xref: /openbmc/qemu/block/mirror.c (revision 464e447a)
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 "sysemu/block-backend.h"
22 #include "qapi/error.h"
23 #include "qapi/qmp/qerror.h"
24 #include "qemu/ratelimit.h"
25 #include "qemu/bitmap.h"
26 
27 #define MAX_IN_FLIGHT 16
28 #define MAX_IO_BYTES (1 << 20) /* 1 Mb */
29 #define DEFAULT_MIRROR_BUF_SIZE (MAX_IN_FLIGHT * MAX_IO_BYTES)
30 
31 /* The mirroring buffer is a list of granularity-sized chunks.
32  * Free chunks are organized in a list.
33  */
34 typedef struct MirrorBuffer {
35     QSIMPLEQ_ENTRY(MirrorBuffer) next;
36 } MirrorBuffer;
37 
38 typedef struct MirrorOp MirrorOp;
39 
40 typedef struct MirrorBlockJob {
41     BlockJob common;
42     BlockBackend *target;
43     BlockDriverState *mirror_top_bs;
44     BlockDriverState *base;
45 
46     /* The name of the graph node to replace */
47     char *replaces;
48     /* The BDS to replace */
49     BlockDriverState *to_replace;
50     /* Used to block operations on the drive-mirror-replace target */
51     Error *replace_blocker;
52     bool is_none_mode;
53     BlockMirrorBackingMode backing_mode;
54     MirrorCopyMode copy_mode;
55     BlockdevOnError on_source_error, on_target_error;
56     bool synced;
57     /* Set when the target is synced (dirty bitmap is clean, nothing
58      * in flight) and the job is running in active mode */
59     bool actively_synced;
60     bool should_complete;
61     int64_t granularity;
62     size_t buf_size;
63     int64_t bdev_length;
64     unsigned long *cow_bitmap;
65     BdrvDirtyBitmap *dirty_bitmap;
66     BdrvDirtyBitmapIter *dbi;
67     uint8_t *buf;
68     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
69     int buf_free_count;
70 
71     uint64_t last_pause_ns;
72     unsigned long *in_flight_bitmap;
73     int in_flight;
74     int64_t bytes_in_flight;
75     QTAILQ_HEAD(MirrorOpList, MirrorOp) ops_in_flight;
76     int ret;
77     bool unmap;
78     int target_cluster_size;
79     int max_iov;
80     bool initial_zeroing_ongoing;
81     int in_active_write_counter;
82     bool prepared;
83 } MirrorBlockJob;
84 
85 typedef struct MirrorBDSOpaque {
86     MirrorBlockJob *job;
87 } MirrorBDSOpaque;
88 
89 struct MirrorOp {
90     MirrorBlockJob *s;
91     QEMUIOVector qiov;
92     int64_t offset;
93     uint64_t bytes;
94 
95     /* The pointee is set by mirror_co_read(), mirror_co_zero(), and
96      * mirror_co_discard() before yielding for the first time */
97     int64_t *bytes_handled;
98 
99     bool is_pseudo_op;
100     bool is_active_write;
101     CoQueue waiting_requests;
102 
103     QTAILQ_ENTRY(MirrorOp) next;
104 };
105 
106 typedef enum MirrorMethod {
107     MIRROR_METHOD_COPY,
108     MIRROR_METHOD_ZERO,
109     MIRROR_METHOD_DISCARD,
110 } MirrorMethod;
111 
112 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
113                                             int error)
114 {
115     s->synced = false;
116     s->actively_synced = false;
117     if (read) {
118         return block_job_error_action(&s->common, s->on_source_error,
119                                       true, error);
120     } else {
121         return block_job_error_action(&s->common, s->on_target_error,
122                                       false, error);
123     }
124 }
125 
126 static void coroutine_fn mirror_wait_on_conflicts(MirrorOp *self,
127                                                   MirrorBlockJob *s,
128                                                   uint64_t offset,
129                                                   uint64_t bytes)
130 {
131     uint64_t self_start_chunk = offset / s->granularity;
132     uint64_t self_end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
133     uint64_t self_nb_chunks = self_end_chunk - self_start_chunk;
134 
135     while (find_next_bit(s->in_flight_bitmap, self_end_chunk,
136                          self_start_chunk) < self_end_chunk &&
137            s->ret >= 0)
138     {
139         MirrorOp *op;
140 
141         QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
142             uint64_t op_start_chunk = op->offset / s->granularity;
143             uint64_t op_nb_chunks = DIV_ROUND_UP(op->offset + op->bytes,
144                                                  s->granularity) -
145                                     op_start_chunk;
146 
147             if (op == self) {
148                 continue;
149             }
150 
151             if (ranges_overlap(self_start_chunk, self_nb_chunks,
152                                op_start_chunk, op_nb_chunks))
153             {
154                 qemu_co_queue_wait(&op->waiting_requests, NULL);
155                 break;
156             }
157         }
158     }
159 }
160 
161 static void coroutine_fn mirror_iteration_done(MirrorOp *op, int ret)
162 {
163     MirrorBlockJob *s = op->s;
164     struct iovec *iov;
165     int64_t chunk_num;
166     int i, nb_chunks;
167 
168     trace_mirror_iteration_done(s, op->offset, op->bytes, ret);
169 
170     s->in_flight--;
171     s->bytes_in_flight -= op->bytes;
172     iov = op->qiov.iov;
173     for (i = 0; i < op->qiov.niov; i++) {
174         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
175         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
176         s->buf_free_count++;
177     }
178 
179     chunk_num = op->offset / s->granularity;
180     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
181 
182     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
183     QTAILQ_REMOVE(&s->ops_in_flight, op, next);
184     if (ret >= 0) {
185         if (s->cow_bitmap) {
186             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
187         }
188         if (!s->initial_zeroing_ongoing) {
189             job_progress_update(&s->common.job, op->bytes);
190         }
191     }
192     qemu_iovec_destroy(&op->qiov);
193 
194     qemu_co_queue_restart_all(&op->waiting_requests);
195     g_free(op);
196 }
197 
198 static void coroutine_fn mirror_write_complete(MirrorOp *op, int ret)
199 {
200     MirrorBlockJob *s = op->s;
201 
202     if (ret < 0) {
203         BlockErrorAction action;
204 
205         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
206         action = mirror_error_action(s, false, -ret);
207         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
208             s->ret = ret;
209         }
210     }
211 
212     mirror_iteration_done(op, ret);
213 }
214 
215 static void coroutine_fn mirror_read_complete(MirrorOp *op, int ret)
216 {
217     MirrorBlockJob *s = op->s;
218 
219     if (ret < 0) {
220         BlockErrorAction action;
221 
222         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
223         action = mirror_error_action(s, true, -ret);
224         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
225             s->ret = ret;
226         }
227 
228         mirror_iteration_done(op, ret);
229         return;
230     }
231 
232     ret = blk_co_pwritev(s->target, op->offset, op->qiov.size, &op->qiov, 0);
233     mirror_write_complete(op, ret);
234 }
235 
236 /* Clip bytes relative to offset to not exceed end-of-file */
237 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s,
238                                         int64_t offset,
239                                         int64_t bytes)
240 {
241     return MIN(bytes, s->bdev_length - offset);
242 }
243 
244 /* Round offset and/or bytes to target cluster if COW is needed, and
245  * return the offset of the adjusted tail against original. */
246 static int mirror_cow_align(MirrorBlockJob *s, int64_t *offset,
247                             uint64_t *bytes)
248 {
249     bool need_cow;
250     int ret = 0;
251     int64_t align_offset = *offset;
252     int64_t align_bytes = *bytes;
253     int max_bytes = s->granularity * s->max_iov;
254 
255     need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap);
256     need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity,
257                           s->cow_bitmap);
258     if (need_cow) {
259         bdrv_round_to_clusters(blk_bs(s->target), *offset, *bytes,
260                                &align_offset, &align_bytes);
261     }
262 
263     if (align_bytes > max_bytes) {
264         align_bytes = max_bytes;
265         if (need_cow) {
266             align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size);
267         }
268     }
269     /* Clipping may result in align_bytes unaligned to chunk boundary, but
270      * that doesn't matter because it's already the end of source image. */
271     align_bytes = mirror_clip_bytes(s, align_offset, align_bytes);
272 
273     ret = align_offset + align_bytes - (*offset + *bytes);
274     *offset = align_offset;
275     *bytes = align_bytes;
276     assert(ret >= 0);
277     return ret;
278 }
279 
280 static inline void coroutine_fn
281 mirror_wait_for_any_operation(MirrorBlockJob *s, bool active)
282 {
283     MirrorOp *op;
284 
285     QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
286         /* Do not wait on pseudo ops, because it may in turn wait on
287          * some other operation to start, which may in fact be the
288          * caller of this function.  Since there is only one pseudo op
289          * at any given time, we will always find some real operation
290          * to wait on. */
291         if (!op->is_pseudo_op && op->is_active_write == active) {
292             qemu_co_queue_wait(&op->waiting_requests, NULL);
293             return;
294         }
295     }
296     abort();
297 }
298 
299 static inline void coroutine_fn
300 mirror_wait_for_free_in_flight_slot(MirrorBlockJob *s)
301 {
302     /* Only non-active operations use up in-flight slots */
303     mirror_wait_for_any_operation(s, false);
304 }
305 
306 /* Perform a mirror copy operation.
