_FIRST(&source->bs->parents) == source &&
1444             QLIST_NEXT(source, next_parent) == NULL)
1445         {
1446             /* Assert that we are back in sync once all active write
1447              * operations are settled.
1448              * Note that we can only assert this if the mirror node
1449              * is the source node's only parent. */
1450             assert(!bdrv_get_dirty_count(op->s->dirty_bitmap));
1451         }
1452     }
1453     bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1454     QTAILQ_REMOVE(&op->s->ops_in_flight, op, next);
1455     qemu_co_queue_restart_all(&op->waiting_requests);
1456     g_free(op);
1457 }
1458 
1459 static int coroutine_fn GRAPH_RDLOCK
1460 bdrv_mirror_top_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
1461                        QEMUIOVector *qiov, BdrvRequestFlags flags)
1462 {
1463     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1464 }
1465 
1466 static int coroutine_fn GRAPH_RDLOCK
1467 bdrv_mirror_top_do_write(BlockDriverState *bs, MirrorMethod method,
1468                          uint64_t offset, uint64_t bytes, QEMUIOVector *qiov,
1469                          int flags)
1470 {
1471     MirrorOp *op = NULL;
1472     MirrorBDSOpaque *s = bs->opaque;
1473     int ret = 0;
1474     bool copy_to_target = false;
1475 
1476     if (s->job) {
1477         copy_to_target = s->job->ret >= 0 &&
1478                          !job_is_cancelled(&s->job->common.job) &&
1479                          s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1480     }
1481 
1482     if (copy_to_target) {
1483         op = active_write_prepare(s->job, offset, bytes);
1484     }
1485 
1486     switch (method) {
1487     case MIRROR_METHOD_COPY:
1488         ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1489         break;
1490 
1491     case MIRROR_METHOD_ZERO:
1492         ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1493         break;
1494 
1495     case MIRROR_METHOD_DISCARD:
1496         ret = bdrv_co_pdiscard(bs->backing, offset, bytes);
1497         break;
1498 
1499     default:
1500         abort();
1501     }
1502 
1503     if (ret < 0) {
1504         goto out;
1505     }
1506 
1507     if (copy_to_target) {
1508         do_sync_target_write(s->job, method, offset, bytes, qiov, flags);
1509     }
1510 
1511 out:
1512     if (copy_to_target) {
1513         active_write_settle(op);
1514     }
1515     return ret;
1516 }
1517 
1518 static int coroutine_fn GRAPH_RDLOCK
1519 bdrv_mirror_top_pwritev(BlockDriverState *bs, int64_t offset, int64_t bytes,
1520                         QEMUIOVector *qiov, BdrvRequestFlags flags)
1521 {
1522     MirrorBDSOpaque *s = bs->opaque;
1523     QEMUIOVector bounce_qiov;
1524     void *bounce_buf;
1525     int ret = 0;
1526     bool copy_to_target = false;
1527 
1528     if (s->job) {
1529         copy_to_target = s->job->ret >= 0 &&
1530                          !job_is_cancelled(&s->job->common.job) &&
1531                          s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1532     }
1533 
1534     if (copy_to_target) {
1535         /* The guest might concurrently modify the data to write; but
1536          * the data on source and destination must match, so we have
1537          * to use a bounce buffer if we are going to write to the
1538          * target now. */
1539         bounce_buf = qemu_blockalign(bs, bytes);
1540         iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes);
1541 
1542         qemu_iovec_init(&bounce_qiov, 1);
1543         qemu_iovec_add(&bounce_qiov, bounce_buf, bytes);
1544         qiov = &bounce_qiov;
1545 
1546         flags &= ~BDRV_REQ_REGISTERED_BUF;
1547     }
1548 
1549     ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov,
1550                                    flags);
1551 
1552     if (copy_to_target) {
1553         qemu_iovec_destroy(&bounce_qiov);
1554         qemu_vfree(bounce_buf);
1555     }
1556 
1557     return ret;
1558 }
1559 
1560 static int coroutine_fn GRAPH_RDLOCK bdrv_mirror_top_flush(BlockDriverState *bs)
1561 {
1562     if (bs->backing == NULL) {
1563         /* we can be here after failed bdrv_append in mirror_start_job */
1564         return 0;
1565     }
1566     return bdrv_co_flush(bs->backing->bs);
1567 }
1568 
1569 static int coroutine_fn GRAPH_RDLOCK
1570 bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs, int64_t offset,
