xref: /openbmc/qemu/block/mirror.c (revision 8acf8650)
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 void coroutine_fn mirror_iteration(MirrorBlockJob *s)
475 {
476     BlockDriverState *source = s->mirror_top_bs->backing->bs;
477     MirrorOp *pseudo_op;
478     int64_t offset;
479     /* At least the first dirty chunk is mirrored in one iteration. */
480     int nb_chunks = 1;
481     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
482     int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES);
483 
484     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
485     offset = bdrv_dirty_iter_next(s->dbi);
486     if (offset < 0) {
487         bdrv_set_dirty_iter(s->dbi, 0);
488         offset = bdrv_dirty_iter_next(s->dbi);
489         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
490         assert(offset >= 0);
491     }
492     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
493 
494     /*
495      * Wait for concurrent requests to @offset.  The next loop will limit the
496      * copied area based on in_flight_bitmap so we only copy an area that does
497      * not overlap with concurrent in-flight requests.  Still, we would like to
498      * copy something, so wait until there are at least no more requests to the
499      * very beginning of the area.
500      */
501     mirror_wait_on_conflicts(NULL, s, offset, 1);
502 
503     job_pause_point(&s->common.job);
504 
505     /* Find the number of consective dirty chunks following the first dirty
506      * one, and wait for in flight requests in them. */
507     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
508     while (nb_chunks * s->granularity < s->buf_size) {
509         int64_t next_dirty;
510         int64_t next_offset = offset + nb_chunks * s->granularity;
511         int64_t next_chunk = next_offset / s->granularity;
512         if (next_offset >= s->bdev_length ||
513             !bdrv_dirty_bitmap_get_locked(s->dirty_bitmap, next_offset)) {
514             break;
515         }
516         if (test_bit(next_chunk, s->in_flight_bitmap)) {
517             break;
518         }
519 
520         next_dirty = bdrv_dirty_iter_next(s->dbi);
521         if (next_dirty > next_offset || next_dirty < 0) {
522             /* The bitmap iterator's cache is stale, refresh it */
523             bdrv_set_dirty_iter(s->dbi, next_offset);
524             next_dirty = bdrv_dirty_iter_next(s->dbi);
525         }
526         assert(next_dirty == next_offset);
527         nb_chunks++;
528     }
529 
530     /* Clear dirty bits before querying the block status, because
531      * calling bdrv_block_status_above could yield - if some blocks are
532      * marked dirty in this window, we need to know.
533      */
534     bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset,
535                                    nb_chunks * s->granularity);
536     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
537 
538     /* Before claiming an area in the in-flight bitmap, we have to
539      * create a MirrorOp for it so that conflicting requests can wait
540      * for it.  mirror_perform() will create the real MirrorOps later,
541      * for now we just create a pseudo operation that will wake up all
542      * conflicting requests once all real operations have been
543      * launched. */
544     pseudo_op = g_new(MirrorOp, 1);
545     *pseudo_op = (MirrorOp){
546         .offset         = offset,
547         .bytes          = nb_chunks * s->granularity,
548         .is_pseudo_op   = true,
549     };
550     qemu_co_queue_init(&pseudo_op->waiting_requests);
551     QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next);
552 
553     bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks);
554     while (nb_chunks > 0 && offset < s->bdev_length) {
555         int ret;
556         int64_t io_bytes;
557         int64_t io_bytes_acct;
558         MirrorMethod mirror_method = MIRROR_METHOD_COPY;
559 
560         assert(!(offset % s->granularity));
561         WITH_GRAPH_RDLOCK_GUARD() {
562             ret = bdrv_block_status_above(source, NULL, offset,
563                                         nb_chunks * s->granularity,
564                                         &io_bytes, NULL, NULL);
565         }
566         if (ret < 0) {
567             io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes);
568         } else if (ret & BDRV_BLOCK_DATA) {
569             io_bytes = MIN(io_bytes, max_io_bytes);
570         }
571 
572         io_bytes -= io_bytes % s->granularity;
573         if (io_bytes < s->granularity) {
574             io_bytes = s->granularity;
575         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
576             int64_t target_offset;
577             int64_t target_bytes;
578             WITH_GRAPH_RDLOCK_GUARD() {
579                 bdrv_round_to_clusters(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     if (s->prepared) {
666         return 0;
667     }
668     s->prepared = true;
669 
670     mirror_top_bs = s->mirror_top_bs;
671     bs_opaque = mirror_top_bs->opaque;
672     src = mirror_top_bs->backing->bs;
673     target_bs = blk_bs(s->target);
674 
675     if (bdrv_chain_contains(src, target_bs)) {
676         bdrv_unfreeze_backing_chain(mirror_top_bs, target_bs);
677     }
678 
679     bdrv_release_dirty_bitmap(s->dirty_bitmap);
680 
681     /* Make sure that the source BDS doesn't go away during bdrv_replace_node,
682      * before we can call bdrv_drained_end */
683     bdrv_ref(src);
684     bdrv_ref(mirror_top_bs);
685     bdrv_ref(target_bs);
686 
687     /*
688      * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
689      * inserting target_bs at s->to_replace, where we might not be able to get
690      * these permissions.
691      */
692     blk_unref(s->target);
693     s->target = NULL;
694 
695     /* We don't access the source any more. Dropping any WRITE/RESIZE is
696      * required before it could become a backing file of target_bs. Not having
697      * these permissions any more means that we can't allow any new requests on
698      * mirror_top_bs from now on, so keep it drained. */
699     bdrv_drained_begin(mirror_top_bs);
700     bs_opaque->stop = true;
701     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
702                              &error_abort);
703     if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
704         BlockDriverState *backing = s->is_none_mode ? src : s->base;
705         BlockDriverState *unfiltered_target = bdrv_skip_filters(target_bs);
706 
707         if (bdrv_cow_bs(unfiltered_target) != backing) {
708             bdrv_set_backing_hd(unfiltered_target, backing, &local_err);
709             if (local_err) {
710                 error_report_err(local_err);
711                 local_err = NULL;
712                 ret = -EPERM;
713             }
714         }
715     } else if (!abort && s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
716         assert(!bdrv_backing_chain_next(target_bs));
717         ret = bdrv_open_backing_file(bdrv_skip_filters(target_bs), NULL,
718                                      "backing", &local_err);
719         if (ret < 0) {
720             error_report_err(local_err);
721             local_err = NULL;
722         }
723     }
724 
725     if (s->to_replace) {
726         replace_aio_context = bdrv_get_aio_context(s->to_replace);
727         aio_context_acquire(replace_aio_context);
728     }
729 
730     if (s->should_complete && !abort) {
731         BlockDriverState *to_replace = s->to_replace ?: src;
732         bool ro = bdrv_is_read_only(to_replace);
733 
734         if (ro != bdrv_is_read_only(target_bs)) {
735             bdrv_reopen_set_read_only(target_bs, ro, NULL);
736         }
737 
738         /* The mirror job has no requests in flight any more, but we need to
739          * drain potential other users of the BDS before changing the graph. */
740         assert(s->in_drain);
741         bdrv_drained_begin(target_bs);
742         /*
743          * Cannot use check_to_replace_node() here, because that would
744          * check for an op blocker on @to_replace, and we have our own
745          * there.