307  *
308  * *op->bytes_handled is set to the number of bytes copied after and
309  * including offset, excluding any bytes copied prior to offset due
310  * to alignment.  This will be op->bytes if no alignment is necessary,
311  * or (new_end - op->offset) if the tail is rounded up or down due to
312  * alignment or buffer limit.
313  */
314 static void coroutine_fn mirror_co_read(void *opaque)
315 {
316     MirrorOp *op = opaque;
317     MirrorBlockJob *s = op->s;
318     int nb_chunks;
319     uint64_t ret;
320     uint64_t max_bytes;
321 
322     max_bytes = s->granularity * s->max_iov;
323 
324     /* We can only handle as much as buf_size at a time. */
325     op->bytes = MIN(s->buf_size, MIN(max_bytes, op->bytes));
326     assert(op->bytes);
327     assert(op->bytes < BDRV_REQUEST_MAX_BYTES);
328     *op->bytes_handled = op->bytes;
329 
330     if (s->cow_bitmap) {
331         *op->bytes_handled += mirror_cow_align(s, &op->offset, &op->bytes);
332     }
333     /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */
334     assert(*op->bytes_handled <= UINT_MAX);
335     assert(op->bytes <= s->buf_size);
336     /* The offset is granularity-aligned because:
337      * 1) Caller passes in aligned values;
338      * 2) mirror_cow_align is used only when target cluster is larger. */
339     assert(QEMU_IS_ALIGNED(op->offset, s->granularity));
340     /* The range is sector-aligned, since bdrv_getlength() rounds up. */
341     assert(QEMU_IS_ALIGNED(op->bytes, BDRV_SECTOR_SIZE));
342     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
343 
344     while (s->buf_free_count < nb_chunks) {
345         trace_mirror_yield_in_flight(s, op->offset, s->in_flight);
346         mirror_wait_for_free_in_flight_slot(s);
347     }
348 
349     /* Now make a QEMUIOVector taking enough granularity-sized chunks
350      * from s->buf_free.
351      */
352     qemu_iovec_init(&op->qiov, nb_chunks);
353     while (nb_chunks-- > 0) {
354         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
355         size_t remaining = op->bytes - op->qiov.size;
356 
357         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
358         s->buf_free_count--;
359         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
360     }
361 
362     /* Copy the dirty cluster.  */
363     s->in_flight++;
364     s->bytes_in_flight += op->bytes;
365     trace_mirror_one_iteration(s, op->offset, op->bytes);
366 
367     ret = bdrv_co_preadv(s->mirror_top_bs->backing, op->offset, op->bytes,
368                          &op->qiov, 0);
369     mirror_read_complete(op, ret);
370 }
371 
372 static void coroutine_fn mirror_co_zero(void *opaque)
373 {
374     MirrorOp *op = opaque;
375     int ret;
376 
377     op->s->in_flight++;
378     op->s->bytes_in_flight += op->bytes;
379     *op->bytes_handled = op->bytes;
380 
381     ret = blk_co_pwrite_zeroes(op->s->target, op->offset, op->bytes,
382                                op->s->unmap ? BDRV_REQ_MAY_UNMAP : 0);
383     mirror_write_complete(op, ret);
384 }
385 
386 static void coroutine_fn mirror_co_discard(void *opaque)
387 {
388     MirrorOp *op = opaque;
389     int ret;
390 
391     op->s->in_flight++;
392     op->s->bytes_in_flight += op->bytes;
393     *op->bytes_handled = op->bytes;
394 
395     ret = blk_co_pdiscard(op->s->target, op->offset, op->bytes);
396     mirror_write_complete(op, ret);
397 }
398 
399 static unsigned mirror_perform(MirrorBlockJob *s, int64_t offset,
400                                unsigned bytes, MirrorMethod mirror_method)
401 {
402     MirrorOp *op;
403     Coroutine *co;
404     int64_t bytes_handled = -1;
405 
406     op = g_new(MirrorOp, 1);
407     *op = (MirrorOp){
408         .s              = s,
409         .offset         = offset,
410         .bytes          = bytes,
411         .bytes_handled  = &bytes_handled,
412     };
413     qemu_co_queue_init(&op->waiting_requests);
414 
415     switch (mirror_method) {
416     case MIRROR_METHOD_COPY:
417         co = qemu_coroutine_create(mirror_co_read, op);
418         break;
419     case MIRROR_METHOD_ZERO:
420         co = qemu_coroutine_create(mirror_co_zero, op);
421         break;
422     case MIRROR_METHOD_DISCARD:
423         co = qemu_coroutine_create(mirror_co_discard, op);
424         break;
425     default:
426         abort();
427     }
428 
429     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
430     qemu_coroutine_enter(co);
431     /* At this point, ownership of op has been moved to the coroutine
432      * and the object may already be freed */
433 
434     /* Assert that this value has been set */
435     assert(bytes_handled >= 0);
436 
437     /* Same assertion as in mirror_co_read() (and for mirror_co_read()
438      * and mirror_co_discard(), bytes_handled == op->bytes, which
439      * is the @bytes parameter given to this function) */
440     assert(bytes_handled <= UINT_MAX);
441     return bytes_handled;
442 }
443 
444 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
445 {
446     BlockDriverState *source = s->mirror_top_bs->backing->bs;
447     MirrorOp *pseudo_op;
448     int64_t offset;
449     uint64_t delay_ns = 0, ret = 0;
450     /* At least the first dirty chunk is mirrored in one iteration. */
451     int nb_chunks = 1;
452     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
453     int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES);
454 
455     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
456     offset = bdrv_dirty_iter_next(s->dbi);
457     if (offset < 0) {
458         bdrv_set_dirty_iter(s->dbi, 0);
459         offset = bdrv_dirty_iter_next(s->dbi);
460         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
461         assert(offset >= 0);
462     }
463     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
464 
465     mirror_wait_on_conflicts(NULL, s, offset, 1);
466 
467     job_pause_point(&s->common.job);
468 
469     /* Find the number of consective dirty chunks following the first dirty
470      * one, and wait for in flight requests in them. */
471     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
472     while (nb_chunks * s->granularity < s->buf_size) {
473         int64_t next_dirty;
474         int64_t next_offset = offset + nb_chunks * s->granularity;
475         int64_t next_chunk = next_offset / s->granularity;
476         if (next_offset >= s->bdev_length ||
477             !bdrv_get_dirty_locked(source, s->dirty_bitmap, next_offset)) {
478             break;
479         }
480         if (test_bit(next_chunk, s->in_flight_bitmap)) {
481             break;
482         }
483 
484         next_dirty = bdrv_dirty_iter_next(s->dbi);
485         if (next_dirty > next_offset || next_dirty < 0) {
486             /* The bitmap iterator's cache is stale, refresh it */
487             bdrv_set_dirty_iter(s->dbi, next_offset);
488             next_dirty = bdrv_dirty_iter_next(s->dbi);
489         }
490         assert(next_dirty == next_offset);
491         nb_chunks++;
492     }
493 
494     /* Clear dirty bits before querying the block status, because
495      * calling bdrv_block_status_above could yield - if some blocks are
496      * marked dirty in this window, we need to know.