1571                               int64_t bytes, BdrvRequestFlags flags)
1572 {
1573     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL,
1574                                     flags);
1575 }
1576 
1577 static int coroutine_fn GRAPH_RDLOCK
1578 bdrv_mirror_top_pdiscard(BlockDriverState *bs, int64_t offset, int64_t bytes)
1579 {
1580     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes,
1581                                     NULL, 0);
1582 }
1583 
1584 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs)
1585 {
1586     if (bs->backing == NULL) {
1587         /* we can be here after failed bdrv_attach_child in
1588          * bdrv_set_backing_hd */
1589         return;
1590     }
1591     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1592             bs->backing->bs->filename);
1593 }
1594 
1595 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1596                                        BdrvChildRole role,
1597                                        BlockReopenQueue *reopen_queue,
1598                                        uint64_t perm, uint64_t shared,
1599                                        uint64_t *nperm, uint64_t *nshared)
1600 {
1601     MirrorBDSOpaque *s = bs->opaque;
1602 
1603     if (s->stop) {
1604         /*
1605          * If the job is to be stopped, we do not need to forward
1606          * anything to the real image.
1607          */
1608         *nperm = 0;
1609         *nshared = BLK_PERM_ALL;
1610         return;
1611     }
1612 
1613     bdrv_default_perms(bs, c, role, reopen_queue,
1614                        perm, shared, nperm, nshared);
1615 
1616     if (s->is_commit) {
1617         /*
1618          * For commit jobs, we cannot take CONSISTENT_READ, because
1619          * that permission is unshared for everything above the base
1620          * node (except for filters on the base node).
1621          * We also have to force-share the WRITE permission, or
1622          * otherwise we would block ourselves at the base node (if
1623          * writes are blocked for a node, they are also blocked for
1624          * its backing file).
1625          * (We could also share RESIZE, because it may be needed for
1626          * the target if its size is less than the top node's; but
1627          * bdrv_default_perms_for_cow() automatically shares RESIZE
1628          * for backing nodes if WRITE is shared, so there is no need
1629          * to do it here.)
1630          */
1631         *nperm &= ~BLK_PERM_CONSISTENT_READ;
1632         *nshared |= BLK_PERM_WRITE;
1633     }
1634 }
1635 
1636 /* Dummy node that provides consistent read to its users without requiring it
1637  * from its backing file and that allows writes on the backing file chain. */
1638 static BlockDriver bdrv_mirror_top = {
1639     .format_name                = "mirror_top",
1640     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1641     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1642     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1643     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1644     .bdrv_co_flush              = bdrv_mirror_top_flush,
1645     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1646     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1647 
1648     .is_filter                  = true,
1649     .filtered_child_is_backing  = true,
1650 };
1651 
1652 static BlockJob *mirror_start_job(
1653                              const char *job_id, BlockDriverState *bs,
1654                              int creation_flags, BlockDriverState *target,
1655                              const char *replaces, int64_t speed,
1656                              uint32_t granularity, int64_t buf_size,
1657                              BlockMirrorBackingMode backing_mode,
1658                              bool zero_target,
1659                              BlockdevOnError on_source_error,
1660                              BlockdevOnError on_target_error,
1661                              bool unmap,
1662                              BlockCompletionFunc *cb,
1663                              void *opaque,
1664                              const BlockJobDriver *driver,
1665                              bool is_none_mode, BlockDriverState *base,
1666                              bool auto_complete, const char *filter_node_name,
1667                              bool is_mirror, MirrorCopyMode copy_mode,
1668                              Error **errp)