746          *
747          * TODO Pull out the writer lock from bdrv_replace_node() to here
748          */
749         bdrv_graph_rdlock_main_loop();
750         if (bdrv_recurse_can_replace(src, to_replace)) {
751             bdrv_replace_node(to_replace, target_bs, &local_err);
752         } else {
753             error_setg(&local_err, "Can no longer replace '%s' by '%s', "
754                        "because it can no longer be guaranteed that doing so "
755                        "would not lead to an abrupt change of visible data",
756                        to_replace->node_name, target_bs->node_name);
757         }
758         bdrv_graph_rdunlock_main_loop();
759         bdrv_drained_end(target_bs);
760         if (local_err) {
761             error_report_err(local_err);
762             ret = -EPERM;
763         }
764     }
765     if (s->to_replace) {
766         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
767         error_free(s->replace_blocker);
768         bdrv_unref(s->to_replace);
769     }
770     if (replace_aio_context) {
771         aio_context_release(replace_aio_context);
772     }
773     g_free(s->replaces);
774     bdrv_unref(target_bs);
775 
776     /*
777      * Remove the mirror filter driver from the graph. Before this, get rid of
778      * the blockers on the intermediate nodes so that the resulting state is
779      * valid.
780      */
781     block_job_remove_all_bdrv(bjob);
782     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
783 
784     bs_opaque->job = NULL;
785 
786     bdrv_drained_end(src);
787     bdrv_drained_end(mirror_top_bs);
788     s->in_drain = false;
789     bdrv_unref(mirror_top_bs);
790     bdrv_unref(src);
791 
792     return ret;
793 }
794 
795 static int mirror_prepare(Job *job)
796 {
797     return mirror_exit_common(job);
798 }
799 
800 static void mirror_abort(Job *job)
801 {
802     int ret = mirror_exit_common(job);
803     assert(ret == 0);
804 }
805 
806 static void coroutine_fn mirror_throttle(MirrorBlockJob *s)
807 {
808     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
809 
810     if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) {
811         s->last_pause_ns = now;
812         job_sleep_ns(&s->common.job, 0);
813     } else {
814         job_pause_point(&s->common.job);
815     }
816 }
817 
818 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
819 {
820     int64_t offset;
821     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
822     BlockDriverState *target_bs = blk_bs(s->target);
823     int ret;
824     int64_t count;
825 
826     if (s->zero_target) {
827         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
828             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
829             return 0;
830         }
831 
832         s->initial_zeroing_ongoing = true;
833         for (offset = 0; offset < s->bdev_length; ) {
834             int bytes = MIN(s->bdev_length - offset,
835                             QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
836 
837             mirror_throttle(s);
838 
839             if (job_is_cancelled(&s->common.job)) {
840                 s->initial_zeroing_ongoing = false;
841                 return 0;
842             }
843 
844             if (s->in_flight >= MAX_IN_FLIGHT) {
845                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
846                                    s->in_flight);
847                 mirror_wait_for_free_in_flight_slot(s);
848                 continue;
849             }
850 
851             mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO);
852             offset += bytes;
853         }
854 
855         mirror_wait_for_all_io(s);
856         s->initial_zeroing_ongoing = false;
857     }
858 
859     /* First part, loop on the sectors and initialize the dirty bitmap.  */
860     for (offset = 0; offset < s->bdev_length; ) {
861         /* Just to make sure we are not exceeding int limit. */
862         int bytes = MIN(s->bdev_length - offset,
863                         QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
864 
865         mirror_throttle(s);
866 
867         if (job_is_cancelled(&s->common.job)) {
868             return 0;
869         }
870 
871         WITH_GRAPH_RDLOCK_GUARD() {
872             ret = bdrv_is_allocated_above(bs, s->base_overlay, true, offset,
873                                           bytes, &count);
874         }
875         if (ret < 0) {
876             return ret;
877         }
878 
879         assert(count);
880         if (ret > 0) {
881             bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
882         }
883         offset += count;
884     }
885     return 0;
886 }
887 
888 /* Called when going out of the streaming phase to flush the bulk of the
889  * data to the medium, or just before completing.