497      */
498     bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset,
499                                    nb_chunks * s->granularity);
500     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
501 
502     /* Before claiming an area in the in-flight bitmap, we have to
503      * create a MirrorOp for it so that conflicting requests can wait
504      * for it.  mirror_perform() will create the real MirrorOps later,
505      * for now we just create a pseudo operation that will wake up all
506      * conflicting requests once all real operations have been
507      * launched. */
508     pseudo_op = g_new(MirrorOp, 1);
509     *pseudo_op = (MirrorOp){
510         .offset         = offset,
511         .bytes          = nb_chunks * s->granularity,
512         .is_pseudo_op   = true,
513     };
514     qemu_co_queue_init(&pseudo_op->waiting_requests);
515     QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next);
516 
517     bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks);
518     while (nb_chunks > 0 && offset < s->bdev_length) {
519         int ret;
520         int64_t io_bytes;
521         int64_t io_bytes_acct;
522         MirrorMethod mirror_method = MIRROR_METHOD_COPY;
523 
524         assert(!(offset % s->granularity));
525         ret = bdrv_block_status_above(source, NULL, offset,
526                                       nb_chunks * s->granularity,
527                                       &io_bytes, NULL, NULL);
528         if (ret < 0) {
529             io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes);
530         } else if (ret & BDRV_BLOCK_DATA) {
531             io_bytes = MIN(io_bytes, max_io_bytes);
532         }
533 
534         io_bytes -= io_bytes % s->granularity;
535         if (io_bytes < s->granularity) {
536             io_bytes = s->granularity;
537         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
538             int64_t target_offset;
539             int64_t target_bytes;
540             bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes,
541                                    &target_offset, &target_bytes);
542             if (target_offset == offset &&
543                 target_bytes == io_bytes) {
544                 mirror_method = ret & BDRV_BLOCK_ZERO ?
545                                     MIRROR_METHOD_ZERO :
546                                     MIRROR_METHOD_DISCARD;
547             }
548         }
549 
550         while (s->in_flight >= MAX_IN_FLIGHT) {
551             trace_mirror_yield_in_flight(s, offset, s->in_flight);
552             mirror_wait_for_free_in_flight_slot(s);
553         }
554 
555         if (s->ret < 0) {
556             ret = 0;
557             goto fail;
558         }
559 
560         io_bytes = mirror_clip_bytes(s, offset, io_bytes);
561         io_bytes = mirror_perform(s, offset, io_bytes, mirror_method);
562         if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) {
563             io_bytes_acct = 0;
564         } else {
565             io_bytes_acct = io_bytes;
566         }
567         assert(io_bytes);
568         offset += io_bytes;
569         nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity);
570         delay_ns = block_job_ratelimit_get_delay(&s->common, io_bytes_acct);
571     }
572 
573     ret = delay_ns;
574 fail:
575     QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next);
576     qemu_co_queue_restart_all(&pseudo_op->waiting_requests);
577     g_free(pseudo_op);
578 
579     return ret;
580 }
581 
582 static void mirror_free_init(MirrorBlockJob *s)
583 {
584     int granularity = s->granularity;
585     size_t buf_size = s->buf_size;
586     uint8_t *buf = s->buf;
587 
588     assert(s->buf_free_count == 0);
589     QSIMPLEQ_INIT(&s->buf_free);
590     while (buf_size != 0) {
591         MirrorBuffer *cur = (MirrorBuffer *)buf;
592         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
593         s->buf_free_count++;
594         buf_size -= granularity;
595         buf += granularity;
596     }
597 }
598 
599 /* This is also used for the .pause callback. There is no matching
600  * mirror_resume() because mirror_run() will begin iterating again
601  * when the job is resumed.
602  */
603 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s)
604 {
605     while (s->in_flight > 0) {
606         mirror_wait_for_free_in_flight_slot(s);
607     }
608 }
609 
610 /**
611  * mirror_exit_common: handle both abort() and prepare() cases.
612  * for .prepare, returns 0 on success and -errno on failure.
613  * for .abort cases, denoted by abort = true, MUST return 0.
614  */
615 static int mirror_exit_common(Job *job)
616 {
617     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
618     BlockJob *bjob = &s->common;
619     MirrorBDSOpaque *bs_opaque = s->mirror_top_bs->opaque;
620     AioContext *replace_aio_context = NULL;
621     BlockDriverState *src = s->mirror_top_bs->backing->bs;
622     BlockDriverState *target_bs = blk_bs(s->target);
623     BlockDriverState *mirror_top_bs = s->mirror_top_bs;
624     Error *local_err = NULL;
625     bool abort = job->ret < 0;
626     int ret = 0;
627 
628     if (s->prepared) {
629         return 0;
630     }
631     s->prepared = true;
632 
633     bdrv_release_dirty_bitmap(src, s->dirty_bitmap);
634 
635     /* Make sure that the source BDS doesn't go away during bdrv_replace_node,
636      * before we can call bdrv_drained_end */
637     bdrv_ref(src);
638     bdrv_ref(mirror_top_bs);
639     bdrv_ref(target_bs);
640 
641     /* Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
642      * inserting target_bs at s->to_replace, where we might not be able to get
643      * these permissions.
