xref: /openbmc/qemu/block/mirror.c (revision 795c40b8)
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 "trace.h"
17 #include "block/blockjob_int.h"
18 #include "block/block_int.h"
19 #include "sysemu/block-backend.h"
20 #include "qapi/error.h"
21 #include "qapi/qmp/qerror.h"
22 #include "qemu/ratelimit.h"
23 #include "qemu/bitmap.h"
24 
25 #define SLICE_TIME    100000000ULL /* ns */
26 #define MAX_IN_FLIGHT 16
27 #define MAX_IO_SECTORS ((1 << 20) >> BDRV_SECTOR_BITS) /* 1 Mb */
28 #define DEFAULT_MIRROR_BUF_SIZE \
29     (MAX_IN_FLIGHT * MAX_IO_SECTORS * BDRV_SECTOR_SIZE)
30 
31 /* The mirroring buffer is a list of granularity-sized chunks.
32  * Free chunks are organized in a list.
33  */
34 typedef struct MirrorBuffer {
35     QSIMPLEQ_ENTRY(MirrorBuffer) next;
36 } MirrorBuffer;
37 
38 typedef struct MirrorBlockJob {
39     BlockJob common;
40     RateLimit limit;
41     BlockBackend *target;
42     BlockDriverState *mirror_top_bs;
43     BlockDriverState *source;
44     BlockDriverState *base;
45 
46     /* The name of the graph node to replace */
47     char *replaces;
48     /* The BDS to replace */
49     BlockDriverState *to_replace;
50     /* Used to block operations on the drive-mirror-replace target */
51     Error *replace_blocker;
52     bool is_none_mode;
53     BlockMirrorBackingMode backing_mode;
54     BlockdevOnError on_source_error, on_target_error;
55     bool synced;
56     bool should_complete;
57     int64_t granularity;
58     size_t buf_size;
59     int64_t bdev_length;
60     unsigned long *cow_bitmap;
61     BdrvDirtyBitmap *dirty_bitmap;
62     BdrvDirtyBitmapIter *dbi;
63     uint8_t *buf;
64     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
65     int buf_free_count;
66 
67     uint64_t last_pause_ns;
68     unsigned long *in_flight_bitmap;
69     int in_flight;
70     int64_t sectors_in_flight;
71     int ret;
72     bool unmap;
73     bool waiting_for_io;
74     int target_cluster_sectors;
75     int max_iov;
76     bool initial_zeroing_ongoing;
77 } MirrorBlockJob;
78 
79 typedef struct MirrorOp {
80     MirrorBlockJob *s;
81     QEMUIOVector qiov;
82     int64_t sector_num;
83     int nb_sectors;
84 } MirrorOp;
85 
86 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
87                                             int error)
88 {
89     s->synced = false;
90     if (read) {
91         return block_job_error_action(&s->common, s->on_source_error,
92                                       true, error);
93     } else {
94         return block_job_error_action(&s->common, s->on_target_error,
95                                       false, error);
96     }
97 }
98 
99 static void mirror_iteration_done(MirrorOp *op, int ret)
100 {
101     MirrorBlockJob *s = op->s;
102     struct iovec *iov;
103     int64_t chunk_num;
104     int i, nb_chunks, sectors_per_chunk;
105 
106     trace_mirror_iteration_done(s, op->sector_num, op->nb_sectors, ret);
107 
108     s->in_flight--;
109     s->sectors_in_flight -= op->nb_sectors;
110     iov = op->qiov.iov;
111     for (i = 0; i < op->qiov.niov; i++) {
112         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
113         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
114         s->buf_free_count++;
115     }
116 
117     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
118     chunk_num = op->sector_num / sectors_per_chunk;
119     nb_chunks = DIV_ROUND_UP(op->nb_sectors, sectors_per_chunk);
120     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
121     if (ret >= 0) {
122         if (s->cow_bitmap) {
123             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
124         }
125         if (!s->initial_zeroing_ongoing) {
126             s->common.offset += (uint64_t)op->nb_sectors * BDRV_SECTOR_SIZE;
127         }
128     }
129     qemu_iovec_destroy(&op->qiov);
130     g_free(op);
131 
132     if (s->waiting_for_io) {
133         qemu_coroutine_enter(s->common.co);
134     }
135 }
136 
137 static void mirror_write_complete(void *opaque, int ret)
138 {
139     MirrorOp *op = opaque;
140     MirrorBlockJob *s = op->s;
141 
142     aio_context_acquire(blk_get_aio_context(s->common.blk));
143     if (ret < 0) {
144         BlockErrorAction action;
145 
146         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->sector_num, op->nb_sectors);
147         action = mirror_error_action(s, false, -ret);
148         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
149             s->ret = ret;
150         }
151     }
152     mirror_iteration_done(op, ret);
153     aio_context_release(blk_get_aio_context(s->common.blk));
154 }
155 
156 static void mirror_read_complete(void *opaque, int ret)
157 {
158     MirrorOp *op = opaque;
159     MirrorBlockJob *s = op->s;
160 
161     aio_context_acquire(blk_get_aio_context(s->common.blk));
162     if (ret < 0) {
