xref: /openbmc/qemu/block/mirror.c (revision 36ebc7db)
1 /*
2  * Image mirroring
3  *
4  * Copyright Red Hat, Inc. 2012
5  *
6  * Authors:
7  *  Paolo Bonzini  <pbonzini@redhat.com>
8  *
9  * This work is licensed under the terms of the GNU LGPL, version 2 or later.
10  * See the COPYING.LIB file in the top-level directory.
11  *
12  */
13 
14 #include "qemu/osdep.h"
15 #include "qemu/cutils.h"
16 #include "qemu/coroutine.h"
17 #include "qemu/range.h"
18 #include "trace.h"
19 #include "block/blockjob_int.h"
20 #include "block/block_int.h"
21 #include "block/dirty-bitmap.h"
22 #include "sysemu/block-backend.h"
23 #include "qapi/error.h"
24 #include "qemu/ratelimit.h"
25 #include "qemu/bitmap.h"
26 #include "qemu/memalign.h"
27 
28 #define MAX_IN_FLIGHT 16
29 #define MAX_IO_BYTES (1 << 20) /* 1 Mb */
30 #define DEFAULT_MIRROR_BUF_SIZE (MAX_IN_FLIGHT * MAX_IO_BYTES)
31 
32 /* The mirroring buffer is a list of granularity-sized chunks.
33  * Free chunks are organized in a list.
34  */
35 typedef struct MirrorBuffer {
36     QSIMPLEQ_ENTRY(MirrorBuffer) next;
37 } MirrorBuffer;
38 
39 typedef struct MirrorOp MirrorOp;
40 
41 typedef struct MirrorBlockJob {
42     BlockJob common;
43     BlockBackend *target;
44     BlockDriverState *mirror_top_bs;
45     BlockDriverState *base;
46     BlockDriverState *base_overlay;
47 
48     /* The name of the graph node to replace */
49     char *replaces;
50     /* The BDS to replace */
51     BlockDriverState *to_replace;
52     /* Used to block operations on the drive-mirror-replace target */
53     Error *replace_blocker;
54     bool is_none_mode;
55     BlockMirrorBackingMode backing_mode;
56     /* Whether the target image requires explicit zero-initialization */
57     bool zero_target;
58     MirrorCopyMode copy_mode;
59     BlockdevOnError on_source_error, on_target_error;
60     /* Set when the target is synced (dirty bitmap is clean, nothing
61      * in flight) and the job is running in active mode */
62     bool actively_synced;
63     bool should_complete;
64     int64_t granularity;
65     size_t buf_size;
66     int64_t bdev_length;
67     unsigned long *cow_bitmap;
68     BdrvDirtyBitmap *dirty_bitmap;
69     BdrvDirtyBitmapIter *dbi;
70     uint8_t *buf;
71     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
72     int buf_free_count;
73 
74     uint64_t last_pause_ns;
75     unsigned long *in_flight_bitmap;
76     unsigned in_flight;
77     int64_t bytes_in_flight;
78     QTAILQ_HEAD(, MirrorOp) ops_in_flight;
79     int ret;
80     bool unmap;
81     int target_cluster_size;
82     int max_iov;
83     bool initial_zeroing_ongoing;
84     int in_active_write_counter;
85     int64_t active_write_bytes_in_flight;
86     bool prepared;
87     bool in_drain;
88 } MirrorBlockJob;
89 
90 typedef struct MirrorBDSOpaque {
91     MirrorBlockJob *job;
92     bool stop;
93     bool is_commit;
94 } MirrorBDSOpaque;
95 
96 struct MirrorOp {
97     MirrorBlockJob *s;
98     QEMUIOVector qiov;
99     int64_t offset;
100     uint64_t bytes;
101 
102     /* The pointee is set by mirror_co_read(), mirror_co_zero(), and
103      * mirror_co_discard() before yielding for the first time */
104     int64_t *bytes_handled;
105 
106     bool is_pseudo_op;
107     bool is_active_write;
108     bool is_in_flight;
109     CoQueue waiting_requests;
110     Coroutine *co;
111     MirrorOp *waiting_for_op;
112 
113     QTAILQ_ENTRY(MirrorOp) next;
114 };
115 
116 typedef enum MirrorMethod {
117     MIRROR_METHOD_COPY,
118     MIRROR_METHOD_ZERO,
119     MIRROR_METHOD_DISCARD,
120 } MirrorMethod;
121 
122 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
123                                             int error)
124 {
125     s->actively_synced = false;
126     if (read) {
127         return block_job_error_action(&s->common, s->on_source_error,
128                                       true, error);
129     } else {
130         return block_job_error_action(&s->common, s->on_target_error,
131                                       false, error);
132     }
133 }
134 
135 static void coroutine_fn mirror_wait_on_conflicts(MirrorOp *self,
136                                                   MirrorBlockJob *s,
137                                                   uint64_t offset,
138                                                   uint64_t bytes)
139 {
140     uint64_t self_start_chunk = offset / s->granularity;
141     uint64_t self_end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
142     uint64_t self_nb_chunks = self_end_chunk - self_start_chunk;
143 
144     while (find_next_bit(s->in_flight_bitmap, self_end_chunk,
145                          self_start_chunk) < self_end_chunk &&
146            s->ret >= 0)
147     {
148         MirrorOp *op;
149 
150         QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
151             uint64_t op_start_chunk = op->offset / s->granularity;
152             uint64_t op_nb_chunks = DIV_ROUND_UP(op->offset + op->bytes,
153                                                  s->granularity) -
154                                     op_start_chunk;
155 
156             if (op == self) {
157                 continue;
158             }
159 
160             if (ranges_overlap(self_start_chunk, self_nb_chunks,
161                                op_start_chunk, op_nb_chunks))
162             {
163                 if (self) {
164                     /*
165                      * If the operation is already (indirectly) waiting for us,
166                      * or will wait for us as soon as it wakes up, then just go
167                      * on (instead of producing a deadlock in the former case).
168                      */
169                     if (op->waiting_for_op) {
170                         continue;
171                     }
172 
173                     self->waiting_for_op = op;
174                 }
175 
176                 qemu_co_queue_wait(&op->waiting_requests, NULL);
177 
178                 if (self) {
179                     self->waiting_for_op = NULL;
180                 }
181 
182                 break;
183             }
184         }
185     }
186 }
187 
188 static void coroutine_fn mirror_iteration_done(MirrorOp *op, int ret)
189 {
190     MirrorBlockJob *s = op->s;
191     struct iovec *iov;
192     int64_t chunk_num;
193     int i, nb_chunks;
194 
195     trace_mirror_iteration_done(s, op->offset, op->bytes, ret);
196 
197     s->in_flight--;
198     s->bytes_in_flight -= op->bytes;
199     iov = op->qiov.iov;
200     for (i = 0; i < op->qiov.niov; i++) {
201         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
202         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
203         s->buf_free_count++;
204     }
205 
206     chunk_num = op->offset / s->granularity;
207     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
208 
209     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
210     QTAILQ_REMOVE(&s->ops_in_flight, op, next);
211     if (ret >= 0) {
212         if (s->cow_bitmap) {
213             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
214         }
215         if (!s->initial_zeroing_ongoing) {
216             job_progress_update(&s->common.job, op->bytes);
217         }
218     }
219     qemu_iovec_destroy(&op->qiov);
220 
221     qemu_co_queue_restart_all(&op->waiting_requests);
222     g_free(op);
223 }
224 
225 static void coroutine_fn mirror_write_complete(MirrorOp *op, int ret)
226 {
227     MirrorBlockJob *s = op->s;
228 
229     if (ret < 0) {
230         BlockErrorAction action;
231 
232         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
233         action = mirror_error_action(s, false, -ret);
234         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
235             s->ret = ret;
236         }
237     }
238 
239     mirror_iteration_done(op, ret);
240 }
241 
242 static void coroutine_fn mirror_read_complete(MirrorOp *op, int ret)
243 {
244     MirrorBlockJob *s = op->s;
245 
246     if (ret < 0) {
247         BlockErrorAction action;
248 
249         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
250         action = mirror_error_action(s, true, -ret);
251         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
252             s->ret = ret;
253         }
254 
255         mirror_iteration_done(op, ret);
256         return;
257     }
258 
259     ret = blk_co_pwritev(s->target, op->offset, op->qiov.size, &op->qiov, 0);
260     mirror_write_complete(op, ret);
261 }
262 
263 /* Clip bytes relative to offset to not exceed end-of-file */
264 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s,
265                                         int64_t offset,
266                                         int64_t bytes)
267 {
268     return MIN(bytes, s->bdev_length - offset);
269 }
270 
271 /* Round offset and/or bytes to target cluster if COW is needed, and
272  * return the offset of the adjusted tail against original. */
273 static int mirror_cow_align(MirrorBlockJob *s, int64_t *offset,
274                             uint64_t *bytes)
275 {
276     bool need_cow;
277     int ret = 0;
278     int64_t align_offset = *offset;
279     int64_t align_bytes = *bytes;
280     int max_bytes = s->granularity * s->max_iov;
281 
282     need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap);
283     need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity,
284                           s->cow_bitmap);
285     if (need_cow) {
286         bdrv_round_to_clusters(blk_bs(s->target), *offset, *bytes,
287                                &align_offset, &align_bytes);
288     }
289 
290     if (align_bytes > max_bytes) {
291         align_bytes = max_bytes;
292         if (need_cow) {
293             align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size);
294         }
295     }
296     /* Clipping may result in align_bytes unaligned to chunk boundary, but
297      * that doesn't matter because it's already the end of source image. */
298     align_bytes = mirror_clip_bytes(s, align_offset, align_bytes);
299 
300     ret = align_offset + align_bytes - (*offset + *bytes);
301     *offset = align_offset;
302     *bytes = align_bytes;
303     assert(ret >= 0);
304     return ret;
305 }
306 
307 static inline void coroutine_fn
308 mirror_wait_for_free_in_flight_slot(MirrorBlockJob *s)
309 {
310     MirrorOp *op;
311 
312     QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
313         /*
314          * Do not wait on pseudo ops, because it may in turn wait on
315          * some other operation to start, which may in fact be the
316          * caller of this function.  Since there is only one pseudo op
317          * at any given time, we will always find some real operation
318          * to wait on.
