xref: /openbmc/qemu/block/mirror.c (revision 83974cf4)
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_BYTES (1 << 20) /* 1 Mb */
28 #define DEFAULT_MIRROR_BUF_SIZE (MAX_IN_FLIGHT * MAX_IO_BYTES)
29 
30 /* The mirroring buffer is a list of granularity-sized chunks.
31  * Free chunks are organized in a list.
32  */
33 typedef struct MirrorBuffer {
34     QSIMPLEQ_ENTRY(MirrorBuffer) next;
35 } MirrorBuffer;
36 
37 typedef struct MirrorBlockJob {
38     BlockJob common;
39     RateLimit limit;
40     BlockBackend *target;
41     BlockDriverState *mirror_top_bs;
42     BlockDriverState *source;
43     BlockDriverState *base;
44 
45     /* The name of the graph node to replace */
46     char *replaces;
47     /* The BDS to replace */
48     BlockDriverState *to_replace;
49     /* Used to block operations on the drive-mirror-replace target */
50     Error *replace_blocker;
51     bool is_none_mode;
52     BlockMirrorBackingMode backing_mode;
53     BlockdevOnError on_source_error, on_target_error;
54     bool synced;
55     bool should_complete;
56     int64_t granularity;
57     size_t buf_size;
58     int64_t bdev_length;
59     unsigned long *cow_bitmap;
60     BdrvDirtyBitmap *dirty_bitmap;
61     BdrvDirtyBitmapIter *dbi;
62     uint8_t *buf;
63     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
64     int buf_free_count;
65 
66     uint64_t last_pause_ns;
67     unsigned long *in_flight_bitmap;
68     int in_flight;
69     int64_t bytes_in_flight;
70     int ret;
71     bool unmap;
72     bool waiting_for_io;
73     int target_cluster_size;
74     int max_iov;
75     bool initial_zeroing_ongoing;
76 } MirrorBlockJob;
77 
78 typedef struct MirrorOp {
79     MirrorBlockJob *s;
80     QEMUIOVector qiov;
81     int64_t offset;
82     uint64_t bytes;
83 } MirrorOp;
84 
85 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
86                                             int error)
87 {
88     s->synced = false;
89     if (read) {
90         return block_job_error_action(&s->common, s->on_source_error,
91                                       true, error);
92     } else {
93         return block_job_error_action(&s->common, s->on_target_error,
94                                       false, error);
95     }
96 }
97 
98 static void mirror_iteration_done(MirrorOp *op, int ret)
99 {
100     MirrorBlockJob *s = op->s;
101     struct iovec *iov;
102     int64_t chunk_num;
103     int i, nb_chunks;
104 
105     trace_mirror_iteration_done(s, op->offset, op->bytes, ret);
106 
107     s->in_flight--;
108     s->bytes_in_flight -= op->bytes;
109     iov = op->qiov.iov;
110     for (i = 0; i < op->qiov.niov; i++) {
111         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
112         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
113         s->buf_free_count++;
114     }
115 
116     chunk_num = op->offset / s->granularity;
117     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
118     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
119     if (ret >= 0) {
120         if (s->cow_bitmap) {
121             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
122         }
123         if (!s->initial_zeroing_ongoing) {
124             s->common.offset += op->bytes;
125         }
126     }
127     qemu_iovec_destroy(&op->qiov);
128     g_free(op);
129 
130     if (s->waiting_for_io) {
131         qemu_coroutine_enter(s->common.co);
132     }
133 }
134 
135 static void mirror_write_complete(void *opaque, int ret)
136 {
137     MirrorOp *op = opaque;
138     MirrorBlockJob *s = op->s;
139 
140     aio_context_acquire(blk_get_aio_context(s->common.blk));
141     if (ret < 0) {
142         BlockErrorAction action;
143 
144         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
145         action = mirror_error_action(s, false, -ret);
146         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
147             s->ret = ret;
148         }
149     }
150     mirror_iteration_done(op, ret);
151     aio_context_release(blk_get_aio_context(s->common.blk));
152 }
153 
154 static void mirror_read_complete(void *opaque, int ret)
155 {
156     MirrorOp *op = opaque;
157     MirrorBlockJob *s = op->s;
158 
159     aio_context_acquire(blk_get_aio_context(s->common.blk));
160     if (ret < 0) {
161         BlockErrorAction action;
162 
163         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
164         action = mirror_error_action(s, true, -ret);
165         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
166             s->ret = ret;
167         }
168 
169         mirror_iteration_done(op, ret);
170     } else {
171         blk_aio_pwritev(s->target, op->offset, &op->qiov,
172                         0, mirror_write_complete, op);
173     }
174     aio_context_release(blk_get_aio_context(s->common.blk));
175 }
176 
177 /* Clip bytes relative to offset to not exceed end-of-file */
178 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s,
179                                         int64_t offset,
180                                         int64_t bytes)
181 {
182     return MIN(bytes, s->bdev_length - offset);
183 }
184 
185 /* Round offset and/or bytes to target cluster if COW is needed, and
186  * return the offset of the adjusted tail against original. */
187 static int mirror_cow_align(MirrorBlockJob *s, int64_t *offset,
188                             uint64_t *bytes)
189 {
190     bool need_cow;
191     int ret = 0;
192     int64_t align_offset = *offset;
193     unsigned int align_bytes = *bytes;
194     int max_bytes = s->granularity * s->max_iov;
195 
196     assert(*bytes < INT_MAX);
197     need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap);
198     need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity,
199                           s->cow_bitmap);
200     if (need_cow) {
201         bdrv_round_to_clusters(blk_bs(s->target), *offset, *bytes,
202                                &align_offset, &align_bytes);
203     }
204 
205     if (align_bytes > max_bytes) {
206         align_bytes = max_bytes;
207         if (need_cow) {
208             align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size);
209         }
210     }
211     /* Clipping may result in align_bytes unaligned to chunk boundary, but
212      * that doesn't matter because it's already the end of source image. */
213     align_bytes = mirror_clip_bytes(s, align_offset, align_bytes);
214 
215     ret = align_offset + align_bytes - (*offset + *bytes);
216     *offset = align_offset;
217     *bytes = align_bytes;
218     assert(ret >= 0);
219     return ret;
220 }
221 
222 static inline void mirror_wait_for_io(MirrorBlockJob *s)
223 {
224     assert(!s->waiting_for_io);
225     s->waiting_for_io = true;
226     qemu_coroutine_yield();
227     s->waiting_for_io = false;
228 }
229 
230 /* Submit async read while handling COW.