163         BlockErrorAction action;
164 
165         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->sector_num, op->nb_sectors);
166         action = mirror_error_action(s, true, -ret);
167         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
168             s->ret = ret;
169         }
170 
171         mirror_iteration_done(op, ret);
172     } else {
173         blk_aio_pwritev(s->target, op->sector_num * BDRV_SECTOR_SIZE, &op->qiov,
174                         0, mirror_write_complete, op);
175     }
176     aio_context_release(blk_get_aio_context(s->common.blk));
177 }
178 
179 static inline void mirror_clip_sectors(MirrorBlockJob *s,
180                                        int64_t sector_num,
181                                        int *nb_sectors)
182 {
183     *nb_sectors = MIN(*nb_sectors,
184                       s->bdev_length / BDRV_SECTOR_SIZE - sector_num);
185 }
186 
187 /* Round sector_num and/or nb_sectors to target cluster if COW is needed, and
188  * return the offset of the adjusted tail sector against original. */
189 static int mirror_cow_align(MirrorBlockJob *s,
190                             int64_t *sector_num,
191                             int *nb_sectors)
192 {
193     bool need_cow;
194     int ret = 0;
195     int chunk_sectors = s->granularity >> BDRV_SECTOR_BITS;
196     int64_t align_sector_num = *sector_num;
197     int align_nb_sectors = *nb_sectors;
198     int max_sectors = chunk_sectors * s->max_iov;
199 
200     need_cow = !test_bit(*sector_num / chunk_sectors, s->cow_bitmap);
201     need_cow |= !test_bit((*sector_num + *nb_sectors - 1) / chunk_sectors,
202                           s->cow_bitmap);
203     if (need_cow) {
204         bdrv_round_sectors_to_clusters(blk_bs(s->target), *sector_num,
205                                        *nb_sectors, &align_sector_num,
206                                        &align_nb_sectors);
207     }
208 
209     if (align_nb_sectors > max_sectors) {
210         align_nb_sectors = max_sectors;
211         if (need_cow) {
212             align_nb_sectors = QEMU_ALIGN_DOWN(align_nb_sectors,
213                                                s->target_cluster_sectors);
214         }
215     }
216     /* Clipping may result in align_nb_sectors unaligned to chunk boundary, but
217      * that doesn't matter because it's already the end of source image. */
218     mirror_clip_sectors(s, align_sector_num, &align_nb_sectors);
219 
220     ret = align_sector_num + align_nb_sectors - (*sector_num + *nb_sectors);
221     *sector_num = align_sector_num;
222     *nb_sectors = align_nb_sectors;
223     assert(ret >= 0);
224     return ret;
225 }
226 
227 static inline void mirror_wait_for_io(MirrorBlockJob *s)
228 {
229     assert(!s->waiting_for_io);
230     s->waiting_for_io = true;
231     qemu_coroutine_yield();
232     s->waiting_for_io = false;
233 }
234 
235 /* Submit async read while handling COW.
236  * Returns: The number of sectors copied after and including sector_num,
237  *          excluding any sectors copied prior to sector_num due to alignment.
238  *          This will be nb_sectors if no alignment is necessary, or
239  *          (new_end - sector_num) if tail is rounded up or down due to
240  *          alignment or buffer limit.
241  */
242 static int mirror_do_read(MirrorBlockJob *s, int64_t sector_num,
243                           int nb_sectors)
244 {
245     BlockBackend *source = s->common.blk;
246     int sectors_per_chunk, nb_chunks;
247     int ret;
248     MirrorOp *op;
249     int max_sectors;
250 
251     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
252     max_sectors = sectors_per_chunk * s->max_iov;
253 
254     /* We can only handle as much as buf_size at a time. */
255     nb_sectors = MIN(s->buf_size >> BDRV_SECTOR_BITS, nb_sectors);
256     nb_sectors = MIN(max_sectors, nb_sectors);
257     assert(nb_sectors);
258     ret = nb_sectors;
259 
260     if (s->cow_bitmap) {
261         ret += mirror_cow_align(s, &sector_num, &nb_sectors);
262     }
263     assert(nb_sectors << BDRV_SECTOR_BITS <= s->buf_size);
264     /* The sector range must meet granularity because:
265      * 1) Caller passes in aligned values;
266      * 2) mirror_cow_align is used only when target cluster is larger. */
267     assert(!(sector_num % sectors_per_chunk));
268     nb_chunks = DIV_ROUND_UP(nb_sectors, sectors_per_chunk);
269 
270     while (s->buf_free_count < nb_chunks) {
271         trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
272         mirror_wait_for_io(s);
273     }
274 
275     /* Allocate a MirrorOp that is used as an AIO callback.  */
276     op = g_new(MirrorOp, 1);
277     op->s = s;
278     op->sector_num = sector_num;
279     op->nb_sectors = nb_sectors;
280 
281     /* Now make a QEMUIOVector taking enough granularity-sized chunks
282      * from s->buf_free.