890  */
891 static int coroutine_fn mirror_flush(MirrorBlockJob *s)
892 {
893     int ret = blk_co_flush(s->target);
894     if (ret < 0) {
895         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
896             s->ret = ret;
897         }
898     }
899     return ret;
900 }
901 
902 static int coroutine_fn mirror_run(Job *job, Error **errp)
903 {
904     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
905     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
906     MirrorBDSOpaque *mirror_top_opaque = s->mirror_top_bs->opaque;
907     BlockDriverState *target_bs = blk_bs(s->target);
908     bool need_drain = true;
909     BlockDeviceIoStatus iostatus;
910     int64_t length;
911     int64_t target_length;
912     BlockDriverInfo bdi;
913     char backing_filename[2]; /* we only need 2 characters because we are only
914                                  checking for a NULL string */
915     int ret = 0;
916 
917     if (job_is_cancelled(&s->common.job)) {
918         goto immediate_exit;
919     }
920 
921     bdrv_graph_co_rdlock();
922     s->bdev_length = bdrv_co_getlength(bs);
923     bdrv_graph_co_rdunlock();
924 
925     if (s->bdev_length < 0) {
926         ret = s->bdev_length;
927         goto immediate_exit;
928     }
929 
930     target_length = blk_co_getlength(s->target);
931     if (target_length < 0) {
932         ret = target_length;
933         goto immediate_exit;
934     }
935 
936     /* Active commit must resize the base image if its size differs from the
937      * active layer. */
938     if (s->base == blk_bs(s->target)) {
939         if (s->bdev_length > target_length) {
940             ret = blk_co_truncate(s->target, s->bdev_length, false,
941                                   PREALLOC_MODE_OFF, 0, NULL);
942             if (ret < 0) {
943                 goto immediate_exit;
944             }
945         }
946     } else if (s->bdev_length != target_length) {
947         error_setg(errp, "Source and target image have different sizes");
948         ret = -EINVAL;
949         goto immediate_exit;
950     }
951 
952     if (s->bdev_length == 0) {
953         /* Transition to the READY state and wait for complete. */
954         job_transition_to_ready(&s->common.job);
955         s->actively_synced = true;
956         while (!job_cancel_requested(&s->common.job) && !s->should_complete) {
957             job_yield(&s->common.job);
958         }
959         goto immediate_exit;
960     }
961 
962     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
963     s->in_flight_bitmap = bitmap_new(length);
964 
965     /* If we have no backing file yet in the destination, we cannot let
966      * the destination do COW.  Instead, we copy sectors around the
967      * dirty data if needed.  We need a bitmap to do that.
968      */
969     bdrv_get_backing_filename(target_bs, backing_filename,
970                               sizeof(backing_filename));
971     bdrv_graph_co_rdlock();
972     if (!bdrv_co_get_info(target_bs, &bdi) && bdi.cluster_size) {
973         s->target_cluster_size = bdi.cluster_size;
974     } else {
975         s->target_cluster_size = BDRV_SECTOR_SIZE;
976     }
977     bdrv_graph_co_rdunlock();
978     if (backing_filename[0] && !bdrv_backing_chain_next(target_bs) &&
979         s->granularity < s->target_cluster_size) {
980         s->buf_size = MAX(s->buf_size, s->target_cluster_size);
981         s->cow_bitmap = bitmap_new(length);
982     }
983     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
984 
985     s->buf = qemu_try_blockalign(bs, s->buf_size);
986     if (s->buf == NULL) {
987         ret = -ENOMEM;
988         goto immediate_exit;
989     }
990 
991     mirror_free_init(s);
992 
993     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
994     if (!s->is_none_mode) {
995         ret = mirror_dirty_init(s);
996         if (ret < 0 || job_is_cancelled(&s->common.job)) {
997             goto immediate_exit;
998         }
999     }
1000 
1001     /*
1002      * Only now the job is fully initialised and mirror_top_bs should start
1003      * accessing it.
1004      */
1005     mirror_top_opaque->job = s;
1006 
1007     assert(!s->dbi);
1008     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
1009     for (;;) {
1010         int64_t cnt, delta;
1011         bool should_complete;
1012 
1013         if (s->ret < 0) {
1014             ret = s->ret;
1015             goto immediate_exit;
1016         }
1017 
1018         job_pause_point(&s->common.job);
1019 
1020         if (job_is_cancelled(&s->common.job)) {
1021             ret = 0;
1022             goto immediate_exit;
1023         }
1024 
1025         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1026         /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
1027          * the number of bytes currently being processed; together those are
1028          * the current remaining operation length */
1029         job_progress_set_remaining(&s->common.job,
1030                                    s->bytes_in_flight + cnt +
1031                                    s->active_write_bytes_in_flight);
1032 
1033         /* Note that even when no rate limit is applied we need to yield
1034          * periodically with no pending I/O so that bdrv_drain_all() returns.
1035          * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
1036          * an error, or when the source is clean, whichever comes first. */
1037         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
1038         WITH_JOB_LOCK_GUARD() {
1039             iostatus = s->common.iostatus;
1040         }
1041         if (delta < BLOCK_JOB_SLICE_TIME &&
1042             iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
1043             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
1044                 (cnt == 0 && s->in_flight > 0)) {
1045                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
1046                 mirror_wait_for_free_in_flight_slot(s);
1047                 continue;
1048             } else if (cnt != 0) {
1049                 mirror_iteration(s);
1050             }
1051         }
1052 
1053         should_complete = false;
1054         if (s->in_flight == 0 && cnt == 0) {
1055             trace_mirror_before_flush(s);
1056             if (!job_is_ready(&s->common.job)) {
1057                 if (mirror_flush(s) < 0) {
1058                     /* Go check s->ret.  */
1059                     continue;
1060                 }
1061                 /* We're out of the streaming phase.  From now on, if the job
1062                  * is cancelled we will actually complete all pending I/O and
1063                  * report completion.  This way, block-job-cancel will leave
1064                  * the target in a consistent state.
1065                  */
1066                 job_transition_to_ready(&s->common.job);
1067                 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) {
1068                     s->actively_synced = true;
1069                 }
1070             }
1071 
1072             should_complete = s->should_complete ||
1073                 job_cancel_requested(&s->common.job);
1074             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1075         }
1076 
1077         if (cnt == 0 && should_complete) {
1078             /* The dirty bitmap is not updated while operations are pending.
1079              * If we're about to exit, wait for pending operations before
1080              * calling bdrv_get_dirty_count(bs), or we may exit while the
1081              * source has dirty data to copy!
1082              *
1083              * Note that I/O can be submitted by the guest while
1084              * mirror_populate runs, so pause it now.  Before deciding
1085              * whether to switch to target check one last time if I/O has
1086              * come in the meanwhile, and if not flush the data to disk.
1087              */
1088             trace_mirror_before_drain(s, cnt);
1089 
1090             s->in_drain = true;
1091             bdrv_drained_begin(bs);
1092 
1093             /* Must be zero because we are drained */
1094             assert(s->in_active_write_counter == 0);
1095 
1096             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1097             if (cnt > 0 || mirror_flush(s) < 0) {
1098                 bdrv_drained_end(bs);
1099                 s->in_drain = false;
1100                 continue;
1101             }
1102 
1103             /* The two disks are in sync.  Exit and report successful
1104              * completion.
