xref: /openbmc/qemu/block/mirror.c (revision efb91426)
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     /*
59      * To be accesssed with atomics. Written only under the BQL (required by the
60      * current implementation of mirror_change()).
61      */
62     MirrorCopyMode copy_mode;
63     BlockdevOnError on_source_error, on_target_error;
64     /*
65      * To be accessed with atomics.
66      *
67      * Set when the target is synced (dirty bitmap is clean, nothing in flight)
68      * and the job is running in active mode.
69      */
70     bool actively_synced;
71     bool should_complete;
72     int64_t granularity;
73     size_t buf_size;
74     int64_t bdev_length;
75     unsigned long *cow_bitmap;
76     BdrvDirtyBitmap *dirty_bitmap;
77     BdrvDirtyBitmapIter *dbi;
78     uint8_t *buf;
79     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
80     int buf_free_count;
81 
82     uint64_t last_pause_ns;
83     unsigned long *in_flight_bitmap;
84     unsigned in_flight;
85     int64_t bytes_in_flight;
86     QTAILQ_HEAD(, MirrorOp) ops_in_flight;
87     int ret;
88     bool unmap;
89     int target_cluster_size;
90     int max_iov;
91     bool initial_zeroing_ongoing;
92     int in_active_write_counter;
93     int64_t active_write_bytes_in_flight;
94     bool prepared;
95     bool in_drain;
96 } MirrorBlockJob;
97 
98 typedef struct MirrorBDSOpaque {
99     MirrorBlockJob *job;
100     bool stop;
101     bool is_commit;
102 } MirrorBDSOpaque;
103 
104 struct MirrorOp {
105     MirrorBlockJob *s;
106     QEMUIOVector qiov;
107     int64_t offset;
108     uint64_t bytes;
109 
110     /* The pointee is set by mirror_co_read(), mirror_co_zero(), and
111      * mirror_co_discard() before yielding for the first time */
112     int64_t *bytes_handled;
113 
114     bool is_pseudo_op;
115     bool is_active_write;
116     bool is_in_flight;
117     CoQueue waiting_requests;
118     Coroutine *co;
119     MirrorOp *waiting_for_op;
120 
121     QTAILQ_ENTRY(MirrorOp) next;
122 };
123 
124 typedef enum MirrorMethod {
125     MIRROR_METHOD_COPY,
126     MIRROR_METHOD_ZERO,
127     MIRROR_METHOD_DISCARD,
128 } MirrorMethod;
129 
130 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
131                                             int error)
132 {
133     qatomic_set(&s->actively_synced, false);
134     if (read) {
135         return block_job_error_action(&s->common, s->on_source_error,
136                                       true, error);
137     } else {
138         return block_job_error_action(&s->common, s->on_target_error,
139                                       false, error);
140     }
141 }
142 
143 static void coroutine_fn mirror_wait_on_conflicts(MirrorOp *self,
144                                                   MirrorBlockJob *s,
145                                                   uint64_t offset,
146                                                   uint64_t bytes)
147 {
148     uint64_t self_start_chunk = offset / s->granularity;
149     uint64_t self_end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
150     uint64_t self_nb_chunks = self_end_chunk - self_start_chunk;
151 
152     while (find_next_bit(s->in_flight_bitmap, self_end_chunk,
153                          self_start_chunk) < self_end_chunk &&
154            s->ret >= 0)
155     {
156         MirrorOp *op;
157 
158         QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
159             uint64_t op_start_chunk = op->offset / s->granularity;
160             uint64_t op_nb_chunks = DIV_ROUND_UP(op->offset + op->bytes,
161                                                  s->granularity) -
162                                     op_start_chunk;
163 
164             if (op == self) {
165                 continue;
166             }
167 
168             if (ranges_overlap(self_start_chunk, self_nb_chunks,
169                                op_start_chunk, op_nb_chunks))
170             {
171                 if (self) {
172                     /*
173                      * If the operation is already (indirectly) waiting for us,
174                      * or will wait for us as soon as it wakes up, then just go
175                      * on (instead of producing a deadlock in the former case).
176                      */
177                     if (op->waiting_for_op) {
178                         continue;
179                     }
180 
181                     self->waiting_for_op = op;
182                 }
183 
184                 qemu_co_queue_wait(&op->waiting_requests, NULL);
185 
186                 if (self) {
187                     self->waiting_for_op = NULL;
188                 }
189 
190                 break;
191             }
192         }
193     }
194 }
195 
196 static void coroutine_fn mirror_iteration_done(MirrorOp *op, int ret)
197 {
198     MirrorBlockJob *s = op->s;
199     struct iovec *iov;
200     int64_t chunk_num;
201     int i, nb_chunks;
202 
203     trace_mirror_iteration_done(s, op->offset, op->bytes, ret);
204 
205     s->in_flight--;
206     s->bytes_in_flight -= op->bytes;
207     iov = op->qiov.iov;
208     for (i = 0; i < op->qiov.niov; i++) {
209         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
210         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
211         s->buf_free_count++;
212     }
213 
214     chunk_num = op->offset / s->granularity;
215     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
216 
217     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
218     QTAILQ_REMOVE(&s->ops_in_flight, op, next);
219     if (ret >= 0) {
220         if (s->cow_bitmap) {
221             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
222         }
223         if (!s->initial_zeroing_ongoing) {
224             job_progress_update(&s->common.job, op->bytes);
225         }
226     }
227     qemu_iovec_destroy(&op->qiov);
228 
229     qemu_co_queue_restart_all(&op->waiting_requests);
230     g_free(op);
231 }
232 
233 static void coroutine_fn mirror_write_complete(MirrorOp *op, int ret)
234 {
235     MirrorBlockJob *s = op->s;
236 
237     if (ret < 0) {
238         BlockErrorAction action;
239 
240         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
241         action = mirror_error_action(s, false, -ret);
242         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
243             s->ret = ret;
244         }
245     }
246 
247     mirror_iteration_done(op, ret);
248 }
249 
250 static void coroutine_fn mirror_read_complete(MirrorOp *op, int ret)
251 {
252     MirrorBlockJob *s = op->s;
253 
254     if (ret < 0) {
255         BlockErrorAction action;
256 
257         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
258         action = mirror_error_action(s, true, -ret);
259         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
260             s->ret = ret;
261         }
262 
263         mirror_iteration_done(op, ret);
264         return;
265     }
266 
267     ret = blk_co_pwritev(s->target, op->offset, op->qiov.size, &op->qiov, 0);
268     mirror_write_complete(op, ret);
269 }
270 
271 /* Clip bytes relative to offset to not exceed end-of-file */
272 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s,
273                                         int64_t offset,
274                                         int64_t bytes)
275 {
276     return MIN(bytes, s->bdev_length - offset);
277 }
278 
279 /* Round offset and/or bytes to target cluster if COW is needed, and
280  * return the offset of the adjusted tail against original. */
281 static int coroutine_fn mirror_cow_align(MirrorBlockJob *s, int64_t *offset,
282                                          uint64_t *bytes)
283 {
284     bool need_cow;
285     int ret = 0;
286     int64_t align_offset = *offset;
287     int64_t align_bytes = *bytes;
288     int max_bytes = s->granularity * s->max_iov;
289 
290     need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap);
291     need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity,
292                           s->cow_bitmap);
293     if (need_cow) {
294         bdrv_round_to_subclusters(blk_bs(s->target), *offset, *bytes,
295                                   &align_offset, &align_bytes);
296     }
297 
298     if (align_bytes > max_bytes) {
299         align_bytes = max_bytes;
300         if (need_cow) {
301             align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size);
302         }
303     }
304     /* Clipping may result in align_bytes unaligned to chunk boundary, but
305      * that doesn't matter because it's already the end of source image. */
306     align_bytes = mirror_clip_bytes(s, align_offset, align_bytes);
307 
308     ret = align_offset + align_bytes - (*offset + *bytes);
309     *offset = align_offset;
310     *bytes = align_bytes;
311     assert(ret >= 0);
312     return ret;
313 }
314 
315 static inline void coroutine_fn
316 mirror_wait_for_free_in_flight_slot(MirrorBlockJob *s)
317 {
318     MirrorOp *op;
319 
320     QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
321         /*
322          * Do not wait on pseudo ops, because it may in turn wait on
323          * some other operation to start, which may in fact be the
324          * caller of this function.  Since there is only one pseudo op
325          * at any given time, we will always find some real operation
326          * to wait on.
327          * Also, do not wait on active operations, because they do not
328          * use up in-flight slots.
329          */
330         if (!op->is_pseudo_op && op->is_in_flight && !op->is_active_write) {
331             qemu_co_queue_wait(&op->waiting_requests, NULL);
332             return;
333         }
334     }
335     abort();
336 }
337 
338 /* Perform a mirror copy operation.