283      */
284     qemu_iovec_init(&op->qiov, nb_chunks);
285     while (nb_chunks-- > 0) {
286         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
287         size_t remaining = nb_sectors * BDRV_SECTOR_SIZE - op->qiov.size;
288 
289         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
290         s->buf_free_count--;
291         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
292     }
293 
294     /* Copy the dirty cluster.  */
295     s->in_flight++;
296     s->sectors_in_flight += nb_sectors;
297     trace_mirror_one_iteration(s, sector_num, nb_sectors);
298 
299     blk_aio_preadv(source, sector_num * BDRV_SECTOR_SIZE, &op->qiov, 0,
300                    mirror_read_complete, op);
301     return ret;
302 }
303 
304 static void mirror_do_zero_or_discard(MirrorBlockJob *s,
305                                       int64_t sector_num,
306                                       int nb_sectors,
307                                       bool is_discard)
308 {
309     MirrorOp *op;
310 
311     /* Allocate a MirrorOp that is used as an AIO callback. The qiov is zeroed
312      * so the freeing in mirror_iteration_done is nop. */
313     op = g_new0(MirrorOp, 1);
314     op->s = s;
315     op->sector_num = sector_num;
316     op->nb_sectors = nb_sectors;
317 
318     s->in_flight++;
319     s->sectors_in_flight += nb_sectors;
320     if (is_discard) {
321         blk_aio_pdiscard(s->target, sector_num << BDRV_SECTOR_BITS,
322                          op->nb_sectors << BDRV_SECTOR_BITS,
323                          mirror_write_complete, op);
324     } else {
325         blk_aio_pwrite_zeroes(s->target, sector_num * BDRV_SECTOR_SIZE,
326                               op->nb_sectors * BDRV_SECTOR_SIZE,
327                               s->unmap ? BDRV_REQ_MAY_UNMAP : 0,
328                               mirror_write_complete, op);
329     }
330 }
331 
332 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
333 {
334     BlockDriverState *source = s->source;
335     int64_t sector_num, first_chunk;
336     uint64_t delay_ns = 0;
337     /* At least the first dirty chunk is mirrored in one iteration. */
338     int nb_chunks = 1;
339     int64_t end = s->bdev_length / BDRV_SECTOR_SIZE;
340     int sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
341     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
342     int max_io_sectors = MAX((s->buf_size >> BDRV_SECTOR_BITS) / MAX_IN_FLIGHT,
343                              MAX_IO_SECTORS);
344 
345     sector_num = bdrv_dirty_iter_next(s->dbi);
346     if (sector_num < 0) {
347         bdrv_set_dirty_iter(s->dbi, 0);
348         sector_num = bdrv_dirty_iter_next(s->dbi);
349         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
350         assert(sector_num >= 0);
351     }
352 
353     first_chunk = sector_num / sectors_per_chunk;
354     while (test_bit(first_chunk, s->in_flight_bitmap)) {
355         trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
356         mirror_wait_for_io(s);
357     }
358 
359     block_job_pause_point(&s->common);
360 
361     /* Find the number of consective dirty chunks following the first dirty
362      * one, and wait for in flight requests in them. */
363     while (nb_chunks * sectors_per_chunk < (s->buf_size >> BDRV_SECTOR_BITS)) {
364         int64_t next_dirty;
365         int64_t next_sector = sector_num + nb_chunks * sectors_per_chunk;
366         int64_t next_chunk = next_sector / sectors_per_chunk;
367         if (next_sector >= end ||
368             !bdrv_get_dirty(source, s->dirty_bitmap, next_sector)) {
369             break;
370         }
371         if (test_bit(next_chunk, s->in_flight_bitmap)) {
372             break;
373         }
374 
375         next_dirty = bdrv_dirty_iter_next(s->dbi);
376         if (next_dirty > next_sector || next_dirty < 0) {
377             /* The bitmap iterator's cache is stale, refresh it */
378             bdrv_set_dirty_iter(s->dbi, next_sector);
379             next_dirty = bdrv_dirty_iter_next(s->dbi);
380         }
381         assert(next_dirty == next_sector);
382         nb_chunks++;
383     }
384 
385     /* Clear dirty bits before querying the block status, because
386      * calling bdrv_get_block_status_above could yield - if some blocks are
387      * marked dirty in this window, we need to know.
388      */
389     bdrv_reset_dirty_bitmap(s->dirty_bitmap, sector_num,
390                             nb_chunks * sectors_per_chunk);
391     bitmap_set(s->in_flight_bitmap, sector_num / sectors_per_chunk, nb_chunks);
392     while (nb_chunks > 0 && sector_num < end) {
393         int64_t ret;
394         int io_sectors, io_sectors_acct;
395         BlockDriverState *file;
396         enum MirrorMethod {
397             MIRROR_METHOD_COPY,
398             MIRROR_METHOD_ZERO,
399             MIRROR_METHOD_DISCARD
400         } mirror_method = MIRROR_METHOD_COPY;
401 
402         assert(!(sector_num % sectors_per_chunk));
403         ret = bdrv_get_block_status_above(source, NULL, sector_num,
404                                           nb_chunks * sectors_per_chunk,
405                                           &io_sectors, &file);
406         if (ret < 0) {
407             io_sectors = MIN(nb_chunks * sectors_per_chunk, max_io_sectors);
408         } else if (ret & BDRV_BLOCK_DATA) {
409             io_sectors = MIN(io_sectors, max_io_sectors);
410         }
411 
412         io_sectors -= io_sectors % sectors_per_chunk;
413         if (io_sectors < sectors_per_chunk) {
414             io_sectors = sectors_per_chunk;
415         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
416             int64_t target_sector_num;
417             int target_nb_sectors;
418             bdrv_round_sectors_to_clusters(blk_bs(s->target), sector_num,
419                                            io_sectors,  &target_sector_num,
420                                            &target_nb_sectors);
421             if (target_sector_num == sector_num &&
422                 target_nb_sectors == io_sectors) {
423                 mirror_method = ret & BDRV_BLOCK_ZERO ?
424                                     MIRROR_METHOD_ZERO :
425                                     MIRROR_METHOD_DISCARD;
426             }
427         }
428 
429         while (s->in_flight >= MAX_IN_FLIGHT) {
430             trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
431             mirror_wait_for_io(s);
432         }
433 
434         if (s->ret < 0) {
435             return 0;
436         }
437 
438         mirror_clip_sectors(s, sector_num, &io_sectors);
439         switch (mirror_method) {
440         case MIRROR_METHOD_COPY:
441             io_sectors = mirror_do_read(s, sector_num, io_sectors);
442             io_sectors_acct = io_sectors;
443             break;
444         case MIRROR_METHOD_ZERO:
445         case MIRROR_METHOD_DISCARD:
446             mirror_do_zero_or_discard(s, sector_num, io_sectors,
447                                       mirror_method == MIRROR_METHOD_DISCARD);
448             if (write_zeroes_ok) {
449                 io_sectors_acct = 0;
450             } else {
451                 io_sectors_acct = io_sectors;
452             }
453             break;
454         default:
455             abort();
456         }
457         assert(io_sectors);
458         sector_num += io_sectors;
459         nb_chunks -= DIV_ROUND_UP(io_sectors, sectors_per_chunk);
460         if (s->common.speed) {
461             delay_ns = ratelimit_calculate_delay(&s->limit, io_sectors_acct);
462         }
463     }
464     return delay_ns;
465 }
466 
467 static void mirror_free_init(MirrorBlockJob *s)
468 {
469     int granularity = s->granularity;
470     size_t buf_size = s->buf_size;
471     uint8_t *buf = s->buf;
472 
473     assert(s->buf_free_count == 0);
474     QSIMPLEQ_INIT(&s->buf_free);
475     while (buf_size != 0) {
476         MirrorBuffer *cur = (MirrorBuffer *)buf;
477         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
478         s->buf_free_count++;
479         buf_size -= granularity;
480         buf += granularity;
481     }
482 }
483 
484 /* This is also used for the .pause callback. There is no matching
485  * mirror_resume() because mirror_run() will begin iterating again
486  * when the job is resumed.
