xref: /openbmc/qemu/block/mirror.c (revision 8ab5e8a5)
1 /*
2  * Image mirroring
3  *
4  * Copyright Red Hat, Inc. 2012
5  *
6  * Authors:
7  *  Paolo Bonzini  <pbonzini@redhat.com>
8  *
9  * This work is licensed under the terms of the GNU LGPL, version 2 or later.
10  * See the COPYING.LIB file in the top-level directory.
11  *
12  */
13 
14 #include "qemu/osdep.h"
15 #include "qemu/cutils.h"
16 #include "qemu/coroutine.h"
17 #include "qemu/range.h"
18 #include "trace.h"
19 #include "block/blockjob_int.h"
20 #include "block/block_int.h"
21 #include "block/dirty-bitmap.h"
22 #include "sysemu/block-backend.h"
23 #include "qapi/error.h"
24 #include "qapi/qmp/qerror.h"
25 #include "qemu/ratelimit.h"
26 #include "qemu/bitmap.h"
27 #include "qemu/memalign.h"
28 
29 #define MAX_IN_FLIGHT 16
30 #define MAX_IO_BYTES (1 << 20) /* 1 Mb */
31 #define DEFAULT_MIRROR_BUF_SIZE (MAX_IN_FLIGHT * MAX_IO_BYTES)
32 
33 /* The mirroring buffer is a list of granularity-sized chunks.
34  * Free chunks are organized in a list.
35  */
36 typedef struct MirrorBuffer {
37     QSIMPLEQ_ENTRY(MirrorBuffer) next;
38 } MirrorBuffer;
39 
40 typedef struct MirrorOp MirrorOp;
41 
42 typedef struct MirrorBlockJob {
43     BlockJob common;
44     BlockBackend *target;
45     BlockDriverState *mirror_top_bs;
46     BlockDriverState *base;
47     BlockDriverState *base_overlay;
48 
49     /* The name of the graph node to replace */
50     char *replaces;
51     /* The BDS to replace */
52     BlockDriverState *to_replace;
53     /* Used to block operations on the drive-mirror-replace target */
54     Error *replace_blocker;
55     bool is_none_mode;
56     BlockMirrorBackingMode backing_mode;
57     /* Whether the target image requires explicit zero-initialization */
58     bool zero_target;
59     MirrorCopyMode copy_mode;
60     BlockdevOnError on_source_error, on_target_error;
61     /* Set when the target is synced (dirty bitmap is clean, nothing
62      * in flight) and the job is running in active mode */
63     bool actively_synced;
64     bool should_complete;
65     int64_t granularity;
66     size_t buf_size;
67     int64_t bdev_length;
68     unsigned long *cow_bitmap;
69     BdrvDirtyBitmap *dirty_bitmap;
70     BdrvDirtyBitmapIter *dbi;
71     uint8_t *buf;
72     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
73     int buf_free_count;
74 
75     uint64_t last_pause_ns;
76     unsigned long *in_flight_bitmap;
77     unsigned in_flight;
78     int64_t bytes_in_flight;
79     QTAILQ_HEAD(, MirrorOp) ops_in_flight;
80     int ret;
81     bool unmap;
82     int target_cluster_size;
83     int max_iov;
84     bool initial_zeroing_ongoing;
85     int in_active_write_counter;
86     int64_t active_write_bytes_in_flight;
87     bool prepared;
88     bool in_drain;
89 } MirrorBlockJob;
90 
91 typedef struct MirrorBDSOpaque {
92     MirrorBlockJob *job;
93     bool stop;
94     bool is_commit;
95 } MirrorBDSOpaque;
96 
97 struct MirrorOp {
98     MirrorBlockJob *s;
99     QEMUIOVector qiov;
100     int64_t offset;
101     uint64_t bytes;
102 
103     /* The pointee is set by mirror_co_read(), mirror_co_zero(), and
104      * mirror_co_discard() before yielding for the first time */
105     int64_t *bytes_handled;
106 
107     bool is_pseudo_op;
108     bool is_active_write;
109     bool is_in_flight;
110     CoQueue waiting_requests;
111     Coroutine *co;
112     MirrorOp *waiting_for_op;
113 
114     QTAILQ_ENTRY(MirrorOp) next;
115 };
116 
117 typedef enum MirrorMethod {
118     MIRROR_METHOD_COPY,
119     MIRROR_METHOD_ZERO,
120     MIRROR_METHOD_DISCARD,
121 } MirrorMethod;
122 
123 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
124                                             int error)
125 {
126     s->actively_synced = false;
127     if (read) {
128         return block_job_error_action(&s->common, s->on_source_error,
129                                       true, error);
130     } else {
131         return block_job_error_action(&s->common, s->on_target_error,
132                                       false, error);
133     }
134 }
135 
136 static void coroutine_fn mirror_wait_on_conflicts(MirrorOp *self,
137                                                   MirrorBlockJob *s,
138                                                   uint64_t offset,
139                                                   uint64_t bytes)
140 {
141     uint64_t self_start_chunk = offset / s->granularity;
142     uint64_t self_end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
143     uint64_t self_nb_chunks = self_end_chunk - self_start_chunk;
144 
145     while (find_next_bit(s->in_flight_bitmap, self_end_chunk,
146                          self_start_chunk) < self_end_chunk &&
147            s->ret >= 0)
148     {
149         MirrorOp *op;
150 
151         QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
152             uint64_t op_start_chunk = op->offset / s->granularity;
153             uint64_t op_nb_chunks = DIV_ROUND_UP(op->offset + op->bytes,
154                                                  s->granularity) -
155                                     op_start_chunk;
156 
157             if (op == self) {
158                 continue;
159             }
160 
161             if (ranges_overlap(self_start_chunk, self_nb_chunks,
162                                op_start_chunk, op_nb_chunks))
163             {
164                 if (self) {
165                     /*
166                      * If the operation is already (indirectly) waiting for us,
167                      * or will wait for us as soon as it wakes up, then just go
168                      * on (instead of producing a deadlock in the former case).
169                      */
170                     if (op->waiting_for_op) {
171                         continue;
172                     }
173 
174                     self->waiting_for_op = op;
175                 }
176 
177                 qemu_co_queue_wait(&op->waiting_requests, NULL);
178 
179                 if (self) {
180                     self->waiting_for_op = NULL;
181                 }
182 
183                 break;
184             }
185         }
186     }
187 }
188 
189 static void coroutine_fn mirror_iteration_done(MirrorOp *op, int ret)
190 {
191     MirrorBlockJob *s = op->s;
192     struct iovec *iov;
193     int64_t chunk_num;
194     int i, nb_chunks;
195 
196     trace_mirror_iteration_done(s, op->offset, op->bytes, ret);
197 
198     s->in_flight--;
199     s->bytes_in_flight -= op->bytes;
200     iov = op->qiov.iov;
201     for (i = 0; i < op->qiov.niov; i++) {
202         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
203         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
204         s->buf_free_count++;
205     }
206 
207     chunk_num = op->offset / s->granularity;
208     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
209 
210     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
211     QTAILQ_REMOVE(&s->ops_in_flight, op, next);
212     if (ret >= 0) {
213         if (s->cow_bitmap) {
214             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
215         }
216         if (!s->initial_zeroing_ongoing) {
217             job_progress_update(&s->common.job, op->bytes);
218         }
219     }
220     qemu_iovec_destroy(&op->qiov);
221 
222     qemu_co_queue_restart_all(&op->waiting_requests);
223     g_free(op);
224 }
225 
226 static void coroutine_fn mirror_write_complete(MirrorOp *op, int ret)
227 {
228     MirrorBlockJob *s = op->s;
229 
230     if (ret < 0) {
231         BlockErrorAction action;
232 
233         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
234         action = mirror_error_action(s, false, -ret);
235         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
236             s->ret = ret;
237         }
238     }
239 
240     mirror_iteration_done(op, ret);
241 }
242 
243 static void coroutine_fn mirror_read_complete(MirrorOp *op, int ret)
244 {
245     MirrorBlockJob *s = op->s;
246 
247     if (ret < 0) {
248         BlockErrorAction action;
249 
250         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
251         action = mirror_error_action(s, true, -ret);
252         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
253             s->ret = ret;
254         }
255 
256         mirror_iteration_done(op, ret);
257         return;
258     }
259 
260     ret = blk_co_pwritev(s->target, op->offset, op->qiov.size, &op->qiov, 0);
261     mirror_write_complete(op, ret);
262 }
263 
264 /* Clip bytes relative to offset to not exceed end-of-file */
265 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s,
266                                         int64_t offset,
267                                         int64_t bytes)
268 {
269     return MIN(bytes, s->bdev_length - offset);
270 }
271 
272 /* Round offset and/or bytes to target cluster if COW is needed, and
273  * return the offset of the adjusted tail against original. */
274 static int mirror_cow_align(MirrorBlockJob *s, int64_t *offset,
275                             uint64_t *bytes)
276 {
277     bool need_cow;
278     int ret = 0;
279     int64_t align_offset = *offset;
280     int64_t align_bytes = *bytes;
281     int max_bytes = s->granularity * s->max_iov;
282 
283     need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap);
284     need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity,
285                           s->cow_bitmap);
286     if (need_cow) {
287         bdrv_round_to_clusters(blk_bs(s->target), *offset, *bytes,
288                                &align_offset, &align_bytes);
289     }
290 
291     if (align_bytes > max_bytes) {
292         align_bytes = max_bytes;
293         if (need_cow) {
294             align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size);
295         }
296     }
297     /* Clipping may result in align_bytes unaligned to chunk boundary, but
298      * that doesn't matter because it's already the end of source image. */
299     align_bytes = mirror_clip_bytes(s, align_offset, align_bytes);
300 
301     ret = align_offset + align_bytes - (*offset + *bytes);
302     *offset = align_offset;
303     *bytes = align_bytes;
304     assert(ret >= 0);
305     return ret;
306 }
307 
308 static inline void coroutine_fn
309 mirror_wait_for_free_in_flight_slot(MirrorBlockJob *s)
310 {
311     MirrorOp *op;
312 
313     QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
314         /*
315          * Do not wait on pseudo ops, because it may in turn wait on
316          * some other operation to start, which may in fact be the
317          * caller of this function.  Since there is only one pseudo op
318          * at any given time, we will always find some real operation
319          * to wait on.