339  *
340  * *op->bytes_handled is set to the number of bytes copied after and
341  * including offset, excluding any bytes copied prior to offset due
342  * to alignment.  This will be op->bytes if no alignment is necessary,
343  * or (new_end - op->offset) if the tail is rounded up or down due to
344  * alignment or buffer limit.
345  */
346 static void coroutine_fn mirror_co_read(void *opaque)
347 {
348     MirrorOp *op = opaque;
349     MirrorBlockJob *s = op->s;
350     int nb_chunks;
351     uint64_t ret;
352     uint64_t max_bytes;
353 
354     max_bytes = s->granularity * s->max_iov;
355 
356     /* We can only handle as much as buf_size at a time. */
357     op->bytes = MIN(s->buf_size, MIN(max_bytes, op->bytes));
358     assert(op->bytes);
359     assert(op->bytes < BDRV_REQUEST_MAX_BYTES);
360     *op->bytes_handled = op->bytes;
361 
362     if (s->cow_bitmap) {
363         *op->bytes_handled += mirror_cow_align(s, &op->offset, &op->bytes);
364     }
365     /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */
366     assert(*op->bytes_handled <= UINT_MAX);
367     assert(op->bytes <= s->buf_size);
368     /* The offset is granularity-aligned because:
369      * 1) Caller passes in aligned values;
370      * 2) mirror_cow_align is used only when target cluster is larger. */
371     assert(QEMU_IS_ALIGNED(op->offset, s->granularity));
372     /* The range is sector-aligned, since bdrv_getlength() rounds up. */
373     assert(QEMU_IS_ALIGNED(op->bytes, BDRV_SECTOR_SIZE));
374     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
375 
376     while (s->buf_free_count < nb_chunks) {
377         trace_mirror_yield_in_flight(s, op->offset, s->in_flight);
378         mirror_wait_for_free_in_flight_slot(s);
379     }
380 
381     /* Now make a QEMUIOVector taking enough granularity-sized chunks
382      * from s->buf_free.
383      */
384     qemu_iovec_init(&op->qiov, nb_chunks);
385     while (nb_chunks-- > 0) {
386         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
387         size_t remaining = op->bytes - op->qiov.size;
388 
389         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
390         s->buf_free_count--;
391         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
392     }
393 
394     /* Copy the dirty cluster.  */
395     s->in_flight++;
396     s->bytes_in_flight += op->bytes;
397     op->is_in_flight = true;
398     trace_mirror_one_iteration(s, op->offset, op->bytes);
399 
400     WITH_GRAPH_RDLOCK_GUARD() {
401         ret = bdrv_co_preadv(s->mirror_top_bs->backing, op->offset, op->bytes,
402                              &op->qiov, 0);
403     }
404     mirror_read_complete(op, ret);
405 }
406 
407 static void coroutine_fn mirror_co_zero(void *opaque)
408 {
409     MirrorOp *op = opaque;
410     int ret;
411 
412     op->s->in_flight++;
413     op->s->bytes_in_flight += op->bytes;
414     *op->bytes_handled = op->bytes;
415     op->is_in_flight = true;
416 
417     ret = blk_co_pwrite_zeroes(op->s->target, op->offset, op->bytes,
418                                op->s->unmap ? BDRV_REQ_MAY_UNMAP : 0);
419     mirror_write_complete(op, ret);
420 }
421 
422 static void coroutine_fn mirror_co_discard(void *opaque)
423 {
424     MirrorOp *op = opaque;
425     int ret;
426 
427     op->s->in_flight++;
428     op->s->bytes_in_flight += op->bytes;
429     *op->bytes_handled = op->bytes;
430     op->is_in_flight = true;
431 
432     ret = blk_co_pdiscard(op->s->target, op->offset, op->bytes);
433     mirror_write_complete(op, ret);
434 }
435 
436 static unsigned mirror_perform(MirrorBlockJob *s, int64_t offset,
437                                unsigned bytes, MirrorMethod mirror_method)
438 {
439     MirrorOp *op;
440     Coroutine *co;
441     int64_t bytes_handled = -1;
442 
443     op = g_new(MirrorOp, 1);
444     *op = (MirrorOp){
445         .s              = s,
446         .offset         = offset,
447         .bytes          = bytes,
448         .bytes_handled  = &bytes_handled,
449     };
450     qemu_co_queue_init(&op->waiting_requests);
451 
452     switch (mirror_method) {
453     case MIRROR_METHOD_COPY:
454         co = qemu_coroutine_create(mirror_co_read, op);
455         break;
456     case MIRROR_METHOD_ZERO:
457         co = qemu_coroutine_create(mirror_co_zero, op);
458         break;
459     case MIRROR_METHOD_DISCARD:
460         co = qemu_coroutine_create(mirror_co_discard, op);
461         break;
462     default:
463         abort();
464     }
465     op->co = co;
466 
467     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
468     qemu_coroutine_enter(co);
469     /* At this point, ownership of op has been moved to the coroutine
470      * and the object may already be freed */
471 
472     /* Assert that this value has been set */
473     assert(bytes_handled >= 0);
474 
475     /* Same assertion as in mirror_co_read() (and for mirror_co_read()
476      * and mirror_co_discard(), bytes_handled == op->bytes, which
477      * is the @bytes parameter given to this function) */
478     assert(bytes_handled <= UINT_MAX);
479     return bytes_handled;
480 }
481 
482 static void coroutine_fn GRAPH_RDLOCK mirror_iteration(MirrorBlockJob *s)
483 {
484     BlockDriverState *source = s->mirror_top_bs->backing->bs;
485     MirrorOp *pseudo_op;
486     int64_t offset;
487     /* At least the first dirty chunk is mirrored in one iteration. */
488     int nb_chunks = 1;
489     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
490     int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES);
491 
492     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
493     offset = bdrv_dirty_iter_next(s->dbi);
494     if (offset < 0) {
495         bdrv_set_dirty_iter(s->dbi, 0);
496         offset = bdrv_dirty_iter_next(s->dbi);
497         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
498         assert(offset >= 0);
499     }
500     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
501 
502     /*
503      * Wait for concurrent requests to @offset.  The next loop will limit the
504      * copied area based on in_flight_bitmap so we only copy an area that does
505      * not overlap with concurrent in-flight requests.  Still, we would like to
506      * copy something, so wait until there are at least no more requests to the
507      * very beginning of the area.
508      */
509     mirror_wait_on_conflicts(NULL, s, offset, 1);
510 
511     job_pause_point(&s->common.job);
512 
513     /* Find the number of consecutive dirty chunks following the first dirty
514      * one, and wait for in flight requests in them. */
515     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
516     while (nb_chunks * s->granularity < s->buf_size) {
517         int64_t next_dirty;
518         int64_t next_offset = offset + nb_chunks * s->granularity;
519         int64_t next_chunk = next_offset / s->granularity;
520         if (next_offset >= s->bdev_length ||
521             !bdrv_dirty_bitmap_get_locked(s->dirty_bitmap, next_offset)) {
522             break;
523         }
524         if (test_bit(next_chunk, s->in_flight_bitmap)) {
525             break;
526         }
527 
528         next_dirty = bdrv_dirty_iter_next(s->dbi);
529         if (next_dirty > next_offset || next_dirty < 0) {
530             /* The bitmap iterator's cache is stale, refresh it */
531             bdrv_set_dirty_iter(s->dbi, next_offset);
532             next_dirty = bdrv_dirty_iter_next(s->dbi);
533         }
534         assert(next_dirty == next_offset);
535         nb_chunks++;
536     }
537 
538     /* Clear dirty bits before querying the block status, because
539      * calling bdrv_block_status_above could yield - if some blocks are
540      * marked dirty in this window, we need to know.
541      */
542     bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset,
543                                    nb_chunks * s->granularity);
544     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
545 
546     /* Before claiming an area in the in-flight bitmap, we have to
547      * create a MirrorOp for it so that conflicting requests can wait
548      * for it.  mirror_perform() will create the real MirrorOps later,
549      * for now we just create a pseudo operation that will wake up all
550      * conflicting requests once all real operations have been
551      * launched. */
552     pseudo_op = g_new(MirrorOp, 1);
553     *pseudo_op = (MirrorOp){
554         .offset         = offset,
555         .bytes          = nb_chunks * s->granularity,
556         .is_pseudo_op   = true,
557     };
558     qemu_co_queue_init(&pseudo_op->waiting_requests);
559     QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next);
560 
561     bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks);
562     while (nb_chunks > 0 && offset < s->bdev_length) {
563         int ret;
564         int64_t io_bytes;
565         int64_t io_bytes_acct;
566         MirrorMethod mirror_method = MIRROR_METHOD_COPY;
567 
568         assert(!(offset % s->granularity));
569         WITH_GRAPH_RDLOCK_GUARD() {
570             ret = bdrv_co_block_status_above(source, NULL, offset,
571                                              nb_chunks * s->granularity,
572                                              &io_bytes, NULL, NULL);
573         }
574         if (ret < 0) {
575             io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes);
576         } else if (ret & BDRV_BLOCK_DATA) {
577             io_bytes = MIN(io_bytes, max_io_bytes);
578         }
579 
580         io_bytes -= io_bytes % s->granularity;
581         if (io_bytes < s->granularity) {
582             io_bytes = s->granularity;
583         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
584             int64_t target_offset;
585             int64_t target_bytes;
586             WITH_GRAPH_RDLOCK_GUARD() {
587                 bdrv_round_to_subclusters(blk_bs(s->target), offset, io_bytes,
588                                           &target_offset, &target_bytes);
589             }
590             if (target_offset == offset &&
591                 target_bytes == io_bytes) {
592                 mirror_method = ret & BDRV_BLOCK_ZERO ?