1105              */
1106             assert(QLIST_EMPTY(&bs->tracked_requests));
1107             need_drain = false;
1108             break;
1109         }
1110 
1111         if (job_is_ready(&s->common.job) && !should_complete) {
1112             if (s->in_flight == 0 && cnt == 0) {
1113                 trace_mirror_before_sleep(s, cnt, job_is_ready(&s->common.job),
1114                                           BLOCK_JOB_SLICE_TIME);
1115                 job_sleep_ns(&s->common.job, BLOCK_JOB_SLICE_TIME);
1116             }
1117         } else {
1118             block_job_ratelimit_sleep(&s->common);
1119         }
1120         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1121     }
1122 
1123 immediate_exit:
1124     if (s->in_flight > 0) {
1125         /* We get here only if something went wrong.  Either the job failed,
1126          * or it was cancelled prematurely so that we do not guarantee that
1127          * the target is a copy of the source.
1128          */
1129         assert(ret < 0 || job_is_cancelled(&s->common.job));
1130         assert(need_drain);
1131         mirror_wait_for_all_io(s);
1132     }
1133 
1134     assert(s->in_flight == 0);
1135     qemu_vfree(s->buf);
1136     g_free(s->cow_bitmap);
1137     g_free(s->in_flight_bitmap);
1138     bdrv_dirty_iter_free(s->dbi);
1139 
1140     if (need_drain) {
1141         s->in_drain = true;
1142         bdrv_drained_begin(bs);
1143     }
1144 
1145     return ret;
1146 }
1147 
1148 static void mirror_complete(Job *job, Error **errp)
1149 {
1150     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1151 
1152     if (!job_is_ready(job)) {
1153         error_setg(errp, "The active block job '%s' cannot be completed",
1154                    job->id);
1155         return;
1156     }
1157 
1158     /* block all operations on to_replace bs */
1159     if (s->replaces) {
1160         AioContext *replace_aio_context;
1161 
1162         s->to_replace = bdrv_find_node(s->replaces);
1163         if (!s->to_replace) {
1164             error_setg(errp, "Node name '%s' not found", s->replaces);
1165             return;
1166         }
1167 
1168         replace_aio_context = bdrv_get_aio_context(s->to_replace);
1169         aio_context_acquire(replace_aio_context);
1170 
1171         /* TODO Translate this into child freeze system. */
1172         error_setg(&s->replace_blocker,
1173                    "block device is in use by block-job-complete");
1174         bdrv_op_block_all(s->to_replace, s->replace_blocker);
1175         bdrv_ref(s->to_replace);
1176 
1177         aio_context_release(replace_aio_context);
1178     }
1179 
1180     s->should_complete = true;
1181 
1182     /* If the job is paused, it will be re-entered when it is resumed */
1183     WITH_JOB_LOCK_GUARD() {
1184         if (!job->paused) {
1185             job_enter_cond_locked(job, NULL);
1186         }
1187     }
1188 }
1189 
1190 static void coroutine_fn mirror_pause(Job *job)
1191 {
1192     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1193 
1194     mirror_wait_for_all_io(s);
1195 }
1196 
1197 static bool mirror_drained_poll(BlockJob *job)
1198 {
1199     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1200 
1201     /* If the job isn't paused nor cancelled, we can't be sure that it won't
1202      * issue more requests. We make an exception if we've reached this point
1203      * from one of our own drain sections, to avoid a deadlock waiting for
1204      * ourselves.
1205      */
1206     WITH_JOB_LOCK_GUARD() {
1207         if (!s->common.job.paused && !job_is_cancelled_locked(&job->job)
1208             && !s->in_drain) {
1209             return true;
1210         }
1211     }
1212 
1213     return !!s->in_flight;
1214 }
1215 
1216 static bool mirror_cancel(Job *job, bool force)
1217 {
1218     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1219     BlockDriverState *target = blk_bs(s->target);
1220 
1221     /*
1222      * Before the job is READY, we treat any cancellation like a
1223      * force-cancellation.
1224      */
1225     force = force || !job_is_ready(job);
1226 
1227     if (force) {
1228         bdrv_cancel_in_flight(target);
1229     }
1230     return force;
1231 }
1232 
1233 static bool commit_active_cancel(Job *job, bool force)
1234 {
1235     /* Same as above in mirror_cancel() */
1236     return force || !job_is_ready(job);
1237 }
1238 
1239 static const BlockJobDriver mirror_job_driver = {
1240     .job_driver = {
1241         .instance_size          = sizeof(MirrorBlockJob),
1242         .job_type               = JOB_TYPE_MIRROR,
1243         .free                   = block_job_free,
1244         .user_resume            = block_job_user_resume,
1245         .run                    = mirror_run,
1246         .prepare                = mirror_prepare,
1247         .abort                  = mirror_abort,
1248         .pause                  = mirror_pause,
1249         .complete               = mirror_complete,
1250         .cancel                 = mirror_cancel,
1251     },
1252     .drained_poll           = mirror_drained_poll,
1253 };
1254 
1255 static const BlockJobDriver commit_active_job_driver = {
1256     .job_driver = {
1257         .instance_size          = sizeof(MirrorBlockJob),
1258         .job_type               = JOB_TYPE_COMMIT,
1259         .free                   = block_job_free,
1260         .user_resume            = block_job_user_resume,
1261         .run                    = mirror_run,
1262         .prepare                = mirror_prepare,
1263         .abort                  = mirror_abort,
1264         .pause                  = mirror_pause,
1265         .complete               = mirror_complete,
1266         .cancel                 = commit_active_cancel,
1267     },
1268     .drained_poll           = mirror_drained_poll,
1269 };
1270 
1271 static void coroutine_fn
1272 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method,
1273                      uint64_t offset, uint64_t bytes,
1274                      QEMUIOVector *qiov, int flags)
1275 {
1276     int ret;
1277     size_t qiov_offset = 0;
1278     int64_t bitmap_offset, bitmap_end;
1279 
1280     if (!QEMU_IS_ALIGNED(offset, job->granularity) &&
1281         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset))
1282     {
1283             /*
1284              * Dirty unaligned padding: ignore it.
1285              *
1286              * Reasoning:
1287              * 1. If we copy it, we can't reset corresponding bit in
1288              *    dirty_bitmap as there may be some "dirty" bytes still not
1289              *    copied.