231  * Returns: The number of bytes copied after and including offset,
232  *          excluding any bytes copied prior to offset due to alignment.
233  *          This will be @bytes if no alignment is necessary, or
234  *          (new_end - offset) if tail is rounded up or down due to
235  *          alignment or buffer limit.
236  */
237 static uint64_t mirror_do_read(MirrorBlockJob *s, int64_t offset,
238                                uint64_t bytes)
239 {
240     BlockBackend *source = s->common.blk;
241     int nb_chunks;
242     uint64_t ret;
243     MirrorOp *op;
244     uint64_t max_bytes;
245 
246     max_bytes = s->granularity * s->max_iov;
247 
248     /* We can only handle as much as buf_size at a time. */
249     bytes = MIN(s->buf_size, MIN(max_bytes, bytes));
250     assert(bytes);
251     assert(bytes < BDRV_REQUEST_MAX_BYTES);
252     ret = bytes;
253 
254     if (s->cow_bitmap) {
255         ret += mirror_cow_align(s, &offset, &bytes);
256     }
257     assert(bytes <= s->buf_size);
258     /* The offset is granularity-aligned because:
259      * 1) Caller passes in aligned values;
260      * 2) mirror_cow_align is used only when target cluster is larger. */
261     assert(QEMU_IS_ALIGNED(offset, s->granularity));
262     /* The range is sector-aligned, since bdrv_getlength() rounds up. */
263     assert(QEMU_IS_ALIGNED(bytes, BDRV_SECTOR_SIZE));
264     nb_chunks = DIV_ROUND_UP(bytes, s->granularity);
265 
266     while (s->buf_free_count < nb_chunks) {
267         trace_mirror_yield_in_flight(s, offset, s->in_flight);
268         mirror_wait_for_io(s);
269     }
270 
271     /* Allocate a MirrorOp that is used as an AIO callback.  */
272     op = g_new(MirrorOp, 1);
273     op->s = s;
274     op->offset = offset;
275     op->bytes = bytes;
276 
277     /* Now make a QEMUIOVector taking enough granularity-sized chunks
278      * from s->buf_free.
279      */
280     qemu_iovec_init(&op->qiov, nb_chunks);
281     while (nb_chunks-- > 0) {
282         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
283         size_t remaining = bytes - op->qiov.size;
284 
285         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
286         s->buf_free_count--;
287         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
288     }
289 
290     /* Copy the dirty cluster.  */
291     s->in_flight++;
292     s->bytes_in_flight += bytes;
293     trace_mirror_one_iteration(s, offset, bytes);
294 
295     blk_aio_preadv(source, offset, &op->qiov, 0, mirror_read_complete, op);
296     return ret;
297 }
298 
299 static void mirror_do_zero_or_discard(MirrorBlockJob *s,
300                                       int64_t offset,
301                                       uint64_t bytes,
302                                       bool is_discard)
303 {
304     MirrorOp *op;
305 
306     /* Allocate a MirrorOp that is used as an AIO callback. The qiov is zeroed
307      * so the freeing in mirror_iteration_done is nop. */
308     op = g_new0(MirrorOp, 1);
309     op->s = s;
310     op->offset = offset;
311     op->bytes = bytes;
312 
313     s->in_flight++;
314     s->bytes_in_flight += bytes;
315     if (is_discard) {
316         blk_aio_pdiscard(s->target, offset,
317                          op->bytes, mirror_write_complete, op);
318     } else {
319         blk_aio_pwrite_zeroes(s->target, offset,
320                               op->bytes, s->unmap ? BDRV_REQ_MAY_UNMAP : 0,
321                               mirror_write_complete, op);
322     }
323 }
324 
325 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
326 {
327     BlockDriverState *source = s->source;
328     int64_t offset, first_chunk;
329     uint64_t delay_ns = 0;
330     /* At least the first dirty chunk is mirrored in one iteration. */
331     int nb_chunks = 1;
332     int sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
333     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
334     int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES);
335 
336     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
337     offset = bdrv_dirty_iter_next(s->dbi);
338     if (offset < 0) {
339         bdrv_set_dirty_iter(s->dbi, 0);
340         offset = bdrv_dirty_iter_next(s->dbi);
341         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
342         assert(offset >= 0);
343     }
344     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
345 
346     first_chunk = offset / s->granularity;
347     while (test_bit(first_chunk, s->in_flight_bitmap)) {
348         trace_mirror_yield_in_flight(s, offset, s->in_flight);
349         mirror_wait_for_io(s);
350     }
351 
352     block_job_pause_point(&s->common);
353 
354     /* Find the number of consective dirty chunks following the first dirty
355      * one, and wait for in flight requests in them. */
356     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
357     while (nb_chunks * s->granularity < s->buf_size) {
358         int64_t next_dirty;
359         int64_t next_offset = offset + nb_chunks * s->granularity;
360         int64_t next_chunk = next_offset / s->granularity;
361         if (next_offset >= s->bdev_length ||
362             !bdrv_get_dirty_locked(source, s->dirty_bitmap, next_offset)) {
363             break;
364         }
365         if (test_bit(next_chunk, s->in_flight_bitmap)) {
366             break;
367         }
368 
369         next_dirty = bdrv_dirty_iter_next(s->dbi);
370         if (next_dirty > next_offset || next_dirty < 0) {
371             /* The bitmap iterator's cache is stale, refresh it */
372             bdrv_set_dirty_iter(s->dbi, next_offset);
373             next_dirty = bdrv_dirty_iter_next(s->dbi);
374         }
375         assert(next_dirty == next_offset);
376         nb_chunks++;
377     }
378 
379     /* Clear dirty bits before querying the block status, because
380      * calling bdrv_get_block_status_above could yield - if some blocks are
381      * marked dirty in this window, we need to know.