593                                     MIRROR_METHOD_ZERO :
594                                     MIRROR_METHOD_DISCARD;
595             }
596         }
597 
598         while (s->in_flight >= MAX_IN_FLIGHT) {
599             trace_mirror_yield_in_flight(s, offset, s->in_flight);
600             mirror_wait_for_free_in_flight_slot(s);
601         }
602 
603         if (s->ret < 0) {
604             ret = 0;
605             goto fail;
606         }
607 
608         io_bytes = mirror_clip_bytes(s, offset, io_bytes);
609         io_bytes = mirror_perform(s, offset, io_bytes, mirror_method);
610         if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) {
611             io_bytes_acct = 0;
612         } else {
613             io_bytes_acct = io_bytes;
614         }
615         assert(io_bytes);
616         offset += io_bytes;
617         nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity);
618         block_job_ratelimit_processed_bytes(&s->common, io_bytes_acct);
619     }
620 
621 fail:
622     QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next);
623     qemu_co_queue_restart_all(&pseudo_op->waiting_requests);
624     g_free(pseudo_op);
625 }
626 
627 static void mirror_free_init(MirrorBlockJob *s)
628 {
629     int granularity = s->granularity;
630     size_t buf_size = s->buf_size;
631     uint8_t *buf = s->buf;
632 
633     assert(s->buf_free_count == 0);
634     QSIMPLEQ_INIT(&s->buf_free);
635     while (buf_size != 0) {
636         MirrorBuffer *cur = (MirrorBuffer *)buf;
637         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
638         s->buf_free_count++;
639         buf_size -= granularity;
640         buf += granularity;
641     }
642 }
643 
644 /* This is also used for the .pause callback. There is no matching
645  * mirror_resume() because mirror_run() will begin iterating again
646  * when the job is resumed.
647  */
648 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s)
649 {
650     while (s->in_flight > 0) {
651         mirror_wait_for_free_in_flight_slot(s);
652     }
653 }
654 
655 /**
656  * mirror_exit_common: handle both abort() and prepare() cases.
657  * for .prepare, returns 0 on success and -errno on failure.
658  * for .abort cases, denoted by abort = true, MUST return 0.
659  */
660 static int mirror_exit_common(Job *job)
661 {
662     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
663     BlockJob *bjob = &s->common;
664     MirrorBDSOpaque *bs_opaque;
665     BlockDriverState *src;
666     BlockDriverState *target_bs;
667     BlockDriverState *mirror_top_bs;
668     Error *local_err = NULL;
669     bool abort = job->ret < 0;
670     int ret = 0;
671 
672     GLOBAL_STATE_CODE();
673 
674     if (s->prepared) {
675         return 0;
676     }
677     s->prepared = true;
678 
679     bdrv_graph_rdlock_main_loop();
680 
681     mirror_top_bs = s->mirror_top_bs;
682     bs_opaque = mirror_top_bs->opaque;
683     src = mirror_top_bs->backing->bs;
684     target_bs = blk_bs(s->target);
685 
686     if (bdrv_chain_contains(src, target_bs)) {
687         bdrv_unfreeze_backing_chain(mirror_top_bs, target_bs);
688     }
689 
690     bdrv_release_dirty_bitmap(s->dirty_bitmap);
691 
692     /* Make sure that the source BDS doesn't go away during bdrv_replace_node,
693      * before we can call bdrv_drained_end */
694     bdrv_ref(src);
695     bdrv_ref(mirror_top_bs);
696     bdrv_ref(target_bs);
697 
698     bdrv_graph_rdunlock_main_loop();
699 
700     /*
701      * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
702      * inserting target_bs at s->to_replace, where we might not be able to get
703      * these permissions.
704      */
705     blk_unref(s->target);
706     s->target = NULL;
707 
708     /* We don't access the source any more. Dropping any WRITE/RESIZE is
709      * required before it could become a backing file of target_bs. Not having
710      * these permissions any more means that we can't allow any new requests on
711      * mirror_top_bs from now on, so keep it drained. */
712     bdrv_drained_begin(mirror_top_bs);
713     bdrv_drained_begin(target_bs);
714     bs_opaque->stop = true;
715 
716     bdrv_graph_rdlock_main_loop();
717     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
718                              &error_abort);
719 
720     if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
721         BlockDriverState *backing = s->is_none_mode ? src : s->base;
722         BlockDriverState *unfiltered_target = bdrv_skip_filters(target_bs);
723 
724         if (bdrv_cow_bs(unfiltered_target) != backing) {
725             bdrv_set_backing_hd(unfiltered_target, backing, &local_err);
726             if (local_err) {
727                 error_report_err(local_err);
728                 local_err = NULL;
729                 ret = -EPERM;
730             }
731         }
732     } else if (!abort && s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
733         assert(!bdrv_backing_chain_next(target_bs));
734         ret = bdrv_open_backing_file(bdrv_skip_filters(target_bs), NULL,
735                                      "backing", &local_err);
736         if (ret < 0) {
737             error_report_err(local_err);
738             local_err = NULL;
739         }
740     }
741     bdrv_graph_rdunlock_main_loop();
742 
743     if (s->should_complete && !abort) {
744         BlockDriverState *to_replace = s->to_replace ?: src;
745         bool ro = bdrv_is_read_only(to_replace);
746 
747         if (ro != bdrv_is_read_only(target_bs)) {
748             bdrv_reopen_set_read_only(target_bs, ro, NULL);
749         }
750 
751         /* The mirror job has no requests in flight any more, but we need to
752          * drain potential other users of the BDS before changing the graph. */
753         assert(s->in_drain);
754         bdrv_drained_begin(to_replace);
755         /*
756          * Cannot use check_to_replace_node() here, because that would
757          * check for an op blocker on @to_replace, and we have our own
758          * there.
759          */
760         bdrv_graph_wrlock();
761         if (bdrv_recurse_can_replace(src, to_replace)) {
762             bdrv_replace_node(to_replace, target_bs, &local_err);
763         } else {
764             error_setg(&local_err, "Can no longer replace '%s' by '%s', "
765                        "because it can no longer be guaranteed that doing so "
766                        "would not lead to an abrupt change of visible data",
767                        to_replace->node_name, target_bs->node_name);
768         }
769         bdrv_graph_wrunlock();
770         bdrv_drained_end(to_replace);
771         if (local_err) {
772             error_report_err(local_err);
773             ret = -EPERM;
774         }
775     }
776     if (s->to_replace) {
777         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
778         error_free(s->replace_blocker);
779         bdrv_unref(s->to_replace);
780     }
781     g_free(s->replaces);
782 
783     /*
784      * Remove the mirror filter driver from the graph. Before this, get rid of
785      * the blockers on the intermediate nodes so that the resulting state is
786      * valid.