1669 {
1670     MirrorBlockJob *s;
1671     MirrorBDSOpaque *bs_opaque;
1672     BlockDriverState *mirror_top_bs;
1673     bool target_is_backing;
1674     uint64_t target_perms, target_shared_perms;
1675     int ret;
1676 
1677     GLOBAL_STATE_CODE();
1678 
1679     if (granularity == 0) {
1680         granularity = bdrv_get_default_bitmap_granularity(target);
1681     }
1682 
1683     assert(is_power_of_2(granularity));
1684 
1685     if (buf_size < 0) {
1686         error_setg(errp, "Invalid parameter 'buf-size'");
1687         return NULL;
1688     }
1689 
1690     if (buf_size == 0) {
1691         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1692     }
1693 
1694     if (bdrv_skip_filters(bs) == bdrv_skip_filters(target)) {
1695         error_setg(errp, "Can't mirror node into itself");
1696         return NULL;
1697     }
1698 
1699     target_is_backing = bdrv_chain_contains(bs, target);
1700 
1701     /* In the case of active commit, add dummy driver to provide consistent
1702      * reads on the top, while disabling it in the intermediate nodes, and make
1703      * the backing chain writable. */
1704     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1705                                          BDRV_O_RDWR, errp);
1706     if (mirror_top_bs == NULL) {
1707         return NULL;
1708     }
1709     if (!filter_node_name) {
1710         mirror_top_bs->implicit = true;
1711     }
1712 
1713     /* So that we can always drop this node */
1714     mirror_top_bs->never_freeze = true;
1715 
1716     mirror_top_bs->total_sectors = bs->total_sectors;
1717     mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
1718     mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
1719                                           BDRV_REQ_NO_FALLBACK;
1720     bs_opaque = g_new0(MirrorBDSOpaque, 1);
1721     mirror_top_bs->opaque = bs_opaque;
1722 
1723     bs_opaque->is_commit = target_is_backing;
1724 
1725     bdrv_drained_begin(bs);
1726     ret = bdrv_append(mirror_top_bs, bs, errp);
1727     bdrv_drained_end(bs);
1728 
1729     if (ret < 0) {
1730         bdrv_unref(mirror_top_bs);
1731         return NULL;
1732     }
1733 
1734     /* Make sure that the source is not resized while the job is running */
1735     s = block_job_create(job_id, driver, NULL, mirror_top_bs,
1736                          BLK_PERM_CONSISTENT_READ,
1737                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1738                          BLK_PERM_WRITE, speed,
1739                          creation_flags, cb, opaque, errp);
1740     if (!s) {
1741         goto fail;
1742     }
1743 
1744     /* The block job now has a reference to this node */
1745     bdrv_unref(mirror_top_bs);
1746 
1747     s->mirror_top_bs = mirror_top_bs;
1748 
1749     /* No resize for the target either; while the mirror is still running, a
1750      * consistent read isn't necessarily possible. We could possibly allow
1751      * writes and graph modifications, though it would likely defeat the
1752      * purpose of a mirror, so leave them blocked for now.
1753      *
1754      * In the case of active commit, things look a bit different, though,
1755      * because the target is an already populated backing file in active use.
1756      * We can allow anything except resize there.*/
1757 
1758     target_perms = BLK_PERM_WRITE;
1759     target_shared_perms = BLK_PERM_WRITE_UNCHANGED;
1760 
1761     if (target_is_backing) {
1762         int64_t bs_size, target_size;
1763         bs_size = bdrv_getlength(bs);
1764         if (bs_size < 0) {
1765             error_setg_errno(errp, -bs_size,
1766                              "Could not inquire top image size");
1767             goto fail;
1768         }
1769 
1770         target_size = bdrv_getlength(target);
1771         if (target_size < 0) {
1772             error_setg_errno(errp, -target_size,
1773                              "Could not inquire base image size");
1774             goto fail;
1775         }
1776 
1777         if (target_size < bs_size) {
1778             target_perms |= BLK_PERM_RESIZE;
1779         }
1780 
1781         target_shared_perms |= BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE;
1782     } else if (bdrv_chain_contains(bs, bdrv_skip_filters(target))) {
1783         /*
1784          * We may want to allow this in the future, but it would
1785          * require taking some extra care.