487  */
488 static void mirror_wait_for_all_io(MirrorBlockJob *s)
489 {
490     while (s->in_flight > 0) {
491         mirror_wait_for_io(s);
492     }
493 }
494 
495 typedef struct {
496     int ret;
497 } MirrorExitData;
498 
499 static void mirror_exit(BlockJob *job, void *opaque)
500 {
501     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
502     MirrorExitData *data = opaque;
503     AioContext *replace_aio_context = NULL;
504     BlockDriverState *src = s->source;
505     BlockDriverState *target_bs = blk_bs(s->target);
506     BlockDriverState *mirror_top_bs = s->mirror_top_bs;
507     Error *local_err = NULL;
508 
509     /* Make sure that the source BDS doesn't go away before we called
510      * block_job_completed(). */
511     bdrv_ref(src);
512     bdrv_ref(mirror_top_bs);
513     bdrv_ref(target_bs);
514 
515     /* Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
516      * inserting target_bs at s->to_replace, where we might not be able to get
517      * these permissions. */
518     blk_unref(s->target);
519     s->target = NULL;
520 
521     /* We don't access the source any more. Dropping any WRITE/RESIZE is
522      * required before it could become a backing file of target_bs. */
523     bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
524                             &error_abort);
525     if (s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
526         BlockDriverState *backing = s->is_none_mode ? src : s->base;
527         if (backing_bs(target_bs) != backing) {
528             bdrv_set_backing_hd(target_bs, backing, &local_err);
529             if (local_err) {
530                 error_report_err(local_err);
531                 data->ret = -EPERM;
532             }
533         }
534     }
535 
536     if (s->to_replace) {
537         replace_aio_context = bdrv_get_aio_context(s->to_replace);
538         aio_context_acquire(replace_aio_context);
539     }
540 
541     if (s->should_complete && data->ret == 0) {
542         BlockDriverState *to_replace = src;
543         if (s->to_replace) {
544             to_replace = s->to_replace;
545         }
546 
547         if (bdrv_get_flags(target_bs) != bdrv_get_flags(to_replace)) {
548             bdrv_reopen(target_bs, bdrv_get_flags(to_replace), NULL);
549         }
550 
551         /* The mirror job has no requests in flight any more, but we need to
552          * drain potential other users of the BDS before changing the graph. */
553         bdrv_drained_begin(target_bs);
554         bdrv_replace_node(to_replace, target_bs, &local_err);
555         bdrv_drained_end(target_bs);
556         if (local_err) {
557             error_report_err(local_err);
558             data->ret = -EPERM;
559         }
560     }
561     if (s->to_replace) {
562         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
563         error_free(s->replace_blocker);
564         bdrv_unref(s->to_replace);
565     }
566     if (replace_aio_context) {
567         aio_context_release(replace_aio_context);
568     }
569     g_free(s->replaces);
570     bdrv_unref(target_bs);
571 
572     /* Remove the mirror filter driver from the graph. Before this, get rid of
573      * the blockers on the intermediate nodes so that the resulting state is
574      * valid. Also give up permissions on mirror_top_bs->backing, which might
575      * block the removal. */
576     block_job_remove_all_bdrv(job);
577     bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
578                             &error_abort);
579     bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
580 
581     /* We just changed the BDS the job BB refers to (with either or both of the
582      * bdrv_replace_node() calls), so switch the BB back so the cleanup does
583      * the right thing. We don't need any permissions any more now. */
584     blk_remove_bs(job->blk);
585     blk_set_perm(job->blk, 0, BLK_PERM_ALL, &error_abort);
586     blk_insert_bs(job->blk, mirror_top_bs, &error_abort);
587 
588     block_job_completed(&s->common, data->ret);
589 
590     g_free(data);
591     bdrv_drained_end(src);
592     bdrv_unref(mirror_top_bs);
593     bdrv_unref(src);
594 }
595 
596 static void mirror_throttle(MirrorBlockJob *s)
597 {
598     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
599 
600     if (now - s->last_pause_ns > SLICE_TIME) {
601         s->last_pause_ns = now;
602         block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, 0);
603     } else {
604         block_job_pause_point(&s->common);
605     }
606 }
607 
608 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
609 {
610     int64_t sector_num, end;
611     BlockDriverState *base = s->base;
612     BlockDriverState *bs = s->source;
613     BlockDriverState *target_bs = blk_bs(s->target);
614     int ret, n;
615 
616     end = s->bdev_length / BDRV_SECTOR_SIZE;
617 
618     if (base == NULL && !bdrv_has_zero_init(target_bs)) {
619         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
620             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, end);
621             return 0;
622         }
623 
624         s->initial_zeroing_ongoing = true;
625         for (sector_num = 0; sector_num < end; ) {
626             int nb_sectors = MIN(end - sector_num,
627                 QEMU_ALIGN_DOWN(INT_MAX, s->granularity) >> BDRV_SECTOR_BITS);
628 
629             mirror_throttle(s);
630 
631             if (block_job_is_cancelled(&s->common)) {
632                 s->initial_zeroing_ongoing = false;
633                 return 0;
634             }
635 
636             if (s->in_flight >= MAX_IN_FLIGHT) {
637                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
638                                    s->in_flight);
639                 mirror_wait_for_io(s);
640                 continue;
641             }
642 
643             mirror_do_zero_or_discard(s, sector_num, nb_sectors, false);
644             sector_num += nb_sectors;
645         }
646 
647         mirror_wait_for_all_io(s);
648         s->initial_zeroing_ongoing = false;
649     }
650 
651     /* First part, loop on the sectors and initialize the dirty bitmap.  */
652     for (sector_num = 0; sector_num < end; ) {
653         /* Just to make sure we are not exceeding int limit. */
654         int nb_sectors = MIN(INT_MAX >> BDRV_SECTOR_BITS,
655                              end - sector_num);
656 
657         mirror_throttle(s);
658 
659         if (block_job_is_cancelled(&s->common)) {
660             return 0;
661         }
662 
663         ret = bdrv_is_allocated_above(bs, base, sector_num, nb_sectors, &n);
664         if (ret < 0) {
665             return ret;
666         }
667 
668         assert(n > 0);
669         if (ret == 1) {
670             bdrv_set_dirty_bitmap(s->dirty_bitmap, sector_num, n);
671         }
672         sector_num += n;
673     }
674     return 0;
675 }
676 
677 /* Called when going out of the streaming phase to flush the bulk of the
678  * data to the medium, or just before completing.