787      */
788     block_job_remove_all_bdrv(bjob);
789     bdrv_graph_wrlock();
790     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
791     bdrv_graph_wrunlock();
792 
793     bdrv_drained_end(target_bs);
794     bdrv_unref(target_bs);
795 
796     bs_opaque->job = NULL;
797 
798     bdrv_drained_end(src);
799     bdrv_drained_end(mirror_top_bs);
800     s->in_drain = false;
801     bdrv_unref(mirror_top_bs);
802     bdrv_unref(src);
803 
804     return ret;
805 }
806 
807 static int mirror_prepare(Job *job)
808 {
809     return mirror_exit_common(job);
810 }
811 
812 static void mirror_abort(Job *job)
813 {
814     int ret = mirror_exit_common(job);
815     assert(ret == 0);
816 }
817 
818 static void coroutine_fn mirror_throttle(MirrorBlockJob *s)
819 {
820     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
821 
822     if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) {
823         s->last_pause_ns = now;
824         job_sleep_ns(&s->common.job, 0);
825     } else {
826         job_pause_point(&s->common.job);
827     }
828 }
829 
830 static int coroutine_fn GRAPH_UNLOCKED mirror_dirty_init(MirrorBlockJob *s)
831 {
832     int64_t offset;
833     BlockDriverState *bs;
834     BlockDriverState *target_bs = blk_bs(s->target);
835     int ret;
836     int64_t count;
837 
838     bdrv_graph_co_rdlock();
839     bs = s->mirror_top_bs->backing->bs;
840     bdrv_graph_co_rdunlock();
841 
842     if (s->zero_target) {
843         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
844             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
845             return 0;
846         }
847 
848         s->initial_zeroing_ongoing = true;
849         for (offset = 0; offset < s->bdev_length; ) {
850             int bytes = MIN(s->bdev_length - offset,
851                             QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
852 
853             mirror_throttle(s);
854 
855             if (job_is_cancelled(&s->common.job)) {
856                 s->initial_zeroing_ongoing = false;
857                 return 0;
858             }
859 
860             if (s->in_flight >= MAX_IN_FLIGHT) {
861                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
862                                    s->in_flight);
863                 mirror_wait_for_free_in_flight_slot(s);
864                 continue;
865             }
866 
867             mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO);
868             offset += bytes;
869         }
870 
871         mirror_wait_for_all_io(s);
872         s->initial_zeroing_ongoing = false;
873     }
874 
875     /* First part, loop on the sectors and initialize the dirty bitmap.  */
876     for (offset = 0; offset < s->bdev_length; ) {
877         /* Just to make sure we are not exceeding int limit. */
878         int bytes = MIN(s->bdev_length - offset,
879                         QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
880 
881         mirror_throttle(s);
882 
883         if (job_is_cancelled(&s->common.job)) {
884             return 0;
885         }
886 
887         WITH_GRAPH_RDLOCK_GUARD() {
888             ret = bdrv_co_is_allocated_above(bs, s->base_overlay, true, offset,
889                                              bytes, &count);
890         }
891         if (ret < 0) {
892             return ret;
893         }
894 
895         assert(count);
896         if (ret > 0) {
897             bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
898         }
899         offset += count;
900     }
901     return 0;
902 }
903 
904 /* Called when going out of the streaming phase to flush the bulk of the
905  * data to the medium, or just before completing.
906  */
907 static int coroutine_fn mirror_flush(MirrorBlockJob *s)
908 {
909     int ret = blk_co_flush(s->target);
910     if (ret < 0) {
911         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
912             s->ret = ret;
913         }
914     }
915     return ret;
916 }
917 
918 static int coroutine_fn mirror_run(Job *job, Error **errp)
919 {
920     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
921     BlockDriverState *bs;
922     MirrorBDSOpaque *mirror_top_opaque = s->mirror_top_bs->opaque;
923     BlockDriverState *target_bs = blk_bs(s->target);
924     bool need_drain = true;
925     BlockDeviceIoStatus iostatus;
926     int64_t length;
927     int64_t target_length;
928     BlockDriverInfo bdi;
929     char backing_filename[2]; /* we only need 2 characters because we are only
930                                  checking for a NULL string */
931     int ret = 0;
932 
933     bdrv_graph_co_rdlock();
934     bs = bdrv_filter_bs(s->mirror_top_bs);
935     bdrv_graph_co_rdunlock();
936 
937     if (job_is_cancelled(&s->common.job)) {
938         goto immediate_exit;
939     }
940 
941     bdrv_graph_co_rdlock();
942     s->bdev_length = bdrv_co_getlength(bs);
943     bdrv_graph_co_rdunlock();
944 
945     if (s->bdev_length < 0) {
946         ret = s->bdev_length;
947         goto immediate_exit;
948     }
949 
950     target_length = blk_co_getlength(s->target);
951     if (target_length < 0) {
952         ret = target_length;
953         goto immediate_exit;
954     }
955 
956     /* Active commit must resize the base image if its size differs from the
957      * active layer. */
958     if (s->base == blk_bs(s->target)) {
959         if (s->bdev_length > target_length) {
960             ret = blk_co_truncate(s->target, s->bdev_length, false,
961                                   PREALLOC_MODE_OFF, 0, NULL);
962             if (ret < 0) {
963                 goto immediate_exit;
964             }
965         }
966     } else if (s->bdev_length != target_length) {
967         error_setg(errp, "Source and target image have different sizes");
968         ret = -EINVAL;
969         goto immediate_exit;
970     }
971 
972     if (s->bdev_length == 0) {
973         /* Transition to the READY state and wait for complete. */
974         job_transition_to_ready(&s->common.job);
975         qatomic_set(&s->actively_synced, true);
976         while (!job_cancel_requested(&s->common.job) && !s->should_complete) {
977             job_yield(&s->common.job);
978         }
979         goto immediate_exit;
980     }
981 
982     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
983     s->in_flight_bitmap = bitmap_new(length);
984 
985     /* If we have no backing file yet in the destination, we cannot let
986      * the destination do COW.  Instead, we copy sectors around the
987      * dirty data if needed.  We need a bitmap to do that.
988      */
989     bdrv_get_backing_filename(target_bs, backing_filename,
990                               sizeof(backing_filename));
991     bdrv_graph_co_rdlock();
992     if (!bdrv_co_get_info(target_bs, &bdi) && bdi.cluster_size) {
993         s->target_cluster_size = bdi.cluster_size;
994     } else {
995         s->target_cluster_size = BDRV_SECTOR_SIZE;
996     }
997     if (backing_filename[0] && !bdrv_backing_chain_next(target_bs) &&
998         s->granularity < s->target_cluster_size) {
999         s->buf_size = MAX(s->buf_size, s->target_cluster_size);
1000         s->cow_bitmap = bitmap_new(length);
1001     }
1002     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
1003     bdrv_graph_co_rdunlock();
1004 
1005     s->buf = qemu_try_blockalign(bs, s->buf_size);
1006     if (s->buf == NULL) {
1007         ret = -ENOMEM;
1008         goto immediate_exit;
1009     }
1010 
1011     mirror_free_init(s);
1012 
1013     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1014     if (!s->is_none_mode) {
1015         ret = mirror_dirty_init(s);
1016         if (ret < 0 || job_is_cancelled(&s->common.job)) {
1017             goto immediate_exit;
1018         }
1019     }
1020 
1021     /*
1022      * Only now the job is fully initialised and mirror_top_bs should start
1023      * accessing it.
1024      */
1025     mirror_top_opaque->job = s;
1026 
1027     assert(!s->dbi);
1028     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
1029     for (;;) {
1030         int64_t cnt, delta;
1031         bool should_complete;
1032 
1033         if (s->ret < 0) {
1034             ret = s->ret;
1035             goto immediate_exit;
1036         }
1037 
1038         job_pause_point(&s->common.job);
1039 
1040         if (job_is_cancelled(&s->common.job)) {
1041             ret = 0;
1042             goto immediate_exit;
1043         }
1044 
1045         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1046         /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
1047          * the number of bytes currently being processed; together those are
1048          * the current remaining operation length */
1049         job_progress_set_remaining(&s->common.job,
1050                                    s->bytes_in_flight + cnt +
1051                                    s->active_write_bytes_in_flight);
1052 
1053         /* Note that even when no rate limit is applied we need to yield
1054          * periodically with no pending I/O so that bdrv_drain_all() returns.
1055          * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
1056          * an error, or when the source is clean, whichever comes first. */
1057         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
1058         WITH_JOB_LOCK_GUARD() {
1059             iostatus = s->common.iostatus;
1060         }
1061         if (delta < BLOCK_JOB_SLICE_TIME &&
1062             iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
1063             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
1064                 (cnt == 0 && s->in_flight > 0)) {
1065                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
1066                 mirror_wait_for_free_in_flight_slot(s);
1067                 continue;
1068             } else if (cnt != 0) {
1069                 bdrv_graph_co_rdlock();
1070                 mirror_iteration(s);
1071                 bdrv_graph_co_rdunlock();
1072             }
1073         }
1074 
1075         should_complete = false;
1076         if (s->in_flight == 0 && cnt == 0) {
1077             trace_mirror_before_flush(s);
1078             if (!job_is_ready(&s->common.job)) {
1079                 if (mirror_flush(s) < 0) {
1080                     /* Go check s->ret.  */
1081                     continue;
1082                 }
1083                 /* We're out of the streaming phase.  From now on, if the job
1084                  * is cancelled we will actually complete all pending I/O and
1085                  * report completion.  This way, block-job-cancel will leave
1086                  * the target in a consistent state.
1087                  */
1088                 job_transition_to_ready(&s->common.job);
1089             }
1090             if (qatomic_read(&s->copy_mode) != MIRROR_COPY_MODE_BACKGROUND) {
1091                 qatomic_set(&s->actively_synced, true);
1092             }
1093 
1094             should_complete = s->should_complete ||
1095                 job_cancel_requested(&s->common.job);
1096             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1097         }
1098 
1099         if (cnt == 0 && should_complete) {
1100             /* The dirty bitmap is not updated while operations are pending.
1101              * If we're about to exit, wait for pending operations before
1102              * calling bdrv_get_dirty_count(bs), or we may exit while the
1103              * source has dirty data to copy!