644      *
645      * Note that blk_unref() alone doesn't necessarily drop permissions because
646      * we might be running nested inside mirror_drain(), which takes an extra
647      * reference, so use an explicit blk_set_perm() first. */
648     blk_set_perm(s->target, 0, BLK_PERM_ALL, &error_abort);
649     blk_unref(s->target);
650     s->target = NULL;
651 
652     /* We don't access the source any more. Dropping any WRITE/RESIZE is
653      * required before it could become a backing file of target_bs. */
654     bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
655                             &error_abort);
656     if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
657         BlockDriverState *backing = s->is_none_mode ? src : s->base;
658         if (backing_bs(target_bs) != backing) {
659             bdrv_set_backing_hd(target_bs, backing, &local_err);
660             if (local_err) {
661                 error_report_err(local_err);
662                 ret = -EPERM;
663             }
664         }
665     }
666 
667     if (s->to_replace) {
668         replace_aio_context = bdrv_get_aio_context(s->to_replace);
669         aio_context_acquire(replace_aio_context);
670     }
671 
672     if (s->should_complete && !abort) {
673         BlockDriverState *to_replace = s->to_replace ?: src;
674         bool ro = bdrv_is_read_only(to_replace);
675 
676         if (ro != bdrv_is_read_only(target_bs)) {
677             bdrv_reopen_set_read_only(target_bs, ro, NULL);
678         }
679 
680         /* The mirror job has no requests in flight any more, but we need to
681          * drain potential other users of the BDS before changing the graph. */
682         bdrv_drained_begin(target_bs);
683         bdrv_replace_node(to_replace, target_bs, &local_err);
684         bdrv_drained_end(target_bs);
685         if (local_err) {
686             error_report_err(local_err);
687             ret = -EPERM;
688         }
689     }
690     if (s->to_replace) {
691         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
692         error_free(s->replace_blocker);
693         bdrv_unref(s->to_replace);
694     }
695     if (replace_aio_context) {
696         aio_context_release(replace_aio_context);
697     }
698     g_free(s->replaces);
699     bdrv_unref(target_bs);
700 
701     /* Remove the mirror filter driver from the graph. Before this, get rid of
702      * the blockers on the intermediate nodes so that the resulting state is
703      * valid. Also give up permissions on mirror_top_bs->backing, which might
704      * block the removal. */
705     block_job_remove_all_bdrv(bjob);
706     bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
707                             &error_abort);
708     bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
709 
710     /* We just changed the BDS the job BB refers to (with either or both of the
711      * bdrv_replace_node() calls), so switch the BB back so the cleanup does
712      * the right thing. We don't need any permissions any more now. */
713     blk_remove_bs(bjob->blk);
714     blk_set_perm(bjob->blk, 0, BLK_PERM_ALL, &error_abort);
715     blk_insert_bs(bjob->blk, mirror_top_bs, &error_abort);
716 
717     bs_opaque->job = NULL;
718 
719     bdrv_drained_end(src);
720     bdrv_unref(mirror_top_bs);
721     bdrv_unref(src);
722 
723     return ret;
724 }
725 
726 static int mirror_prepare(Job *job)
727 {
728     return mirror_exit_common(job);
729 }
730 
731 static void mirror_abort(Job *job)
732 {
733     int ret = mirror_exit_common(job);
734     assert(ret == 0);
735 }
736 
737 static void coroutine_fn mirror_throttle(MirrorBlockJob *s)
738 {
739     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
740 
741     if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) {
742         s->last_pause_ns = now;
743         job_sleep_ns(&s->common.job, 0);
744     } else {
745         job_pause_point(&s->common.job);
746     }
747 }
748 
749 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
750 {
751     int64_t offset;
752     BlockDriverState *base = s->base;
753     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
754     BlockDriverState *target_bs = blk_bs(s->target);
755     int ret;
756     int64_t count;
757 
758     if (base == NULL && !bdrv_has_zero_init(target_bs)) {
759         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
760             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
761             return 0;
762         }
763 
764         s->initial_zeroing_ongoing = true;
765         for (offset = 0; offset < s->bdev_length; ) {
766             int bytes = MIN(s->bdev_length - offset,
767                             QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
768 
769             mirror_throttle(s);
770 
771             if (job_is_cancelled(&s->common.job)) {
772                 s->initial_zeroing_ongoing = false;
773                 return 0;
774             }
775 
776             if (s->in_flight >= MAX_IN_FLIGHT) {
777                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
778                                    s->in_flight);
779                 mirror_wait_for_free_in_flight_slot(s);
780                 continue;
781             }
782 
783             mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO);
784             offset += bytes;
785         }
786 
787         mirror_wait_for_all_io(s);
788         s->initial_zeroing_ongoing = false;
789     }
790 
791     /* First part, loop on the sectors and initialize the dirty bitmap.  */
792     for (offset = 0; offset < s->bdev_length; ) {
793         /* Just to make sure we are not exceeding int limit. */
794         int bytes = MIN(s->bdev_length - offset,
795                         QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
796 
797         mirror_throttle(s);
798 
799         if (job_is_cancelled(&s->common.job)) {
800             return 0;
801         }
802 
803         ret = bdrv_is_allocated_above(bs, base, offset, bytes, &count);
804         if (ret < 0) {
805             return ret;
806         }
807 
808         assert(count);
809         if (ret == 1) {
810             bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
811         }
812         offset += count;
813     }
814     return 0;
815 }
816 
817 /* Called when going out of the streaming phase to flush the bulk of the
818  * data to the medium, or just before completing.
819  */
820 static int mirror_flush(MirrorBlockJob *s)
821 {
822     int ret = blk_flush(s->target);
823     if (ret < 0) {
824         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
825             s->ret = ret;
826         }
827     }
828     return ret;
829 }
830 
831 static int coroutine_fn mirror_run(Job *job, Error **errp)
832 {
833     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
834     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
835     BlockDriverState *target_bs = blk_bs(s->target);
836     bool need_drain = true;
837     int64_t length;
838     BlockDriverInfo bdi;
839     char backing_filename[2]; /* we only need 2 characters because we are only
840                                  checking for a NULL string */
841     int ret = 0;
842 
843     if (job_is_cancelled(&s->common.job)) {
844         goto immediate_exit;
845     }
846 
847     s->bdev_length = bdrv_getlength(bs);
848     if (s->bdev_length < 0) {
849         ret = s->bdev_length;
850         goto immediate_exit;
851     }
852 
853     /* Active commit must resize the base image if its size differs from the
854      * active layer. */
855     if (s->base == blk_bs(s->target)) {
856         int64_t base_length;
857 
858         base_length = blk_getlength(s->target);
859         if (base_length < 0) {
860             ret = base_length;
861             goto immediate_exit;
862         }
863 
864         if (s->bdev_length > base_length) {
865             ret = blk_truncate(s->target, s->bdev_length, PREALLOC_MODE_OFF,
866                                NULL);
867             if (ret < 0) {
868                 goto immediate_exit;
869             }
870         }
871     }
872 
873     if (s->bdev_length == 0) {
874         /* Transition to the READY state and wait for complete. */
875         job_transition_to_ready(&s->common.job);
876         s->synced = true;
877         s->actively_synced = true;
878         while (!job_is_cancelled(&s->common.job) && !s->should_complete) {
879             job_yield(&s->common.job);
880         }
881         s->common.job.cancelled = false;
882         goto immediate_exit;
883     }
884 
885     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
886     s->in_flight_bitmap = bitmap_new(length);
887 
888     /* If we have no backing file yet in the destination, we cannot let
889      * the destination do COW.  Instead, we copy sectors around the
890      * dirty data if needed.  We need a bitmap to do that.