320          * Also, do not wait on active operations, because they do not
321          * use up in-flight slots.
322          */
323         if (!op->is_pseudo_op && op->is_in_flight && !op->is_active_write) {
324             qemu_co_queue_wait(&op->waiting_requests, NULL);
325             return;
326         }
327     }
328     abort();
329 }
330 
331 /* Perform a mirror copy operation.
332  *
333  * *op->bytes_handled is set to the number of bytes copied after and
334  * including offset, excluding any bytes copied prior to offset due
335  * to alignment.  This will be op->bytes if no alignment is necessary,
336  * or (new_end - op->offset) if the tail is rounded up or down due to
337  * alignment or buffer limit.
338  */
339 static void coroutine_fn mirror_co_read(void *opaque)
340 {
341     MirrorOp *op = opaque;
342     MirrorBlockJob *s = op->s;
343     int nb_chunks;
344     uint64_t ret;
345     uint64_t max_bytes;
346 
347     max_bytes = s->granularity * s->max_iov;
348 
349     /* We can only handle as much as buf_size at a time. */
350     op->bytes = MIN(s->buf_size, MIN(max_bytes, op->bytes));
351     assert(op->bytes);
352     assert(op->bytes < BDRV_REQUEST_MAX_BYTES);
353     *op->bytes_handled = op->bytes;
354 
355     if (s->cow_bitmap) {
356         *op->bytes_handled += mirror_cow_align(s, &op->offset, &op->bytes);
357     }
358     /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */
359     assert(*op->bytes_handled <= UINT_MAX);
360     assert(op->bytes <= s->buf_size);
361     /* The offset is granularity-aligned because:
362      * 1) Caller passes in aligned values;
363      * 2) mirror_cow_align is used only when target cluster is larger. */
364     assert(QEMU_IS_ALIGNED(op->offset, s->granularity));
365     /* The range is sector-aligned, since bdrv_getlength() rounds up. */
366     assert(QEMU_IS_ALIGNED(op->bytes, BDRV_SECTOR_SIZE));
367     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
368 
369     while (s->buf_free_count < nb_chunks) {
370         trace_mirror_yield_in_flight(s, op->offset, s->in_flight);
371         mirror_wait_for_free_in_flight_slot(s);
372     }
373 
374     /* Now make a QEMUIOVector taking enough granularity-sized chunks
375      * from s->buf_free.
376      */
377     qemu_iovec_init(&op->qiov, nb_chunks);
378     while (nb_chunks-- > 0) {
379         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
380         size_t remaining = op->bytes - op->qiov.size;
381 
382         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
383         s->buf_free_count--;
384         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
385     }
386 
387     /* Copy the dirty cluster.  */
388     s->in_flight++;
389     s->bytes_in_flight += op->bytes;
390     op->is_in_flight = true;
391     trace_mirror_one_iteration(s, op->offset, op->bytes);
392 
393     ret = bdrv_co_preadv(s->mirror_top_bs->backing, op->offset, op->bytes,
394                          &op->qiov, 0);
395     mirror_read_complete(op, ret);
396 }
397 
398 static void coroutine_fn mirror_co_zero(void *opaque)
399 {
400     MirrorOp *op = opaque;
401     int ret;
402 
403     op->s->in_flight++;
404     op->s->bytes_in_flight += op->bytes;
405     *op->bytes_handled = op->bytes;
406     op->is_in_flight = true;
407 
408     ret = blk_co_pwrite_zeroes(op->s->target, op->offset, op->bytes,
409                                op->s->unmap ? BDRV_REQ_MAY_UNMAP : 0);
410     mirror_write_complete(op, ret);
411 }
412 
413 static void coroutine_fn mirror_co_discard(void *opaque)
414 {
415     MirrorOp *op = opaque;
416     int ret;
417 
418     op->s->in_flight++;
419     op->s->bytes_in_flight += op->bytes;
420     *op->bytes_handled = op->bytes;
421     op->is_in_flight = true;
422 
423     ret = blk_co_pdiscard(op->s->target, op->offset, op->bytes);
424     mirror_write_complete(op, ret);
425 }
426 
427 static unsigned mirror_perform(MirrorBlockJob *s, int64_t offset,
428                                unsigned bytes, MirrorMethod mirror_method)
429 {
430     MirrorOp *op;
431     Coroutine *co;
432     int64_t bytes_handled = -1;
433 
434     op = g_new(MirrorOp, 1);
435     *op = (MirrorOp){
436         .s              = s,
437         .offset         = offset,
438         .bytes          = bytes,
439         .bytes_handled  = &bytes_handled,
440     };
441     qemu_co_queue_init(&op->waiting_requests);
442 
443     switch (mirror_method) {
444     case MIRROR_METHOD_COPY:
445         co = qemu_coroutine_create(mirror_co_read, op);
446         break;
447     case MIRROR_METHOD_ZERO:
448         co = qemu_coroutine_create(mirror_co_zero, op);
449         break;
450     case MIRROR_METHOD_DISCARD:
451         co = qemu_coroutine_create(mirror_co_discard, op);
452         break;
453     default:
454         abort();
455     }
456     op->co = co;
457 
458     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
459     qemu_coroutine_enter(co);
460     /* At this point, ownership of op has been moved to the coroutine
461      * and the object may already be freed */
462 
463     /* Assert that this value has been set */
464     assert(bytes_handled >= 0);
465 
466     /* Same assertion as in mirror_co_read() (and for mirror_co_read()
467      * and mirror_co_discard(), bytes_handled == op->bytes, which
468      * is the @bytes parameter given to this function) */
469     assert(bytes_handled <= UINT_MAX);
470     return bytes_handled;
471 }
472 
473 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
474 {
475     BlockDriverState *source = s->mirror_top_bs->backing->bs;
476     MirrorOp *pseudo_op;
477     int64_t offset;
478     uint64_t delay_ns = 0, ret = 0;
479     /* At least the first dirty chunk is mirrored in one iteration. */
480     int nb_chunks = 1;
481     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
482     int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES);
483 
484     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
485     offset = bdrv_dirty_iter_next(s->dbi);
486     if (offset < 0) {
487         bdrv_set_dirty_iter(s->dbi, 0);
488         offset = bdrv_dirty_iter_next(s->dbi);
489         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
490         assert(offset >= 0);
491     }
492     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
493 
494     /*
495      * Wait for concurrent requests to @offset.  The next loop will limit the
496      * copied area based on in_flight_bitmap so we only copy an area that does
497      * not overlap with concurrent in-flight requests.  Still, we would like to
498      * copy something, so wait until there are at least no more requests to the
499      * very beginning of the area.