1104              *
1105              * Note that I/O can be submitted by the guest while
1106              * mirror_populate runs, so pause it now.  Before deciding
1107              * whether to switch to target check one last time if I/O has
1108              * come in the meanwhile, and if not flush the data to disk.
1109              */
1110             trace_mirror_before_drain(s, cnt);
1111 
1112             s->in_drain = true;
1113             bdrv_drained_begin(bs);
1114 
1115             /* Must be zero because we are drained */
1116             assert(s->in_active_write_counter == 0);
1117 
1118             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1119             if (cnt > 0 || mirror_flush(s) < 0) {
1120                 bdrv_drained_end(bs);
1121                 s->in_drain = false;
1122                 continue;
1123             }
1124 
1125             /* The two disks are in sync.  Exit and report successful
1126              * completion.
1127              */
1128             assert(QLIST_EMPTY(&bs->tracked_requests));
1129             need_drain = false;
1130             break;
1131         }
1132 
1133         if (job_is_ready(&s->common.job) && !should_complete) {
1134             if (s->in_flight == 0 && cnt == 0) {
1135                 trace_mirror_before_sleep(s, cnt, job_is_ready(&s->common.job),
1136                                           BLOCK_JOB_SLICE_TIME);
1137                 job_sleep_ns(&s->common.job, BLOCK_JOB_SLICE_TIME);
1138             }
1139         } else {
1140             block_job_ratelimit_sleep(&s->common);
1141         }
1142         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1143     }
1144 
1145 immediate_exit:
1146     if (s->in_flight > 0) {
1147         /* We get here only if something went wrong.  Either the job failed,
1148          * or it was cancelled prematurely so that we do not guarantee that
1149          * the target is a copy of the source.
1150          */
1151         assert(ret < 0 || job_is_cancelled(&s->common.job));
1152         assert(need_drain);
1153         mirror_wait_for_all_io(s);
1154     }
1155 
1156     assert(s->in_flight == 0);
1157     qemu_vfree(s->buf);
1158     g_free(s->cow_bitmap);
1159     g_free(s->in_flight_bitmap);
1160     bdrv_dirty_iter_free(s->dbi);
1161 
1162     if (need_drain) {
1163         s->in_drain = true;
1164         bdrv_drained_begin(bs);
1165     }
1166 
1167     return ret;
1168 }
1169 
1170 static void mirror_complete(Job *job, Error **errp)
1171 {
1172     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1173 
1174     if (!job_is_ready(job)) {
1175         error_setg(errp, "The active block job '%s' cannot be completed",
1176                    job->id);
1177         return;
1178     }
1179 
1180     /* block all operations on to_replace bs */
1181     if (s->replaces) {
1182         s->to_replace = bdrv_find_node(s->replaces);
1183         if (!s->to_replace) {
1184             error_setg(errp, "Node name '%s' not found", s->replaces);
1185             return;
1186         }
1187 
1188         /* TODO Translate this into child freeze system. */
1189         error_setg(&s->replace_blocker,
1190                    "block device is in use by block-job-complete");
1191         bdrv_op_block_all(s->to_replace, s->replace_blocker);
1192         bdrv_ref(s->to_replace);
1193     }
1194 
1195     s->should_complete = true;
1196 
1197     /* If the job is paused, it will be re-entered when it is resumed */
1198     WITH_JOB_LOCK_GUARD() {
1199         if (!job->paused) {
1200             job_enter_cond_locked(job, NULL);
1201         }
1202     }
1203 }
1204 
1205 static void coroutine_fn mirror_pause(Job *job)
1206 {
1207     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1208 
1209     mirror_wait_for_all_io(s);
1210 }
1211 
1212 static bool mirror_drained_poll(BlockJob *job)
1213 {
1214     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1215 
1216     /* If the job isn't paused nor cancelled, we can't be sure that it won't
1217      * issue more requests. We make an exception if we've reached this point
1218      * from one of our own drain sections, to avoid a deadlock waiting for
1219      * ourselves.
1220      */
1221     WITH_JOB_LOCK_GUARD() {
1222         if (!s->common.job.paused && !job_is_cancelled_locked(&job->job)
1223             && !s->in_drain) {
1224             return true;
1225         }
1226     }
1227 
1228     return !!s->in_flight;
1229 }
1230 
1231 static bool mirror_cancel(Job *job, bool force)
1232 {
1233     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1234     BlockDriverState *target = blk_bs(s->target);
1235 
1236     /*
1237      * Before the job is READY, we treat any cancellation like a
1238      * force-cancellation.
1239      */
1240     force = force || !job_is_ready(job);
1241 
1242     if (force) {
1243         bdrv_cancel_in_flight(target);
1244     }
1245     return force;
1246 }
1247 
1248 static bool commit_active_cancel(Job *job, bool force)
1249 {
1250     /* Same as above in mirror_cancel() */
1251     return force || !job_is_ready(job);
1252 }
1253 
1254 static void mirror_change(BlockJob *job, BlockJobChangeOptions *opts,
1255                           Error **errp)
1256 {
1257     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1258     BlockJobChangeOptionsMirror *change_opts = &opts->u.mirror;
1259     MirrorCopyMode current;
1260 
1261     /*
1262      * The implementation relies on the fact that copy_mode is only written
1263      * under the BQL. Otherwise, further synchronization would be required.
1264      */
1265 
1266     GLOBAL_STATE_CODE();
1267 
1268     if (qatomic_read(&s->copy_mode) == change_opts->copy_mode) {
1269         return;
1270     }
1271 
1272     if (change_opts->copy_mode != MIRROR_COPY_MODE_WRITE_BLOCKING) {
1273         error_setg(errp, "Change to copy mode '%s' is not implemented",
1274                    MirrorCopyMode_str(change_opts->copy_mode));
1275         return;
1276     }
1277 
1278     current = qatomic_cmpxchg(&s->copy_mode, MIRROR_COPY_MODE_BACKGROUND,
1279                               change_opts->copy_mode);
1280     if (current != MIRROR_COPY_MODE_BACKGROUND) {
1281         error_setg(errp, "Expected current copy mode '%s', got '%s'",
1282                    MirrorCopyMode_str(MIRROR_COPY_MODE_BACKGROUND),
1283                    MirrorCopyMode_str(current));
1284     }
1285 }
1286 
1287 static void mirror_query(BlockJob *job, BlockJobInfo *info)
1288 {
1289     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1290 
1291     info->u.mirror = (BlockJobInfoMirror) {
1292         .actively_synced = qatomic_read(&s->actively_synced),
1293     };
1294 }
1295 
1296 static const BlockJobDriver mirror_job_driver = {
1297     .job_driver = {
1298         .instance_size          = sizeof(MirrorBlockJob),
1299         .job_type               = JOB_TYPE_MIRROR,
1300         .free                   = block_job_free,
1301         .user_resume            = block_job_user_resume,
1302         .run                    = mirror_run,
1303         .prepare                = mirror_prepare,
1304         .abort                  = mirror_abort,
1305         .pause                  = mirror_pause,
1306         .complete               = mirror_complete,
1307         .cancel                 = mirror_cancel,
1308     },
1309     .drained_poll           = mirror_drained_poll,
1310     .change                 = mirror_change,
1311     .query                  = mirror_query,
1312 };
1313 
1314 static const BlockJobDriver commit_active_job_driver = {
1315     .job_driver = {
1316         .instance_size          = sizeof(MirrorBlockJob),
1317         .job_type               = JOB_TYPE_COMMIT,
1318         .free                   = block_job_free,
1319         .user_resume            = block_job_user_resume,
1320         .run                    = mirror_run,
1321         .prepare                = mirror_prepare,
1322         .abort                  = mirror_abort,
1323         .pause                  = mirror_pause,
1324         .complete               = mirror_complete,
1325         .cancel                 = commit_active_cancel,
1326     },
1327     .drained_poll           = mirror_drained_poll,
1328 };
1329 
1330 static void coroutine_fn
1331 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method,
1332                      uint64_t offset, uint64_t bytes,
1333                      QEMUIOVector *qiov, int flags)
1334 {
1335     int ret;
1336     size_t qiov_offset = 0;
1337     int64_t bitmap_offset, bitmap_end;
1338 
1339     if (!QEMU_IS_ALIGNED(offset, job->granularity) &&
1340         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset))
1341     {
1342             /*
1343              * Dirty unaligned padding: ignore it.
1344              *
1345              * Reasoning:
1346              * 1. If we copy it, we can't reset corresponding bit in
1347              *    dirty_bitmap as there may be some "dirty" bytes still not
1348              *    copied.
1349              * 2. It's already dirty, so skipping it we don't diverge mirror
1350              *    progress.
1351              *
1352              * Note, that because of this, guest write may have no contribution
1353              * into mirror converge, but that's not bad, as we have background
1354              * process of mirroring. If under some bad circumstances (high guest
1355              * IO load) background process starve, we will not converge anyway,
1356              * even if each write will contribute, as guest is not guaranteed to
1357              * rewrite the whole disk.