319          * Also, do not wait on active operations, because they do not
320          * use up in-flight slots.
321          */
322         if (!op->is_pseudo_op && op->is_in_flight && !op->is_active_write) {
323             qemu_co_queue_wait(&op->waiting_requests, NULL);
324             return;
325         }
326     }
327     abort();
328 }
329 
330 /* Perform a mirror copy operation.
331  *
332  * *op->bytes_handled is set to the number of bytes copied after and
333  * including offset, excluding any bytes copied prior to offset due
334  * to alignment.  This will be op->bytes if no alignment is necessary,
335  * or (new_end - op->offset) if the tail is rounded up or down due to
336  * alignment or buffer limit.
337  */
338 static void coroutine_fn mirror_co_read(void *opaque)
339 {
340     MirrorOp *op = opaque;
341     MirrorBlockJob *s = op->s;
342     int nb_chunks;
343     uint64_t ret;
344     uint64_t max_bytes;
345 
346     max_bytes = s->granularity * s->max_iov;
347 
348     /* We can only handle as much as buf_size at a time. */
349     op->bytes = MIN(s->buf_size, MIN(max_bytes, op->bytes));
350     assert(op->bytes);
351     assert(op->bytes < BDRV_REQUEST_MAX_BYTES);
352     *op->bytes_handled = op->bytes;
353 
354     if (s->cow_bitmap) {
355         *op->bytes_handled += mirror_cow_align(s, &op->offset, &op->bytes);
356     }
357     /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */
358     assert(*op->bytes_handled <= UINT_MAX);
359     assert(op->bytes <= s->buf_size);
360     /* The offset is granularity-aligned because:
361      * 1) Caller passes in aligned values;
362      * 2) mirror_cow_align is used only when target cluster is larger. */
363     assert(QEMU_IS_ALIGNED(op->offset, s->granularity));
364     /* The range is sector-aligned, since bdrv_getlength() rounds up. */
365     assert(QEMU_IS_ALIGNED(op->bytes, BDRV_SECTOR_SIZE));
366     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
367 
368     while (s->buf_free_count < nb_chunks) {
369         trace_mirror_yield_in_flight(s, op->offset, s->in_flight);
370         mirror_wait_for_free_in_flight_slot(s);
371     }
372 
373     /* Now make a QEMUIOVector taking enough granularity-sized chunks
374      * from s->buf_free.
375      */
376     qemu_iovec_init(&op->qiov, nb_chunks);
377     while (nb_chunks-- > 0) {
378         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
379         size_t remaining = op->bytes - op->qiov.size;
380 
381         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
382         s->buf_free_count--;
383         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
384     }
385 
386     /* Copy the dirty cluster.  */
387     s->in_flight++;
388     s->bytes_in_flight += op->bytes;
389     op->is_in_flight = true;
390     trace_mirror_one_iteration(s, op->offset, op->bytes);
391 
392     ret = bdrv_co_preadv(s->mirror_top_bs->backing, op->offset, op->bytes,
393                          &op->qiov, 0);
394     mirror_read_complete(op, ret);
395 }
396 
397 static void coroutine_fn mirror_co_zero(void *opaque)
398 {
399     MirrorOp *op = opaque;
400     int ret;
401 
402     op->s->in_flight++;
403     op->s->bytes_in_flight += op->bytes;
404     *op->bytes_handled = op->bytes;
405     op->is_in_flight = true;
406 
407     ret = blk_co_pwrite_zeroes(op->s->target, op->offset, op->bytes,
408                                op->s->unmap ? BDRV_REQ_MAY_UNMAP : 0);
409     mirror_write_complete(op, ret);
410 }
411 
412 static void coroutine_fn mirror_co_discard(void *opaque)
413 {
414     MirrorOp *op = opaque;
415     int ret;
416 
417     op->s->in_flight++;
418     op->s->bytes_in_flight += op->bytes;
419     *op->bytes_handled = op->bytes;
420     op->is_in_flight = true;
421 
422     ret = blk_co_pdiscard(op->s->target, op->offset, op->bytes);
423     mirror_write_complete(op, ret);
424 }
425 
426 static unsigned mirror_perform(MirrorBlockJob *s, int64_t offset,
427                                unsigned bytes, MirrorMethod mirror_method)
428 {
429     MirrorOp *op;
430     Coroutine *co;
431     int64_t bytes_handled = -1;
432 
433     op = g_new(MirrorOp, 1);
434     *op = (MirrorOp){
435         .s              = s,
436         .offset         = offset,
437         .bytes          = bytes,
438         .bytes_handled  = &bytes_handled,
439     };
440     qemu_co_queue_init(&op->waiting_requests);
441 
442     switch (mirror_method) {
443     case MIRROR_METHOD_COPY:
444         co = qemu_coroutine_create(mirror_co_read, op);
445         break;
446     case MIRROR_METHOD_ZERO:
447         co = qemu_coroutine_create(mirror_co_zero, op);
448         break;
449     case MIRROR_METHOD_DISCARD:
450         co = qemu_coroutine_create(mirror_co_discard, op);
451         break;
452     default:
453         abort();
454     }
455     op->co = co;
456 
457     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
458     qemu_coroutine_enter(co);
459     /* At this point, ownership of op has been moved to the coroutine
460      * and the object may already be freed */
461 
462     /* Assert that this value has been set */
463     assert(bytes_handled >= 0);
464 
465     /* Same assertion as in mirror_co_read() (and for mirror_co_read()
466      * and mirror_co_discard(), bytes_handled == op->bytes, which
467      * is the @bytes parameter given to this function) */
468     assert(bytes_handled <= UINT_MAX);
469     return bytes_handled;
470 }
471 
472 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
473 {
474     BlockDriverState *source = s->mirror_top_bs->backing->bs;
475     MirrorOp *pseudo_op;
476     int64_t offset;
477     uint64_t delay_ns = 0, ret = 0;
478     /* At least the first dirty chunk is mirrored in one iteration. */
479     int nb_chunks = 1;
480     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
481     int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES);
482 
483     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
484     offset = bdrv_dirty_iter_next(s->dbi);
485     if (offset < 0) {
486         bdrv_set_dirty_iter(s->dbi, 0);
487         offset = bdrv_dirty_iter_next(s->dbi);
488         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
489         assert(offset >= 0);
490     }
491     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
492 
493     /*
494      * Wait for concurrent requests to @offset.  The next loop will limit the
495      * copied area based on in_flight_bitmap so we only copy an area that does
496      * not overlap with concurrent in-flight requests.  Still, we would like to
497      * copy something, so wait until there are at least no more requests to the
498      * very beginning of the area.