500      */
501     mirror_wait_on_conflicts(NULL, s, offset, 1);
502 
503     job_pause_point(&s->common.job);
504 
505     /* Find the number of consective dirty chunks following the first dirty
506      * one, and wait for in flight requests in them. */
507     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
508     while (nb_chunks * s->granularity < s->buf_size) {
509         int64_t next_dirty;
510         int64_t next_offset = offset + nb_chunks * s->granularity;
511         int64_t next_chunk = next_offset / s->granularity;
512         if (next_offset >= s->bdev_length ||
513             !bdrv_dirty_bitmap_get_locked(s->dirty_bitmap, next_offset)) {
514             break;
515         }
516         if (test_bit(next_chunk, s->in_flight_bitmap)) {
517             break;
518         }
519 
520         next_dirty = bdrv_dirty_iter_next(s->dbi);
521         if (next_dirty > next_offset || next_dirty < 0) {
522             /* The bitmap iterator's cache is stale, refresh it */
523             bdrv_set_dirty_iter(s->dbi, next_offset);
524             next_dirty = bdrv_dirty_iter_next(s->dbi);
525         }
526         assert(next_dirty == next_offset);
527         nb_chunks++;
528     }
529 
530     /* Clear dirty bits before querying the block status, because
531      * calling bdrv_block_status_above could yield - if some blocks are
532      * marked dirty in this window, we need to know.
533      */
534     bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset,
535                                    nb_chunks * s->granularity);
536     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
537 
538     /* Before claiming an area in the in-flight bitmap, we have to
539      * create a MirrorOp for it so that conflicting requests can wait
540      * for it.  mirror_perform() will create the real MirrorOps later,
541      * for now we just create a pseudo operation that will wake up all
542      * conflicting requests once all real operations have been
543      * launched. */
544     pseudo_op = g_new(MirrorOp, 1);
545     *pseudo_op = (MirrorOp){
546         .offset         = offset,
547         .bytes          = nb_chunks * s->granularity,
548         .is_pseudo_op   = true,
549     };
550     qemu_co_queue_init(&pseudo_op->waiting_requests);
551     QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next);
552 
553     bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks);
554     while (nb_chunks > 0 && offset < s->bdev_length) {
555         int ret;
556         int64_t io_bytes;
557         int64_t io_bytes_acct;
558         MirrorMethod mirror_method = MIRROR_METHOD_COPY;
559 
560         assert(!(offset % s->granularity));
561         ret = bdrv_block_status_above(source, NULL, offset,
562                                       nb_chunks * s->granularity,
563                                       &io_bytes, NULL, NULL);
564         if (ret < 0) {
565             io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes);
566         } else if (ret & BDRV_BLOCK_DATA) {
567             io_bytes = MIN(io_bytes, max_io_bytes);
568         }
569 
570         io_bytes -= io_bytes % s->granularity;
571         if (io_bytes < s->granularity) {
572             io_bytes = s->granularity;
573         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
574             int64_t target_offset;
575             int64_t target_bytes;
576             bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes,
577                                    &target_offset, &target_bytes);
578             if (target_offset == offset &&
579                 target_bytes == io_bytes) {
580                 mirror_method = ret & BDRV_BLOCK_ZERO ?
581                                     MIRROR_METHOD_ZERO :
582                                     MIRROR_METHOD_DISCARD;
583             }
584         }
585 
586         while (s->in_flight >= MAX_IN_FLIGHT) {
587             trace_mirror_yield_in_flight(s, offset, s->in_flight);
588             mirror_wait_for_free_in_flight_slot(s);
589         }
590 
591         if (s->ret < 0) {
592             ret = 0;
593             goto fail;
594         }
595 
596         io_bytes = mirror_clip_bytes(s, offset, io_bytes);
597         io_bytes = mirror_perform(s, offset, io_bytes, mirror_method);
598         if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) {
599             io_bytes_acct = 0;
600         } else {
601             io_bytes_acct = io_bytes;
602         }
603         assert(io_bytes);
604         offset += io_bytes;
605         nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity);
606         delay_ns = block_job_ratelimit_get_delay(&s->common, io_bytes_acct);
607     }
608 
609     ret = delay_ns;
610 fail:
611     QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next);
612     qemu_co_queue_restart_all(&pseudo_op->waiting_requests);
613     g_free(pseudo_op);
614 
615     return ret;
616 }
617 
618 static void mirror_free_init(MirrorBlockJob *s)
619 {
620     int granularity = s->granularity;
621     size_t buf_size = s->buf_size;
622     uint8_t *buf = s->buf;
623 
624     assert(s->buf_free_count == 0);
625     QSIMPLEQ_INIT(&s->buf_free);
626     while (buf_size != 0) {
627         MirrorBuffer *cur = (MirrorBuffer *)buf;
628         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
629         s->buf_free_count++;
630         buf_size -= granularity;
631         buf += granularity;
632     }
633 }
634 
635 /* This is also used for the .pause callback. There is no matching
636  * mirror_resume() because mirror_run() will begin iterating again
637  * when the job is resumed.
638  */
639 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s)
640 {
641     while (s->in_flight > 0) {
642         mirror_wait_for_free_in_flight_slot(s);
643     }
644 }
645 
646 /**
647  * mirror_exit_common: handle both abort() and prepare() cases.
648  * for .prepare, returns 0 on success and -errno on failure.
649  * for .abort cases, denoted by abort = true, MUST return 0.
650  */
651 static int mirror_exit_common(Job *job)
652 {
653     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
654     BlockJob *bjob = &s->common;
655     MirrorBDSOpaque *bs_opaque;
656     AioContext *replace_aio_context = NULL;
657     BlockDriverState *src;
658     BlockDriverState *target_bs;
659     BlockDriverState *mirror_top_bs;
660     Error *local_err = NULL;
661     bool abort = job->ret < 0;
662     int ret = 0;
663 
664     if (s->prepared) {
665         return 0;
666     }
667     s->prepared = true;
668 
669     mirror_top_bs = s->mirror_top_bs;
670     bs_opaque = mirror_top_bs->opaque;
671     src = mirror_top_bs->backing->bs;
672     target_bs = blk_bs(s->target);
673 
674     if (bdrv_chain_contains(src, target_bs)) {
675         bdrv_unfreeze_backing_chain(mirror_top_bs, target_bs);
676     }
677 
678     bdrv_release_dirty_bitmap(s->dirty_bitmap);
679 
680     /* Make sure that the source BDS doesn't go away during bdrv_replace_node,
681      * before we can call bdrv_drained_end */
682     bdrv_ref(src);
683     bdrv_ref(mirror_top_bs);
684     bdrv_ref(target_bs);
685 
686     /*
687      * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
688      * inserting target_bs at s->to_replace, where we might not be able to get
689      * these permissions.
690      */
691     blk_unref(s->target);
692     s->target = NULL;
693 
694     /* We don't access the source any more. Dropping any WRITE/RESIZE is
695      * required before it could become a backing file of target_bs. Not having
696      * these permissions any more means that we can't allow any new requests on
697      * mirror_top_bs from now on, so keep it drained. */
698     bdrv_drained_begin(mirror_top_bs);
699     bs_opaque->stop = true;
700     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
701                              &error_abort);
702     if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
703         BlockDriverState *backing = s->is_none_mode ? src : s->base;
704         BlockDriverState *unfiltered_target = bdrv_skip_filters(target_bs);
705 
706         if (bdrv_cow_bs(unfiltered_target) != backing) {
707             bdrv_set_backing_hd(unfiltered_target, backing, &local_err);
708             if (local_err) {
709                 error_report_err(local_err);
710                 local_err = NULL;
711                 ret = -EPERM;
712             }
713         }
714     } else if (!abort && s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
715         assert(!bdrv_backing_chain_next(target_bs));
716         ret = bdrv_open_backing_file(bdrv_skip_filters(target_bs), NULL,
717                                      "backing", &local_err);
718         if (ret < 0) {
719             error_report_err(local_err);
720             local_err = NULL;
721         }
722     }
723 
724     if (s->to_replace) {
725         replace_aio_context = bdrv_get_aio_context(s->to_replace);
726         aio_context_acquire(replace_aio_context);
727     }
728 
729     if (s->should_complete && !abort) {
730         BlockDriverState *to_replace = s->to_replace ?: src;
731         bool ro = bdrv_is_read_only(to_replace);
732 
733         if (ro != bdrv_is_read_only(target_bs)) {
734             bdrv_reopen_set_read_only(target_bs, ro, NULL);
735         }
736 
737         /* The mirror job has no requests in flight any more, but we need to
738          * drain potential other users of the BDS before changing the graph. */
739         assert(s->in_drain);
740         bdrv_drained_begin(target_bs);
741         /*
742          * Cannot use check_to_replace_node() here, because that would
743          * check for an op blocker on @to_replace, and we have our own
744          * there.