1786          */
1787         error_setg(errp, "Cannot mirror to a filter on top of a node in the "
1788                    "source's backing chain");
1789         goto fail;
1790     }
1791 
1792     s->target = blk_new(s->common.job.aio_context,
1793                         target_perms, target_shared_perms);
1794     ret = blk_insert_bs(s->target, target, errp);
1795     if (ret < 0) {
1796         goto fail;
1797     }
1798     if (is_mirror) {
1799         /* XXX: Mirror target could be a NBD server of target QEMU in the case
1800          * of non-shared block migration. To allow migration completion, we
1801          * have to allow "inactivate" of the target BB.  When that happens, we
1802          * know the job is drained, and the vcpus are stopped, so no write
1803          * operation will be performed. Block layer already has assertions to
1804          * ensure that. */
1805         blk_set_force_allow_inactivate(s->target);
1806     }
1807     blk_set_allow_aio_context_change(s->target, true);
1808     blk_set_disable_request_queuing(s->target, true);
1809 
1810     s->replaces = g_strdup(replaces);
1811     s->on_source_error = on_source_error;
1812     s->on_target_error = on_target_error;
1813     s->is_none_mode = is_none_mode;
1814     s->backing_mode = backing_mode;
1815     s->zero_target = zero_target;
1816     s->copy_mode = copy_mode;
1817     s->base = base;
1818     s->base_overlay = bdrv_find_overlay(bs, base);
1819     s->granularity = granularity;
1820     s->buf_size = ROUND_UP(buf_size, granularity);
1821     s->unmap = unmap;
1822     if (auto_complete) {
1823         s->should_complete = true;
1824     }
1825 
1826     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1827     if (!s->dirty_bitmap) {
1828         goto fail;
1829     }
1830     if (s->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING) {
1831         bdrv_disable_dirty_bitmap(s->dirty_bitmap);
1832     }
1833 
1834     ret = block_job_add_bdrv(&s->common, "source", bs, 0,
1835                              BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE |
1836                              BLK_PERM_CONSISTENT_READ,
1837                              errp);
1838     if (ret < 0) {
1839         goto fail;
1840     }
1841 
1842     /* Required permissions are already taken with blk_new() */
1843     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1844                        &error_abort);
1845 
1846     /* In commit_active_start() all intermediate nodes disappear, so
1847      * any jobs in them must be blocked */
1848     if (target_is_backing) {
1849         BlockDriverState *iter, *filtered_target;
1850         uint64_t iter_shared_perms;
1851 
1852         /*
1853          * The topmost node with
1854          * bdrv_skip_filters(filtered_target) == bdrv_skip_filters(target)
1855          */
1856         filtered_target = bdrv_cow_bs(bdrv_find_overlay(bs, target));
1857 
1858         assert(bdrv_skip_filters(filtered_target) ==
1859                bdrv_skip_filters(target));
1860 
1861         /*
1862          * XXX BLK_PERM_WRITE needs to be allowed so we don't block
1863          * ourselves at s->base (if writes are blocked for a node, they are
1864          * also blocked for its backing file). The other options would be a
1865          * second filter driver above s->base (== target).
1866          */
1867         iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE;
1868 
1869         for (iter = bdrv_filter_or_cow_bs(bs); iter != target;
1870              iter = bdrv_filter_or_cow_bs(iter))
1871         {
1872             if (iter == filtered_target) {
1873                 /*
1874                  * From here on, all nodes are filters on the base.
1875                  * This allows us to share BLK_PERM_CONSISTENT_READ.