891      */
892     bdrv_get_backing_filename(target_bs, backing_filename,
893                               sizeof(backing_filename));
894     if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) {
895         s->target_cluster_size = bdi.cluster_size;
896     } else {
897         s->target_cluster_size = BDRV_SECTOR_SIZE;
898     }
899     if (backing_filename[0] && !target_bs->backing &&
900         s->granularity < s->target_cluster_size) {
901         s->buf_size = MAX(s->buf_size, s->target_cluster_size);
902         s->cow_bitmap = bitmap_new(length);
903     }
904     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
905 
906     s->buf = qemu_try_blockalign(bs, s->buf_size);
907     if (s->buf == NULL) {
908         ret = -ENOMEM;
909         goto immediate_exit;
910     }
911 
912     mirror_free_init(s);
913 
914     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
915     if (!s->is_none_mode) {
916         ret = mirror_dirty_init(s);
917         if (ret < 0 || job_is_cancelled(&s->common.job)) {
918             goto immediate_exit;
919         }
920     }
921 
922     assert(!s->dbi);
923     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
924     for (;;) {
925         uint64_t delay_ns = 0;
926         int64_t cnt, delta;
927         bool should_complete;
928 
929         /* Do not start passive operations while there are active
930          * writes in progress */
931         while (s->in_active_write_counter) {
932             mirror_wait_for_any_operation(s, true);
933         }
934 
935         if (s->ret < 0) {
936             ret = s->ret;
937             goto immediate_exit;
938         }
939 
940         job_pause_point(&s->common.job);
941 
942         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
943         /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
944          * the number of bytes currently being processed; together those are
945          * the current remaining operation length */
946         job_progress_set_remaining(&s->common.job, s->bytes_in_flight + cnt);
947 
948         /* Note that even when no rate limit is applied we need to yield
949          * periodically with no pending I/O so that bdrv_drain_all() returns.
950          * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
951          * an error, or when the source is clean, whichever comes first. */
952         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
953         if (delta < BLOCK_JOB_SLICE_TIME &&
954             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
955             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
956                 (cnt == 0 && s->in_flight > 0)) {
957                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
958                 mirror_wait_for_free_in_flight_slot(s);
959                 continue;
960             } else if (cnt != 0) {
961                 delay_ns = mirror_iteration(s);
962             }
963         }
964 
965         should_complete = false;
966         if (s->in_flight == 0 && cnt == 0) {
967             trace_mirror_before_flush(s);
968             if (!s->synced) {
969                 if (mirror_flush(s) < 0) {
970                     /* Go check s->ret.  */
971                     continue;
972                 }
973                 /* We're out of the streaming phase.  From now on, if the job
974                  * is cancelled we will actually complete all pending I/O and
975                  * report completion.  This way, block-job-cancel will leave
976                  * the target in a consistent state.
977                  */
978                 job_transition_to_ready(&s->common.job);
979                 s->synced = true;
980                 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) {
981                     s->actively_synced = true;
982                 }
983             }
984 
985             should_complete = s->should_complete ||
986                 job_is_cancelled(&s->common.job);
987             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
988         }
989 
990         if (cnt == 0 && should_complete) {
991             /* The dirty bitmap is not updated while operations are pending.
992              * If we're about to exit, wait for pending operations before
993              * calling bdrv_get_dirty_count(bs), or we may exit while the
994              * source has dirty data to copy!
995              *
996              * Note that I/O can be submitted by the guest while
997              * mirror_populate runs, so pause it now.  Before deciding
998              * whether to switch to target check one last time if I/O has
999              * come in the meanwhile, and if not flush the data to disk.
1000              */
1001             trace_mirror_before_drain(s, cnt);
1002 
1003             bdrv_drained_begin(bs);
1004             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1005             if (cnt > 0 || mirror_flush(s) < 0) {
1006                 bdrv_drained_end(bs);
1007                 continue;
1008             }
1009 
1010             /* The two disks are in sync.  Exit and report successful
1011              * completion.
1012              */
1013             assert(QLIST_EMPTY(&bs->tracked_requests));
1014             s->common.job.cancelled = false;
1015             need_drain = false;
1016             break;
1017         }
1018 
1019         ret = 0;
1020 
1021         if (s->synced && !should_complete) {
1022             delay_ns = (s->in_flight == 0 &&
1023                         cnt == 0 ? BLOCK_JOB_SLICE_TIME : 0);
1024         }
1025         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
1026         job_sleep_ns(&s->common.job, delay_ns);
1027         if (job_is_cancelled(&s->common.job) &&
1028             (!s->synced || s->common.job.force_cancel))
1029         {
1030             break;
1031         }
1032         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1033     }
1034 
1035 immediate_exit:
1036     if (s->in_flight > 0) {
1037         /* We get here only if something went wrong.  Either the job failed,
1038          * or it was cancelled prematurely so that we do not guarantee that
1039          * the target is a copy of the source.
1040          */
1041         assert(ret < 0 || ((s->common.job.force_cancel || !s->synced) &&
1042                job_is_cancelled(&s->common.job)));
1043         assert(need_drain);
1044         mirror_wait_for_all_io(s);
1045     }
1046 
1047     assert(s->in_flight == 0);
1048     qemu_vfree(s->buf);
1049     g_free(s->cow_bitmap);
1050     g_free(s->in_flight_bitmap);
1051     bdrv_dirty_iter_free(s->dbi);
1052 
1053     if (need_drain) {
1054         bdrv_drained_begin(bs);
1055     }
1056 
1057     return ret;
1058 }
1059 
1060 static void mirror_complete(Job *job, Error **errp)
1061 {
1062     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1063     BlockDriverState *target;
1064 
1065     target = blk_bs(s->target);
1066 
1067     if (!s->synced) {
1068         error_setg(errp, "The active block job '%s' cannot be completed",
1069                    job->id);
1070         return;
1071     }
1072 
1073     if (s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
1074         int ret;
1075 
1076         assert(!target->backing);
1077         ret = bdrv_open_backing_file(target, NULL, "backing", errp);
1078         if (ret < 0) {
1079             return;
1080         }
1081     }
1082 
1083     /* block all operations on to_replace bs */
1084     if (s->replaces) {
1085         AioContext *replace_aio_context;
1086 
1087         s->to_replace = bdrv_find_node(s->replaces);
1088         if (!s->to_replace) {
1089             error_setg(errp, "Node name '%s' not found", s->replaces);
1090             return;
1091         }
1092 
1093         replace_aio_context = bdrv_get_aio_context(s->to_replace);
1094         aio_context_acquire(replace_aio_context);
1095 
1096         /* TODO Translate this into permission system. Current definition of
1097          * GRAPH_MOD would require to request it for the parents; they might
1098          * not even be BlockDriverStates, however, so a BdrvChild can't address
1099          * them. May need redefinition of GRAPH_MOD. */
1100         error_setg(&s->replace_blocker,
1101                    "block device is in use by block-job-complete");
1102         bdrv_op_block_all(s->to_replace, s->replace_blocker);
1103         bdrv_ref(s->to_replace);
1104 
1105         aio_context_release(replace_aio_context);
1106     }
1107 
1108     s->should_complete = true;
1109     job_enter(job);
1110 }
1111 
1112 static void coroutine_fn mirror_pause(Job *job)
1113 {
1114     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1115 
1116     mirror_wait_for_all_io(s);
1117 }
1118 
1119 static bool mirror_drained_poll(BlockJob *job)
1120 {
1121     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1122     return !!s->in_flight;
1123 }
1124 
1125 static void mirror_attached_aio_context(BlockJob *job, AioContext *new_context)
1126 {
1127     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1128 
1129     blk_set_aio_context(s->target, new_context);
1130 }
1131 
1132 static void mirror_drain(BlockJob *job)
1133 {
1134     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1135 
1136     /* Need to keep a reference in case blk_drain triggers execution
1137      * of mirror_complete...