499      */
500     mirror_wait_on_conflicts(NULL, s, offset, 1);
501 
502     job_pause_point(&s->common.job);
503 
504     /* Find the number of consective dirty chunks following the first dirty
505      * one, and wait for in flight requests in them. */
506     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
507     while (nb_chunks * s->granularity < s->buf_size) {
508         int64_t next_dirty;
509         int64_t next_offset = offset + nb_chunks * s->granularity;
510         int64_t next_chunk = next_offset / s->granularity;
511         if (next_offset >= s->bdev_length ||
512             !bdrv_dirty_bitmap_get_locked(s->dirty_bitmap, next_offset)) {
513             break;
514         }
515         if (test_bit(next_chunk, s->in_flight_bitmap)) {
516             break;
517         }
518 
519         next_dirty = bdrv_dirty_iter_next(s->dbi);
520         if (next_dirty > next_offset || next_dirty < 0) {
521             /* The bitmap iterator's cache is stale, refresh it */
522             bdrv_set_dirty_iter(s->dbi, next_offset);
523             next_dirty = bdrv_dirty_iter_next(s->dbi);
524         }
525         assert(next_dirty == next_offset);
526         nb_chunks++;
527     }
528 
529     /* Clear dirty bits before querying the block status, because
530      * calling bdrv_block_status_above could yield - if some blocks are
531      * marked dirty in this window, we need to know.
532      */
533     bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset,
534                                    nb_chunks * s->granularity);
535     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
536 
537     /* Before claiming an area in the in-flight bitmap, we have to
538      * create a MirrorOp for it so that conflicting requests can wait
539      * for it.  mirror_perform() will create the real MirrorOps later,
540      * for now we just create a pseudo operation that will wake up all
541      * conflicting requests once all real operations have been
542      * launched. */
543     pseudo_op = g_new(MirrorOp, 1);
544     *pseudo_op = (MirrorOp){
545         .offset         = offset,
546         .bytes          = nb_chunks * s->granularity,
547         .is_pseudo_op   = true,
548     };
549     qemu_co_queue_init(&pseudo_op->waiting_requests);
550     QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next);
551 
552     bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks);
553     while (nb_chunks > 0 && offset < s->bdev_length) {
554         int ret;
555         int64_t io_bytes;
556         int64_t io_bytes_acct;
557         MirrorMethod mirror_method = MIRROR_METHOD_COPY;
558 
559         assert(!(offset % s->granularity));
560         ret = bdrv_block_status_above(source, NULL, offset,
561                                       nb_chunks * s->granularity,
562                                       &io_bytes, NULL, NULL);
563         if (ret < 0) {
564             io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes);
565         } else if (ret & BDRV_BLOCK_DATA) {
566             io_bytes = MIN(io_bytes, max_io_bytes);
567         }
568 
569         io_bytes -= io_bytes % s->granularity;
570         if (io_bytes < s->granularity) {
571             io_bytes = s->granularity;
572         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
573             int64_t target_offset;
574             int64_t target_bytes;
575             bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes,
576                                    &target_offset, &target_bytes);
577             if (target_offset == offset &&
578                 target_bytes == io_bytes) {
579                 mirror_method = ret & BDRV_BLOCK_ZERO ?
580                                     MIRROR_METHOD_ZERO :
581                                     MIRROR_METHOD_DISCARD;
582             }
583         }
584 
585         while (s->in_flight >= MAX_IN_FLIGHT) {
586             trace_mirror_yield_in_flight(s, offset, s->in_flight);
587             mirror_wait_for_free_in_flight_slot(s);
588         }
589 
590         if (s->ret < 0) {
591             ret = 0;
592             goto fail;
593         }
594 
595         io_bytes = mirror_clip_bytes(s, offset, io_bytes);
596         io_bytes = mirror_perform(s, offset, io_bytes, mirror_method);
597         if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) {
598             io_bytes_acct = 0;
599         } else {
600             io_bytes_acct = io_bytes;
601         }
602         assert(io_bytes);
603         offset += io_bytes;
604         nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity);
605         delay_ns = block_job_ratelimit_get_delay(&s->common, io_bytes_acct);
606     }
607 
608     ret = delay_ns;
609 fail:
610     QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next);
611     qemu_co_queue_restart_all(&pseudo_op->waiting_requests);
612     g_free(pseudo_op);
613 
614     return ret;
615 }
616 
617 static void mirror_free_init(MirrorBlockJob *s)
618 {
619     int granularity = s->granularity;
620     size_t buf_size = s->buf_size;
621     uint8_t *buf = s->buf;
622 
623     assert(s->buf_free_count == 0);
624     QSIMPLEQ_INIT(&s->buf_free);
625     while (buf_size != 0) {
626         MirrorBuffer *cur = (MirrorBuffer *)buf;
627         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
628         s->buf_free_count++;
629         buf_size -= granularity;
630         buf += granularity;
631     }
632 }
633 
634 /* This is also used for the .pause callback. There is no matching
635  * mirror_resume() because mirror_run() will begin iterating again
636  * when the job is resumed.
637  */
638 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s)
639 {
640     while (s->in_flight > 0) {
641         mirror_wait_for_free_in_flight_slot(s);
642     }
643 }
644 
645 /**
646  * mirror_exit_common: handle both abort() and prepare() cases.
647  * for .prepare, returns 0 on success and -errno on failure.
648  * for .abort cases, denoted by abort = true, MUST return 0.
649  */
650 static int mirror_exit_common(Job *job)
651 {
652     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
653     BlockJob *bjob = &s->common;
654     MirrorBDSOpaque *bs_opaque;
655     AioContext *replace_aio_context = NULL;
656     BlockDriverState *src;
657     BlockDriverState *target_bs;
658     BlockDriverState *mirror_top_bs;
659     Error *local_err = NULL;
660     bool abort = job->ret < 0;
661     int ret = 0;
662 
663     if (s->prepared) {
664         return 0;
665     }
666     s->prepared = true;
667 
668     mirror_top_bs = s->mirror_top_bs;
669     bs_opaque = mirror_top_bs->opaque;
670     src = mirror_top_bs->backing->bs;
671     target_bs = blk_bs(s->target);
672 
673     if (bdrv_chain_contains(src, target_bs)) {
674         bdrv_unfreeze_backing_chain(mirror_top_bs, target_bs);
675     }
676 
677     bdrv_release_dirty_bitmap(s->dirty_bitmap);
678 
679     /* Make sure that the source BDS doesn't go away during bdrv_replace_node,
680      * before we can call bdrv_drained_end */
681     bdrv_ref(src);
682     bdrv_ref(mirror_top_bs);
683     bdrv_ref(target_bs);
684 
685     /*
686      * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
687      * inserting target_bs at s->to_replace, where we might not be able to get
688      * these permissions.
689      */
690     blk_unref(s->target);
691     s->target = NULL;
692 
693     /* We don't access the source any more. Dropping any WRITE/RESIZE is
694      * required before it could become a backing file of target_bs. Not having
695      * these permissions any more means that we can't allow any new requests on
696      * mirror_top_bs from now on, so keep it drained. */
697     bdrv_drained_begin(mirror_top_bs);
698     bs_opaque->stop = true;
699     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
700                              &error_abort);
701     if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
702         BlockDriverState *backing = s->is_none_mode ? src : s->base;
703         BlockDriverState *unfiltered_target = bdrv_skip_filters(target_bs);
704 
705         if (bdrv_cow_bs(unfiltered_target) != backing) {
706             bdrv_set_backing_hd(unfiltered_target, backing, &local_err);
707             if (local_err) {
708                 error_report_err(local_err);
709                 local_err = NULL;
710                 ret = -EPERM;
711             }
712         }
713     } else if (!abort && s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
714         assert(!bdrv_backing_chain_next(target_bs));
715         ret = bdrv_open_backing_file(bdrv_skip_filters(target_bs), NULL,
716                                      "backing", &local_err);
717         if (ret < 0) {
718             error_report_err(local_err);
719             local_err = NULL;
720         }
721     }
722 
723     if (s->to_replace) {
724         replace_aio_context = bdrv_get_aio_context(s->to_replace);
725         aio_context_acquire(replace_aio_context);
726     }
727 
728     if (s->should_complete && !abort) {
729         BlockDriverState *to_replace = s->to_replace ?: src;
730         bool ro = bdrv_is_read_only(to_replace);
731 
732         if (ro != bdrv_is_read_only(target_bs)) {
733             bdrv_reopen_set_read_only(target_bs, ro, NULL);
734         }
735 
736         /* The mirror job has no requests in flight any more, but we need to
737          * drain potential other users of the BDS before changing the graph. */
738         assert(s->in_drain);
739         bdrv_drained_begin(target_bs);
740         /*
741          * Cannot use check_to_replace_node() here, because that would
742          * check for an op blocker on @to_replace, and we have our own
743          * there.