679  */
680 static int mirror_flush(MirrorBlockJob *s)
681 {
682     int ret = blk_flush(s->target);
683     if (ret < 0) {
684         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
685             s->ret = ret;
686         }
687     }
688     return ret;
689 }
690 
691 static void coroutine_fn mirror_run(void *opaque)
692 {
693     MirrorBlockJob *s = opaque;
694     MirrorExitData *data;
695     BlockDriverState *bs = s->source;
696     BlockDriverState *target_bs = blk_bs(s->target);
697     bool need_drain = true;
698     int64_t length;
699     BlockDriverInfo bdi;
700     char backing_filename[2]; /* we only need 2 characters because we are only
701                                  checking for a NULL string */
702     int ret = 0;
703     int target_cluster_size = BDRV_SECTOR_SIZE;
704 
705     if (block_job_is_cancelled(&s->common)) {
706         goto immediate_exit;
707     }
708 
709     s->bdev_length = bdrv_getlength(bs);
710     if (s->bdev_length < 0) {
711         ret = s->bdev_length;
712         goto immediate_exit;
713     }
714 
715     /* Active commit must resize the base image if its size differs from the
716      * active layer. */
717     if (s->base == blk_bs(s->target)) {
718         int64_t base_length;
719 
720         base_length = blk_getlength(s->target);
721         if (base_length < 0) {
722             ret = base_length;
723             goto immediate_exit;
724         }
725 
726         if (s->bdev_length > base_length) {
727             ret = blk_truncate(s->target, s->bdev_length, NULL);
728             if (ret < 0) {
729                 goto immediate_exit;
730             }
731         }
732     }
733 
734     if (s->bdev_length == 0) {
735         /* Report BLOCK_JOB_READY and wait for complete. */
736         block_job_event_ready(&s->common);
737         s->synced = true;
738         while (!block_job_is_cancelled(&s->common) && !s->should_complete) {
739             block_job_yield(&s->common);
740         }
741         s->common.cancelled = false;
742         goto immediate_exit;
743     }
744 
745     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
746     s->in_flight_bitmap = bitmap_new(length);
747 
748     /* If we have no backing file yet in the destination, we cannot let
749      * the destination do COW.  Instead, we copy sectors around the
750      * dirty data if needed.  We need a bitmap to do that.
751      */
752     bdrv_get_backing_filename(target_bs, backing_filename,
753                               sizeof(backing_filename));
754     if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) {
755         target_cluster_size = bdi.cluster_size;
756     }
757     if (backing_filename[0] && !target_bs->backing
758         && s->granularity < target_cluster_size) {
759         s->buf_size = MAX(s->buf_size, target_cluster_size);
760         s->cow_bitmap = bitmap_new(length);
761     }
762     s->target_cluster_sectors = target_cluster_size >> BDRV_SECTOR_BITS;
763     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
764 
765     s->buf = qemu_try_blockalign(bs, s->buf_size);
766     if (s->buf == NULL) {
767         ret = -ENOMEM;
768         goto immediate_exit;
769     }
770 
771     mirror_free_init(s);
772 
773     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
774     if (!s->is_none_mode) {
775         ret = mirror_dirty_init(s);
776         if (ret < 0 || block_job_is_cancelled(&s->common)) {
777             goto immediate_exit;
778         }
779     }
780 
781     assert(!s->dbi);
782     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap, 0);
783     for (;;) {
784         uint64_t delay_ns = 0;
785         int64_t cnt, delta;
786         bool should_complete;
787 
788         if (s->ret < 0) {
789             ret = s->ret;
790             goto immediate_exit;
791         }
792 
793         block_job_pause_point(&s->common);
794 
795         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
796         /* s->common.offset contains the number of bytes already processed so
797          * far, cnt is the number of dirty sectors remaining and
798          * s->sectors_in_flight is the number of sectors currently being
799          * processed; together those are the current total operation length */
800         s->common.len = s->common.offset +
801                         (cnt + s->sectors_in_flight) * BDRV_SECTOR_SIZE;
802 
803         /* Note that even when no rate limit is applied we need to yield
804          * periodically with no pending I/O so that bdrv_drain_all() returns.
805          * We do so every SLICE_TIME nanoseconds, or when there is an error,
806          * or when the source is clean, whichever comes first.