382      */
383     bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset,
384                                    nb_chunks * s->granularity);
385     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
386 
387     bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks);
388     while (nb_chunks > 0 && offset < s->bdev_length) {
389         int64_t ret;
390         int io_sectors;
391         unsigned int io_bytes;
392         int64_t io_bytes_acct;
393         BlockDriverState *file;
394         enum MirrorMethod {
395             MIRROR_METHOD_COPY,
396             MIRROR_METHOD_ZERO,
397             MIRROR_METHOD_DISCARD
398         } mirror_method = MIRROR_METHOD_COPY;
399 
400         assert(!(offset % s->granularity));
401         ret = bdrv_get_block_status_above(source, NULL,
402                                           offset >> BDRV_SECTOR_BITS,
403                                           nb_chunks * sectors_per_chunk,
404                                           &io_sectors, &file);
405         io_bytes = io_sectors * BDRV_SECTOR_SIZE;
406         if (ret < 0) {
407             io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes);
408         } else if (ret & BDRV_BLOCK_DATA) {
409             io_bytes = MIN(io_bytes, max_io_bytes);
410         }
411 
412         io_bytes -= io_bytes % s->granularity;
413         if (io_bytes < s->granularity) {
414             io_bytes = s->granularity;
415         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
416             int64_t target_offset;
417             unsigned int target_bytes;
418             bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes,
419                                    &target_offset, &target_bytes);
420             if (target_offset == offset &&
421                 target_bytes == io_bytes) {
422                 mirror_method = ret & BDRV_BLOCK_ZERO ?
423                                     MIRROR_METHOD_ZERO :
424                                     MIRROR_METHOD_DISCARD;
425             }
426         }
427 
428         while (s->in_flight >= MAX_IN_FLIGHT) {
429             trace_mirror_yield_in_flight(s, offset, s->in_flight);
430             mirror_wait_for_io(s);
431         }
432 
433         if (s->ret < 0) {
434             return 0;
435         }
436 
437         io_bytes = mirror_clip_bytes(s, offset, io_bytes);
438         switch (mirror_method) {
439         case MIRROR_METHOD_COPY:
440             io_bytes = io_bytes_acct = mirror_do_read(s, offset, io_bytes);
441             break;
442         case MIRROR_METHOD_ZERO:
443         case MIRROR_METHOD_DISCARD:
444             mirror_do_zero_or_discard(s, offset, io_bytes,
445                                       mirror_method == MIRROR_METHOD_DISCARD);
446             if (write_zeroes_ok) {
447                 io_bytes_acct = 0;
448             } else {
449                 io_bytes_acct = io_bytes;
450             }
451             break;
452         default:
453             abort();
454         }
455         assert(io_bytes);
456         offset += io_bytes;
457         nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity);
458         if (s->common.speed) {
459             delay_ns = ratelimit_calculate_delay(&s->limit, io_bytes_acct);
460         }
461     }
462     return delay_ns;
463 }
464 
465 static void mirror_free_init(MirrorBlockJob *s)
466 {
467     int granularity = s->granularity;
468     size_t buf_size = s->buf_size;
469     uint8_t *buf = s->buf;
470 
471     assert(s->buf_free_count == 0);
472     QSIMPLEQ_INIT(&s->buf_free);
473     while (buf_size != 0) {
474         MirrorBuffer *cur = (MirrorBuffer *)buf;
475         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
476         s->buf_free_count++;
477         buf_size -= granularity;
478         buf += granularity;
479     }
480 }
481 
482 /* This is also used for the .pause callback. There is no matching
483  * mirror_resume() because mirror_run() will begin iterating again
484  * when the job is resumed.