744          */
745         if (bdrv_recurse_can_replace(src, to_replace)) {
746             bdrv_replace_node(to_replace, target_bs, &local_err);
747         } else {
748             error_setg(&local_err, "Can no longer replace '%s' by '%s', "
749                        "because it can no longer be guaranteed that doing so "
750                        "would not lead to an abrupt change of visible data",
751                        to_replace->node_name, target_bs->node_name);
752         }
753         bdrv_drained_end(target_bs);
754         if (local_err) {
755             error_report_err(local_err);
756             ret = -EPERM;
757         }
758     }
759     if (s->to_replace) {
760         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
761         error_free(s->replace_blocker);
762         bdrv_unref(s->to_replace);
763     }
764     if (replace_aio_context) {
765         aio_context_release(replace_aio_context);
766     }
767     g_free(s->replaces);
768     bdrv_unref(target_bs);
769 
770     /*
771      * Remove the mirror filter driver from the graph. Before this, get rid of
772      * the blockers on the intermediate nodes so that the resulting state is
773      * valid.
774      */
775     block_job_remove_all_bdrv(bjob);
776     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
777 
778     bs_opaque->job = NULL;
779 
780     bdrv_drained_end(src);
781     bdrv_drained_end(mirror_top_bs);
782     s->in_drain = false;
783     bdrv_unref(mirror_top_bs);
784     bdrv_unref(src);
785 
786     return ret;
787 }
788 
789 static int mirror_prepare(Job *job)
790 {
791     return mirror_exit_common(job);
792 }
793 
794 static void mirror_abort(Job *job)
795 {
796     int ret = mirror_exit_common(job);
797     assert(ret == 0);
798 }
799 
800 static void coroutine_fn mirror_throttle(MirrorBlockJob *s)
801 {
802     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
803 
804     if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) {
805         s->last_pause_ns = now;
806         job_sleep_ns(&s->common.job, 0);
807     } else {
808         job_pause_point(&s->common.job);
809     }
810 }
811 
812 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
813 {
814     int64_t offset;
815     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
816     BlockDriverState *target_bs = blk_bs(s->target);
817     int ret;
818     int64_t count;
819 
820     if (s->zero_target) {
821         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
822             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
823             return 0;
824         }
825 
826         s->initial_zeroing_ongoing = true;
827         for (offset = 0; offset < s->bdev_length; ) {
828             int bytes = MIN(s->bdev_length - offset,
829                             QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
830 
831             mirror_throttle(s);
832 
833             if (job_is_cancelled(&s->common.job)) {
834                 s->initial_zeroing_ongoing = false;
835                 return 0;
836             }
837 
838             if (s->in_flight >= MAX_IN_FLIGHT) {
839                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
840                                    s->in_flight);
841                 mirror_wait_for_free_in_flight_slot(s);
842                 continue;
843             }
844 
845             mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO);
846             offset += bytes;
847         }
848 
849         mirror_wait_for_all_io(s);
850         s->initial_zeroing_ongoing = false;
851     }
852 
853     /* First part, loop on the sectors and initialize the dirty bitmap.  */
854     for (offset = 0; offset < s->bdev_length; ) {
855         /* Just to make sure we are not exceeding int limit. */
856         int bytes = MIN(s->bdev_length - offset,
857                         QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
858 
859         mirror_throttle(s);
860 
861         if (job_is_cancelled(&s->common.job)) {
862             return 0;
863         }
864 
865         ret = bdrv_is_allocated_above(bs, s->base_overlay, true, offset, bytes,
866                                       &count);
867         if (ret < 0) {
868             return ret;
869         }
870 
871         assert(count);
872         if (ret > 0) {
873             bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
874         }
875         offset += count;
876     }
877     return 0;
878 }
879 
880 /* Called when going out of the streaming phase to flush the bulk of the
881  * data to the medium, or just before completing.
882  */
883 static int mirror_flush(MirrorBlockJob *s)
884 {
885     int ret = blk_flush(s->target);
886     if (ret < 0) {
887         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
888             s->ret = ret;
889         }
890     }
891     return ret;
892 }
893 
894 static int coroutine_fn mirror_run(Job *job, Error **errp)
895 {
896     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
897     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
898     BlockDriverState *target_bs = blk_bs(s->target);
899     bool need_drain = true;
900     BlockDeviceIoStatus iostatus;
901     int64_t length;
902     int64_t target_length;
903     BlockDriverInfo bdi;
904     char backing_filename[2]; /* we only need 2 characters because we are only
905                                  checking for a NULL string */
906     int ret = 0;
907 
908     if (job_is_cancelled(&s->common.job)) {
909         goto immediate_exit;
910     }
911 
912     s->bdev_length = bdrv_co_getlength(bs);
913     if (s->bdev_length < 0) {
914         ret = s->bdev_length;
915         goto immediate_exit;
916     }
917 
918     target_length = blk_co_getlength(s->target);
919     if (target_length < 0) {
920         ret = target_length;
921         goto immediate_exit;
922     }
923 
924     /* Active commit must resize the base image if its size differs from the
925      * active layer. */
926     if (s->base == blk_bs(s->target)) {
927         if (s->bdev_length > target_length) {
928             ret = blk_co_truncate(s->target, s->bdev_length, false,
929                                   PREALLOC_MODE_OFF, 0, NULL);
930             if (ret < 0) {
931                 goto immediate_exit;
932             }
933         }
934     } else if (s->bdev_length != target_length) {
935         error_setg(errp, "Source and target image have different sizes");
936         ret = -EINVAL;
937         goto immediate_exit;
938     }
939 
940     if (s->bdev_length == 0) {
941         /* Transition to the READY state and wait for complete. */
942         job_transition_to_ready(&s->common.job);
943         s->actively_synced = true;
944         while (!job_cancel_requested(&s->common.job) && !s->should_complete) {
945             job_yield(&s->common.job);
946         }
947         goto immediate_exit;
948     }
949 
950     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
951     s->in_flight_bitmap = bitmap_new(length);
952 
953     /* If we have no backing file yet in the destination, we cannot let
954      * the destination do COW.  Instead, we copy sectors around the
955      * dirty data if needed.  We need a bitmap to do that.
956      */
957     bdrv_get_backing_filename(target_bs, backing_filename,
958                               sizeof(backing_filename));
959     if (!bdrv_co_get_info(target_bs, &bdi) && bdi.cluster_size) {
960         s->target_cluster_size = bdi.cluster_size;
961     } else {
962         s->target_cluster_size = BDRV_SECTOR_SIZE;
963     }
964     if (backing_filename[0] && !bdrv_backing_chain_next(target_bs) &&
965         s->granularity < s->target_cluster_size) {
966         s->buf_size = MAX(s->buf_size, s->target_cluster_size);
967         s->cow_bitmap = bitmap_new(length);
968     }
969     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
970 
971     s->buf = qemu_try_blockalign(bs, s->buf_size);
972     if (s->buf == NULL) {
973         ret = -ENOMEM;
974         goto immediate_exit;
975     }
976 
977     mirror_free_init(s);
978 
979     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
980     if (!s->is_none_mode) {
981         ret = mirror_dirty_init(s);
982         if (ret < 0 || job_is_cancelled(&s->common.job)) {
983             goto immediate_exit;
984         }
985     }
986 
987     assert(!s->dbi);
988     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
989     for (;;) {
990         uint64_t delay_ns = 0;
991         int64_t cnt, delta;
992         bool should_complete;
993 
994         if (s->ret < 0) {
995             ret = s->ret;
996             goto immediate_exit;
997         }
998 
999         job_pause_point(&s->common.job);
1000 
1001         if (job_is_cancelled(&s->common.job)) {
1002             ret = 0;
1003             goto immediate_exit;
1004         }
1005 
1006         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1007         /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
1008          * the number of bytes currently being processed; together those are
1009          * the current remaining operation length */
1010         job_progress_set_remaining(&s->common.job,
1011                                    s->bytes_in_flight + cnt +
1012                                    s->active_write_bytes_in_flight);
1013 
1014         /* Note that even when no rate limit is applied we need to yield
1015          * periodically with no pending I/O so that bdrv_drain_all() returns.