745          */
746         if (bdrv_recurse_can_replace(src, to_replace)) {
747             bdrv_replace_node(to_replace, target_bs, &local_err);
748         } else {
749             error_setg(&local_err, "Can no longer replace '%s' by '%s', "
750                        "because it can no longer be guaranteed that doing so "
751                        "would not lead to an abrupt change of visible data",
752                        to_replace->node_name, target_bs->node_name);
753         }
754         bdrv_drained_end(target_bs);
755         if (local_err) {
756             error_report_err(local_err);
757             ret = -EPERM;
758         }
759     }
760     if (s->to_replace) {
761         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
762         error_free(s->replace_blocker);
763         bdrv_unref(s->to_replace);
764     }
765     if (replace_aio_context) {
766         aio_context_release(replace_aio_context);
767     }
768     g_free(s->replaces);
769     bdrv_unref(target_bs);
770 
771     /*
772      * Remove the mirror filter driver from the graph. Before this, get rid of
773      * the blockers on the intermediate nodes so that the resulting state is
774      * valid.
775      */
776     block_job_remove_all_bdrv(bjob);
777     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
778 
779     bs_opaque->job = NULL;
780 
781     bdrv_drained_end(src);
782     bdrv_drained_end(mirror_top_bs);
783     s->in_drain = false;
784     bdrv_unref(mirror_top_bs);
785     bdrv_unref(src);
786 
787     return ret;
788 }
789 
790 static int mirror_prepare(Job *job)
791 {
792     return mirror_exit_common(job);
793 }
794 
795 static void mirror_abort(Job *job)
796 {
797     int ret = mirror_exit_common(job);
798     assert(ret == 0);
799 }
800 
801 static void coroutine_fn mirror_throttle(MirrorBlockJob *s)
802 {
803     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
804 
805     if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) {
806         s->last_pause_ns = now;
807         job_sleep_ns(&s->common.job, 0);
808     } else {
809         job_pause_point(&s->common.job);
810     }
811 }
812 
813 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
814 {
815     int64_t offset;
816     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
817     BlockDriverState *target_bs = blk_bs(s->target);
818     int ret;
819     int64_t count;
820 
821     if (s->zero_target) {
822         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
823             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
824             return 0;
825         }
826 
827         s->initial_zeroing_ongoing = true;
828         for (offset = 0; offset < s->bdev_length; ) {
829             int bytes = MIN(s->bdev_length - offset,
830                             QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
831 
832             mirror_throttle(s);
833 
834             if (job_is_cancelled(&s->common.job)) {
835                 s->initial_zeroing_ongoing = false;
836                 return 0;
837             }
838 
839             if (s->in_flight >= MAX_IN_FLIGHT) {
840                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
841                                    s->in_flight);
842                 mirror_wait_for_free_in_flight_slot(s);
843                 continue;
844             }
845 
846             mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO);
847             offset += bytes;
848         }
849 
850         mirror_wait_for_all_io(s);
851         s->initial_zeroing_ongoing = false;
852     }
853 
854     /* First part, loop on the sectors and initialize the dirty bitmap.  */
855     for (offset = 0; offset < s->bdev_length; ) {
856         /* Just to make sure we are not exceeding int limit. */
857         int bytes = MIN(s->bdev_length - offset,
858                         QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
859 
860         mirror_throttle(s);
861 
862         if (job_is_cancelled(&s->common.job)) {
863             return 0;
864         }
865 
866         ret = bdrv_is_allocated_above(bs, s->base_overlay, true, offset, bytes,
867                                       &count);
868         if (ret < 0) {
869             return ret;
870         }
871 
872         assert(count);
873         if (ret > 0) {
874             bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
875         }
876         offset += count;
877     }
878     return 0;
879 }
880 
881 /* Called when going out of the streaming phase to flush the bulk of the
882  * data to the medium, or just before completing.
883  */
884 static int mirror_flush(MirrorBlockJob *s)
885 {
886     int ret = blk_flush(s->target);
887     if (ret < 0) {
888         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
889             s->ret = ret;
890         }
891     }
892     return ret;
893 }
894 
895 static int coroutine_fn mirror_run(Job *job, Error **errp)
896 {
897     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
898     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
899     BlockDriverState *target_bs = blk_bs(s->target);
900     bool need_drain = true;
901     BlockDeviceIoStatus iostatus;
902     int64_t length;
903     int64_t target_length;
904     BlockDriverInfo bdi;
905     char backing_filename[2]; /* we only need 2 characters because we are only
906                                  checking for a NULL string */
907     int ret = 0;
908 
909     if (job_is_cancelled(&s->common.job)) {
910         goto immediate_exit;
911     }
912 
913     s->bdev_length = bdrv_co_getlength(bs);
914     if (s->bdev_length < 0) {
915         ret = s->bdev_length;
916         goto immediate_exit;
917     }
918 
919     target_length = blk_co_getlength(s->target);
920     if (target_length < 0) {
921         ret = target_length;
922         goto immediate_exit;
923     }
924 
925     /* Active commit must resize the base image if its size differs from the
926      * active layer. */
927     if (s->base == blk_bs(s->target)) {
928         if (s->bdev_length > target_length) {
929             ret = blk_co_truncate(s->target, s->bdev_length, false,
930                                   PREALLOC_MODE_OFF, 0, NULL);
931             if (ret < 0) {
932                 goto immediate_exit;
933             }
934         }
935     } else if (s->bdev_length != target_length) {
936         error_setg(errp, "Source and target image have different sizes");
937         ret = -EINVAL;
938         goto immediate_exit;
939     }
940 
941     if (s->bdev_length == 0) {
942         /* Transition to the READY state and wait for complete. */
943         job_transition_to_ready(&s->common.job);
944         s->actively_synced = true;
945         while (!job_cancel_requested(&s->common.job) && !s->should_complete) {
946             job_yield(&s->common.job);
947         }
948         goto immediate_exit;
949     }
950 
951     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
952     s->in_flight_bitmap = bitmap_new(length);
953 
954     /* If we have no backing file yet in the destination, we cannot let
955      * the destination do COW.  Instead, we copy sectors around the
956      * dirty data if needed.  We need a bitmap to do that.
957      */
958     bdrv_get_backing_filename(target_bs, backing_filename,
959                               sizeof(backing_filename));
960     if (!bdrv_co_get_info(target_bs, &bdi) && bdi.cluster_size) {
961         s->target_cluster_size = bdi.cluster_size;
962     } else {
963         s->target_cluster_size = BDRV_SECTOR_SIZE;
964     }
965     if (backing_filename[0] && !bdrv_backing_chain_next(target_bs) &&
966         s->granularity < s->target_cluster_size) {
967         s->buf_size = MAX(s->buf_size, s->target_cluster_size);
968         s->cow_bitmap = bitmap_new(length);
969     }
970     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
971 
972     s->buf = qemu_try_blockalign(bs, s->buf_size);
973     if (s->buf == NULL) {
974         ret = -ENOMEM;
975         goto immediate_exit;
976     }
977 
978     mirror_free_init(s);
979 
980     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
981     if (!s->is_none_mode) {
982         ret = mirror_dirty_init(s);
983         if (ret < 0 || job_is_cancelled(&s->common.job)) {
984             goto immediate_exit;
985         }
986     }
987 
988     assert(!s->dbi);
989     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
990     for (;;) {
991         uint64_t delay_ns = 0;
992         int64_t cnt, delta;
993         bool should_complete;
994 
995         if (s->ret < 0) {
996             ret = s->ret;
997             goto immediate_exit;
998         }
999 
1000         job_pause_point(&s->common.job);
1001 
1002         if (job_is_cancelled(&s->common.job)) {
1003             ret = 0;
1004             goto immediate_exit;
1005         }
1006 
1007         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1008         /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
1009          * the number of bytes currently being processed; together those are
1010          * the current remaining operation length */
1011         job_progress_set_remaining(&s->common.job,
1012                                    s->bytes_in_flight + cnt +
1013                                    s->active_write_bytes_in_flight);
1014 
1015         /* Note that even when no rate limit is applied we need to yield
1016          * periodically with no pending I/O so that bdrv_drain_all() returns.