485  */
486 static void mirror_wait_for_all_io(MirrorBlockJob *s)
487 {
488     while (s->in_flight > 0) {
489         mirror_wait_for_io(s);
490     }
491 }
492 
493 typedef struct {
494     int ret;
495 } MirrorExitData;
496 
497 static void mirror_exit(BlockJob *job, void *opaque)
498 {
499     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
500     MirrorExitData *data = opaque;
501     AioContext *replace_aio_context = NULL;
502     BlockDriverState *src = s->source;
503     BlockDriverState *target_bs = blk_bs(s->target);
504     BlockDriverState *mirror_top_bs = s->mirror_top_bs;
505     Error *local_err = NULL;
506 
507     bdrv_release_dirty_bitmap(src, s->dirty_bitmap);
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      *
519      * Note that blk_unref() alone doesn't necessarily drop permissions because
520      * we might be running nested inside mirror_drain(), which takes an extra
521      * reference, so use an explicit blk_set_perm() first. */
522     blk_set_perm(s->target, 0, BLK_PERM_ALL, &error_abort);
523     blk_unref(s->target);
524     s->target = NULL;
525 
526     /* We don't access the source any more. Dropping any WRITE/RESIZE is
527      * required before it could become a backing file of target_bs. */
528     bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
529                             &error_abort);
530     if (s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
531         BlockDriverState *backing = s->is_none_mode ? src : s->base;
532         if (backing_bs(target_bs) != backing) {
533             bdrv_set_backing_hd(target_bs, backing, &local_err);
534             if (local_err) {
535                 error_report_err(local_err);
536                 data->ret = -EPERM;
537             }
538         }
539     }
540 
541     if (s->to_replace) {
542         replace_aio_context = bdrv_get_aio_context(s->to_replace);
543         aio_context_acquire(replace_aio_context);
544     }
545 
546     if (s->should_complete && data->ret == 0) {
547         BlockDriverState *to_replace = src;
548         if (s->to_replace) {
549             to_replace = s->to_replace;
550         }
551 
552         if (bdrv_get_flags(target_bs) != bdrv_get_flags(to_replace)) {
553             bdrv_reopen(target_bs, bdrv_get_flags(to_replace), NULL);
554         }
555 
556         /* The mirror job has no requests in flight any more, but we need to
557          * drain potential other users of the BDS before changing the graph. */
558         bdrv_drained_begin(target_bs);
559         bdrv_replace_node(to_replace, target_bs, &local_err);
560         bdrv_drained_end(target_bs);
561         if (local_err) {
562             error_report_err(local_err);
563             data->ret = -EPERM;
564         }
565     }
566     if (s->to_replace) {
567         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
568         error_free(s->replace_blocker);
569         bdrv_unref(s->to_replace);
570     }
571     if (replace_aio_context) {
572         aio_context_release(replace_aio_context);
573     }
574     g_free(s->replaces);
575     bdrv_unref(target_bs);
576 
577     /* Remove the mirror filter driver from the graph. Before this, get rid of
578      * the blockers on the intermediate nodes so that the resulting state is
579      * valid. Also give up permissions on mirror_top_bs->backing, which might
580      * block the removal. */
581     block_job_remove_all_bdrv(job);
582     bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
583                             &error_abort);
584     bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
585 
586     /* We just changed the BDS the job BB refers to (with either or both of the
587      * bdrv_replace_node() calls), so switch the BB back so the cleanup does
588      * the right thing. We don't need any permissions any more now. */
589     blk_remove_bs(job->blk);
590     blk_set_perm(job->blk, 0, BLK_PERM_ALL, &error_abort);
591     blk_insert_bs(job->blk, mirror_top_bs, &error_abort);
592 
593     block_job_completed(&s->common, data->ret);
594 
595     g_free(data);
596     bdrv_drained_end(src);
597     bdrv_unref(mirror_top_bs);
598     bdrv_unref(src);
599 }
600 
601 static void mirror_throttle(MirrorBlockJob *s)
602 {
603     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
604 
605     if (now - s->last_pause_ns > SLICE_TIME) {
606         s->last_pause_ns = now;
607         block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, 0);
608     } else {
609         block_job_pause_point(&s->common);
610     }
611 }
612 
613 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
614 {
615     int64_t offset;
616     BlockDriverState *base = s->base;
617     BlockDriverState *bs = s->source;
618     BlockDriverState *target_bs = blk_bs(s->target);
619     int ret;
620     int64_t count;
621 
622     if (base == NULL && !bdrv_has_zero_init(target_bs)) {
623         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
624             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
625             return 0;
626         }
627 
628         s->initial_zeroing_ongoing = true;
629         for (offset = 0; offset < s->bdev_length; ) {
630             int bytes = MIN(s->bdev_length - offset,
631                             QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
632 
633             mirror_throttle(s);
634 
635             if (block_job_is_cancelled(&s->common)) {
636                 s->initial_zeroing_ongoing = false;
637                 return 0;
638             }
639 
640             if (s->in_flight >= MAX_IN_FLIGHT) {
641                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
642                                    s->in_flight);
643                 mirror_wait_for_io(s);
644                 continue;
645             }
646 
647             mirror_do_zero_or_discard(s, offset, bytes, false);
648             offset += bytes;
649         }
650 
651         mirror_wait_for_all_io(s);
652         s->initial_zeroing_ongoing = false;
653     }
654 
655     /* First part, loop on the sectors and initialize the dirty bitmap.  */
656     for (offset = 0; offset < s->bdev_length; ) {
657         /* Just to make sure we are not exceeding int limit. */
658         int bytes = MIN(s->bdev_length - offset,
659                         QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
660 
661         mirror_throttle(s);
662 
663         if (block_job_is_cancelled(&s->common)) {
664             return 0;
665         }
666 
667         ret = bdrv_is_allocated_above(bs, base, offset, bytes, &count);
668         if (ret < 0) {
669             return ret;
670         }
671 
672         assert(count);
673         if (ret == 1) {
674             bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
675         }
676         offset += count;
677     }
678     return 0;
679 }
680 
681 /* Called when going out of the streaming phase to flush the bulk of the
682  * data to the medium, or just before completing.