1358              */
1359             qiov_offset = QEMU_ALIGN_UP(offset, job->granularity) - offset;
1360             if (bytes <= qiov_offset) {
1361                 /* nothing to do after shrink */
1362                 return;
1363             }
1364             offset += qiov_offset;
1365             bytes -= qiov_offset;
1366     }
1367 
1368     if (!QEMU_IS_ALIGNED(offset + bytes, job->granularity) &&
1369         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset + bytes - 1))
1370     {
1371         uint64_t tail = (offset + bytes) % job->granularity;
1372 
1373         if (bytes <= tail) {
1374             /* nothing to do after shrink */
1375             return;
1376         }
1377         bytes -= tail;
1378     }
1379 
1380     /*
1381      * Tails are either clean or shrunk, so for bitmap resetting
1382      * we safely align the range down.
1383      */
1384     bitmap_offset = QEMU_ALIGN_UP(offset, job->granularity);
1385     bitmap_end = QEMU_ALIGN_DOWN(offset + bytes, job->granularity);
1386     if (bitmap_offset < bitmap_end) {
1387         bdrv_reset_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1388                                 bitmap_end - bitmap_offset);
1389     }
1390 
1391     job_progress_increase_remaining(&job->common.job, bytes);
1392     job->active_write_bytes_in_flight += bytes;
1393 
1394     switch (method) {
1395     case MIRROR_METHOD_COPY:
1396         ret = blk_co_pwritev_part(job->target, offset, bytes,
1397                                   qiov, qiov_offset, flags);
1398         break;
1399 
1400     case MIRROR_METHOD_ZERO:
1401         assert(!qiov);
1402         ret = blk_co_pwrite_zeroes(job->target, offset, bytes, flags);
1403         break;
1404 
1405     case MIRROR_METHOD_DISCARD:
1406         assert(!qiov);
1407         ret = blk_co_pdiscard(job->target, offset, bytes);
1408         break;
1409 
1410     default:
1411         abort();
1412     }
1413 
1414     job->active_write_bytes_in_flight -= bytes;
1415     if (ret >= 0) {
1416         job_progress_update(&job->common.job, bytes);
1417     } else {
1418         BlockErrorAction action;
1419 
1420         /*
1421          * We failed, so we should mark dirty the whole area, aligned up.
1422          * Note that we don't care about shrunk tails if any: they were dirty
1423          * at function start, and they must be still dirty, as we've locked
1424          * the region for in-flight op.
1425          */
1426         bitmap_offset = QEMU_ALIGN_DOWN(offset, job->granularity);
1427         bitmap_end = QEMU_ALIGN_UP(offset + bytes, job->granularity);
1428         bdrv_set_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1429                               bitmap_end - bitmap_offset);
1430         qatomic_set(&job->actively_synced, false);
1431 
1432         action = mirror_error_action(job, false, -ret);
1433         if (action == BLOCK_ERROR_ACTION_REPORT) {
1434             if (!job->ret) {
1435                 job->ret = ret;
1436             }
1437         }
1438     }
1439 }
1440 
1441 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s,
1442                                                    uint64_t offset,
1443                                                    uint64_t bytes)
1444 {
1445     MirrorOp *op;
1446     uint64_t start_chunk = offset / s->granularity;
1447     uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
1448 
1449     op = g_new(MirrorOp, 1);
1450     *op = (MirrorOp){
1451         .s                  = s,
1452         .offset             = offset,
1453         .bytes              = bytes,
1454         .is_active_write    = true,
1455         .is_in_flight       = true,
1456         .co                 = qemu_coroutine_self(),
1457     };
1458     qemu_co_queue_init(&op->waiting_requests);
1459     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
1460 
1461     s->in_active_write_counter++;
1462 
1463     /*
1464      * Wait for concurrent requests affecting the area.  If there are already
1465      * running requests that are copying off now-to-be stale data in the area,
1466      * we must wait for them to finish before we begin writing fresh data to the
1467      * target so that the write operations appear in the correct order.
1468      * Note that background requests (see mirror_iteration()) in contrast only
1469      * wait for conflicting requests at the start of the dirty area, and then
1470      * (based on the in_flight_bitmap) truncate the area to copy so it will not
1471      * conflict with any requests beyond that.  For active writes, however, we
1472      * cannot truncate that area.  The request from our parent must be blocked
1473      * until the area is copied in full.  Therefore, we must wait for the whole
1474      * area to become free of concurrent requests.
1475      */
1476     mirror_wait_on_conflicts(op, s, offset, bytes);
1477 
1478     bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1479 
1480     return op;
1481 }
1482 
1483 static void coroutine_fn GRAPH_RDLOCK active_write_settle(MirrorOp *op)
1484 {
1485     uint64_t start_chunk = op->offset / op->s->granularity;
1486     uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes,
1487                                       op->s->granularity);
1488 
1489     if (!--op->s->in_active_write_counter &&
1490         qatomic_read(&op->s->actively_synced)) {
1491         BdrvChild *source = op->s->mirror_top_bs->backing;
1492 
1493         if (QLIST_FIRST(&source->bs->parents) == source &&
1494             QLIST_NEXT(source, next_parent) == NULL)
1495         {
1496             /* Assert that we are back in sync once all active write
1497              * operations are settled.
1498              * Note that we can only assert this if the mirror node
1499              * is the source node's only parent. */
1500             assert(!bdrv_get_dirty_count(op->s->dirty_bitmap));
1501         }
1502     }
1503     bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1504     QTAILQ_REMOVE(&op->s->ops_in_flight, op, next);
1505     qemu_co_queue_restart_all(&op->waiting_requests);
1506     g_free(op);
1507 }
1508 
1509 static int coroutine_fn GRAPH_RDLOCK
1510 bdrv_mirror_top_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
1511                        QEMUIOVector *qiov, BdrvRequestFlags flags)
1512 {
1513     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1514 }
1515 
1516 static bool should_copy_to_target(MirrorBDSOpaque *s)
1517 {
1518     return s->job && s->job->ret >= 0 &&
1519         !job_is_cancelled(&s->job->common.job) &&
1520         qatomic_read(&s->job->copy_mode) == MIRROR_COPY_MODE_WRITE_BLOCKING;
1521 }
1522 
1523 static int coroutine_fn GRAPH_RDLOCK
1524 bdrv_mirror_top_do_write(BlockDriverState *bs, MirrorMethod method,
1525                          bool copy_to_target, uint64_t offset, uint64_t bytes,
1526                          QEMUIOVector *qiov, int flags)
1527 {
1528     MirrorOp *op = NULL;
1529     MirrorBDSOpaque *s = bs->opaque;
1530     int ret = 0;
1531 
1532     if (copy_to_target) {
1533         op = active_write_prepare(s->job, offset, bytes);
1534     }
1535 
1536     switch (method) {
1537     case MIRROR_METHOD_COPY:
1538         ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1539         break;
1540 
1541     case MIRROR_METHOD_ZERO:
1542         ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1543         break;
1544 
1545     case MIRROR_METHOD_DISCARD:
1546         ret = bdrv_co_pdiscard(bs->backing, offset, bytes);
1547         break;
1548 
1549     default:
1550         abort();
1551     }
1552 
1553     if (!copy_to_target && s->job && s->job->dirty_bitmap) {
1554         qatomic_set(&s->job->actively_synced, false);
1555         bdrv_set_dirty_bitmap(s->job->dirty_bitmap, offset, bytes);
1556     }
1557 
1558     if (ret < 0) {
1559         goto out;
1560     }
1561 
1562     if (copy_to_target) {
1563         do_sync_target_write(s->job, method, offset, bytes, qiov, flags);
1564     }
1565 
1566 out:
1567     if (copy_to_target) {
1568         active_write_settle(op);
1569     }
1570     return ret;
1571 }
1572 
1573 static int coroutine_fn GRAPH_RDLOCK
1574 bdrv_mirror_top_pwritev(BlockDriverState *bs, int64_t offset, int64_t bytes,
1575                         QEMUIOVector *qiov, BdrvRequestFlags flags)
1576 {
1577     QEMUIOVector bounce_qiov;
1578     void *bounce_buf;
1579     int ret = 0;
1580     bool copy_to_target = should_copy_to_target(bs->opaque);
1581 
1582     if (copy_to_target) {
1583         /* The guest might concurrently modify the data to write; but
1584          * the data on source and destination must match, so we have
1585          * to use a bounce buffer if we are going to write to the
1586          * target now. */
1587         bounce_buf = qemu_blockalign(bs, bytes);
1588         iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes);
1589 
1590         qemu_iovec_init(&bounce_qiov, 1);
1591         qemu_iovec_add(&bounce_qiov, bounce_buf, bytes);
1592         qiov = &bounce_qiov;
1593 
1594         flags &= ~BDRV_REQ_REGISTERED_BUF;
1595     }
1596 
1597     ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, copy_to_target,
1598                                    offset, bytes, qiov, flags);
1599 
1600     if (copy_to_target) {
1601         qemu_iovec_destroy(&bounce_qiov);
1602         qemu_vfree(bounce_buf);
1603     }
1604 
1605     return ret;
1606 }
1607 
1608 static int coroutine_fn GRAPH_RDLOCK bdrv_mirror_top_flush(BlockDriverState *bs)
1609 {
1610     if (bs->backing == NULL) {
1611         /* we can be here after failed bdrv_append in mirror_start_job */
1612         return 0;
1613     }
1614     return bdrv_co_flush(bs->backing->bs);
1615 }
1616 
1617 static int coroutine_fn GRAPH_RDLOCK
1618 bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs, int64_t offset,
1619                               int64_t bytes, BdrvRequestFlags flags)
1620 {
1621     bool copy_to_target = should_copy_to_target(bs->opaque);
1622     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, copy_to_target,
1623                                     offset, bytes, NULL, flags);
1624 }
1625 
1626 static int coroutine_fn GRAPH_RDLOCK
1627 bdrv_mirror_top_pdiscard(BlockDriverState *bs, int64_t offset, int64_t bytes)
1628 {
1629     bool copy_to_target = should_copy_to_target(bs->opaque);
1630     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, copy_to_target,
1631                                     offset, bytes, NULL, 0);
1632 }
1633 
1634 static void GRAPH_RDLOCK bdrv_mirror_top_refresh_filename(BlockDriverState *bs)
1635 {
1636     if (bs->backing == NULL) {
1637         /* we can be here after failed bdrv_attach_child in
1638          * bdrv_set_backing_hd */
1639         return;
1640     }
1641     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1642             bs->backing->bs->filename);
1643 }
1644 
1645 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1646                                        BdrvChildRole role,
1647                                        BlockReopenQueue *reopen_queue,
1648                                        uint64_t perm, uint64_t shared,
1649                                        uint64_t *nperm, uint64_t *nshared)
1650 {
1651     MirrorBDSOpaque *s = bs->opaque;
1652 
1653     if (s->stop) {
1654         /*
1655          * If the job is to be stopped, we do not need to forward
1656          * anything to the real image.