1290              * 2. It's already dirty, so skipping it we don't diverge mirror
1291              *    progress.
1292              *
1293              * Note, that because of this, guest write may have no contribution
1294              * into mirror converge, but that's not bad, as we have background
1295              * process of mirroring. If under some bad circumstances (high guest
1296              * IO load) background process starve, we will not converge anyway,
1297              * even if each write will contribute, as guest is not guaranteed to
1298              * rewrite the whole disk.
1299              */
1300             qiov_offset = QEMU_ALIGN_UP(offset, job->granularity) - offset;
1301             if (bytes <= qiov_offset) {
1302                 /* nothing to do after shrink */
1303                 return;
1304             }
1305             offset += qiov_offset;
1306             bytes -= qiov_offset;
1307     }
1308 
1309     if (!QEMU_IS_ALIGNED(offset + bytes, job->granularity) &&
1310         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset + bytes - 1))
1311     {
1312         uint64_t tail = (offset + bytes) % job->granularity;
1313 
1314         if (bytes <= tail) {
1315             /* nothing to do after shrink */
1316             return;
1317         }
1318         bytes -= tail;
1319     }
1320 
1321     /*
1322      * Tails are either clean or shrunk, so for bitmap resetting
1323      * we safely align the range down.
1324      */
1325     bitmap_offset = QEMU_ALIGN_UP(offset, job->granularity);
1326     bitmap_end = QEMU_ALIGN_DOWN(offset + bytes, job->granularity);
1327     if (bitmap_offset < bitmap_end) {
1328         bdrv_reset_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1329                                 bitmap_end - bitmap_offset);
1330     }
1331 
1332     job_progress_increase_remaining(&job->common.job, bytes);
1333     job->active_write_bytes_in_flight += bytes;
1334 
1335     switch (method) {
1336     case MIRROR_METHOD_COPY:
1337         ret = blk_co_pwritev_part(job->target, offset, bytes,
1338                                   qiov, qiov_offset, flags);
1339         break;
1340 
1341     case MIRROR_METHOD_ZERO:
1342         assert(!qiov);
1343         ret = blk_co_pwrite_zeroes(job->target, offset, bytes, flags);
1344         break;
1345 
1346     case MIRROR_METHOD_DISCARD:
1347         assert(!qiov);
1348         ret = blk_co_pdiscard(job->target, offset, bytes);
1349         break;
1350 
1351     default:
1352         abort();
1353     }
1354 
1355     job->active_write_bytes_in_flight -= bytes;
1356     if (ret >= 0) {
1357         job_progress_update(&job->common.job, bytes);
1358     } else {
1359         BlockErrorAction action;
1360 
1361         /*
1362          * We failed, so we should mark dirty the whole area, aligned up.
1363          * Note that we don't care about shrunk tails if any: they were dirty
1364          * at function start, and they must be still dirty, as we've locked
1365          * the region for in-flight op.
1366          */
1367         bitmap_offset = QEMU_ALIGN_DOWN(offset, job->granularity);
1368         bitmap_end = QEMU_ALIGN_UP(offset + bytes, job->granularity);
1369         bdrv_set_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1370                               bitmap_end - bitmap_offset);
1371         job->actively_synced = false;
1372 
1373         action = mirror_error_action(job, false, -ret);
1374         if (action == BLOCK_ERROR_ACTION_REPORT) {
1375             if (!job->ret) {
1376                 job->ret = ret;
1377             }
1378         }
1379     }
1380 }
1381 
1382 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s,
1383                                                    uint64_t offset,
1384                                                    uint64_t bytes)
1385 {
1386     MirrorOp *op;
1387     uint64_t start_chunk = offset / s->granularity;
1388     uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
1389 
1390     op = g_new(MirrorOp, 1);
1391     *op = (MirrorOp){
1392         .s                  = s,
1393         .offset             = offset,
1394         .bytes              = bytes,
1395         .is_active_write    = true,
1396         .is_in_flight       = true,
1397         .co                 = qemu_coroutine_self(),
1398     };
1399     qemu_co_queue_init(&op->waiting_requests);
1400     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
1401 
1402     s->in_active_write_counter++;
1403 
1404     /*
1405      * Wait for concurrent requests affecting the area.  If there are already
1406      * running requests that are copying off now-to-be stale data in the area,
1407      * we must wait for them to finish before we begin writing fresh data to the
1408      * target so that the write operations appear in the correct order.
1409      * Note that background requests (see mirror_iteration()) in contrast only
1410      * wait for conflicting requests at the start of the dirty area, and then
1411      * (based on the in_flight_bitmap) truncate the area to copy so it will not
1412      * conflict with any requests beyond that.  For active writes, however, we
1413      * cannot truncate that area.  The request from our parent must be blocked
1414      * until the area is copied in full.  Therefore, we must wait for the whole
1415      * area to become free of concurrent requests.
1416      */
1417     mirror_wait_on_conflicts(op, s, offset, bytes);
1418 
1419     bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1420 
1421     return op;
1422 }
1423 
1424 static void coroutine_fn GRAPH_RDLOCK active_write_settle(MirrorOp *op)
1425 {
1426     uint64_t start_chunk = op->offset / op->s->granularity;
1427     uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes,
1428                                       op->s->granularity);
1429 
1430     if (!--op->s->in_active_write_counter && op->s->actively_synced) {
1431         BdrvChild *source = op->s->mirror_top_bs->backing;
1432 
1433         if (QLIST_FIRST(&source->bs->parents) == source &&
1434             QLIST_NEXT(source, next_parent) == NULL)
1435         {
1436             /* Assert that we are back in sync once all active write
1437              * operations are settled.