1876                  */
1877                 iter_shared_perms |= BLK_PERM_CONSISTENT_READ;
1878             }
1879 
1880             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1881                                      iter_shared_perms, errp);
1882             if (ret < 0) {
1883                 goto fail;
1884             }
1885         }
1886 
1887         if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) {
1888             goto fail;
1889         }
1890     }
1891 
1892     QTAILQ_INIT(&s->ops_in_flight);
1893 
1894     trace_mirror_start(bs, s, opaque);
1895     job_start(&s->common.job);
1896 
1897     return &s->common;
1898 
1899 fail:
1900     if (s) {
1901         /* Make sure this BDS does not go away until we have completed the graph
1902          * changes below */
1903         bdrv_ref(mirror_top_bs);
1904 
1905         g_free(s->replaces);
1906         blk_unref(s->target);
1907         bs_opaque->job = NULL;
1908         if (s->dirty_bitmap) {
1909             bdrv_release_dirty_bitmap(s->dirty_bitmap);
1910         }
1911         job_early_fail(&s->common.job);
1912     }
1913 
1914     bs_opaque->stop = true;
1915     bdrv_graph_rdlock_main_loop();
1916     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
1917                              &error_abort);
1918     bdrv_graph_rdunlock_main_loop();
1919     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
1920 
1921     bdrv_unref(mirror_top_bs);
1922 
1923     return NULL;
1924 }
1925 
1926 void mirror_start(const char *job_id, BlockDriverState *bs,
1927                   BlockDriverState *target, const char *replaces,
1928                   int creation_flags, int64_t speed,
1929                   uint32_t granularity, int64_t buf_size,
1930                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1931                   bool zero_target,
1932                   BlockdevOnError on_source_error,
1933                   BlockdevOnError on_target_error,
1934                   bool unmap, const char *filter_node_name,
1935                   MirrorCopyMode copy_mode, Error **errp)
1936 {
1937     bool is_none_mode;
1938     BlockDriverState *base;
1939 
1940     GLOBAL_STATE_CODE();
1941 
1942     if ((mode == MIRROR_SYNC_MODE_INCREMENTAL) ||
1943         (mode == MIRROR_SYNC_MODE_BITMAP)) {
1944         error_setg(errp, "Sync mode '%s' not supported",
1945                    MirrorSyncMode_str(mode));
1946         return;
1947     }
1948     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1949     base = mode == MIRROR_SYNC_MODE_TOP ? bdrv_backing_chain_next(bs) : NULL;
1950     mirror_start_job(job_id, bs, creation_flags, target, replaces,
1951                      speed, granularity, buf_size, backing_mode, zero_target,
1952                      on_source_error, on_target_error, unmap, NULL, NULL,
1953                      &mirror_job_driver, is_none_mode, base, false,
1954                      filter_node_name, true, copy_mode, errp);
1955 }
1956 
1957 BlockJob *commit_active_start(const char *job_id, BlockDriverState *bs,
1958                               BlockDriverState *base, int creation_flags,
1959                               int64_t speed, BlockdevOnError on_error,
1960                               const char *filter_node_name,
1961                               BlockCompletionFunc *cb, void *opaque,
1962                               bool auto_complete, Error **errp)
1963 {
1964     bool base_read_only;
1965     BlockJob *job;
1966 
1967     GLOBAL_STATE_CODE();
1968 
1969     base_read_only = bdrv_is_read_only(base);
1970 
1971     if (base_read_only) {
1972         if (bdrv_reopen_set_read_only(base, false, errp) < 0) {
1973             return NULL;
1974         }
1975     }
1976 
1977     job = mirror_start_job(
1978                      job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1979                      MIRROR_LEAVE_BACKING_CHAIN, false,
1980                      on_error, on_error, true, cb, opaque,
1981                      &commit_active_job_driver, false, base, auto_complete,
1982                      filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND,
1983                      errp);
1984     if (!job) {
1985         goto error_restore_flags;
1986     }
1987 
1988     return job;
1989 
1990 error_restore_flags:
1991     /* ignore error and errp for bdrv_reopen, because we want to propagate
1992      * the original error */
1993     if (base_read_only) {
1994         bdrv_reopen_set_read_only(base, true, NULL);
1995     }
1996     return NULL;
1997 }
1998