1657          */
1658         *nperm = 0;
1659         *nshared = BLK_PERM_ALL;
1660         return;
1661     }
1662 
1663     bdrv_default_perms(bs, c, role, reopen_queue,
1664                        perm, shared, nperm, nshared);
1665 
1666     if (s->is_commit) {
1667         /*
1668          * For commit jobs, we cannot take CONSISTENT_READ, because
1669          * that permission is unshared for everything above the base
1670          * node (except for filters on the base node).
1671          * We also have to force-share the WRITE permission, or
1672          * otherwise we would block ourselves at the base node (if
1673          * writes are blocked for a node, they are also blocked for
1674          * its backing file).
1675          * (We could also share RESIZE, because it may be needed for
1676          * the target if its size is less than the top node's; but
1677          * bdrv_default_perms_for_cow() automatically shares RESIZE
1678          * for backing nodes if WRITE is shared, so there is no need
1679          * to do it here.)
1680          */
1681         *nperm &= ~BLK_PERM_CONSISTENT_READ;
1682         *nshared |= BLK_PERM_WRITE;
1683     }
1684 }
1685 
1686 /* Dummy node that provides consistent read to its users without requiring it
1687  * from its backing file and that allows writes on the backing file chain. */
1688 static BlockDriver bdrv_mirror_top = {
1689     .format_name                = "mirror_top",
1690     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1691     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1692     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1693     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1694     .bdrv_co_flush              = bdrv_mirror_top_flush,
1695     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1696     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1697 
1698     .is_filter                  = true,
1699     .filtered_child_is_backing  = true,
1700 };
1701 
1702 static BlockJob *mirror_start_job(
1703                              const char *job_id, BlockDriverState *bs,
1704                              int creation_flags, BlockDriverState *target,
1705                              const char *replaces, int64_t speed,
1706                              uint32_t granularity, int64_t buf_size,
1707                              BlockMirrorBackingMode backing_mode,
1708                              bool zero_target,
1709                              BlockdevOnError on_source_error,
1710                              BlockdevOnError on_target_error,
1711                              bool unmap,
1712                              BlockCompletionFunc *cb,
1713                              void *opaque,
1714                              const BlockJobDriver *driver,
1715                              bool is_none_mode, BlockDriverState *base,
1716                              bool auto_complete, const char *filter_node_name,
1717                              bool is_mirror, MirrorCopyMode copy_mode,
1718                              Error **errp)
1719 {
1720     MirrorBlockJob *s;
1721     MirrorBDSOpaque *bs_opaque;
1722     BlockDriverState *mirror_top_bs;
1723     bool target_is_backing;
1724     uint64_t target_perms, target_shared_perms;
1725     int ret;
1726 
1727     GLOBAL_STATE_CODE();
1728 
1729     if (granularity == 0) {
1730         granularity = bdrv_get_default_bitmap_granularity(target);
1731     }
1732 
1733     assert(is_power_of_2(granularity));
1734 
1735     if (buf_size < 0) {
1736         error_setg(errp, "Invalid parameter 'buf-size'");
1737         return NULL;
1738     }
1739 
1740     if (buf_size == 0) {
1741         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1742     }
1743 
1744     bdrv_graph_rdlock_main_loop();
1745     if (bdrv_skip_filters(bs) == bdrv_skip_filters(target)) {
1746         error_setg(errp, "Can't mirror node into itself");
1747         bdrv_graph_rdunlock_main_loop();
1748         return NULL;
1749     }
1750 
1751     target_is_backing = bdrv_chain_contains(bs, target);
1752     bdrv_graph_rdunlock_main_loop();
1753 
1754     /* In the case of active commit, add dummy driver to provide consistent
1755      * reads on the top, while disabling it in the intermediate nodes, and make
1756      * the backing chain writable. */
1757     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1758                                          BDRV_O_RDWR, errp);
1759     if (mirror_top_bs == NULL) {
1760         return NULL;
1761     }
1762     if (!filter_node_name) {
1763         mirror_top_bs->implicit = true;
1764     }
1765 
1766     /* So that we can always drop this node */
1767     mirror_top_bs->never_freeze = true;
1768 
1769     mirror_top_bs->total_sectors = bs->total_sectors;
1770     mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
1771     mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
1772                                           BDRV_REQ_NO_FALLBACK;
1773     bs_opaque = g_new0(MirrorBDSOpaque, 1);
1774     mirror_top_bs->opaque = bs_opaque;
1775 
1776     bs_opaque->is_commit = target_is_backing;
1777 
1778     bdrv_drained_begin(bs);
1779     ret = bdrv_append(mirror_top_bs, bs, errp);
1780     bdrv_drained_end(bs);
1781 
1782     if (ret < 0) {
1783         bdrv_unref(mirror_top_bs);
1784         return NULL;
1785     }
1786 
1787     /* Make sure that the source is not resized while the job is running */
1788     s = block_job_create(job_id, driver, NULL, mirror_top_bs,
1789                          BLK_PERM_CONSISTENT_READ,
1790                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1791                          BLK_PERM_WRITE, speed,
1792                          creation_flags, cb, opaque, errp);
1793     if (!s) {
1794         goto fail;
1795     }
1796 
1797     /* The block job now has a reference to this node */
1798     bdrv_unref(mirror_top_bs);
1799 
1800     s->mirror_top_bs = mirror_top_bs;
1801 
1802     /* No resize for the target either; while the mirror is still running, a
1803      * consistent read isn't necessarily possible. We could possibly allow
1804      * writes and graph modifications, though it would likely defeat the
1805      * purpose of a mirror, so leave them blocked for now.
1806      *
1807      * In the case of active commit, things look a bit different, though,
1808      * because the target is an already populated backing file in active use.
1809      * We can allow anything except resize there.*/
1810 
1811     target_perms = BLK_PERM_WRITE;
1812     target_shared_perms = BLK_PERM_WRITE_UNCHANGED;
1813 
1814     if (target_is_backing) {
1815         int64_t bs_size, target_size;
1816         bs_size = bdrv_getlength(bs);
1817         if (bs_size < 0) {
1818             error_setg_errno(errp, -bs_size,
1819                              "Could not inquire top image size");
1820             goto fail;
1821         }
1822 
1823         target_size = bdrv_getlength(target);
1824         if (target_size < 0) {
1825             error_setg_errno(errp, -target_size,
1826                              "Could not inquire base image size");
1827             goto fail;
1828         }
1829 
1830         if (target_size < bs_size) {
1831             target_perms |= BLK_PERM_RESIZE;
1832         }
1833 
1834         target_shared_perms |= BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE;
1835     } else {
1836         bdrv_graph_rdlock_main_loop();
1837         if (bdrv_chain_contains(bs, bdrv_skip_filters(target))) {
1838             /*
1839              * We may want to allow this in the future, but it would
1840              * require taking some extra care.