807          */
808         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
809         if (delta < SLICE_TIME &&
810             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
811             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
812                 (cnt == 0 && s->in_flight > 0)) {
813                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
814                 mirror_wait_for_io(s);
815                 continue;
816             } else if (cnt != 0) {
817                 delay_ns = mirror_iteration(s);
818             }
819         }
820 
821         should_complete = false;
822         if (s->in_flight == 0 && cnt == 0) {
823             trace_mirror_before_flush(s);
824             if (!s->synced) {
825                 if (mirror_flush(s) < 0) {
826                     /* Go check s->ret.  */
827                     continue;
828                 }
829                 /* We're out of the streaming phase.  From now on, if the job
830                  * is cancelled we will actually complete all pending I/O and
831                  * report completion.  This way, block-job-cancel will leave
832                  * the target in a consistent state.
833                  */
834                 block_job_event_ready(&s->common);
835                 s->synced = true;
836             }
837 
838             should_complete = s->should_complete ||
839                 block_job_is_cancelled(&s->common);
840             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
841         }
842 
843         if (cnt == 0 && should_complete) {
844             /* The dirty bitmap is not updated while operations are pending.
845              * If we're about to exit, wait for pending operations before
846              * calling bdrv_get_dirty_count(bs), or we may exit while the
847              * source has dirty data to copy!
848              *
849              * Note that I/O can be submitted by the guest while
850              * mirror_populate runs, so pause it now.  Before deciding
851              * whether to switch to target check one last time if I/O has
852              * come in the meanwhile, and if not flush the data to disk.
853              */
854             trace_mirror_before_drain(s, cnt);
855 
856             bdrv_drained_begin(bs);
857             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
858             if (cnt > 0 || mirror_flush(s) < 0) {
859                 bdrv_drained_end(bs);
860                 continue;
861             }
862 
863             /* The two disks are in sync.  Exit and report successful
864              * completion.
865              */
866             assert(QLIST_EMPTY(&bs->tracked_requests));
867             s->common.cancelled = false;
868             need_drain = false;
869             break;
870         }
871 
872         ret = 0;
873         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
874         if (!s->synced) {
875             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
876             if (block_job_is_cancelled(&s->common)) {
877                 break;
878             }
879         } else if (!should_complete) {
880             delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0);
881             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
882         }
883         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
884     }
885 
886 immediate_exit:
887     if (s->in_flight > 0) {
888         /* We get here only if something went wrong.  Either the job failed,
889          * or it was cancelled prematurely so that we do not guarantee that
890          * the target is a copy of the source.
891          */
892         assert(ret < 0 || (!s->synced && block_job_is_cancelled(&s->common)));
893         assert(need_drain);
894         mirror_wait_for_all_io(s);
895     }
896 
897     assert(s->in_flight == 0);
898     qemu_vfree(s->buf);
899     g_free(s->cow_bitmap);
900     g_free(s->in_flight_bitmap);
901     bdrv_dirty_iter_free(s->dbi);
902     bdrv_release_dirty_bitmap(bs, s->dirty_bitmap);
903 
904     data = g_malloc(sizeof(*data));
905     data->ret = ret;
906 
907     if (need_drain) {
908         bdrv_drained_begin(bs);
909     }
910     block_job_defer_to_main_loop(&s->common, mirror_exit, data);
911 }
912 
913 static void mirror_set_speed(BlockJob *job, int64_t speed, Error **errp)
914 {
915     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
916 
917     if (speed < 0) {
918         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
919         return;
920     }
921     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
922 }
923 
924 static void mirror_complete(BlockJob *job, Error **errp)
925 {
926     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
927     BlockDriverState *target;
928 
929     target = blk_bs(s->target);
930 
931     if (!s->synced) {
932         error_setg(errp, "The active block job '%s' cannot be completed",
933                    job->id);
934         return;
935     }
936 
937     if (s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
938         int ret;
939 
940         assert(!target->backing);
941         ret = bdrv_open_backing_file(target, NULL, "backing", errp);
942         if (ret < 0) {
943             return;
944         }
945     }
946 
947     /* block all operations on to_replace bs */
948     if (s->replaces) {
949         AioContext *replace_aio_context;
950 
951         s->to_replace = bdrv_find_node(s->replaces);
952         if (!s->to_replace) {
953             error_setg(errp, "Node name '%s' not found", s->replaces);
954             return;
955         }
956 
957         replace_aio_context = bdrv_get_aio_context(s->to_replace);
958         aio_context_acquire(replace_aio_context);
959 
960         /* TODO Translate this into permission system. Current definition of
961          * GRAPH_MOD would require to request it for the parents; they might
962          * not even be BlockDriverStates, however, so a BdrvChild can't address
963          * them. May need redefinition of GRAPH_MOD. */
964         error_setg(&s->replace_blocker,
965                    "block device is in use by block-job-complete");
966         bdrv_op_block_all(s->to_replace, s->replace_blocker);
967         bdrv_ref(s->to_replace);
968 
969         aio_context_release(replace_aio_context);
970     }
971 
972     s->should_complete = true;
973     block_job_enter(&s->common);
974 }
975 
976 static void mirror_pause(BlockJob *job)
977 {
978     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
979 
980     mirror_wait_for_all_io(s);
981 }
982 
983 static void mirror_attached_aio_context(BlockJob *job, AioContext *new_context)
984 {
985     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
986 
987     blk_set_aio_context(s->target, new_context);
988 }
989 
990 static void mirror_drain(BlockJob *job)
991 {
992     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
993 
994     /* Need to keep a reference in case blk_drain triggers execution
995      * of mirror_complete...