683  */
684 static int mirror_flush(MirrorBlockJob *s)
685 {
686     int ret = blk_flush(s->target);
687     if (ret < 0) {
688         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
689             s->ret = ret;
690         }
691     }
692     return ret;
693 }
694 
695 static void coroutine_fn mirror_run(void *opaque)
696 {
697     MirrorBlockJob *s = opaque;
698     MirrorExitData *data;
699     BlockDriverState *bs = s->source;
700     BlockDriverState *target_bs = blk_bs(s->target);
701     bool need_drain = true;
702     int64_t length;
703     BlockDriverInfo bdi;
704     char backing_filename[2]; /* we only need 2 characters because we are only
705                                  checking for a NULL string */
706     int ret = 0;
707 
708     if (block_job_is_cancelled(&s->common)) {
709         goto immediate_exit;
710     }
711 
712     s->bdev_length = bdrv_getlength(bs);
713     if (s->bdev_length < 0) {
714         ret = s->bdev_length;
715         goto immediate_exit;
716     }
717 
718     /* Active commit must resize the base image if its size differs from the
719      * active layer. */
720     if (s->base == blk_bs(s->target)) {
721         int64_t base_length;
722 
723         base_length = blk_getlength(s->target);
724         if (base_length < 0) {
725             ret = base_length;
726             goto immediate_exit;
727         }
728 
729         if (s->bdev_length > base_length) {
730             ret = blk_truncate(s->target, s->bdev_length, PREALLOC_MODE_OFF,
731                                NULL);
732             if (ret < 0) {
733                 goto immediate_exit;
734             }
735         }
736     }
737 
738     if (s->bdev_length == 0) {
739         /* Report BLOCK_JOB_READY and wait for complete. */
740         block_job_event_ready(&s->common);
741         s->synced = true;
742         while (!block_job_is_cancelled(&s->common) && !s->should_complete) {
743             block_job_yield(&s->common);
744         }
745         s->common.cancelled = false;
746         goto immediate_exit;
747     }
748 
749     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
750     s->in_flight_bitmap = bitmap_new(length);
751 
752     /* If we have no backing file yet in the destination, we cannot let
753      * the destination do COW.  Instead, we copy sectors around the
754      * dirty data if needed.  We need a bitmap to do that.
755      */
756     bdrv_get_backing_filename(target_bs, backing_filename,
757                               sizeof(backing_filename));
758     if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) {
759         s->target_cluster_size = bdi.cluster_size;
760     } else {
761         s->target_cluster_size = BDRV_SECTOR_SIZE;
762     }
763     if (backing_filename[0] && !target_bs->backing &&
764         s->granularity < s->target_cluster_size) {
765         s->buf_size = MAX(s->buf_size, s->target_cluster_size);
766         s->cow_bitmap = bitmap_new(length);
767     }
768     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
769 
770     s->buf = qemu_try_blockalign(bs, s->buf_size);
771     if (s->buf == NULL) {
772         ret = -ENOMEM;
773         goto immediate_exit;
774     }
775 
776     mirror_free_init(s);
777 
778     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
779     if (!s->is_none_mode) {
780         ret = mirror_dirty_init(s);
781         if (ret < 0 || block_job_is_cancelled(&s->common)) {
782             goto immediate_exit;
783         }
784     }
785 
786     assert(!s->dbi);
787     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
788     for (;;) {
789         uint64_t delay_ns = 0;
790         int64_t cnt, delta;
791         bool should_complete;
792 
793         if (s->ret < 0) {
794             ret = s->ret;
795             goto immediate_exit;
796         }
797 
798         block_job_pause_point(&s->common);
799 
800         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
801         /* s->common.offset contains the number of bytes already processed so
802          * far, cnt is the number of dirty bytes remaining and
803          * s->bytes_in_flight is the number of bytes currently being
804          * processed; together those are the current total operation length */
805         s->common.len = s->common.offset + s->bytes_in_flight + cnt;
806 
807         /* Note that even when no rate limit is applied we need to yield
808          * periodically with no pending I/O so that bdrv_drain_all() returns.
809          * We do so every SLICE_TIME nanoseconds, or when there is an error,
810          * or when the source is clean, whichever comes first.
811          */
812         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
813         if (delta < SLICE_TIME &&
814             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
815             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
816                 (cnt == 0 && s->in_flight > 0)) {
817                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
818                 mirror_wait_for_io(s);
819                 continue;
820             } else if (cnt != 0) {
821                 delay_ns = mirror_iteration(s);
822             }
823         }
824 
825         should_complete = false;
826         if (s->in_flight == 0 && cnt == 0) {
827             trace_mirror_before_flush(s);
828             if (!s->synced) {
829                 if (mirror_flush(s) < 0) {
830                     /* Go check s->ret.  */
831                     continue;
832                 }
833                 /* We're out of the streaming phase.  From now on, if the job
834                  * is cancelled we will actually complete all pending I/O and
835                  * report completion.  This way, block-job-cancel will leave
836                  * the target in a consistent state.
837                  */
838                 block_job_event_ready(&s->common);
839                 s->synced = true;
840             }
841 
842             should_complete = s->should_complete ||
843                 block_job_is_cancelled(&s->common);
844             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
845         }
846 
847         if (cnt == 0 && should_complete) {
848             /* The dirty bitmap is not updated while operations are pending.
849              * If we're about to exit, wait for pending operations before
850              * calling bdrv_get_dirty_count(bs), or we may exit while the
851              * source has dirty data to copy!
852              *
853              * Note that I/O can be submitted by the guest while
854              * mirror_populate runs, so pause it now.  Before deciding
855              * whether to switch to target check one last time if I/O has
856              * come in the meanwhile, and if not flush the data to disk.