1017          * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
1018          * an error, or when the source is clean, whichever comes first. */
1019         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
1020         WITH_JOB_LOCK_GUARD() {
1021             iostatus = s->common.iostatus;
1022         }
1023         if (delta < BLOCK_JOB_SLICE_TIME &&
1024             iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
1025             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
1026                 (cnt == 0 && s->in_flight > 0)) {
1027                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
1028                 mirror_wait_for_free_in_flight_slot(s);
1029                 continue;
1030             } else if (cnt != 0) {
1031                 delay_ns = mirror_iteration(s);
1032             }
1033         }
1034 
1035         should_complete = false;
1036         if (s->in_flight == 0 && cnt == 0) {
1037             trace_mirror_before_flush(s);
1038             if (!job_is_ready(&s->common.job)) {
1039                 if (mirror_flush(s) < 0) {
1040                     /* Go check s->ret.  */
1041                     continue;
1042                 }
1043                 /* We're out of the streaming phase.  From now on, if the job
1044                  * is cancelled we will actually complete all pending I/O and
1045                  * report completion.  This way, block-job-cancel will leave
1046                  * the target in a consistent state.
1047                  */
1048                 job_transition_to_ready(&s->common.job);
1049                 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) {
1050                     s->actively_synced = true;
1051                 }
1052             }
1053 
1054             should_complete = s->should_complete ||
1055                 job_cancel_requested(&s->common.job);
1056             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1057         }
1058 
1059         if (cnt == 0 && should_complete) {
1060             /* The dirty bitmap is not updated while operations are pending.
1061              * If we're about to exit, wait for pending operations before
1062              * calling bdrv_get_dirty_count(bs), or we may exit while the
1063              * source has dirty data to copy!
1064              *
1065              * Note that I/O can be submitted by the guest while
1066              * mirror_populate runs, so pause it now.  Before deciding
1067              * whether to switch to target check one last time if I/O has
1068              * come in the meanwhile, and if not flush the data to disk.
1069              */
1070             trace_mirror_before_drain(s, cnt);
1071 
1072             s->in_drain = true;
1073             bdrv_drained_begin(bs);
1074 
1075             /* Must be zero because we are drained */
1076             assert(s->in_active_write_counter == 0);
1077 
1078             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1079             if (cnt > 0 || mirror_flush(s) < 0) {
1080                 bdrv_drained_end(bs);
1081                 s->in_drain = false;
1082                 continue;
1083             }
1084 
1085             /* The two disks are in sync.  Exit and report successful
1086              * completion.
1087              */
1088             assert(QLIST_EMPTY(&bs->tracked_requests));
1089             need_drain = false;
1090             break;
1091         }
1092 
1093         if (job_is_ready(&s->common.job) && !should_complete) {
1094             delay_ns = (s->in_flight == 0 &&
1095                         cnt == 0 ? BLOCK_JOB_SLICE_TIME : 0);
1096         }
1097         trace_mirror_before_sleep(s, cnt, job_is_ready(&s->common.job),
1098                                   delay_ns);
1099         job_sleep_ns(&s->common.job, delay_ns);
1100         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1101     }
1102 
1103 immediate_exit:
1104     if (s->in_flight > 0) {
1105         /* We get here only if something went wrong.  Either the job failed,
1106          * or it was cancelled prematurely so that we do not guarantee that
1107          * the target is a copy of the source.
1108          */
1109         assert(ret < 0 || job_is_cancelled(&s->common.job));
1110         assert(need_drain);
1111         mirror_wait_for_all_io(s);
1112     }
1113 
1114     assert(s->in_flight == 0);
1115     qemu_vfree(s->buf);
1116     g_free(s->cow_bitmap);
1117     g_free(s->in_flight_bitmap);
1118     bdrv_dirty_iter_free(s->dbi);
1119 
1120     if (need_drain) {
1121         s->in_drain = true;
1122         bdrv_drained_begin(bs);
1123     }
1124 
1125     return ret;
1126 }
1127 
1128 static void mirror_complete(Job *job, Error **errp)
1129 {
1130     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1131 
1132     if (!job_is_ready(job)) {
1133         error_setg(errp, "The active block job '%s' cannot be completed",
1134                    job->id);
1135         return;
1136     }
1137 
1138     /* block all operations on to_replace bs */
1139     if (s->replaces) {
1140         AioContext *replace_aio_context;
1141 
1142         s->to_replace = bdrv_find_node(s->replaces);
1143         if (!s->to_replace) {
1144             error_setg(errp, "Node name '%s' not found", s->replaces);
1145             return;
1146         }
1147 
1148         replace_aio_context = bdrv_get_aio_context(s->to_replace);
1149         aio_context_acquire(replace_aio_context);
1150 
1151         /* TODO Translate this into child freeze system. */
1152         error_setg(&s->replace_blocker,
1153                    "block device is in use by block-job-complete");
1154         bdrv_op_block_all(s->to_replace, s->replace_blocker);
1155         bdrv_ref(s->to_replace);
1156 
1157         aio_context_release(replace_aio_context);
1158     }
1159 
1160     s->should_complete = true;
1161 
1162     /* If the job is paused, it will be re-entered when it is resumed */
1163     WITH_JOB_LOCK_GUARD() {
1164         if (!job->paused) {
1165             job_enter_cond_locked(job, NULL);
1166         }
1167     }
1168 }
1169 
1170 static void coroutine_fn mirror_pause(Job *job)
1171 {
1172     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1173 
1174     mirror_wait_for_all_io(s);
1175 }
1176 
1177 static bool mirror_drained_poll(BlockJob *job)
1178 {
1179     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1180 
1181     /* If the job isn't paused nor cancelled, we can't be sure that it won't
1182      * issue more requests. We make an exception if we've reached this point
1183      * from one of our own drain sections, to avoid a deadlock waiting for
1184      * ourselves.
1185      */
1186     WITH_JOB_LOCK_GUARD() {
1187         if (!s->common.job.paused && !job_is_cancelled_locked(&job->job)
1188             && !s->in_drain) {
1189             return true;
1190         }
1191     }
1192 
1193     return !!s->in_flight;
1194 }
1195 
1196 static bool mirror_cancel(Job *job, bool force)
1197 {
1198     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1199     BlockDriverState *target = blk_bs(s->target);
1200 
1201     /*
1202      * Before the job is READY, we treat any cancellation like a
1203      * force-cancellation.
1204      */
1205     force = force || !job_is_ready(job);
1206 
1207     if (force) {
1208         bdrv_cancel_in_flight(target);
1209     }
1210     return force;
1211 }
1212 
1213 static bool commit_active_cancel(Job *job, bool force)
1214 {
1215     /* Same as above in mirror_cancel() */
1216     return force || !job_is_ready(job);
1217 }
1218 
1219 static const BlockJobDriver mirror_job_driver = {
1220     .job_driver = {
1221         .instance_size          = sizeof(MirrorBlockJob),
1222         .job_type               = JOB_TYPE_MIRROR,
1223         .free                   = block_job_free,
1224         .user_resume            = block_job_user_resume,
1225         .run                    = mirror_run,
1226         .prepare                = mirror_prepare,
1227         .abort                  = mirror_abort,
1228         .pause                  = mirror_pause,
1229         .complete               = mirror_complete,
1230         .cancel                 = mirror_cancel,
1231     },
1232     .drained_poll           = mirror_drained_poll,
1233 };
1234 
1235 static const BlockJobDriver commit_active_job_driver = {
1236     .job_driver = {
1237         .instance_size          = sizeof(MirrorBlockJob),
1238         .job_type               = JOB_TYPE_COMMIT,
1239         .free                   = block_job_free,
1240         .user_resume            = block_job_user_resume,
1241         .run                    = mirror_run,
1242         .prepare                = mirror_prepare,
1243         .abort                  = mirror_abort,
1244         .pause                  = mirror_pause,
1245         .complete               = mirror_complete,
1246         .cancel                 = commit_active_cancel,
1247     },
1248     .drained_poll           = mirror_drained_poll,
1249 };
1250 
1251 static void coroutine_fn
1252 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method,
1253                      uint64_t offset, uint64_t bytes,
1254                      QEMUIOVector *qiov, int flags)
1255 {
1256     int ret;
1257     size_t qiov_offset = 0;
1258     int64_t bitmap_offset, bitmap_end;
1259 
1260     if (!QEMU_IS_ALIGNED(offset, job->granularity) &&
1261         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset))
1262     {
1263             /*
1264              * Dirty unaligned padding: ignore it.