1138      */
1139     if (s->target) {
1140         BlockBackend *target = s->target;
1141         blk_ref(target);
1142         blk_drain(target);
1143         blk_unref(target);
1144     }
1145 }
1146 
1147 static const BlockJobDriver mirror_job_driver = {
1148     .job_driver = {
1149         .instance_size          = sizeof(MirrorBlockJob),
1150         .job_type               = JOB_TYPE_MIRROR,
1151         .free                   = block_job_free,
1152         .user_resume            = block_job_user_resume,
1153         .drain                  = block_job_drain,
1154         .run                    = mirror_run,
1155         .prepare                = mirror_prepare,
1156         .abort                  = mirror_abort,
1157         .pause                  = mirror_pause,
1158         .complete               = mirror_complete,
1159     },
1160     .drained_poll           = mirror_drained_poll,
1161     .attached_aio_context   = mirror_attached_aio_context,
1162     .drain                  = mirror_drain,
1163 };
1164 
1165 static const BlockJobDriver commit_active_job_driver = {
1166     .job_driver = {
1167         .instance_size          = sizeof(MirrorBlockJob),
1168         .job_type               = JOB_TYPE_COMMIT,
1169         .free                   = block_job_free,
1170         .user_resume            = block_job_user_resume,
1171         .drain                  = block_job_drain,
1172         .run                    = mirror_run,
1173         .prepare                = mirror_prepare,
1174         .abort                  = mirror_abort,
1175         .pause                  = mirror_pause,
1176         .complete               = mirror_complete,
1177     },
1178     .drained_poll           = mirror_drained_poll,
1179     .attached_aio_context   = mirror_attached_aio_context,
1180     .drain                  = mirror_drain,
1181 };
1182 
1183 static void coroutine_fn
1184 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method,
1185                      uint64_t offset, uint64_t bytes,
1186                      QEMUIOVector *qiov, int flags)
1187 {
1188     BdrvDirtyBitmapIter *iter;
1189     QEMUIOVector target_qiov;
1190     uint64_t dirty_offset;
1191     int dirty_bytes;
1192 
1193     if (qiov) {
1194         qemu_iovec_init(&target_qiov, qiov->niov);
1195     }
1196 
1197     iter = bdrv_dirty_iter_new(job->dirty_bitmap);
1198     bdrv_set_dirty_iter(iter, offset);
1199 
1200     while (true) {
1201         bool valid_area;
1202         int ret;
1203 
1204         bdrv_dirty_bitmap_lock(job->dirty_bitmap);
1205         valid_area = bdrv_dirty_iter_next_area(iter, offset + bytes,
1206                                                &dirty_offset, &dirty_bytes);
1207         if (!valid_area) {
1208             bdrv_dirty_bitmap_unlock(job->dirty_bitmap);
1209             break;
1210         }
1211 
1212         bdrv_reset_dirty_bitmap_locked(job->dirty_bitmap,
1213                                        dirty_offset, dirty_bytes);
1214         bdrv_dirty_bitmap_unlock(job->dirty_bitmap);
1215 
1216         job_progress_increase_remaining(&job->common.job, dirty_bytes);
1217 
1218         assert(dirty_offset - offset <= SIZE_MAX);
1219         if (qiov) {
1220             qemu_iovec_reset(&target_qiov);
1221             qemu_iovec_concat(&target_qiov, qiov,
1222                               dirty_offset - offset, dirty_bytes);
1223         }
1224 
1225         switch (method) {
1226         case MIRROR_METHOD_COPY:
1227             ret = blk_co_pwritev(job->target, dirty_offset, dirty_bytes,
1228                                  qiov ? &target_qiov : NULL, flags);
1229             break;
1230 
1231         case MIRROR_METHOD_ZERO:
1232             assert(!qiov);
1233             ret = blk_co_pwrite_zeroes(job->target, dirty_offset, dirty_bytes,
1234                                        flags);
1235             break;
1236 
1237         case MIRROR_METHOD_DISCARD:
1238             assert(!qiov);
1239             ret = blk_co_pdiscard(job->target, dirty_offset, dirty_bytes);
1240             break;
1241 
1242         default:
1243             abort();
1244         }
1245 
1246         if (ret >= 0) {
1247             job_progress_update(&job->common.job, dirty_bytes);
1248         } else {
1249             BlockErrorAction action;
1250 
1251             bdrv_set_dirty_bitmap(job->dirty_bitmap, dirty_offset, dirty_bytes);
1252             job->actively_synced = false;
1253 
1254             action = mirror_error_action(job, false, -ret);
1255             if (action == BLOCK_ERROR_ACTION_REPORT) {
1256                 if (!job->ret) {
1257                     job->ret = ret;
1258                 }
1259                 break;
1260             }
1261         }
1262     }
1263 
1264     bdrv_dirty_iter_free(iter);
1265     if (qiov) {
1266         qemu_iovec_destroy(&target_qiov);
1267     }
1268 }
1269 
1270 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s,
1271                                                    uint64_t offset,
1272                                                    uint64_t bytes)
1273 {
1274     MirrorOp *op;
1275     uint64_t start_chunk = offset / s->granularity;
1276     uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
1277 
1278     op = g_new(MirrorOp, 1);
1279     *op = (MirrorOp){
1280         .s                  = s,
1281         .offset             = offset,
1282         .bytes              = bytes,
1283         .is_active_write    = true,
1284     };
1285     qemu_co_queue_init(&op->waiting_requests);
1286     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
1287 
1288     s->in_active_write_counter++;
1289 
1290     mirror_wait_on_conflicts(op, s, offset, bytes);
1291 
1292     bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1293 
1294     return op;
1295 }
1296 
1297 static void coroutine_fn active_write_settle(MirrorOp *op)
1298 {
1299     uint64_t start_chunk = op->offset / op->s->granularity;
1300     uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes,
1301                                       op->s->granularity);
1302 
1303     if (!--op->s->in_active_write_counter && op->s->actively_synced) {
1304         BdrvChild *source = op->s->mirror_top_bs->backing;
1305 
1306         if (QLIST_FIRST(&source->bs->parents) == source &&
1307             QLIST_NEXT(source, next_parent) == NULL)
1308         {
1309             /* Assert that we are back in sync once all active write
1310              * operations are settled.