857              */
858             trace_mirror_before_drain(s, cnt);
859 
860             bdrv_drained_begin(bs);
861             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
862             if (cnt > 0 || mirror_flush(s) < 0) {
863                 bdrv_drained_end(bs);
864                 continue;
865             }
866 
867             /* The two disks are in sync.  Exit and report successful
868              * completion.
869              */
870             assert(QLIST_EMPTY(&bs->tracked_requests));
871             s->common.cancelled = false;
872             need_drain = false;
873             break;
874         }
875 
876         ret = 0;
877         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
878         if (!s->synced) {
879             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
880             if (block_job_is_cancelled(&s->common)) {
881                 break;
882             }
883         } else if (!should_complete) {
884             delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0);
885             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
886         }
887         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
888     }
889 
890 immediate_exit:
891     if (s->in_flight > 0) {
892         /* We get here only if something went wrong.  Either the job failed,
893          * or it was cancelled prematurely so that we do not guarantee that
894          * the target is a copy of the source.
895          */
896         assert(ret < 0 || (!s->synced && block_job_is_cancelled(&s->common)));
897         assert(need_drain);
898         mirror_wait_for_all_io(s);
899     }
900 
901     assert(s->in_flight == 0);
902     qemu_vfree(s->buf);
903     g_free(s->cow_bitmap);
904     g_free(s->in_flight_bitmap);
905     bdrv_dirty_iter_free(s->dbi);
906 
907     data = g_malloc(sizeof(*data));
908     data->ret = ret;
909 
910     if (need_drain) {
911         bdrv_drained_begin(bs);
912     }
913     block_job_defer_to_main_loop(&s->common, mirror_exit, data);
914 }
915 
916 static void mirror_set_speed(BlockJob *job, int64_t speed, Error **errp)
917 {
918     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
919 
920     if (speed < 0) {
921         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
922         return;
923     }
924     ratelimit_set_speed(&s->limit, speed, SLICE_TIME);
925 }
926 
927 static void mirror_complete(BlockJob *job, Error **errp)
928 {
929     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
930     BlockDriverState *target;
931 
932     target = blk_bs(s->target);
933 
934     if (!s->synced) {
935         error_setg(errp, "The active block job '%s' cannot be completed",
936                    job->id);
937         return;
938     }
939 
940     if (s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
941         int ret;
942 
943         assert(!target->backing);
944         ret = bdrv_open_backing_file(target, NULL, "backing", errp);
945         if (ret < 0) {
946             return;
947         }
948     }
949 
950     /* block all operations on to_replace bs */
951     if (s->replaces) {
952         AioContext *replace_aio_context;
953 
954         s->to_replace = bdrv_find_node(s->replaces);
955         if (!s->to_replace) {
956             error_setg(errp, "Node name '%s' not found", s->replaces);
957             return;
958         }
959 
960         replace_aio_context = bdrv_get_aio_context(s->to_replace);
961         aio_context_acquire(replace_aio_context);
962 
963         /* TODO Translate this into permission system. Current definition of
964          * GRAPH_MOD would require to request it for the parents; they might
965          * not even be BlockDriverStates, however, so a BdrvChild can't address
966          * them. May need redefinition of GRAPH_MOD. */
967         error_setg(&s->replace_blocker,
968                    "block device is in use by block-job-complete");
969         bdrv_op_block_all(s->to_replace, s->replace_blocker);
970         bdrv_ref(s->to_replace);
971 
972         aio_context_release(replace_aio_context);
973     }
974 
975     s->should_complete = true;
976     block_job_enter(&s->common);
977 }
978 
979 static void mirror_pause(BlockJob *job)
980 {
981     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
982 
983     mirror_wait_for_all_io(s);
984 }
985 
986 static void mirror_attached_aio_context(BlockJob *job, AioContext *new_context)
987 {
988     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
989 
990     blk_set_aio_context(s->target, new_context);
991 }
992 
993 static void mirror_drain(BlockJob *job)
994 {
995     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
996 
997     /* Need to keep a reference in case blk_drain triggers execution
998      * of mirror_complete...