1438              * Note that we can only assert this if the mirror node
1439              * is the source node's only parent. */
1440             assert(!bdrv_get_dirty_count(op->s->dirty_bitmap));
1441         }
1442     }
1443     bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1444     QTAILQ_REMOVE(&op->s->ops_in_flight, op, next);
1445     qemu_co_queue_restart_all(&op->waiting_requests);
1446     g_free(op);
1447 }
1448 
1449 static int coroutine_fn GRAPH_RDLOCK
1450 bdrv_mirror_top_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
1451                        QEMUIOVector *qiov, BdrvRequestFlags flags)
1452 {
1453     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1454 }
1455 
1456 static int coroutine_fn GRAPH_RDLOCK
1457 bdrv_mirror_top_do_write(BlockDriverState *bs, MirrorMethod method,
1458                          uint64_t offset, uint64_t bytes, QEMUIOVector *qiov,
1459                          int flags)
1460 {
1461     MirrorOp *op = NULL;
1462     MirrorBDSOpaque *s = bs->opaque;
1463     int ret = 0;
1464     bool copy_to_target = false;
1465 
1466     if (s->job) {
1467         copy_to_target = s->job->ret >= 0 &&
1468                          !job_is_cancelled(&s->job->common.job) &&
1469                          s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1470     }
1471 
1472     if (copy_to_target) {
1473         op = active_write_prepare(s->job, offset, bytes);
1474     }
1475 
1476     switch (method) {
1477     case MIRROR_METHOD_COPY:
1478         ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1479         break;
1480 
1481     case MIRROR_METHOD_ZERO:
1482         ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1483         break;
1484 
1485     case MIRROR_METHOD_DISCARD:
1486         ret = bdrv_co_pdiscard(bs->backing, offset, bytes);
1487         break;
1488 
1489     default:
1490         abort();
1491     }
1492 
1493     if (ret < 0) {
1494         goto out;
1495     }
1496 
1497     if (copy_to_target) {
1498         do_sync_target_write(s->job, method, offset, bytes, qiov, flags);
1499     }
1500 
1501 out:
1502     if (copy_to_target) {
1503         active_write_settle(op);
1504     }
1505     return ret;
1506 }
1507 
1508 static int coroutine_fn GRAPH_RDLOCK
1509 bdrv_mirror_top_pwritev(BlockDriverState *bs, int64_t offset, int64_t bytes,
1510                         QEMUIOVector *qiov, BdrvRequestFlags flags)
1511 {
1512     MirrorBDSOpaque *s = bs->opaque;
1513     QEMUIOVector bounce_qiov;
1514     void *bounce_buf;
1515     int ret = 0;
1516     bool copy_to_target = false;
1517 
1518     if (s->job) {
1519         copy_to_target = s->job->ret >= 0 &&
1520                          !job_is_cancelled(&s->job->common.job) &&
1521                          s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1522     }
1523 
1524     if (copy_to_target) {
1525         /* The guest might concurrently modify the data to write; but
1526          * the data on source and destination must match, so we have
1527          * to use a bounce buffer if we are going to write to the
1528          * target now. */
1529         bounce_buf = qemu_blockalign(bs, bytes);
1530         iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes);
1531 
1532         qemu_iovec_init(&bounce_qiov, 1);
1533         qemu_iovec_add(&bounce_qiov, bounce_buf, bytes);
1534         qiov = &bounce_qiov;
1535 
1536         flags &= ~BDRV_REQ_REGISTERED_BUF;
1537     }
1538 
1539     ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov,
1540                                    flags);
1541 
1542     if (copy_to_target) {
1543         qemu_iovec_destroy(&bounce_qiov);
1544         qemu_vfree(bounce_buf);
1545     }
1546 
1547     return ret;
1548 }
1549 
1550 static int coroutine_fn GRAPH_RDLOCK bdrv_mirror_top_flush(BlockDriverState *bs)
1551 {
1552     if (bs->backing == NULL) {
1553         /* we can be here after failed bdrv_append in mirror_start_job */
1554         return 0;
1555     }
1556     return bdrv_co_flush(bs->backing->bs);
1557 }
1558 
1559 static int coroutine_fn GRAPH_RDLOCK
1560 bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs, int64_t offset,
1561                               int64_t bytes, BdrvRequestFlags flags)
1562 {
1563     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL,
1564                                     flags);
1565 }
1566 
1567 static int coroutine_fn GRAPH_RDLOCK
1568 bdrv_mirror_top_pdiscard(BlockDriverState *bs, int64_t offset, int64_t bytes)
1569 {
1570     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes,
1571                                     NULL, 0);
1572 }
1573 
1574 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs)
1575 {
1576     if (bs->backing == NULL) {
1577         /* we can be here after failed bdrv_attach_child in
1578          * bdrv_set_backing_hd */
1579         return;
1580     }
1581     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1582             bs->backing->bs->filename);
1583 }
1584 
1585 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1586                                        BdrvChildRole role,
1587                                        BlockReopenQueue *reopen_queue,
1588                                        uint64_t perm, uint64_t shared,
1589                                        uint64_t *nperm, uint64_t *nshared)
1590 {
1591     MirrorBDSOpaque *s = bs->opaque;
1592 
1593     if (s->stop) {
1594         /*
1595          * If the job is to be stopped, we do not need to forward
1596          * anything to the real image.
1597          */
1598         *nperm = 0;
1599         *nshared = BLK_PERM_ALL;
1600         return;
1601     }
1602 
1603     bdrv_default_perms(bs, c, role, reopen_queue,
1604                        perm, shared, nperm, nshared);
1605 
1606     if (s->is_commit) {
1607         /*
1608          * For commit jobs, we cannot take CONSISTENT_READ, because
1609          * that permission is unshared for everything above the base
1610          * node (except for filters on the base node).
1611          * We also have to force-share the WRITE permission, or
1612          * otherwise we would block ourselves at the base node (if
1613          * writes are blocked for a node, they are also blocked for
1614          * its backing file).
1615          * (We could also share RESIZE, because it may be needed for
1616          * the target if its size is less than the top node's; but
1617          * bdrv_default_perms_for_cow() automatically shares RESIZE
1618          * for backing nodes if WRITE is shared, so there is no need
1619          * to do it here.)