1311              * Note that we can only assert this if the mirror node
1312              * is the source node's only parent. */
1313             assert(!bdrv_get_dirty_count(op->s->dirty_bitmap));
1314         }
1315     }
1316     bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1317     QTAILQ_REMOVE(&op->s->ops_in_flight, op, next);
1318     qemu_co_queue_restart_all(&op->waiting_requests);
1319     g_free(op);
1320 }
1321 
1322 static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs,
1323     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1324 {
1325     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1326 }
1327 
1328 static int coroutine_fn bdrv_mirror_top_do_write(BlockDriverState *bs,
1329     MirrorMethod method, uint64_t offset, uint64_t bytes, QEMUIOVector *qiov,
1330     int flags)
1331 {
1332     MirrorOp *op = NULL;
1333     MirrorBDSOpaque *s = bs->opaque;
1334     int ret = 0;
1335     bool copy_to_target;
1336 
1337     copy_to_target = s->job->ret >= 0 &&
1338                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1339 
1340     if (copy_to_target) {
1341         op = active_write_prepare(s->job, offset, bytes);
1342     }
1343 
1344     switch (method) {
1345     case MIRROR_METHOD_COPY:
1346         ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1347         break;
1348 
1349     case MIRROR_METHOD_ZERO:
1350         ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1351         break;
1352 
1353     case MIRROR_METHOD_DISCARD:
1354         ret = bdrv_co_pdiscard(bs->backing, offset, bytes);
1355         break;
1356 
1357     default:
1358         abort();
1359     }
1360 
1361     if (ret < 0) {
1362         goto out;
1363     }
1364 
1365     if (copy_to_target) {
1366         do_sync_target_write(s->job, method, offset, bytes, qiov, flags);
1367     }
1368 
1369 out:
1370     if (copy_to_target) {
1371         active_write_settle(op);
1372     }
1373     return ret;
1374 }
1375 
1376 static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs,
1377     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1378 {
1379     MirrorBDSOpaque *s = bs->opaque;
1380     QEMUIOVector bounce_qiov;
1381     void *bounce_buf;
1382     int ret = 0;
1383     bool copy_to_target;
1384 
1385     copy_to_target = s->job->ret >= 0 &&
1386                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1387 
1388     if (copy_to_target) {
1389         /* The guest might concurrently modify the data to write; but
1390          * the data on source and destination must match, so we have
1391          * to use a bounce buffer if we are going to write to the
1392          * target now. */
1393         bounce_buf = qemu_blockalign(bs, bytes);
1394         iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes);
1395 
1396         qemu_iovec_init(&bounce_qiov, 1);
1397         qemu_iovec_add(&bounce_qiov, bounce_buf, bytes);
1398         qiov = &bounce_qiov;
1399     }
1400 
1401     ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov,
1402                                    flags);
1403 
1404     if (copy_to_target) {
1405         qemu_iovec_destroy(&bounce_qiov);
1406         qemu_vfree(bounce_buf);
1407     }
1408 
1409     return ret;
1410 }
1411 
1412 static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs)
1413 {
1414     if (bs->backing == NULL) {
1415         /* we can be here after failed bdrv_append in mirror_start_job */
1416         return 0;
1417     }
1418     return bdrv_co_flush(bs->backing->bs);
1419 }
1420 
1421 static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs,
1422     int64_t offset, int bytes, BdrvRequestFlags flags)
1423 {
1424     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL,
1425                                     flags);
1426 }
1427 
1428 static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs,
1429     int64_t offset, int bytes)
1430 {
1431     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes,
1432                                     NULL, 0);
1433 }
1434 
1435 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs, QDict *opts)
1436 {
1437     if (bs->backing == NULL) {
1438         /* we can be here after failed bdrv_attach_child in
1439          * bdrv_set_backing_hd */
1440         return;
1441     }
1442     bdrv_refresh_filename(bs->backing->bs);
1443     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1444             bs->backing->bs->filename);
1445 }
1446 
1447 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1448                                        const BdrvChildRole *role,
1449                                        BlockReopenQueue *reopen_queue,
1450                                        uint64_t perm, uint64_t shared,
1451                                        uint64_t *nperm, uint64_t *nshared)
1452 {
1453     /* Must be able to forward guest writes to the real image */
1454     *nperm = 0;
1455     if (perm & BLK_PERM_WRITE) {
1456         *nperm |= BLK_PERM_WRITE;
1457     }
1458 
1459     *nshared = BLK_PERM_ALL;
1460 }
1461 
1462 /* Dummy node that provides consistent read to its users without requiring it
1463  * from its backing file and that allows writes on the backing file chain. */
1464 static BlockDriver bdrv_mirror_top = {
1465     .format_name                = "mirror_top",
1466     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1467     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1468     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1469     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1470     .bdrv_co_flush              = bdrv_mirror_top_flush,
1471     .bdrv_co_block_status       = bdrv_co_block_status_from_backing,
1472     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1473     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1474 };
1475 
1476 static void mirror_start_job(const char *job_id, BlockDriverState *bs,
1477                              int creation_flags, BlockDriverState *target,
1478                              const char *replaces, int64_t speed,
1479                              uint32_t granularity, int64_t buf_size,
1480                              BlockMirrorBackingMode backing_mode,
1481                              BlockdevOnError on_source_error,
1482                              BlockdevOnError on_target_error,
1483                              bool unmap,
1484                              BlockCompletionFunc *cb,
1485                              void *opaque,
1486                              const BlockJobDriver *driver,
1487                              bool is_none_mode, BlockDriverState *base,
1488                              bool auto_complete, const char *filter_node_name,
1489                              bool is_mirror, MirrorCopyMode copy_mode,
1490                              Error **errp)
1491 {
1492     MirrorBlockJob *s;
1493     MirrorBDSOpaque *bs_opaque;
1494     BlockDriverState *mirror_top_bs;
1495     bool target_graph_mod;
1496     bool target_is_backing;
1497     Error *local_err = NULL;
1498     int ret;
1499 
1500     if (granularity == 0) {
1501         granularity = bdrv_get_default_bitmap_granularity(target);
1502     }
1503 
1504     assert(is_power_of_2(granularity));
1505 
1506     if (buf_size < 0) {
1507         error_setg(errp, "Invalid parameter 'buf-size'");
1508         return;
1509     }
1510 
1511     if (buf_size == 0) {
1512         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1513     }
1514 
1515     if (bs == target) {
1516         error_setg(errp, "Can't mirror node into itself");
1517         return;
1518     }
1519 
1520     /* In the case of active commit, add dummy driver to provide consistent
1521      * reads on the top, while disabling it in the intermediate nodes, and make
1522      * the backing chain writable. */
1523     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1524                                          BDRV_O_RDWR, errp);
1525     if (mirror_top_bs == NULL) {
1526         return;
1527     }
1528     if (!filter_node_name) {
1529         mirror_top_bs->implicit = true;
1530     }
1531     mirror_top_bs->total_sectors = bs->total_sectors;
1532     mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
1533     mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED;
1534     bs_opaque = g_new0(MirrorBDSOpaque, 1);
1535     mirror_top_bs->opaque = bs_opaque;
1536     bdrv_set_aio_context(mirror_top_bs, bdrv_get_aio_context(bs));
1537 
1538     /* bdrv_append takes ownership of the mirror_top_bs reference, need to keep
1539      * it alive until block_job_create() succeeds even if bs has no parent. */
1540     bdrv_ref(mirror_top_bs);
1541     bdrv_drained_begin(bs);
1542     bdrv_append(mirror_top_bs, bs, &local_err);
1543     bdrv_drained_end(bs);
1544 
1545     if (local_err) {
1546         bdrv_unref(mirror_top_bs);
1547         error_propagate(errp, local_err);
1548         return;
1549     }
1550 
1551     /* Make sure that the source is not resized while the job is running */
1552     s = block_job_create(job_id, driver, NULL, mirror_top_bs,
1553                          BLK_PERM_CONSISTENT_READ,
1554                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1555                          BLK_PERM_WRITE | BLK_PERM_GRAPH_MOD, speed,
1556                          creation_flags, cb, opaque, errp);
1557     if (!s) {
1558         goto fail;
1559     }
1560     bs_opaque->job = s;
1561 
1562     /* The block job now has a reference to this node */
1563     bdrv_unref(mirror_top_bs);
1564 
1565     s->mirror_top_bs = mirror_top_bs;
1566 
1567     /* No resize for the target either; while the mirror is still running, a
1568      * consistent read isn't necessarily possible. We could possibly allow
1569      * writes and graph modifications, though it would likely defeat the
1570      * purpose of a mirror, so leave them blocked for now.