999      */
1000     if (s->target) {
1001         BlockBackend *target = s->target;
1002         blk_ref(target);
1003         blk_drain(target);
1004         blk_unref(target);
1005     }
1006 }
1007 
1008 static const BlockJobDriver mirror_job_driver = {
1009     .instance_size          = sizeof(MirrorBlockJob),
1010     .job_type               = BLOCK_JOB_TYPE_MIRROR,
1011     .set_speed              = mirror_set_speed,
1012     .start                  = mirror_run,
1013     .complete               = mirror_complete,
1014     .pause                  = mirror_pause,
1015     .attached_aio_context   = mirror_attached_aio_context,
1016     .drain                  = mirror_drain,
1017 };
1018 
1019 static const BlockJobDriver commit_active_job_driver = {
1020     .instance_size          = sizeof(MirrorBlockJob),
1021     .job_type               = BLOCK_JOB_TYPE_COMMIT,
1022     .set_speed              = mirror_set_speed,
1023     .start                  = mirror_run,
1024     .complete               = mirror_complete,
1025     .pause                  = mirror_pause,
1026     .attached_aio_context   = mirror_attached_aio_context,
1027     .drain                  = mirror_drain,
1028 };
1029 
1030 static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs,
1031     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1032 {
1033     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1034 }
1035 
1036 static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs,
1037     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1038 {
1039     return bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1040 }
1041 
1042 static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs)
1043 {
1044     if (bs->backing == NULL) {
1045         /* we can be here after failed bdrv_append in mirror_start_job */
1046         return 0;
1047     }
1048     return bdrv_co_flush(bs->backing->bs);
1049 }
1050 
1051 static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs,
1052     int64_t offset, int bytes, BdrvRequestFlags flags)
1053 {
1054     return bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1055 }
1056 
1057 static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs,
1058     int64_t offset, int bytes)
1059 {
1060     return bdrv_co_pdiscard(bs->backing->bs, offset, bytes);
1061 }
1062 
1063 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs, QDict *opts)
1064 {
1065     if (bs->backing == NULL) {
1066         /* we can be here after failed bdrv_attach_child in
1067          * bdrv_set_backing_hd */
1068         return;
1069     }
1070     bdrv_refresh_filename(bs->backing->bs);
1071     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1072             bs->backing->bs->filename);
1073 }
1074 
1075 static void bdrv_mirror_top_close(BlockDriverState *bs)
1076 {
1077 }
1078 
1079 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1080                                        const BdrvChildRole *role,
1081                                        BlockReopenQueue *reopen_queue,
1082                                        uint64_t perm, uint64_t shared,
1083                                        uint64_t *nperm, uint64_t *nshared)
1084 {
1085     /* Must be able to forward guest writes to the real image */
1086     *nperm = 0;
1087     if (perm & BLK_PERM_WRITE) {
1088         *nperm |= BLK_PERM_WRITE;
1089     }
1090 
1091     *nshared = BLK_PERM_ALL;
1092 }
1093 
1094 /* Dummy node that provides consistent read to its users without requiring it
1095  * from its backing file and that allows writes on the backing file chain. */
1096 static BlockDriver bdrv_mirror_top = {
1097     .format_name                = "mirror_top",
1098     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1099     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1100     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1101     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1102     .bdrv_co_flush              = bdrv_mirror_top_flush,
1103     .bdrv_co_get_block_status   = bdrv_co_get_block_status_from_backing,
1104     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1105     .bdrv_close                 = bdrv_mirror_top_close,
1106     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1107 };
1108 
1109 static void mirror_start_job(const char *job_id, BlockDriverState *bs,
1110                              int creation_flags, BlockDriverState *target,
1111                              const char *replaces, int64_t speed,
1112                              uint32_t granularity, int64_t buf_size,
1113                              BlockMirrorBackingMode backing_mode,
1114                              BlockdevOnError on_source_error,
1115                              BlockdevOnError on_target_error,
1116                              bool unmap,
1117                              BlockCompletionFunc *cb,
1118                              void *opaque,
1119                              const BlockJobDriver *driver,
1120                              bool is_none_mode, BlockDriverState *base,
1121                              bool auto_complete, const char *filter_node_name,
1122                              bool is_mirror,
1123                              Error **errp)
1124 {
1125     MirrorBlockJob *s;
1126     BlockDriverState *mirror_top_bs;
1127     bool target_graph_mod;
1128     bool target_is_backing;
1129     Error *local_err = NULL;
1130     int ret;
1131 
1132     if (granularity == 0) {
1133         granularity = bdrv_get_default_bitmap_granularity(target);
1134     }
1135 
1136     assert ((granularity & (granularity - 1)) == 0);
1137     /* Granularity must be large enough for sector-based dirty bitmap */
1138     assert(granularity >= BDRV_SECTOR_SIZE);
1139 
1140     if (buf_size < 0) {
1141         error_setg(errp, "Invalid parameter 'buf-size'");
1142         return;
1143     }
1144 
1145     if (buf_size == 0) {
1146         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1147     }
1148 
1149     /* In the case of active commit, add dummy driver to provide consistent
1150      * reads on the top, while disabling it in the intermediate nodes, and make
1151      * the backing chain writable. */
1152     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1153                                          BDRV_O_RDWR, errp);
1154     if (mirror_top_bs == NULL) {
1155         return;
1156     }
1157     if (!filter_node_name) {
1158         mirror_top_bs->implicit = true;
1159     }
1160     mirror_top_bs->total_sectors = bs->total_sectors;
1161     bdrv_set_aio_context(mirror_top_bs, bdrv_get_aio_context(bs));
1162 
1163     /* bdrv_append takes ownership of the mirror_top_bs reference, need to keep
1164      * it alive until block_job_create() succeeds even if bs has no parent. */
1165     bdrv_ref(mirror_top_bs);
1166     bdrv_drained_begin(bs);
1167     bdrv_append(mirror_top_bs, bs, &local_err);
1168     bdrv_drained_end(bs);
1169 
1170     if (local_err) {
1171         bdrv_unref(mirror_top_bs);
1172         error_propagate(errp, local_err);
1173         return;
1174     }
1175 
1176     /* Make sure that the source is not resized while the job is running */
1177     s = block_job_create(job_id, driver, mirror_top_bs,
1178                          BLK_PERM_CONSISTENT_READ,
1179                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1180                          BLK_PERM_WRITE | BLK_PERM_GRAPH_MOD, speed,
1181                          creation_flags, cb, opaque, errp);
1182     if (!s) {
1183         goto fail;
1184     }
1185     /* The block job now has a reference to this node */
1186     bdrv_unref(mirror_top_bs);
1187 
1188     s->source = bs;
1189     s->mirror_top_bs = mirror_top_bs;
1190 
1191     /* No resize for the target either; while the mirror is still running, a
1192      * consistent read isn't necessarily possible. We could possibly allow
1193      * writes and graph modifications, though it would likely defeat the
1194      * purpose of a mirror, so leave them blocked for now.
1195      *
1196      * In the case of active commit, things look a bit different, though,
1197      * because the target is an already populated backing file in active use.