1016          * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
1017          * an error, or when the source is clean, whichever comes first. */
1018         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
1019         WITH_JOB_LOCK_GUARD() {
1020             iostatus = s->common.iostatus;
1021         }
1022         if (delta < BLOCK_JOB_SLICE_TIME &&
1023             iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
1024             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
1025                 (cnt == 0 && s->in_flight > 0)) {
1026                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
1027                 mirror_wait_for_free_in_flight_slot(s);
1028                 continue;
1029             } else if (cnt != 0) {
1030                 delay_ns = mirror_iteration(s);
1031             }
1032         }
1033 
1034         should_complete = false;
1035         if (s->in_flight == 0 && cnt == 0) {
1036             trace_mirror_before_flush(s);
1037             if (!job_is_ready(&s->common.job)) {
1038                 if (mirror_flush(s) < 0) {
1039                     /* Go check s->ret.  */
1040                     continue;
1041                 }
1042                 /* We're out of the streaming phase.  From now on, if the job
1043                  * is cancelled we will actually complete all pending I/O and
1044                  * report completion.  This way, block-job-cancel will leave
1045                  * the target in a consistent state.
1046                  */
1047                 job_transition_to_ready(&s->common.job);
1048                 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) {
1049                     s->actively_synced = true;
1050                 }
1051             }
1052 
1053             should_complete = s->should_complete ||
1054                 job_cancel_requested(&s->common.job);
1055             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1056         }
1057 
1058         if (cnt == 0 && should_complete) {
1059             /* The dirty bitmap is not updated while operations are pending.
1060              * If we're about to exit, wait for pending operations before
1061              * calling bdrv_get_dirty_count(bs), or we may exit while the
1062              * source has dirty data to copy!
1063              *
1064              * Note that I/O can be submitted by the guest while
1065              * mirror_populate runs, so pause it now.  Before deciding
1066              * whether to switch to target check one last time if I/O has
1067              * come in the meanwhile, and if not flush the data to disk.
1068              */
1069             trace_mirror_before_drain(s, cnt);
1070 
1071             s->in_drain = true;
1072             bdrv_drained_begin(bs);
1073 
1074             /* Must be zero because we are drained */
1075             assert(s->in_active_write_counter == 0);
1076 
1077             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1078             if (cnt > 0 || mirror_flush(s) < 0) {
1079                 bdrv_drained_end(bs);
1080                 s->in_drain = false;
1081                 continue;
1082             }
1083 
1084             /* The two disks are in sync.  Exit and report successful
1085              * completion.
1086              */
1087             assert(QLIST_EMPTY(&bs->tracked_requests));
1088             need_drain = false;
1089             break;
1090         }
1091 
1092         if (job_is_ready(&s->common.job) && !should_complete) {
1093             delay_ns = (s->in_flight == 0 &&
1094                         cnt == 0 ? BLOCK_JOB_SLICE_TIME : 0);
1095         }
1096         trace_mirror_before_sleep(s, cnt, job_is_ready(&s->common.job),
1097                                   delay_ns);
1098         job_sleep_ns(&s->common.job, delay_ns);
1099         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1100     }
1101 
1102 immediate_exit:
1103     if (s->in_flight > 0) {
1104         /* We get here only if something went wrong.  Either the job failed,
1105          * or it was cancelled prematurely so that we do not guarantee that
1106          * the target is a copy of the source.
1107          */
1108         assert(ret < 0 || job_is_cancelled(&s->common.job));
1109         assert(need_drain);
1110         mirror_wait_for_all_io(s);
1111     }
1112 
1113     assert(s->in_flight == 0);
1114     qemu_vfree(s->buf);
1115     g_free(s->cow_bitmap);
1116     g_free(s->in_flight_bitmap);
1117     bdrv_dirty_iter_free(s->dbi);
1118 
1119     if (need_drain) {
1120         s->in_drain = true;
1121         bdrv_drained_begin(bs);
1122     }
1123 
1124     return ret;
1125 }
1126 
1127 static void mirror_complete(Job *job, Error **errp)
1128 {
1129     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1130 
1131     if (!job_is_ready(job)) {
1132         error_setg(errp, "The active block job '%s' cannot be completed",
1133                    job->id);
1134         return;
1135     }
1136 
1137     /* block all operations on to_replace bs */
1138     if (s->replaces) {
1139         AioContext *replace_aio_context;
1140 
1141         s->to_replace = bdrv_find_node(s->replaces);
1142         if (!s->to_replace) {
1143             error_setg(errp, "Node name '%s' not found", s->replaces);
1144             return;
1145         }
1146 
1147         replace_aio_context = bdrv_get_aio_context(s->to_replace);
1148         aio_context_acquire(replace_aio_context);
1149 
1150         /* TODO Translate this into child freeze system. */
1151         error_setg(&s->replace_blocker,
1152                    "block device is in use by block-job-complete");
1153         bdrv_op_block_all(s->to_replace, s->replace_blocker);
1154         bdrv_ref(s->to_replace);
1155 
1156         aio_context_release(replace_aio_context);
1157     }
1158 
1159     s->should_complete = true;
1160 
1161     /* If the job is paused, it will be re-entered when it is resumed */
1162     WITH_JOB_LOCK_GUARD() {
1163         if (!job->paused) {
1164             job_enter_cond_locked(job, NULL);
1165         }
1166     }
1167 }
1168 
1169 static void coroutine_fn mirror_pause(Job *job)
1170 {
1171     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1172 
1173     mirror_wait_for_all_io(s);
1174 }
1175 
1176 static bool mirror_drained_poll(BlockJob *job)
1177 {
1178     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1179 
1180     /* If the job isn't paused nor cancelled, we can't be sure that it won't
1181      * issue more requests. We make an exception if we've reached this point
1182      * from one of our own drain sections, to avoid a deadlock waiting for
1183      * ourselves.
1184      */
1185     WITH_JOB_LOCK_GUARD() {
1186         if (!s->common.job.paused && !job_is_cancelled_locked(&job->job)
1187             && !s->in_drain) {
1188             return true;
1189         }
1190     }
1191 
1192     return !!s->in_flight;
1193 }
1194 
1195 static bool mirror_cancel(Job *job, bool force)
1196 {
1197     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1198     BlockDriverState *target = blk_bs(s->target);
1199 
1200     /*
1201      * Before the job is READY, we treat any cancellation like a
1202      * force-cancellation.
1203      */
1204     force = force || !job_is_ready(job);
1205 
1206     if (force) {
1207         bdrv_cancel_in_flight(target);
1208     }
1209     return force;
1210 }
1211 
1212 static bool commit_active_cancel(Job *job, bool force)
1213 {
1214     /* Same as above in mirror_cancel() */
1215     return force || !job_is_ready(job);
1216 }
1217 
1218 static const BlockJobDriver mirror_job_driver = {
1219     .job_driver = {
1220         .instance_size          = sizeof(MirrorBlockJob),
1221         .job_type               = JOB_TYPE_MIRROR,
1222         .free                   = block_job_free,
1223         .user_resume            = block_job_user_resume,
1224         .run                    = mirror_run,
1225         .prepare                = mirror_prepare,
1226         .abort                  = mirror_abort,
1227         .pause                  = mirror_pause,
1228         .complete               = mirror_complete,
1229         .cancel                 = mirror_cancel,
1230     },
1231     .drained_poll           = mirror_drained_poll,
1232 };
1233 
1234 static const BlockJobDriver commit_active_job_driver = {
1235     .job_driver = {
1236         .instance_size          = sizeof(MirrorBlockJob),
1237         .job_type               = JOB_TYPE_COMMIT,
1238         .free                   = block_job_free,
1239         .user_resume            = block_job_user_resume,
1240         .run                    = mirror_run,
1241         .prepare                = mirror_prepare,
1242         .abort                  = mirror_abort,
1243         .pause                  = mirror_pause,
1244         .complete               = mirror_complete,
1245         .cancel                 = commit_active_cancel,
1246     },
1247     .drained_poll           = mirror_drained_poll,
1248 };
1249 
1250 static void coroutine_fn
1251 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method,
1252                      uint64_t offset, uint64_t bytes,
1253                      QEMUIOVector *qiov, int flags)
1254 {
1255     int ret;
1256     size_t qiov_offset = 0;
1257     int64_t bitmap_offset, bitmap_end;
1258 
1259     if (!QEMU_IS_ALIGNED(offset, job->granularity) &&
1260         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset))
1261     {
1262             /*
1263              * Dirty unaligned padding: ignore it.