1571      *
1572      * In the case of active commit, things look a bit different, though,
1573      * because the target is an already populated backing file in active use.
1574      * We can allow anything except resize there.*/
1575     target_is_backing = bdrv_chain_contains(bs, target);
1576     target_graph_mod = (backing_mode != MIRROR_LEAVE_BACKING_CHAIN);
1577     s->target = blk_new(BLK_PERM_WRITE | BLK_PERM_RESIZE |
1578                         (target_graph_mod ? BLK_PERM_GRAPH_MOD : 0),
1579                         BLK_PERM_WRITE_UNCHANGED |
1580                         (target_is_backing ? BLK_PERM_CONSISTENT_READ |
1581                                              BLK_PERM_WRITE |
1582                                              BLK_PERM_GRAPH_MOD : 0));
1583     ret = blk_insert_bs(s->target, target, errp);
1584     if (ret < 0) {
1585         goto fail;
1586     }
1587     if (is_mirror) {
1588         /* XXX: Mirror target could be a NBD server of target QEMU in the case
1589          * of non-shared block migration. To allow migration completion, we
1590          * have to allow "inactivate" of the target BB.  When that happens, we
1591          * know the job is drained, and the vcpus are stopped, so no write
1592          * operation will be performed. Block layer already has assertions to
1593          * ensure that. */
1594         blk_set_force_allow_inactivate(s->target);
1595     }
1596 
1597     s->replaces = g_strdup(replaces);
1598     s->on_source_error = on_source_error;
1599     s->on_target_error = on_target_error;
1600     s->is_none_mode = is_none_mode;
1601     s->backing_mode = backing_mode;
1602     s->copy_mode = copy_mode;
1603     s->base = base;
1604     s->granularity = granularity;
1605     s->buf_size = ROUND_UP(buf_size, granularity);
1606     s->unmap = unmap;
1607     if (auto_complete) {
1608         s->should_complete = true;
1609     }
1610 
1611     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1612     if (!s->dirty_bitmap) {
1613         goto fail;
1614     }
1615 
1616     /* Required permissions are already taken with blk_new() */
1617     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1618                        &error_abort);
1619 
1620     /* In commit_active_start() all intermediate nodes disappear, so
1621      * any jobs in them must be blocked */
1622     if (target_is_backing) {
1623         BlockDriverState *iter;
1624         for (iter = backing_bs(bs); iter != target; iter = backing_bs(iter)) {
1625             /* XXX BLK_PERM_WRITE needs to be allowed so we don't block
1626              * ourselves at s->base (if writes are blocked for a node, they are
1627              * also blocked for its backing file). The other options would be a
1628              * second filter driver above s->base (== target). */
1629             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1630                                      BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
1631                                      errp);
1632             if (ret < 0) {
1633                 goto fail;
1634             }
1635         }
1636     }
1637 
1638     QTAILQ_INIT(&s->ops_in_flight);
1639 
1640     trace_mirror_start(bs, s, opaque);
1641     job_start(&s->common.job);
1642     return;
1643 
1644 fail:
1645     if (s) {
1646         /* Make sure this BDS does not go away until we have completed the graph
1647          * changes below */
1648         bdrv_ref(mirror_top_bs);
1649 
1650         g_free(s->replaces);
1651         blk_unref(s->target);
1652         bs_opaque->job = NULL;
1653         job_early_fail(&s->common.job);
1654     }
1655 
1656     bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
1657                             &error_abort);
1658     bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
1659 
1660     bdrv_unref(mirror_top_bs);
1661 }
1662 
1663 void mirror_start(const char *job_id, BlockDriverState *bs,
1664                   BlockDriverState *target, const char *replaces,
1665                   int creation_flags, int64_t speed,
1666                   uint32_t granularity, int64_t buf_size,
1667                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1668                   BlockdevOnError on_source_error,
1669                   BlockdevOnError on_target_error,
1670                   bool unmap, const char *filter_node_name,
1671                   MirrorCopyMode copy_mode, Error **errp)
1672 {
1673     bool is_none_mode;
1674     BlockDriverState *base;
1675 
1676     if (mode == MIRROR_SYNC_MODE_INCREMENTAL) {
1677         error_setg(errp, "Sync mode 'incremental' not supported");
1678         return;
1679     }
1680     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1681     base = mode == MIRROR_SYNC_MODE_TOP ? backing_bs(bs) : NULL;
1682     mirror_start_job(job_id, bs, creation_flags, target, replaces,
1683                      speed, granularity, buf_size, backing_mode,
1684                      on_source_error, on_target_error, unmap, NULL, NULL,
1685                      &mirror_job_driver, is_none_mode, base, false,
1686                      filter_node_name, true, copy_mode, errp);
1687 }
1688 
1689 void commit_active_start(const char *job_id, BlockDriverState *bs,
1690                          BlockDriverState *base, int creation_flags,
1691                          int64_t speed, BlockdevOnError on_error,
1692                          const char *filter_node_name,
1693                          BlockCompletionFunc *cb, void *opaque,
1694                          bool auto_complete, Error **errp)
1695 {
1696     bool base_read_only;
1697     Error *local_err = NULL;
1698 
1699     base_read_only = bdrv_is_read_only(base);
1700 
1701     if (base_read_only) {
1702         if (bdrv_reopen_set_read_only(base, false, errp) < 0) {
1703             return;
1704         }
1705     }
1706 
1707     mirror_start_job(job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1708                      MIRROR_LEAVE_BACKING_CHAIN,
1709                      on_error, on_error, true, cb, opaque,
1710                      &commit_active_job_driver, false, base, auto_complete,
1711                      filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND,
1712                      &local_err);
1713     if (local_err) {
1714         error_propagate(errp, local_err);
1715         goto error_restore_flags;
1716     }
1717 
1718     return;
1719 
1720 error_restore_flags:
1721     /* ignore error and errp for bdrv_reopen, because we want to propagate
1722      * the original error */
1723     if (base_read_only) {
1724         bdrv_reopen_set_read_only(base, true, NULL);
1725     }
1726     return;
1727 }
1728