1198      * We can allow anything except resize there.*/
1199     target_is_backing = bdrv_chain_contains(bs, target);
1200     target_graph_mod = (backing_mode != MIRROR_LEAVE_BACKING_CHAIN);
1201     s->target = blk_new(BLK_PERM_WRITE | BLK_PERM_RESIZE |
1202                         (target_graph_mod ? BLK_PERM_GRAPH_MOD : 0),
1203                         BLK_PERM_WRITE_UNCHANGED |
1204                         (target_is_backing ? BLK_PERM_CONSISTENT_READ |
1205                                              BLK_PERM_WRITE |
1206                                              BLK_PERM_GRAPH_MOD : 0));
1207     ret = blk_insert_bs(s->target, target, errp);
1208     if (ret < 0) {
1209         goto fail;
1210     }
1211     if (is_mirror) {
1212         /* XXX: Mirror target could be a NBD server of target QEMU in the case
1213          * of non-shared block migration. To allow migration completion, we
1214          * have to allow "inactivate" of the target BB.  When that happens, we
1215          * know the job is drained, and the vcpus are stopped, so no write
1216          * operation will be performed. Block layer already has assertions to
1217          * ensure that. */
1218         blk_set_force_allow_inactivate(s->target);
1219     }
1220 
1221     s->replaces = g_strdup(replaces);
1222     s->on_source_error = on_source_error;
1223     s->on_target_error = on_target_error;
1224     s->is_none_mode = is_none_mode;
1225     s->backing_mode = backing_mode;
1226     s->base = base;
1227     s->granularity = granularity;
1228     s->buf_size = ROUND_UP(buf_size, granularity);
1229     s->unmap = unmap;
1230     if (auto_complete) {
1231         s->should_complete = true;
1232     }
1233 
1234     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1235     if (!s->dirty_bitmap) {
1236         goto fail;
1237     }
1238 
1239     /* Required permissions are already taken with blk_new() */
1240     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1241                        &error_abort);
1242 
1243     /* In commit_active_start() all intermediate nodes disappear, so
1244      * any jobs in them must be blocked */
1245     if (target_is_backing) {
1246         BlockDriverState *iter;
1247         for (iter = backing_bs(bs); iter != target; iter = backing_bs(iter)) {
1248             /* XXX BLK_PERM_WRITE needs to be allowed so we don't block
1249              * ourselves at s->base (if writes are blocked for a node, they are
1250              * also blocked for its backing file). The other options would be a
1251              * second filter driver above s->base (== target). */
1252             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1253                                      BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
1254                                      errp);
1255             if (ret < 0) {
1256                 goto fail;
1257             }
1258         }
1259     }
1260 
1261     trace_mirror_start(bs, s, opaque);
1262     block_job_start(&s->common);
1263     return;
1264 
1265 fail:
1266     if (s) {
1267         /* Make sure this BDS does not go away until we have completed the graph
1268          * changes below */
1269         bdrv_ref(mirror_top_bs);
1270 
1271         g_free(s->replaces);
1272         blk_unref(s->target);
1273         block_job_early_fail(&s->common);
1274     }
1275 
1276     bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
1277                             &error_abort);
1278     bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
1279 
1280     bdrv_unref(mirror_top_bs);
1281 }
1282 
1283 void mirror_start(const char *job_id, BlockDriverState *bs,
1284                   BlockDriverState *target, const char *replaces,
1285                   int64_t speed, uint32_t granularity, int64_t buf_size,
1286                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1287                   BlockdevOnError on_source_error,
1288                   BlockdevOnError on_target_error,
1289                   bool unmap, const char *filter_node_name, Error **errp)
1290 {
1291     bool is_none_mode;
1292     BlockDriverState *base;
1293 
1294     if (mode == MIRROR_SYNC_MODE_INCREMENTAL) {
1295         error_setg(errp, "Sync mode 'incremental' not supported");
1296         return;
1297     }
1298     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1299     base = mode == MIRROR_SYNC_MODE_TOP ? backing_bs(bs) : NULL;
1300     mirror_start_job(job_id, bs, BLOCK_JOB_DEFAULT, target, replaces,
1301                      speed, granularity, buf_size, backing_mode,
1302                      on_source_error, on_target_error, unmap, NULL, NULL,
1303                      &mirror_job_driver, is_none_mode, base, false,
1304                      filter_node_name, true, errp);
1305 }
1306 
1307 void commit_active_start(const char *job_id, BlockDriverState *bs,
1308                          BlockDriverState *base, int creation_flags,
1309                          int64_t speed, BlockdevOnError on_error,
1310                          const char *filter_node_name,
1311                          BlockCompletionFunc *cb, void *opaque,
1312                          bool auto_complete, Error **errp)
1313 {
1314     int orig_base_flags;
1315     Error *local_err = NULL;
1316 
1317     orig_base_flags = bdrv_get_flags(base);
1318 
1319     if (bdrv_reopen(base, bs->open_flags, errp)) {
1320         return;
1321     }
1322 
1323     mirror_start_job(job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1324                      MIRROR_LEAVE_BACKING_CHAIN,
1325                      on_error, on_error, true, cb, opaque,
1326                      &commit_active_job_driver, false, base, auto_complete,
1327                      filter_node_name, false, &local_err);
1328     if (local_err) {
1329         error_propagate(errp, local_err);
1330         goto error_restore_flags;
1331     }
1332 
1333     return;
1334 
1335 error_restore_flags:
1336     /* ignore error and errp for bdrv_reopen, because we want to propagate
1337      * the original error */
1338     bdrv_reopen(base, orig_base_flags, NULL);
1339     return;
1340 }
1341