1264              *
1265              * Reasoning:
1266              * 1. If we copy it, we can't reset corresponding bit in
1267              *    dirty_bitmap as there may be some "dirty" bytes still not
1268              *    copied.
1269              * 2. It's already dirty, so skipping it we don't diverge mirror
1270              *    progress.
1271              *
1272              * Note, that because of this, guest write may have no contribution
1273              * into mirror converge, but that's not bad, as we have background
1274              * process of mirroring. If under some bad circumstances (high guest
1275              * IO load) background process starve, we will not converge anyway,
1276              * even if each write will contribute, as guest is not guaranteed to
1277              * rewrite the whole disk.
1278              */
1279             qiov_offset = QEMU_ALIGN_UP(offset, job->granularity) - offset;
1280             if (bytes <= qiov_offset) {
1281                 /* nothing to do after shrink */
1282                 return;
1283             }
1284             offset += qiov_offset;
1285             bytes -= qiov_offset;
1286     }
1287 
1288     if (!QEMU_IS_ALIGNED(offset + bytes, job->granularity) &&
1289         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset + bytes - 1))
1290     {
1291         uint64_t tail = (offset + bytes) % job->granularity;
1292 
1293         if (bytes <= tail) {
1294             /* nothing to do after shrink */
1295             return;
1296         }
1297         bytes -= tail;
1298     }
1299 
1300     /*
1301      * Tails are either clean or shrunk, so for bitmap resetting
1302      * we safely align the range down.
1303      */
1304     bitmap_offset = QEMU_ALIGN_UP(offset, job->granularity);
1305     bitmap_end = QEMU_ALIGN_DOWN(offset + bytes, job->granularity);
1306     if (bitmap_offset < bitmap_end) {
1307         bdrv_reset_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1308                                 bitmap_end - bitmap_offset);
1309     }
1310 
1311     job_progress_increase_remaining(&job->common.job, bytes);
1312     job->active_write_bytes_in_flight += bytes;
1313 
1314     switch (method) {
1315     case MIRROR_METHOD_COPY:
1316         ret = blk_co_pwritev_part(job->target, offset, bytes,
1317                                   qiov, qiov_offset, flags);
1318         break;
1319 
1320     case MIRROR_METHOD_ZERO:
1321         assert(!qiov);
1322         ret = blk_co_pwrite_zeroes(job->target, offset, bytes, flags);
1323         break;
1324 
1325     case MIRROR_METHOD_DISCARD:
1326         assert(!qiov);
1327         ret = blk_co_pdiscard(job->target, offset, bytes);
1328         break;
1329 
1330     default:
1331         abort();
1332     }
1333 
1334     job->active_write_bytes_in_flight -= bytes;
1335     if (ret >= 0) {
1336         job_progress_update(&job->common.job, bytes);
1337     } else {
1338         BlockErrorAction action;
1339 
1340         /*
1341          * We failed, so we should mark dirty the whole area, aligned up.
1342          * Note that we don't care about shrunk tails if any: they were dirty
1343          * at function start, and they must be still dirty, as we've locked
1344          * the region for in-flight op.
1345          */
1346         bitmap_offset = QEMU_ALIGN_DOWN(offset, job->granularity);
1347         bitmap_end = QEMU_ALIGN_UP(offset + bytes, job->granularity);
1348         bdrv_set_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1349                               bitmap_end - bitmap_offset);
1350         job->actively_synced = false;
1351 
1352         action = mirror_error_action(job, false, -ret);
1353         if (action == BLOCK_ERROR_ACTION_REPORT) {
1354             if (!job->ret) {
1355                 job->ret = ret;
1356             }
1357         }
1358     }
1359 }
1360 
1361 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s,
1362                                                    uint64_t offset,
1363                                                    uint64_t bytes)
1364 {
1365     MirrorOp *op;
1366     uint64_t start_chunk = offset / s->granularity;
1367     uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
1368 
1369     op = g_new(MirrorOp, 1);
1370     *op = (MirrorOp){
1371         .s                  = s,
1372         .offset             = offset,
1373         .bytes              = bytes,
1374         .is_active_write    = true,
1375         .is_in_flight       = true,
1376         .co                 = qemu_coroutine_self(),
1377     };
1378     qemu_co_queue_init(&op->waiting_requests);
1379     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
1380 
1381     s->in_active_write_counter++;
1382 
1383     /*
1384      * Wait for concurrent requests affecting the area.  If there are already
1385      * running requests that are copying off now-to-be stale data in the area,
1386      * we must wait for them to finish before we begin writing fresh data to the
1387      * target so that the write operations appear in the correct order.
1388      * Note that background requests (see mirror_iteration()) in contrast only
1389      * wait for conflicting requests at the start of the dirty area, and then
1390      * (based on the in_flight_bitmap) truncate the area to copy so it will not
1391      * conflict with any requests beyond that.  For active writes, however, we
1392      * cannot truncate that area.  The request from our parent must be blocked
1393      * until the area is copied in full.  Therefore, we must wait for the whole
1394      * area to become free of concurrent requests.
1395      */
1396     mirror_wait_on_conflicts(op, s, offset, bytes);
1397 
1398     bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1399 
1400     return op;
1401 }
1402 
1403 static void coroutine_fn active_write_settle(MirrorOp *op)
1404 {
1405     uint64_t start_chunk = op->offset / op->s->granularity;
1406     uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes,
1407                                       op->s->granularity);
1408 
1409     if (!--op->s->in_active_write_counter && op->s->actively_synced) {
1410         BdrvChild *source = op->s->mirror_top_bs->backing;
1411 
1412         if (QLIST_FIRST(&source->bs->parents) == source &&
1413             QLIST_NEXT(source, next_parent) == NULL)
1414         {
1415             /* Assert that we are back in sync once all active write
1416              * operations are settled.