1265              *
1266              * Reasoning:
1267              * 1. If we copy it, we can't reset corresponding bit in
1268              *    dirty_bitmap as there may be some "dirty" bytes still not
1269              *    copied.
1270              * 2. It's already dirty, so skipping it we don't diverge mirror
1271              *    progress.
1272              *
1273              * Note, that because of this, guest write may have no contribution
1274              * into mirror converge, but that's not bad, as we have background
1275              * process of mirroring. If under some bad circumstances (high guest
1276              * IO load) background process starve, we will not converge anyway,
1277              * even if each write will contribute, as guest is not guaranteed to
1278              * rewrite the whole disk.
1279              */
1280             qiov_offset = QEMU_ALIGN_UP(offset, job->granularity) - offset;
1281             if (bytes <= qiov_offset) {
1282                 /* nothing to do after shrink */
1283                 return;
1284             }
1285             offset += qiov_offset;
1286             bytes -= qiov_offset;
1287     }
1288 
1289     if (!QEMU_IS_ALIGNED(offset + bytes, job->granularity) &&
1290         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset + bytes - 1))
1291     {
1292         uint64_t tail = (offset + bytes) % job->granularity;
1293 
1294         if (bytes <= tail) {
1295             /* nothing to do after shrink */
1296             return;
1297         }
1298         bytes -= tail;
1299     }
1300 
1301     /*
1302      * Tails are either clean or shrunk, so for bitmap resetting
1303      * we safely align the range down.
1304      */
1305     bitmap_offset = QEMU_ALIGN_UP(offset, job->granularity);
1306     bitmap_end = QEMU_ALIGN_DOWN(offset + bytes, job->granularity);
1307     if (bitmap_offset < bitmap_end) {
1308         bdrv_reset_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1309                                 bitmap_end - bitmap_offset);
1310     }
1311 
1312     job_progress_increase_remaining(&job->common.job, bytes);
1313     job->active_write_bytes_in_flight += bytes;
1314 
1315     switch (method) {
1316     case MIRROR_METHOD_COPY:
1317         ret = blk_co_pwritev_part(job->target, offset, bytes,
1318                                   qiov, qiov_offset, flags);
1319         break;
1320 
1321     case MIRROR_METHOD_ZERO:
1322         assert(!qiov);
1323         ret = blk_co_pwrite_zeroes(job->target, offset, bytes, flags);
1324         break;
1325 
1326     case MIRROR_METHOD_DISCARD:
1327         assert(!qiov);
1328         ret = blk_co_pdiscard(job->target, offset, bytes);
1329         break;
1330 
1331     default:
1332         abort();
1333     }
1334 
1335     job->active_write_bytes_in_flight -= bytes;
1336     if (ret >= 0) {
1337         job_progress_update(&job->common.job, bytes);
1338     } else {
1339         BlockErrorAction action;
1340 
1341         /*
1342          * We failed, so we should mark dirty the whole area, aligned up.
1343          * Note that we don't care about shrunk tails if any: they were dirty
1344          * at function start, and they must be still dirty, as we've locked
1345          * the region for in-flight op.
1346          */
1347         bitmap_offset = QEMU_ALIGN_DOWN(offset, job->granularity);
1348         bitmap_end = QEMU_ALIGN_UP(offset + bytes, job->granularity);
1349         bdrv_set_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1350                               bitmap_end - bitmap_offset);
1351         job->actively_synced = false;
1352 
1353         action = mirror_error_action(job, false, -ret);
1354         if (action == BLOCK_ERROR_ACTION_REPORT) {
1355             if (!job->ret) {
1356                 job->ret = ret;
1357             }
1358         }
1359     }
1360 }
1361 
1362 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s,
1363                                                    uint64_t offset,
1364                                                    uint64_t bytes)
1365 {
1366     MirrorOp *op;
1367     uint64_t start_chunk = offset / s->granularity;
1368     uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
1369 
1370     op = g_new(MirrorOp, 1);
1371     *op = (MirrorOp){
1372         .s                  = s,
1373         .offset             = offset,
1374         .bytes              = bytes,
1375         .is_active_write    = true,
1376         .is_in_flight       = true,
1377         .co                 = qemu_coroutine_self(),
1378     };
1379     qemu_co_queue_init(&op->waiting_requests);
1380     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
1381 
1382     s->in_active_write_counter++;
1383 
1384     /*
1385      * Wait for concurrent requests affecting the area.  If there are already
1386      * running requests that are copying off now-to-be stale data in the area,
1387      * we must wait for them to finish before we begin writing fresh data to the
1388      * target so that the write operations appear in the correct order.
1389      * Note that background requests (see mirror_iteration()) in contrast only
1390      * wait for conflicting requests at the start of the dirty area, and then
1391      * (based on the in_flight_bitmap) truncate the area to copy so it will not
1392      * conflict with any requests beyond that.  For active writes, however, we
1393      * cannot truncate that area.  The request from our parent must be blocked
1394      * until the area is copied in full.  Therefore, we must wait for the whole
1395      * area to become free of concurrent requests.
1396      */
1397     mirror_wait_on_conflicts(op, s, offset, bytes);
1398 
1399     bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1400 
1401     return op;
1402 }
1403 
1404 static void coroutine_fn active_write_settle(MirrorOp *op)
1405 {
1406     uint64_t start_chunk = op->offset / op->s->granularity;
1407     uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes,
1408                                       op->s->granularity);
1409 
1410     if (!--op->s->in_active_write_counter && op->s->actively_synced) {
1411         BdrvChild *source = op->s->mirror_top_bs->backing;
1412 
1413         if (QLIST_FIRST(&source->bs->parents) == source &&
1414             QLIST_NEXT(source, next_parent) == NULL)
1415         {
1416             /* Assert that we are back in sync once all active write
1417              * operations are settled.