996      */
997     if (s->target) {
998         BlockBackend *target = s->target;
999         blk_ref(target);
1000         blk_drain(target);
1001         blk_unref(target);
1002     }
1003 }
1004 
1005 static const BlockJobDriver mirror_job_driver = {
1006     .instance_size          = sizeof(MirrorBlockJob),
1007     .job_type               = BLOCK_JOB_TYPE_MIRROR,
1008     .set_speed              = mirror_set_speed,
1009     .start                  = mirror_run,
1010     .complete               = mirror_complete,
1011     .pause                  = mirror_pause,
1012     .attached_aio_context   = mirror_attached_aio_context,
1013     .drain                  = mirror_drain,
1014 };
1015 
1016 static const BlockJobDriver commit_active_job_driver = {
1017     .instance_size          = sizeof(MirrorBlockJob),
1018     .job_type               = BLOCK_JOB_TYPE_COMMIT,
1019     .set_speed              = mirror_set_speed,
1020     .start                  = mirror_run,
1021     .complete               = mirror_complete,
1022     .pause                  = mirror_pause,
1023     .attached_aio_context   = mirror_attached_aio_context,
1024     .drain                  = mirror_drain,
1025 };
1026 
1027 static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs,
1028     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1029 {
1030     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1031 }
1032 
1033 static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs,
1034     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1035 {
1036     return bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1037 }
1038 
1039 static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs)
1040 {
1041     return bdrv_co_flush(bs->backing->bs);
1042 }
1043 
1044 static int64_t coroutine_fn bdrv_mirror_top_get_block_status(
1045     BlockDriverState *bs, int64_t sector_num, int nb_sectors, int *pnum,
1046     BlockDriverState **file)
1047 {
1048     *pnum = nb_sectors;
1049     *file = bs->backing->bs;
1050     return BDRV_BLOCK_RAW | BDRV_BLOCK_OFFSET_VALID | BDRV_BLOCK_DATA |
1051            (sector_num << BDRV_SECTOR_BITS);
1052 }
1053 
1054 static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs,
1055     int64_t offset, int count, BdrvRequestFlags flags)
1056 {
1057     return bdrv_co_pwrite_zeroes(bs->backing, offset, count, flags);
1058 }
1059 
1060 static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs,
1061     int64_t offset, int count)
1062 {
1063     return bdrv_co_pdiscard(bs->backing->bs, offset, count);
1064 }
1065 
1066 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs, QDict *opts)
1067 {
1068     bdrv_refresh_filename(bs->backing->bs);
1069     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1070             bs->backing->bs->filename);
1071 }
1072 
1073 static void bdrv_mirror_top_close(BlockDriverState *bs)
1074 {
1075 }
1076 
1077 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1078                                        const BdrvChildRole *role,
1079                                        uint64_t perm, uint64_t shared,
1080                                        uint64_t *nperm, uint64_t *nshared)
1081 {
1082     /* Must be able to forward guest writes to the real image */
1083     *nperm = 0;
1084     if (perm & BLK_PERM_WRITE) {
1085         *nperm |= BLK_PERM_WRITE;
1086     }
1087 
1088     *nshared = BLK_PERM_ALL;
1089 }
1090 
1091 /* Dummy node that provides consistent read to its users without requiring it
1092  * from its backing file and that allows writes on the backing file chain. */
1093 static BlockDriver bdrv_mirror_top = {
1094     .format_name                = "mirror_top",
1095     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1096     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1097     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1098     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1099     .bdrv_co_flush              = bdrv_mirror_top_flush,
1100     .bdrv_co_get_block_status   = bdrv_mirror_top_get_block_status,
1101     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1102     .bdrv_close                 = bdrv_mirror_top_close,
1103     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1104 };
1105 
1106 static void mirror_start_job(const char *job_id, BlockDriverState *bs,
1107                              int creation_flags, BlockDriverState *target,
1108                              const char *replaces, int64_t speed,
1109                              uint32_t granularity, int64_t buf_size,
1110                              BlockMirrorBackingMode backing_mode,
1111                              BlockdevOnError on_source_error,
1112                              BlockdevOnError on_target_error,
1113                              bool unmap,
1114                              BlockCompletionFunc *cb,
1115                              void *opaque,
1116                              const BlockJobDriver *driver,
1117                              bool is_none_mode, BlockDriverState *base,
1118                              bool auto_complete, const char *filter_node_name,
1119                              Error **errp)
1120 {
1121     MirrorBlockJob *s;
1122     BlockDriverState *mirror_top_bs;
1123     bool target_graph_mod;
1124     bool target_is_backing;
1125     Error *local_err = NULL;
1126     int ret;
1127 
1128     if (granularity == 0) {
1129         granularity = bdrv_get_default_bitmap_granularity(target);
1130     }
1131 
1132     assert ((granularity & (granularity - 1)) == 0);
1133 
1134     if (buf_size < 0) {
1135         error_setg(errp, "Invalid parameter 'buf-size'");
1136         return;
1137     }
1138 
1139     if (buf_size == 0) {
1140         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1141     }
1142 
1143     /* In the case of active commit, add dummy driver to provide consistent
1144      * reads on the top, while disabling it in the intermediate nodes, and make
1145      * the backing chain writable. */
1146     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1147                                          BDRV_O_RDWR, errp);
1148     if (mirror_top_bs == NULL) {
1149         return;
1150     }
1151     mirror_top_bs->total_sectors = bs->total_sectors;
1152     bdrv_set_aio_context(mirror_top_bs, bdrv_get_aio_context(bs));
1153 
1154     /* bdrv_append takes ownership of the mirror_top_bs reference, need to keep
1155      * it alive until block_job_create() succeeds even if bs has no parent. */
1156     bdrv_ref(mirror_top_bs);
1157     bdrv_drained_begin(bs);
1158     bdrv_append(mirror_top_bs, bs, &local_err);
1159     bdrv_drained_end(bs);
1160 
1161     if (local_err) {
1162         bdrv_unref(mirror_top_bs);
1163         error_propagate(errp, local_err);
1164         return;
1165     }
1166 
1167     /* Make sure that the source is not resized while the job is running */
1168     s = block_job_create(job_id, driver, mirror_top_bs,
1169                          BLK_PERM_CONSISTENT_READ,
1170                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1171                          BLK_PERM_WRITE | BLK_PERM_GRAPH_MOD, speed,
1172                          creation_flags, cb, opaque, errp);
1173     if (!s) {
1174         goto fail;
1175     }
1176     /* The block job now has a reference to this node */
1177     bdrv_unref(mirror_top_bs);
1178 
1179     s->source = bs;
1180     s->mirror_top_bs = mirror_top_bs;
1181 
1182     /* No resize for the target either; while the mirror is still running, a
1183      * consistent read isn't necessarily possible. We could possibly allow
1184      * writes and graph modifications, though it would likely defeat the
1185      * purpose of a mirror, so leave them blocked for now.