1841              */
1842             error_setg(errp, "Cannot mirror to a filter on top of a node in "
1843                        "the source's backing chain");
1844             bdrv_graph_rdunlock_main_loop();
1845             goto fail;
1846         }
1847         bdrv_graph_rdunlock_main_loop();
1848     }
1849 
1850     s->target = blk_new(s->common.job.aio_context,
1851                         target_perms, target_shared_perms);
1852     ret = blk_insert_bs(s->target, target, errp);
1853     if (ret < 0) {
1854         goto fail;
1855     }
1856     if (is_mirror) {
1857         /* XXX: Mirror target could be a NBD server of target QEMU in the case
1858          * of non-shared block migration. To allow migration completion, we
1859          * have to allow "inactivate" of the target BB.  When that happens, we
1860          * know the job is drained, and the vcpus are stopped, so no write
1861          * operation will be performed. Block layer already has assertions to
1862          * ensure that. */
1863         blk_set_force_allow_inactivate(s->target);
1864     }
1865     blk_set_allow_aio_context_change(s->target, true);
1866     blk_set_disable_request_queuing(s->target, true);
1867 
1868     bdrv_graph_rdlock_main_loop();
1869     s->replaces = g_strdup(replaces);
1870     s->on_source_error = on_source_error;
1871     s->on_target_error = on_target_error;
1872     s->is_none_mode = is_none_mode;
1873     s->backing_mode = backing_mode;
1874     s->zero_target = zero_target;
1875     qatomic_set(&s->copy_mode, copy_mode);
1876     s->base = base;
1877     s->base_overlay = bdrv_find_overlay(bs, base);
1878     s->granularity = granularity;
1879     s->buf_size = ROUND_UP(buf_size, granularity);
1880     s->unmap = unmap;
1881     if (auto_complete) {
1882         s->should_complete = true;
1883     }
1884     bdrv_graph_rdunlock_main_loop();
1885 
1886     s->dirty_bitmap = bdrv_create_dirty_bitmap(s->mirror_top_bs, granularity,
1887                                                NULL, errp);
1888     if (!s->dirty_bitmap) {
1889         goto fail;
1890     }
1891 
1892     /*
1893      * The dirty bitmap is set by bdrv_mirror_top_do_write() when not in active
1894      * mode.
1895      */
1896     bdrv_disable_dirty_bitmap(s->dirty_bitmap);
1897 
1898     bdrv_graph_wrlock();
1899     ret = block_job_add_bdrv(&s->common, "source", bs, 0,
1900                              BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE |
1901                              BLK_PERM_CONSISTENT_READ,
1902                              errp);
1903     if (ret < 0) {
1904         bdrv_graph_wrunlock();
1905         goto fail;
1906     }
1907 
1908     /* Required permissions are already taken with blk_new() */
1909     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1910                        &error_abort);
1911 
1912     /* In commit_active_start() all intermediate nodes disappear, so
1913      * any jobs in them must be blocked */
1914     if (target_is_backing) {
1915         BlockDriverState *iter, *filtered_target;
1916         uint64_t iter_shared_perms;
1917 
1918         /*
1919          * The topmost node with
1920          * bdrv_skip_filters(filtered_target) == bdrv_skip_filters(target)
1921          */
1922         filtered_target = bdrv_cow_bs(bdrv_find_overlay(bs, target));
1923 
1924         assert(bdrv_skip_filters(filtered_target) ==
1925                bdrv_skip_filters(target));
1926 
1927         /*
1928          * XXX BLK_PERM_WRITE needs to be allowed so we don't block
1929          * ourselves at s->base (if writes are blocked for a node, they are
1930          * also blocked for its backing file). The other options would be a
1931          * second filter driver above s->base (== target).
1932          */
1933         iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE;
1934 
1935         for (iter = bdrv_filter_or_cow_bs(bs); iter != target;
1936              iter = bdrv_filter_or_cow_bs(iter))
1937         {
1938             if (iter == filtered_target) {
1939                 /*
1940                  * From here on, all nodes are filters on the base.
1941                  * This allows us to share BLK_PERM_CONSISTENT_READ.
1942                  */
1943                 iter_shared_perms |= BLK_PERM_CONSISTENT_READ;
1944             }
1945 
1946             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1947                                      iter_shared_perms, errp);
1948             if (ret < 0) {
1949                 bdrv_graph_wrunlock();
1950                 goto fail;
1951             }
1952         }
1953 
1954         if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) {
1955             bdrv_graph_wrunlock();
1956             goto fail;
1957         }
1958     }
1959     bdrv_graph_wrunlock();
1960 
1961     QTAILQ_INIT(&s->ops_in_flight);
1962 
1963     trace_mirror_start(bs, s, opaque);
1964     job_start(&s->common.job);
1965 
1966     return &s->common;
1967 
1968 fail:
1969     if (s) {
1970         /* Make sure this BDS does not go away until we have completed the graph
1971          * changes below */
1972         bdrv_ref(mirror_top_bs);
1973 
1974         g_free(s->replaces);
1975         blk_unref(s->target);
1976         bs_opaque->job = NULL;
1977         if (s->dirty_bitmap) {
1978             bdrv_release_dirty_bitmap(s->dirty_bitmap);
1979         }
1980         job_early_fail(&s->common.job);
1981     }
1982 
1983     bs_opaque->stop = true;
1984     bdrv_drained_begin(bs);
1985     bdrv_graph_wrlock();
1986     assert(mirror_top_bs->backing->bs == bs);
1987     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
1988                              &error_abort);
1989     bdrv_replace_node(mirror_top_bs, bs, &error_abort);
1990     bdrv_graph_wrunlock();
1991     bdrv_drained_end(bs);
1992 
1993     bdrv_unref(mirror_top_bs);
1994 
1995     return NULL;
1996 }
1997 
1998 void mirror_start(const char *job_id, BlockDriverState *bs,
1999                   BlockDriverState *target, const char *replaces,
2000                   int creation_flags, int64_t speed,
2001                   uint32_t granularity, int64_t buf_size,
2002                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
2003                   bool zero_target,
2004                   BlockdevOnError on_source_error,
2005                   BlockdevOnError on_target_error,
2006                   bool unmap, const char *filter_node_name,
2007                   MirrorCopyMode copy_mode, Error **errp)
2008 {
2009     bool is_none_mode;
2010     BlockDriverState *base;
2011 
2012     GLOBAL_STATE_CODE();
2013 
2014     if ((mode == MIRROR_SYNC_MODE_INCREMENTAL) ||
2015         (mode == MIRROR_SYNC_MODE_BITMAP)) {
2016         error_setg(errp, "Sync mode '%s' not supported",
2017                    MirrorSyncMode_str(mode));
2018         return;
2019     }
2020 
2021     bdrv_graph_rdlock_main_loop();
2022     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
2023     base = mode == MIRROR_SYNC_MODE_TOP ? bdrv_backing_chain_next(bs) : NULL;
2024     bdrv_graph_rdunlock_main_loop();
2025 
2026     mirror_start_job(job_id, bs, creation_flags, target, replaces,
2027                      speed, granularity, buf_size, backing_mode, zero_target,
2028                      on_source_error, on_target_error, unmap, NULL, NULL,
2029                      &mirror_job_driver, is_none_mode, base, false,
2030                      filter_node_name, true, copy_mode, errp);
2031 }
2032 
2033 BlockJob *commit_active_start(const char *job_id, BlockDriverState *bs,
2034                               BlockDriverState *base, int creation_flags,
2035                               int64_t speed, BlockdevOnError on_error,
2036                               const char *filter_node_name,
2037                               BlockCompletionFunc *cb, void *opaque,
2038                               bool auto_complete, Error **errp)
2039 {
2040     bool base_read_only;
2041     BlockJob *job;
2042 
2043     GLOBAL_STATE_CODE();
2044 
2045     base_read_only = bdrv_is_read_only(base);
2046 
2047     if (base_read_only) {
2048         if (bdrv_reopen_set_read_only(base, false, errp) < 0) {
2049             return NULL;
2050         }
2051     }
2052 
2053     job = mirror_start_job(
2054                      job_id, bs, creation_flags, base, NULL, speed, 0, 0,
2055                      MIRROR_LEAVE_BACKING_CHAIN, false,
2056                      on_error, on_error, true, cb, opaque,
2057                      &commit_active_job_driver, false, base, auto_complete,
2058                      filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND,
2059                      errp);
2060     if (!job) {
2061         goto error_restore_flags;
2062     }
2063 
2064     return job;
2065 
2066 error_restore_flags:
2067     /* ignore error and errp for bdrv_reopen, because we want to propagate
2068      * the original error */
2069     if (base_read_only) {
2070         bdrv_reopen_set_read_only(base, true, NULL);
2071     }
2072     return NULL;
2073 }
2074