1620          */
1621         *nperm &= ~BLK_PERM_CONSISTENT_READ;
1622         *nshared |= BLK_PERM_WRITE;
1623     }
1624 }
1625 
1626 /* Dummy node that provides consistent read to its users without requiring it
1627  * from its backing file and that allows writes on the backing file chain. */
1628 static BlockDriver bdrv_mirror_top = {
1629     .format_name                = "mirror_top",
1630     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1631     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1632     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1633     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1634     .bdrv_co_flush              = bdrv_mirror_top_flush,
1635     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1636     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1637 
1638     .is_filter                  = true,
1639     .filtered_child_is_backing  = true,
1640 };
1641 
1642 static BlockJob *mirror_start_job(
1643                              const char *job_id, BlockDriverState *bs,
1644                              int creation_flags, BlockDriverState *target,
1645                              const char *replaces, int64_t speed,
1646                              uint32_t granularity, int64_t buf_size,
1647                              BlockMirrorBackingMode backing_mode,
1648                              bool zero_target,
1649                              BlockdevOnError on_source_error,
1650                              BlockdevOnError on_target_error,
1651                              bool unmap,
1652                              BlockCompletionFunc *cb,
1653                              void *opaque,
1654                              const BlockJobDriver *driver,
1655                              bool is_none_mode, BlockDriverState *base,
1656                              bool auto_complete, const char *filter_node_name,
1657                              bool is_mirror, MirrorCopyMode copy_mode,
1658                              Error **errp)
1659 {
1660     MirrorBlockJob *s;
1661     MirrorBDSOpaque *bs_opaque;
1662     BlockDriverState *mirror_top_bs;
1663     bool target_is_backing;
1664     uint64_t target_perms, target_shared_perms;
1665     int ret;
1666 
1667     if (granularity == 0) {
1668         granularity = bdrv_get_default_bitmap_granularity(target);
1669     }
1670 
1671     assert(is_power_of_2(granularity));
1672 
1673     if (buf_size < 0) {
1674         error_setg(errp, "Invalid parameter 'buf-size'");
1675         return NULL;
1676     }
1677 
1678     if (buf_size == 0) {
1679         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1680     }
1681 
1682     if (bdrv_skip_filters(bs) == bdrv_skip_filters(target)) {
1683         error_setg(errp, "Can't mirror node into itself");
1684         return NULL;
1685     }
1686 
1687     target_is_backing = bdrv_chain_contains(bs, target);
1688 
1689     /* In the case of active commit, add dummy driver to provide consistent
1690      * reads on the top, while disabling it in the intermediate nodes, and make
1691      * the backing chain writable. */
1692     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1693                                          BDRV_O_RDWR, errp);
1694     if (mirror_top_bs == NULL) {
1695         return NULL;
1696     }
1697     if (!filter_node_name) {
1698         mirror_top_bs->implicit = true;
1699     }
1700 
1701     /* So that we can always drop this node */
1702     mirror_top_bs->never_freeze = true;
1703 
1704     mirror_top_bs->total_sectors = bs->total_sectors;
1705     mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
1706     mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
1707                                           BDRV_REQ_NO_FALLBACK;
1708     bs_opaque = g_new0(MirrorBDSOpaque, 1);
1709     mirror_top_bs->opaque = bs_opaque;
1710 
1711     bs_opaque->is_commit = target_is_backing;
1712 
1713     bdrv_drained_begin(bs);
1714     ret = bdrv_append(mirror_top_bs, bs, errp);
1715     bdrv_drained_end(bs);
1716 
1717     if (ret < 0) {
1718         bdrv_unref(mirror_top_bs);
1719         return NULL;
1720     }
1721 
1722     /* Make sure that the source is not resized while the job is running */
1723     s = block_job_create(job_id, driver, NULL, mirror_top_bs,
1724                          BLK_PERM_CONSISTENT_READ,
1725                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1726                          BLK_PERM_WRITE, speed,
1727                          creation_flags, cb, opaque, errp);
1728     if (!s) {
1729         goto fail;
1730     }
1731 
1732     /* The block job now has a reference to this node */
1733     bdrv_unref(mirror_top_bs);
1734 
1735     s->mirror_top_bs = mirror_top_bs;
1736 
1737     /* No resize for the target either; while the mirror is still running, a
1738      * consistent read isn't necessarily possible. We could possibly allow
1739      * writes and graph modifications, though it would likely defeat the
1740      * purpose of a mirror, so leave them blocked for now.
1741      *
1742      * In the case of active commit, things look a bit different, though,
1743      * because the target is an already populated backing file in active use.
1744      * We can allow anything except resize there.*/
1745 
1746     target_perms = BLK_PERM_WRITE;
1747     target_shared_perms = BLK_PERM_WRITE_UNCHANGED;
1748 
1749     if (target_is_backing) {
1750         int64_t bs_size, target_size;
1751         bs_size = bdrv_getlength(bs);
1752         if (bs_size < 0) {
1753             error_setg_errno(errp, -bs_size,
1754                              "Could not inquire top image size");
1755             goto fail;
1756         }
1757 
1758         target_size = bdrv_getlength(target);
1759         if (target_size < 0) {
1760             error_setg_errno(errp, -target_size,
1761                              "Could not inquire base image size");
1762             goto fail;
1763         }
1764 
1765         if (target_size < bs_size) {
1766             target_perms |= BLK_PERM_RESIZE;
1767         }
1768 
1769         target_shared_perms |= BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE;
1770     } else if (bdrv_chain_contains(bs, bdrv_skip_filters(target))) {
1771         /*
1772          * We may want to allow this in the future, but it would
1773          * require taking some extra care.