1417              * Note that we can only assert this if the mirror node
1418              * is the source node's only parent. */
1419             assert(!bdrv_get_dirty_count(op->s->dirty_bitmap));
1420         }
1421     }
1422     bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1423     QTAILQ_REMOVE(&op->s->ops_in_flight, op, next);
1424     qemu_co_queue_restart_all(&op->waiting_requests);
1425     g_free(op);
1426 }
1427 
1428 static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs,
1429     int64_t offset, int64_t bytes, QEMUIOVector *qiov, BdrvRequestFlags flags)
1430 {
1431     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1432 }
1433 
1434 static int coroutine_fn bdrv_mirror_top_do_write(BlockDriverState *bs,
1435     MirrorMethod method, uint64_t offset, uint64_t bytes, QEMUIOVector *qiov,
1436     int flags)
1437 {
1438     MirrorOp *op = NULL;
1439     MirrorBDSOpaque *s = bs->opaque;
1440     int ret = 0;
1441     bool copy_to_target = false;
1442 
1443     if (s->job) {
1444         copy_to_target = s->job->ret >= 0 &&
1445                          !job_is_cancelled(&s->job->common.job) &&
1446                          s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1447     }
1448 
1449     if (copy_to_target) {
1450         op = active_write_prepare(s->job, offset, bytes);
1451     }
1452 
1453     switch (method) {
1454     case MIRROR_METHOD_COPY:
1455         ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1456         break;
1457 
1458     case MIRROR_METHOD_ZERO:
1459         ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1460         break;
1461 
1462     case MIRROR_METHOD_DISCARD:
1463         ret = bdrv_co_pdiscard(bs->backing, offset, bytes);
1464         break;
1465 
1466     default:
1467         abort();
1468     }
1469 
1470     if (ret < 0) {
1471         goto out;
1472     }
1473 
1474     if (copy_to_target) {
1475         do_sync_target_write(s->job, method, offset, bytes, qiov, flags);
1476     }
1477 
1478 out:
1479     if (copy_to_target) {
1480         active_write_settle(op);
1481     }
1482     return ret;
1483 }
1484 
1485 static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs,
1486     int64_t offset, int64_t bytes, QEMUIOVector *qiov, BdrvRequestFlags flags)
1487 {
1488     MirrorBDSOpaque *s = bs->opaque;
1489     QEMUIOVector bounce_qiov;
1490     void *bounce_buf;
1491     int ret = 0;
1492     bool copy_to_target = false;
1493 
1494     if (s->job) {
1495         copy_to_target = s->job->ret >= 0 &&
1496                          !job_is_cancelled(&s->job->common.job) &&
1497                          s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1498     }
1499 
1500     if (copy_to_target) {
1501         /* The guest might concurrently modify the data to write; but
1502          * the data on source and destination must match, so we have
1503          * to use a bounce buffer if we are going to write to the
1504          * target now. */
1505         bounce_buf = qemu_blockalign(bs, bytes);
1506         iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes);
1507 
1508         qemu_iovec_init(&bounce_qiov, 1);
1509         qemu_iovec_add(&bounce_qiov, bounce_buf, bytes);
1510         qiov = &bounce_qiov;
1511 
1512         flags &= ~BDRV_REQ_REGISTERED_BUF;
1513     }
1514 
1515     ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov,
1516                                    flags);
1517 
1518     if (copy_to_target) {
1519         qemu_iovec_destroy(&bounce_qiov);
1520         qemu_vfree(bounce_buf);
1521     }
1522 
1523     return ret;
1524 }
1525 
1526 static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs)
1527 {
1528     if (bs->backing == NULL) {
1529         /* we can be here after failed bdrv_append in mirror_start_job */
1530         return 0;
1531     }
1532     return bdrv_co_flush(bs->backing->bs);
1533 }
1534 
1535 static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs,
1536     int64_t offset, int64_t bytes, BdrvRequestFlags flags)
1537 {
1538     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL,
1539                                     flags);
1540 }
1541 
1542 static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs,
1543     int64_t offset, int64_t bytes)
1544 {
1545     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes,
1546                                     NULL, 0);
1547 }
1548 
1549 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs)
1550 {
1551     if (bs->backing == NULL) {
1552         /* we can be here after failed bdrv_attach_child in
1553          * bdrv_set_backing_hd */
1554         return;
1555     }
1556     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1557             bs->backing->bs->filename);
1558 }
1559 
1560 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1561                                        BdrvChildRole role,
1562                                        BlockReopenQueue *reopen_queue,
1563                                        uint64_t perm, uint64_t shared,
1564                                        uint64_t *nperm, uint64_t *nshared)
1565 {
1566     MirrorBDSOpaque *s = bs->opaque;
1567 
1568     if (s->stop) {
1569         /*
1570          * If the job is to be stopped, we do not need to forward
1571          * anything to the real image.
1572          */
1573         *nperm = 0;
1574         *nshared = BLK_PERM_ALL;
1575         return;
1576     }
1577 
1578     bdrv_default_perms(bs, c, role, reopen_queue,
1579                        perm, shared, nperm, nshared);
1580 
1581     if (s->is_commit) {
1582         /*
1583          * For commit jobs, we cannot take CONSISTENT_READ, because
1584          * that permission is unshared for everything above the base
1585          * node (except for filters on the base node).
1586          * We also have to force-share the WRITE permission, or
1587          * otherwise we would block ourselves at the base node (if
1588          * writes are blocked for a node, they are also blocked for
1589          * its backing file).
1590          * (We could also share RESIZE, because it may be needed for
1591          * the target if its size is less than the top node's; but
1592          * bdrv_default_perms_for_cow() automatically shares RESIZE
1593          * for backing nodes if WRITE is shared, so there is no need
1594          * to do it here.)
1595          */
1596         *nperm &= ~BLK_PERM_CONSISTENT_READ;
1597         *nshared |= BLK_PERM_WRITE;
1598     }
1599 }
1600 
1601 /* Dummy node that provides consistent read to its users without requiring it
1602  * from its backing file and that allows writes on the backing file chain. */
1603 static BlockDriver bdrv_mirror_top = {
1604     .format_name                = "mirror_top",
1605     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1606     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1607     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1608     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1609     .bdrv_co_flush              = bdrv_mirror_top_flush,
1610     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1611     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1612 
1613     .is_filter                  = true,
1614     .filtered_child_is_backing  = true,
1615 };
1616 
1617 static BlockJob *mirror_start_job(
1618                              const char *job_id, BlockDriverState *bs,
1619                              int creation_flags, BlockDriverState *target,
1620                              const char *replaces, int64_t speed,
1621                              uint32_t granularity, int64_t buf_size,
1622                              BlockMirrorBackingMode backing_mode,
1623                              bool zero_target,
1624                              BlockdevOnError on_source_error,
1625                              BlockdevOnError on_target_error,
1626                              bool unmap,
1627                              BlockCompletionFunc *cb,
1628                              void *opaque,
1629                              const BlockJobDriver *driver,
1630                              bool is_none_mode, BlockDriverState *base,
1631                              bool auto_complete, const char *filter_node_name,
1632                              bool is_mirror, MirrorCopyMode copy_mode,
1633                              Error **errp)
1634 {
1635     MirrorBlockJob *s;
1636     MirrorBDSOpaque *bs_opaque;
1637     BlockDriverState *mirror_top_bs;
1638     bool target_is_backing;
1639     uint64_t target_perms, target_shared_perms;
1640     int ret;
1641 
1642     if (granularity == 0) {
1643         granularity = bdrv_get_default_bitmap_granularity(target);
1644     }
1645 
1646     assert(is_power_of_2(granularity));
1647 
1648     if (buf_size < 0) {
1649         error_setg(errp, "Invalid parameter 'buf-size'");
1650         return NULL;
1651     }
1652 
1653     if (buf_size == 0) {
1654         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1655     }
1656 
1657     if (bdrv_skip_filters(bs) == bdrv_skip_filters(target)) {
1658         error_setg(errp, "Can't mirror node into itself");
1659         return NULL;
1660     }
1661 
1662     target_is_backing = bdrv_chain_contains(bs, target);
1663 
1664     /* In the case of active commit, add dummy driver to provide consistent
1665      * reads on the top, while disabling it in the intermediate nodes, and make
1666      * the backing chain writable. */
1667     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1668                                          BDRV_O_RDWR, errp);
1669     if (mirror_top_bs == NULL) {
1670         return NULL;
1671     }
1672     if (!filter_node_name) {
1673         mirror_top_bs->implicit = true;
1674     }
1675 
1676     /* So that we can always drop this node */
1677     mirror_top_bs->never_freeze = true;
1678 
1679     mirror_top_bs->total_sectors = bs->total_sectors;
1680     mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
1681     mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
1682                                           BDRV_REQ_NO_FALLBACK;
1683     bs_opaque = g_new0(MirrorBDSOpaque, 1);
1684     mirror_top_bs->opaque = bs_opaque;
1685 
1686     bs_opaque->is_commit = target_is_backing;
1687 
1688     bdrv_drained_begin(bs);
1689     ret = bdrv_append(mirror_top_bs, bs, errp);
1690     bdrv_drained_end(bs);
1691 
1692     if (ret < 0) {
1693         bdrv_unref(mirror_top_bs);
1694         return NULL;
1695     }
1696 
1697     /* Make sure that the source is not resized while the job is running */
1698     s = block_job_create(job_id, driver, NULL, mirror_top_bs,
1699                          BLK_PERM_CONSISTENT_READ,
1700                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1701                          BLK_PERM_WRITE, speed,
1702                          creation_flags, cb, opaque, errp);
1703     if (!s) {
1704         goto fail;
1705     }
1706     bs_opaque->job = s;
1707 
1708     /* The block job now has a reference to this node */
1709     bdrv_unref(mirror_top_bs);
1710 
1711     s->mirror_top_bs = mirror_top_bs;
1712 
1713     /* No resize for the target either; while the mirror is still running, a
1714      * consistent read isn't necessarily possible. We could possibly allow
1715      * writes and graph modifications, though it would likely defeat the
1716      * purpose of a mirror, so leave them blocked for now.
1717      *
1718      * In the case of active commit, things look a bit different, though,
1719      * because the target is an already populated backing file in active use.