1418              * Note that we can only assert this if the mirror node
1419              * is the source node's only parent. */
1420             assert(!bdrv_get_dirty_count(op->s->dirty_bitmap));
1421         }
1422     }
1423     bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1424     QTAILQ_REMOVE(&op->s->ops_in_flight, op, next);
1425     qemu_co_queue_restart_all(&op->waiting_requests);
1426     g_free(op);
1427 }
1428 
1429 static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs,
1430     int64_t offset, int64_t bytes, QEMUIOVector *qiov, BdrvRequestFlags flags)
1431 {
1432     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1433 }
1434 
1435 static int coroutine_fn bdrv_mirror_top_do_write(BlockDriverState *bs,
1436     MirrorMethod method, uint64_t offset, uint64_t bytes, QEMUIOVector *qiov,
1437     int flags)
1438 {
1439     MirrorOp *op = NULL;
1440     MirrorBDSOpaque *s = bs->opaque;
1441     int ret = 0;
1442     bool copy_to_target = false;
1443 
1444     if (s->job) {
1445         copy_to_target = s->job->ret >= 0 &&
1446                          !job_is_cancelled(&s->job->common.job) &&
1447                          s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1448     }
1449 
1450     if (copy_to_target) {
1451         op = active_write_prepare(s->job, offset, bytes);
1452     }
1453 
1454     switch (method) {
1455     case MIRROR_METHOD_COPY:
1456         ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1457         break;
1458 
1459     case MIRROR_METHOD_ZERO:
1460         ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1461         break;
1462 
1463     case MIRROR_METHOD_DISCARD:
1464         ret = bdrv_co_pdiscard(bs->backing, offset, bytes);
1465         break;
1466 
1467     default:
1468         abort();
1469     }
1470 
1471     if (ret < 0) {
1472         goto out;
1473     }
1474 
1475     if (copy_to_target) {
1476         do_sync_target_write(s->job, method, offset, bytes, qiov, flags);
1477     }
1478 
1479 out:
1480     if (copy_to_target) {
1481         active_write_settle(op);
1482     }
1483     return ret;
1484 }
1485 
1486 static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs,
1487     int64_t offset, int64_t bytes, QEMUIOVector *qiov, BdrvRequestFlags flags)
1488 {
1489     MirrorBDSOpaque *s = bs->opaque;
1490     QEMUIOVector bounce_qiov;
1491     void *bounce_buf;
1492     int ret = 0;
1493     bool copy_to_target = false;
1494 
1495     if (s->job) {
1496         copy_to_target = s->job->ret >= 0 &&
1497                          !job_is_cancelled(&s->job->common.job) &&
1498                          s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1499     }
1500 
1501     if (copy_to_target) {
1502         /* The guest might concurrently modify the data to write; but
1503          * the data on source and destination must match, so we have
1504          * to use a bounce buffer if we are going to write to the
1505          * target now. */
1506         bounce_buf = qemu_blockalign(bs, bytes);
1507         iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes);
1508 
1509         qemu_iovec_init(&bounce_qiov, 1);
1510         qemu_iovec_add(&bounce_qiov, bounce_buf, bytes);
1511         qiov = &bounce_qiov;
1512 
1513         flags &= ~BDRV_REQ_REGISTERED_BUF;
1514     }
1515 
1516     ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov,
1517                                    flags);
1518 
1519     if (copy_to_target) {
1520         qemu_iovec_destroy(&bounce_qiov);
1521         qemu_vfree(bounce_buf);
1522     }
1523 
1524     return ret;
1525 }
1526 
1527 static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs)
1528 {
1529     if (bs->backing == NULL) {
1530         /* we can be here after failed bdrv_append in mirror_start_job */
1531         return 0;
1532     }
1533     return bdrv_co_flush(bs->backing->bs);
1534 }
1535 
1536 static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs,
1537     int64_t offset, int64_t bytes, BdrvRequestFlags flags)
1538 {
1539     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL,
1540                                     flags);
1541 }
1542 
1543 static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs,
1544     int64_t offset, int64_t bytes)
1545 {
1546     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes,
1547                                     NULL, 0);
1548 }
1549 
1550 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs)
1551 {
1552     if (bs->backing == NULL) {
1553         /* we can be here after failed bdrv_attach_child in
1554          * bdrv_set_backing_hd */
1555         return;
1556     }
1557     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1558             bs->backing->bs->filename);
1559 }
1560 
1561 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1562                                        BdrvChildRole role,
1563                                        BlockReopenQueue *reopen_queue,
1564                                        uint64_t perm, uint64_t shared,
1565                                        uint64_t *nperm, uint64_t *nshared)
1566 {
1567     MirrorBDSOpaque *s = bs->opaque;
1568 
1569     if (s->stop) {
1570         /*
1571          * If the job is to be stopped, we do not need to forward
1572          * anything to the real image.
1573          */
1574         *nperm = 0;
1575         *nshared = BLK_PERM_ALL;
1576         return;
1577     }
1578 
1579     bdrv_default_perms(bs, c, role, reopen_queue,
1580                        perm, shared, nperm, nshared);
1581 
1582     if (s->is_commit) {
1583         /*
1584          * For commit jobs, we cannot take CONSISTENT_READ, because
1585          * that permission is unshared for everything above the base
1586          * node (except for filters on the base node).
1587          * We also have to force-share the WRITE permission, or
1588          * otherwise we would block ourselves at the base node (if
1589          * writes are blocked for a node, they are also blocked for
1590          * its backing file).
1591          * (We could also share RESIZE, because it may be needed for
1592          * the target if its size is less than the top node's; but
1593          * bdrv_default_perms_for_cow() automatically shares RESIZE
1594          * for backing nodes if WRITE is shared, so there is no need
1595          * to do it here.)
1596          */
1597         *nperm &= ~BLK_PERM_CONSISTENT_READ;
1598         *nshared |= BLK_PERM_WRITE;
1599     }
1600 }
1601 
1602 /* Dummy node that provides consistent read to its users without requiring it
1603  * from its backing file and that allows writes on the backing file chain. */
1604 static BlockDriver bdrv_mirror_top = {
1605     .format_name                = "mirror_top",
1606     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1607     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1608     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1609     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1610     .bdrv_co_flush              = bdrv_mirror_top_flush,
1611     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1612     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1613 
1614     .is_filter                  = true,
1615     .filtered_child_is_backing  = true,
1616 };
1617 
1618 static BlockJob *mirror_start_job(
1619                              const char *job_id, BlockDriverState *bs,
1620                              int creation_flags, BlockDriverState *target,
1621                              const char *replaces, int64_t speed,
1622                              uint32_t granularity, int64_t buf_size,
1623                              BlockMirrorBackingMode backing_mode,
1624                              bool zero_target,
1625                              BlockdevOnError on_source_error,
1626                              BlockdevOnError on_target_error,
1627                              bool unmap,
1628                              BlockCompletionFunc *cb,
1629                              void *opaque,
1630                              const BlockJobDriver *driver,
1631                              bool is_none_mode, BlockDriverState *base,
1632                              bool auto_complete, const char *filter_node_name,
1633                              bool is_mirror, MirrorCopyMode copy_mode,
1634                              Error **errp)
1635 {
1636     MirrorBlockJob *s;
1637     MirrorBDSOpaque *bs_opaque;
1638     BlockDriverState *mirror_top_bs;
1639     bool target_is_backing;
1640     uint64_t target_perms, target_shared_perms;
1641     int ret;
1642 
1643     if (granularity == 0) {
1644         granularity = bdrv_get_default_bitmap_granularity(target);
1645     }
1646 
1647     assert(is_power_of_2(granularity));
1648 
1649     if (buf_size < 0) {
1650         error_setg(errp, "Invalid parameter 'buf-size'");
1651         return NULL;
1652     }
1653 
1654     if (buf_size == 0) {
1655         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1656     }
1657 
1658     if (bdrv_skip_filters(bs) == bdrv_skip_filters(target)) {
1659         error_setg(errp, "Can't mirror node into itself");
1660         return NULL;
1661     }
1662 
1663     target_is_backing = bdrv_chain_contains(bs, target);
1664 
1665     /* In the case of active commit, add dummy driver to provide consistent
1666      * reads on the top, while disabling it in the intermediate nodes, and make
1667      * the backing chain writable. */
1668     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1669                                          BDRV_O_RDWR, errp);
1670     if (mirror_top_bs == NULL) {
1671         return NULL;
1672     }
1673     if (!filter_node_name) {
1674         mirror_top_bs->implicit = true;
1675     }
1676 
1677     /* So that we can always drop this node */
1678     mirror_top_bs->never_freeze = true;
1679 
1680     mirror_top_bs->total_sectors = bs->total_sectors;
1681     mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
1682     mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
1683                                           BDRV_REQ_NO_FALLBACK;
1684     bs_opaque = g_new0(MirrorBDSOpaque, 1);
1685     mirror_top_bs->opaque = bs_opaque;
1686 
1687     bs_opaque->is_commit = target_is_backing;
1688 
1689     bdrv_drained_begin(bs);
1690     ret = bdrv_append(mirror_top_bs, bs, errp);
1691     bdrv_drained_end(bs);
1692 
1693     if (ret < 0) {
1694         bdrv_unref(mirror_top_bs);
1695         return NULL;
1696     }
1697 
1698     /* Make sure that the source is not resized while the job is running */
1699     s = block_job_create(job_id, driver, NULL, mirror_top_bs,
1700                          BLK_PERM_CONSISTENT_READ,
1701                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1702                          BLK_PERM_WRITE, speed,
1703                          creation_flags, cb, opaque, errp);
1704     if (!s) {
1705         goto fail;
1706     }
1707     bs_opaque->job = s;
1708 
1709     /* The block job now has a reference to this node */
1710     bdrv_unref(mirror_top_bs);
1711 
1712     s->mirror_top_bs = mirror_top_bs;
1713 
1714     /* No resize for the target either; while the mirror is still running, a
1715      * consistent read isn't necessarily possible. We could possibly allow
1716      * writes and graph modifications, though it would likely defeat the
1717      * purpose of a mirror, so leave them blocked for now.
1718      *
1719      * In the case of active commit, things look a bit different, though,
1720      * because the target is an already populated backing file in active use.