1774          */
1775         error_setg(errp, "Cannot mirror to a filter on top of a node in the "
1776                    "source's backing chain");
1777         goto fail;
1778     }
1779 
1780     s->target = blk_new(s->common.job.aio_context,
1781                         target_perms, target_shared_perms);
1782     ret = blk_insert_bs(s->target, target, errp);
1783     if (ret < 0) {
1784         goto fail;
1785     }
1786     if (is_mirror) {
1787         /* XXX: Mirror target could be a NBD server of target QEMU in the case
1788          * of non-shared block migration. To allow migration completion, we
1789          * have to allow "inactivate" of the target BB.  When that happens, we
1790          * know the job is drained, and the vcpus are stopped, so no write
1791          * operation will be performed. Block layer already has assertions to
1792          * ensure that. */
1793         blk_set_force_allow_inactivate(s->target);
1794     }
1795     blk_set_allow_aio_context_change(s->target, true);
1796     blk_set_disable_request_queuing(s->target, true);
1797 
1798     s->replaces = g_strdup(replaces);
1799     s->on_source_error = on_source_error;
1800     s->on_target_error = on_target_error;
1801     s->is_none_mode = is_none_mode;
1802     s->backing_mode = backing_mode;
1803     s->zero_target = zero_target;
1804     s->copy_mode = copy_mode;
1805     s->base = base;
1806     s->base_overlay = bdrv_find_overlay(bs, base);
1807     s->granularity = granularity;
1808     s->buf_size = ROUND_UP(buf_size, granularity);
1809     s->unmap = unmap;
1810     if (auto_complete) {
1811         s->should_complete = true;
1812     }
1813 
1814     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1815     if (!s->dirty_bitmap) {
1816         goto fail;
1817     }
1818     if (s->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING) {
1819         bdrv_disable_dirty_bitmap(s->dirty_bitmap);
1820     }
1821 
1822     ret = block_job_add_bdrv(&s->common, "source", bs, 0,
1823                              BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE |
1824                              BLK_PERM_CONSISTENT_READ,
1825                              errp);
1826     if (ret < 0) {
1827         goto fail;
1828     }
1829 
1830     /* Required permissions are already taken with blk_new() */
1831     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1832                        &error_abort);
1833 
1834     /* In commit_active_start() all intermediate nodes disappear, so
1835      * any jobs in them must be blocked */
1836     if (target_is_backing) {
1837         BlockDriverState *iter, *filtered_target;
1838         uint64_t iter_shared_perms;
1839 
1840         /*
1841          * The topmost node with
1842          * bdrv_skip_filters(filtered_target) == bdrv_skip_filters(target)
1843          */
1844         filtered_target = bdrv_cow_bs(bdrv_find_overlay(bs, target));
1845 
1846         assert(bdrv_skip_filters(filtered_target) ==
1847                bdrv_skip_filters(target));
1848 
1849         /*
1850          * XXX BLK_PERM_WRITE needs to be allowed so we don't block
1851          * ourselves at s->base (if writes are blocked for a node, they are
1852          * also blocked for its backing file). The other options would be a
1853          * second filter driver above s->base (== target).
1854          */
1855         iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE;
1856 
1857         for (iter = bdrv_filter_or_cow_bs(bs); iter != target;
1858              iter = bdrv_filter_or_cow_bs(iter))
1859         {
1860             if (iter == filtered_target) {
1861                 /*
1862                  * From here on, all nodes are filters on the base.
1863                  * This allows us to share BLK_PERM_CONSISTENT_READ.
1864                  */
1865                 iter_shared_perms |= BLK_PERM_CONSISTENT_READ;
1866             }
1867 
1868             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1869                                      iter_shared_perms, errp);
1870             if (ret < 0) {
1871                 goto fail;
1872             }
1873         }
1874 
1875         if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) {
1876             goto fail;
1877         }
1878     }
1879 
1880     QTAILQ_INIT(&s->ops_in_flight);
1881 
1882     trace_mirror_start(bs, s, opaque);
1883     job_start(&s->common.job);
1884 
1885     return &s->common;
1886 
1887 fail:
1888     if (s) {
1889         /* Make sure this BDS does not go away until we have completed the graph
1890          * changes below */
1891         bdrv_ref(mirror_top_bs);
1892 
1893         g_free(s->replaces);
1894         blk_unref(s->target);
1895         bs_opaque->job = NULL;
1896         if (s->dirty_bitmap) {
1897             bdrv_release_dirty_bitmap(s->dirty_bitmap);
1898         }
1899         job_early_fail(&s->common.job);
1900     }
1901 
1902     bs_opaque->stop = true;
1903     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
1904                              &error_abort);
1905     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
1906 
1907     bdrv_unref(mirror_top_bs);
1908 
1909     return NULL;
1910 }
1911 
1912 void mirror_start(const char *job_id, BlockDriverState *bs,
1913                   BlockDriverState *target, const char *replaces,
1914                   int creation_flags, int64_t speed,
1915                   uint32_t granularity, int64_t buf_size,
1916                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1917                   bool zero_target,
1918                   BlockdevOnError on_source_error,
1919                   BlockdevOnError on_target_error,
1920                   bool unmap, const char *filter_node_name,
1921                   MirrorCopyMode copy_mode, Error **errp)
1922 {
1923     bool is_none_mode;
1924     BlockDriverState *base;
1925 
1926     GLOBAL_STATE_CODE();
1927 
1928     if ((mode == MIRROR_SYNC_MODE_INCREMENTAL) ||
1929         (mode == MIRROR_SYNC_MODE_BITMAP)) {
1930         error_setg(errp, "Sync mode '%s' not supported",
1931                    MirrorSyncMode_str(mode));
1932         return;
1933     }
1934     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1935     base = mode == MIRROR_SYNC_MODE_TOP ? bdrv_backing_chain_next(bs) : NULL;
1936     mirror_start_job(job_id, bs, creation_flags, target, replaces,
1937                      speed, granularity, buf_size, backing_mode, zero_target,
1938                      on_source_error, on_target_error, unmap, NULL, NULL,
1939                      &mirror_job_driver, is_none_mode, base, false,
1940                      filter_node_name, true, copy_mode, errp);
1941 }
1942 
1943 BlockJob *commit_active_start(const char *job_id, BlockDriverState *bs,
1944                               BlockDriverState *base, int creation_flags,
1945                               int64_t speed, BlockdevOnError on_error,
1946                               const char *filter_node_name,
1947                               BlockCompletionFunc *cb, void *opaque,
1948                               bool auto_complete, Error **errp)
1949 {
1950     bool base_read_only;
1951     BlockJob *job;
1952 
1953     GLOBAL_STATE_CODE();
1954 
1955     base_read_only = bdrv_is_read_only(base);
1956 
1957     if (base_read_only) {
1958         if (bdrv_reopen_set_read_only(base, false, errp) < 0) {
1959             return NULL;
1960         }
1961     }
1962 
1963     job = mirror_start_job(
1964                      job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1965                      MIRROR_LEAVE_BACKING_CHAIN, false,
1966                      on_error, on_error, true, cb, opaque,
1967                      &commit_active_job_driver, false, base, auto_complete,
1968                      filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND,
1969                      errp);
1970     if (!job) {
1971         goto error_restore_flags;
1972     }
1973 
1974     return job;
1975 
1976 error_restore_flags:
1977     /* ignore error and errp for bdrv_reopen, because we want to propagate
1978      * the original error */
1979     if (base_read_only) {
1980         bdrv_reopen_set_read_only(base, true, NULL);
1981     }
1982     return NULL;
1983 }
1984