1720      * We can allow anything except resize there.*/
1721 
1722     target_perms = BLK_PERM_WRITE;
1723     target_shared_perms = BLK_PERM_WRITE_UNCHANGED;
1724 
1725     if (target_is_backing) {
1726         int64_t bs_size, target_size;
1727         bs_size = bdrv_getlength(bs);
1728         if (bs_size < 0) {
1729             error_setg_errno(errp, -bs_size,
1730                              "Could not inquire top image size");
1731             goto fail;
1732         }
1733 
1734         target_size = bdrv_getlength(target);
1735         if (target_size < 0) {
1736             error_setg_errno(errp, -target_size,
1737                              "Could not inquire base image size");
1738             goto fail;
1739         }
1740 
1741         if (target_size < bs_size) {
1742             target_perms |= BLK_PERM_RESIZE;
1743         }
1744 
1745         target_shared_perms |= BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE;
1746     } else if (bdrv_chain_contains(bs, bdrv_skip_filters(target))) {
1747         /*
1748          * We may want to allow this in the future, but it would
1749          * require taking some extra care.
1750          */
1751         error_setg(errp, "Cannot mirror to a filter on top of a node in the "
1752                    "source's backing chain");
1753         goto fail;
1754     }
1755 
1756     s->target = blk_new(s->common.job.aio_context,
1757                         target_perms, target_shared_perms);
1758     ret = blk_insert_bs(s->target, target, errp);
1759     if (ret < 0) {
1760         goto fail;
1761     }
1762     if (is_mirror) {
1763         /* XXX: Mirror target could be a NBD server of target QEMU in the case
1764          * of non-shared block migration. To allow migration completion, we
1765          * have to allow "inactivate" of the target BB.  When that happens, we
1766          * know the job is drained, and the vcpus are stopped, so no write
1767          * operation will be performed. Block layer already has assertions to
1768          * ensure that. */
1769         blk_set_force_allow_inactivate(s->target);
1770     }
1771     blk_set_allow_aio_context_change(s->target, true);
1772     blk_set_disable_request_queuing(s->target, true);
1773 
1774     s->replaces = g_strdup(replaces);
1775     s->on_source_error = on_source_error;
1776     s->on_target_error = on_target_error;
1777     s->is_none_mode = is_none_mode;
1778     s->backing_mode = backing_mode;
1779     s->zero_target = zero_target;
1780     s->copy_mode = copy_mode;
1781     s->base = base;
1782     s->base_overlay = bdrv_find_overlay(bs, base);
1783     s->granularity = granularity;
1784     s->buf_size = ROUND_UP(buf_size, granularity);
1785     s->unmap = unmap;
1786     if (auto_complete) {
1787         s->should_complete = true;
1788     }
1789 
1790     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1791     if (!s->dirty_bitmap) {
1792         goto fail;
1793     }
1794     if (s->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING) {
1795         bdrv_disable_dirty_bitmap(s->dirty_bitmap);
1796     }
1797 
1798     ret = block_job_add_bdrv(&s->common, "source", bs, 0,
1799                              BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE |
1800                              BLK_PERM_CONSISTENT_READ,
1801                              errp);
1802     if (ret < 0) {
1803         goto fail;
1804     }
1805 
1806     /* Required permissions are already taken with blk_new() */
1807     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1808                        &error_abort);
1809 
1810     /* In commit_active_start() all intermediate nodes disappear, so
1811      * any jobs in them must be blocked */
1812     if (target_is_backing) {
1813         BlockDriverState *iter, *filtered_target;
1814         uint64_t iter_shared_perms;
1815 
1816         /*
1817          * The topmost node with
1818          * bdrv_skip_filters(filtered_target) == bdrv_skip_filters(target)
1819          */
1820         filtered_target = bdrv_cow_bs(bdrv_find_overlay(bs, target));
1821 
1822         assert(bdrv_skip_filters(filtered_target) ==
1823                bdrv_skip_filters(target));
1824 
1825         /*
1826          * XXX BLK_PERM_WRITE needs to be allowed so we don't block
1827          * ourselves at s->base (if writes are blocked for a node, they are
1828          * also blocked for its backing file). The other options would be a
1829          * second filter driver above s->base (== target).
1830          */
1831         iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE;
1832 
1833         for (iter = bdrv_filter_or_cow_bs(bs); iter != target;
1834              iter = bdrv_filter_or_cow_bs(iter))
1835         {
1836             if (iter == filtered_target) {
1837                 /*
1838                  * From here on, all nodes are filters on the base.
1839                  * This allows us to share BLK_PERM_CONSISTENT_READ.
1840                  */
1841                 iter_shared_perms |= BLK_PERM_CONSISTENT_READ;
1842             }
1843 
1844             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1845                                      iter_shared_perms, errp);
1846             if (ret < 0) {
1847                 goto fail;
1848             }
1849         }
1850 
1851         if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) {
1852             goto fail;
1853         }
1854     }
1855 
1856     QTAILQ_INIT(&s->ops_in_flight);
1857 
1858     trace_mirror_start(bs, s, opaque);
1859     job_start(&s->common.job);
1860 
1861     return &s->common;
1862 
1863 fail:
1864     if (s) {
1865         /* Make sure this BDS does not go away until we have completed the graph
1866          * changes below */
1867         bdrv_ref(mirror_top_bs);
1868 
1869         g_free(s->replaces);
1870         blk_unref(s->target);
1871         bs_opaque->job = NULL;
1872         if (s->dirty_bitmap) {
1873             bdrv_release_dirty_bitmap(s->dirty_bitmap);
1874         }
1875         job_early_fail(&s->common.job);
1876     }
1877 
1878     bs_opaque->stop = true;
1879     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
1880                              &error_abort);
1881     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
1882 
1883     bdrv_unref(mirror_top_bs);
1884 
1885     return NULL;
1886 }
1887 
1888 void mirror_start(const char *job_id, BlockDriverState *bs,
1889                   BlockDriverState *target, const char *replaces,
1890                   int creation_flags, int64_t speed,
1891                   uint32_t granularity, int64_t buf_size,
1892                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1893                   bool zero_target,
1894                   BlockdevOnError on_source_error,
1895                   BlockdevOnError on_target_error,
1896                   bool unmap, const char *filter_node_name,
1897                   MirrorCopyMode copy_mode, Error **errp)
1898 {
1899     bool is_none_mode;
1900     BlockDriverState *base;
1901 
1902     GLOBAL_STATE_CODE();
1903 
1904     if ((mode == MIRROR_SYNC_MODE_INCREMENTAL) ||
1905         (mode == MIRROR_SYNC_MODE_BITMAP)) {
1906         error_setg(errp, "Sync mode '%s' not supported",
1907                    MirrorSyncMode_str(mode));
1908         return;
1909     }
1910     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1911     base = mode == MIRROR_SYNC_MODE_TOP ? bdrv_backing_chain_next(bs) : NULL;
1912     mirror_start_job(job_id, bs, creation_flags, target, replaces,
1913                      speed, granularity, buf_size, backing_mode, zero_target,
1914                      on_source_error, on_target_error, unmap, NULL, NULL,
1915                      &mirror_job_driver, is_none_mode, base, false,
1916                      filter_node_name, true, copy_mode, errp);
1917 }
1918 
1919 BlockJob *commit_active_start(const char *job_id, BlockDriverState *bs,
1920                               BlockDriverState *base, int creation_flags,
1921                               int64_t speed, BlockdevOnError on_error,
1922                               const char *filter_node_name,
1923                               BlockCompletionFunc *cb, void *opaque,
1924                               bool auto_complete, Error **errp)
1925 {
1926     bool base_read_only;
1927     BlockJob *job;
1928 
1929     GLOBAL_STATE_CODE();
1930 
1931     base_read_only = bdrv_is_read_only(base);
1932 
1933     if (base_read_only) {
1934         if (bdrv_reopen_set_read_only(base, false, errp) < 0) {
1935             return NULL;
1936         }
1937     }
1938 
1939     job = mirror_start_job(
1940                      job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1941                      MIRROR_LEAVE_BACKING_CHAIN, false,
1942                      on_error, on_error, true, cb, opaque,
1943                      &commit_active_job_driver, false, base, auto_complete,
1944                      filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND,
1945                      errp);
1946     if (!job) {
1947         goto error_restore_flags;
1948     }
1949 
1950     return job;
1951 
1952 error_restore_flags:
1953     /* ignore error and errp for bdrv_reopen, because we want to propagate
1954      * the original error */
1955     if (base_read_only) {
1956         bdrv_reopen_set_read_only(base, true, NULL);
1957     }
1958     return NULL;
1959 }
1960