1186      *
1187      * In the case of active commit, things look a bit different, though,
1188      * because the target is an already populated backing file in active use.
1189      * We can allow anything except resize there.*/
1190     target_is_backing = bdrv_chain_contains(bs, target);
1191     target_graph_mod = (backing_mode != MIRROR_LEAVE_BACKING_CHAIN);
1192     s->target = blk_new(BLK_PERM_WRITE | BLK_PERM_RESIZE |
1193                         (target_graph_mod ? BLK_PERM_GRAPH_MOD : 0),
1194                         BLK_PERM_WRITE_UNCHANGED |
1195                         (target_is_backing ? BLK_PERM_CONSISTENT_READ |
1196                                              BLK_PERM_WRITE |
1197                                              BLK_PERM_GRAPH_MOD : 0));
1198     ret = blk_insert_bs(s->target, target, errp);
1199     if (ret < 0) {
1200         goto fail;
1201     }
1202 
1203     s->replaces = g_strdup(replaces);
1204     s->on_source_error = on_source_error;
1205     s->on_target_error = on_target_error;
1206     s->is_none_mode = is_none_mode;
1207     s->backing_mode = backing_mode;
1208     s->base = base;
1209     s->granularity = granularity;
1210     s->buf_size = ROUND_UP(buf_size, granularity);
1211     s->unmap = unmap;
1212     if (auto_complete) {
1213         s->should_complete = true;
1214     }
1215 
1216     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1217     if (!s->dirty_bitmap) {
1218         goto fail;
1219     }
1220 
1221     /* Required permissions are already taken with blk_new() */
1222     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1223                        &error_abort);
1224 
1225     /* In commit_active_start() all intermediate nodes disappear, so
1226      * any jobs in them must be blocked */
1227     if (target_is_backing) {
1228         BlockDriverState *iter;
1229         for (iter = backing_bs(bs); iter != target; iter = backing_bs(iter)) {
1230             /* XXX BLK_PERM_WRITE needs to be allowed so we don't block
1231              * ourselves at s->base (if writes are blocked for a node, they are
1232              * also blocked for its backing file). The other options would be a
1233              * second filter driver above s->base (== target). */
1234             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1235                                      BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
1236                                      errp);
1237             if (ret < 0) {
1238                 goto fail;
1239             }
1240         }
1241     }
1242 
1243     trace_mirror_start(bs, s, opaque);
1244     block_job_start(&s->common);
1245     return;
1246 
1247 fail:
1248     if (s) {
1249         /* Make sure this BDS does not go away until we have completed the graph
1250          * changes below */
1251         bdrv_ref(mirror_top_bs);
1252 
1253         g_free(s->replaces);
1254         blk_unref(s->target);
1255         block_job_unref(&s->common);
1256     }
1257 
1258     bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
1259                             &error_abort);
1260     bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
1261 
1262     bdrv_unref(mirror_top_bs);
1263 }
1264 
1265 void mirror_start(const char *job_id, BlockDriverState *bs,
1266                   BlockDriverState *target, const char *replaces,
1267                   int64_t speed, uint32_t granularity, int64_t buf_size,
1268                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1269                   BlockdevOnError on_source_error,
1270                   BlockdevOnError on_target_error,
1271                   bool unmap, const char *filter_node_name, Error **errp)
1272 {
1273     bool is_none_mode;
1274     BlockDriverState *base;
1275 
1276     if (mode == MIRROR_SYNC_MODE_INCREMENTAL) {
1277         error_setg(errp, "Sync mode 'incremental' not supported");
1278         return;
1279     }
1280     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1281     base = mode == MIRROR_SYNC_MODE_TOP ? backing_bs(bs) : NULL;
1282     mirror_start_job(job_id, bs, BLOCK_JOB_DEFAULT, target, replaces,
1283                      speed, granularity, buf_size, backing_mode,
1284                      on_source_error, on_target_error, unmap, NULL, NULL,
1285                      &mirror_job_driver, is_none_mode, base, false,
1286                      filter_node_name, errp);
1287 }
1288 
1289 void commit_active_start(const char *job_id, BlockDriverState *bs,
1290                          BlockDriverState *base, int creation_flags,
1291                          int64_t speed, BlockdevOnError on_error,
1292                          const char *filter_node_name,
1293                          BlockCompletionFunc *cb, void *opaque,
1294                          bool auto_complete, Error **errp)
1295 {
1296     int orig_base_flags;
1297     Error *local_err = NULL;
1298 
1299     orig_base_flags = bdrv_get_flags(base);
1300 
1301     if (bdrv_reopen(base, bs->open_flags, errp)) {
1302         return;
1303     }
1304 
1305     mirror_start_job(job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1306                      MIRROR_LEAVE_BACKING_CHAIN,
1307                      on_error, on_error, true, cb, opaque,
1308                      &commit_active_job_driver, false, base, auto_complete,
1309                      filter_node_name, &local_err);
1310     if (local_err) {
1311         error_propagate(errp, local_err);
1312         goto error_restore_flags;
1313     }
1314 
1315     return;
1316 
1317 error_restore_flags:
1318     /* ignore error and errp for bdrv_reopen, because we want to propagate
1319      * the original error */
1320     bdrv_reopen(base, orig_base_flags, NULL);
1321     return;
1322 }
1323