1721      * We can allow anything except resize there.*/
1722 
1723     target_perms = BLK_PERM_WRITE;
1724     target_shared_perms = BLK_PERM_WRITE_UNCHANGED;
1725 
1726     if (target_is_backing) {
1727         int64_t bs_size, target_size;
1728         bs_size = bdrv_getlength(bs);
1729         if (bs_size < 0) {
1730             error_setg_errno(errp, -bs_size,
1731                              "Could not inquire top image size");
1732             goto fail;
1733         }
1734 
1735         target_size = bdrv_getlength(target);
1736         if (target_size < 0) {
1737             error_setg_errno(errp, -target_size,
1738                              "Could not inquire base image size");
1739             goto fail;
1740         }
1741 
1742         if (target_size < bs_size) {
1743             target_perms |= BLK_PERM_RESIZE;
1744         }
1745 
1746         target_shared_perms |= BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE;
1747     } else if (bdrv_chain_contains(bs, bdrv_skip_filters(target))) {
1748         /*
1749          * We may want to allow this in the future, but it would
1750          * require taking some extra care.
1751          */
1752         error_setg(errp, "Cannot mirror to a filter on top of a node in the "
1753                    "source's backing chain");
1754         goto fail;
1755     }
1756 
1757     s->target = blk_new(s->common.job.aio_context,
1758                         target_perms, target_shared_perms);
1759     ret = blk_insert_bs(s->target, target, errp);
1760     if (ret < 0) {
1761         goto fail;
1762     }
1763     if (is_mirror) {
1764         /* XXX: Mirror target could be a NBD server of target QEMU in the case
1765          * of non-shared block migration. To allow migration completion, we
1766          * have to allow "inactivate" of the target BB.  When that happens, we
1767          * know the job is drained, and the vcpus are stopped, so no write
1768          * operation will be performed. Block layer already has assertions to
1769          * ensure that. */
1770         blk_set_force_allow_inactivate(s->target);
1771     }
1772     blk_set_allow_aio_context_change(s->target, true);
1773     blk_set_disable_request_queuing(s->target, true);
1774 
1775     s->replaces = g_strdup(replaces);
1776     s->on_source_error = on_source_error;
1777     s->on_target_error = on_target_error;
1778     s->is_none_mode = is_none_mode;
1779     s->backing_mode = backing_mode;
1780     s->zero_target = zero_target;
1781     s->copy_mode = copy_mode;
1782     s->base = base;
1783     s->base_overlay = bdrv_find_overlay(bs, base);
1784     s->granularity = granularity;
1785     s->buf_size = ROUND_UP(buf_size, granularity);
1786     s->unmap = unmap;
1787     if (auto_complete) {
1788         s->should_complete = true;
1789     }
1790 
1791     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1792     if (!s->dirty_bitmap) {
1793         goto fail;
1794     }
1795     if (s->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING) {
1796         bdrv_disable_dirty_bitmap(s->dirty_bitmap);
1797     }
1798 
1799     ret = block_job_add_bdrv(&s->common, "source", bs, 0,
1800                              BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE |
1801                              BLK_PERM_CONSISTENT_READ,
1802                              errp);
1803     if (ret < 0) {
1804         goto fail;
1805     }
1806 
1807     /* Required permissions are already taken with blk_new() */
1808     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1809                        &error_abort);
1810 
1811     /* In commit_active_start() all intermediate nodes disappear, so
1812      * any jobs in them must be blocked */
1813     if (target_is_backing) {
1814         BlockDriverState *iter, *filtered_target;
1815         uint64_t iter_shared_perms;
1816 
1817         /*
1818          * The topmost node with
1819          * bdrv_skip_filters(filtered_target) == bdrv_skip_filters(target)
1820          */
1821         filtered_target = bdrv_cow_bs(bdrv_find_overlay(bs, target));
1822 
1823         assert(bdrv_skip_filters(filtered_target) ==
1824                bdrv_skip_filters(target));
1825 
1826         /*
1827          * XXX BLK_PERM_WRITE needs to be allowed so we don't block
1828          * ourselves at s->base (if writes are blocked for a node, they are
1829          * also blocked for its backing file). The other options would be a
1830          * second filter driver above s->base (== target).
1831          */
1832         iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE;
1833 
1834         for (iter = bdrv_filter_or_cow_bs(bs); iter != target;
1835              iter = bdrv_filter_or_cow_bs(iter))
1836         {
1837             if (iter == filtered_target) {
1838                 /*
1839                  * From here on, all nodes are filters on the base.
1840                  * This allows us to share BLK_PERM_CONSISTENT_READ.
1841                  */
1842                 iter_shared_perms |= BLK_PERM_CONSISTENT_READ;
1843             }
1844 
1845             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1846                                      iter_shared_perms, errp);
1847             if (ret < 0) {
1848                 goto fail;
1849             }
1850         }
1851 
1852         if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) {
1853             goto fail;
1854         }
1855     }
1856 
1857     QTAILQ_INIT(&s->ops_in_flight);
1858 
1859     trace_mirror_start(bs, s, opaque);
1860     job_start(&s->common.job);
1861 
1862     return &s->common;
1863 
1864 fail:
1865     if (s) {
1866         /* Make sure this BDS does not go away until we have completed the graph
1867          * changes below */
1868         bdrv_ref(mirror_top_bs);
1869 
1870         g_free(s->replaces);
1871         blk_unref(s->target);
1872         bs_opaque->job = NULL;
1873         if (s->dirty_bitmap) {
1874             bdrv_release_dirty_bitmap(s->dirty_bitmap);
1875         }
1876         job_early_fail(&s->common.job);
1877     }
1878 
1879     bs_opaque->stop = true;
1880     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
1881                              &error_abort);
1882     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
1883 
1884     bdrv_unref(mirror_top_bs);
1885 
1886     return NULL;
1887 }
1888 
1889 void mirror_start(const char *job_id, BlockDriverState *bs,
1890                   BlockDriverState *target, const char *replaces,
1891                   int creation_flags, int64_t speed,
1892                   uint32_t granularity, int64_t buf_size,
1893                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1894                   bool zero_target,
1895                   BlockdevOnError on_source_error,
1896                   BlockdevOnError on_target_error,
1897                   bool unmap, const char *filter_node_name,
1898                   MirrorCopyMode copy_mode, Error **errp)
1899 {
1900     bool is_none_mode;
1901     BlockDriverState *base;
1902 
1903     GLOBAL_STATE_CODE();
1904 
1905     if ((mode == MIRROR_SYNC_MODE_INCREMENTAL) ||
1906         (mode == MIRROR_SYNC_MODE_BITMAP)) {
1907         error_setg(errp, "Sync mode '%s' not supported",
1908                    MirrorSyncMode_str(mode));
1909         return;
1910     }
1911     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1912     base = mode == MIRROR_SYNC_MODE_TOP ? bdrv_backing_chain_next(bs) : NULL;
1913     mirror_start_job(job_id, bs, creation_flags, target, replaces,
1914                      speed, granularity, buf_size, backing_mode, zero_target,
1915                      on_source_error, on_target_error, unmap, NULL, NULL,
1916                      &mirror_job_driver, is_none_mode, base, false,
1917                      filter_node_name, true, copy_mode, errp);
1918 }
1919 
1920 BlockJob *commit_active_start(const char *job_id, BlockDriverState *bs,
1921                               BlockDriverState *base, int creation_flags,
1922                               int64_t speed, BlockdevOnError on_error,
1923                               const char *filter_node_name,
1924                               BlockCompletionFunc *cb, void *opaque,
1925                               bool auto_complete, Error **errp)
1926 {
1927     bool base_read_only;
1928     BlockJob *job;
1929 
1930     GLOBAL_STATE_CODE();
1931 
1932     base_read_only = bdrv_is_read_only(base);
1933 
1934     if (base_read_only) {
1935         if (bdrv_reopen_set_read_only(base, false, errp) < 0) {
1936             return NULL;
1937         }
1938     }
1939 
1940     job = mirror_start_job(
1941                      job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1942                      MIRROR_LEAVE_BACKING_CHAIN, false,
1943                      on_error, on_error, true, cb, opaque,
1944                      &commit_active_job_driver, false, base, auto_complete,
1945                      filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND,
1946                      errp);
1947     if (!job) {
1948         goto error_restore_flags;
1949     }
1950 
1951     return job;
1952 
1953 error_restore_flags:
1954     /* ignore error and errp for bdrv_reopen, because we want to propagate
1955      * the original error */
1956     if (base_read_only) {
1957         bdrv_reopen_set_read_only(base, true, NULL);
1958     }
1959     return NULL;
1960 }
1961