xref: /openbmc/qemu/block/mirror.c (revision 9e9b10c6)
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     /* Whether the target image requires explicit zero-initialization */
55     bool zero_target;
56     MirrorCopyMode copy_mode;
57     BlockdevOnError on_source_error, on_target_error;
58     bool synced;
59     /* Set when the target is synced (dirty bitmap is clean, nothing
60      * in flight) and the job is running in active mode */
61     bool actively_synced;
62     bool should_complete;
63     int64_t granularity;
64     size_t buf_size;
65     int64_t bdev_length;
66     unsigned long *cow_bitmap;
67     BdrvDirtyBitmap *dirty_bitmap;
68     BdrvDirtyBitmapIter *dbi;
69     uint8_t *buf;
70     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
71     int buf_free_count;
72 
73     uint64_t last_pause_ns;
74     unsigned long *in_flight_bitmap;
75     int in_flight;
76     int64_t bytes_in_flight;
77     QTAILQ_HEAD(, MirrorOp) ops_in_flight;
78     int ret;
79     bool unmap;
80     int target_cluster_size;
81     int max_iov;
82     bool initial_zeroing_ongoing;
83     int in_active_write_counter;
84     bool prepared;
85     bool in_drain;
86 } MirrorBlockJob;
87 
88 typedef struct MirrorBDSOpaque {
89     MirrorBlockJob *job;
90     bool stop;
91 } MirrorBDSOpaque;
92 
93 struct MirrorOp {
94     MirrorBlockJob *s;
95     QEMUIOVector qiov;
96     int64_t offset;
97     uint64_t bytes;
98 
99     /* The pointee is set by mirror_co_read(), mirror_co_zero(), and
100      * mirror_co_discard() before yielding for the first time */
101     int64_t *bytes_handled;
102 
103     bool is_pseudo_op;
104     bool is_active_write;
105     CoQueue waiting_requests;
106 
107     QTAILQ_ENTRY(MirrorOp) next;
108 };
109 
110 typedef enum MirrorMethod {
111     MIRROR_METHOD_COPY,
112     MIRROR_METHOD_ZERO,
113     MIRROR_METHOD_DISCARD,
114 } MirrorMethod;
115 
116 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
117                                             int error)
118 {
119     s->synced = false;
120     s->actively_synced = false;
121     if (read) {
122         return block_job_error_action(&s->common, s->on_source_error,
123                                       true, error);
124     } else {
125         return block_job_error_action(&s->common, s->on_target_error,
126                                       false, error);
127     }
128 }
129 
130 static void coroutine_fn mirror_wait_on_conflicts(MirrorOp *self,
131                                                   MirrorBlockJob *s,
132                                                   uint64_t offset,
133                                                   uint64_t bytes)
134 {
135     uint64_t self_start_chunk = offset / s->granularity;
136     uint64_t self_end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
137     uint64_t self_nb_chunks = self_end_chunk - self_start_chunk;
138 
139     while (find_next_bit(s->in_flight_bitmap, self_end_chunk,
140                          self_start_chunk) < self_end_chunk &&
141            s->ret >= 0)
142     {
143         MirrorOp *op;
144 
145         QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
146             uint64_t op_start_chunk = op->offset / s->granularity;
147             uint64_t op_nb_chunks = DIV_ROUND_UP(op->offset + op->bytes,
148                                                  s->granularity) -
149                                     op_start_chunk;
150 
151             if (op == self) {
152                 continue;
153             }
154 
155             if (ranges_overlap(self_start_chunk, self_nb_chunks,
156                                op_start_chunk, op_nb_chunks))
157             {
158                 qemu_co_queue_wait(&op->waiting_requests, NULL);
159                 break;
160             }
161         }
162     }
163 }
164 
165 static void coroutine_fn mirror_iteration_done(MirrorOp *op, int ret)
166 {
167     MirrorBlockJob *s = op->s;
168     struct iovec *iov;
169     int64_t chunk_num;
170     int i, nb_chunks;
171 
172     trace_mirror_iteration_done(s, op->offset, op->bytes, ret);
173 
174     s->in_flight--;
175     s->bytes_in_flight -= op->bytes;
176     iov = op->qiov.iov;
177     for (i = 0; i < op->qiov.niov; i++) {
178         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
179         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
180         s->buf_free_count++;
181     }
182 
183     chunk_num = op->offset / s->granularity;
184     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
185 
186     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
187     QTAILQ_REMOVE(&s->ops_in_flight, op, next);
188     if (ret >= 0) {
189         if (s->cow_bitmap) {
190             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
191         }
192         if (!s->initial_zeroing_ongoing) {
193             job_progress_update(&s->common.job, op->bytes);
194         }
195     }
196     qemu_iovec_destroy(&op->qiov);
197 
198     qemu_co_queue_restart_all(&op->waiting_requests);
199     g_free(op);
200 }
201 
202 static void coroutine_fn mirror_write_complete(MirrorOp *op, int ret)
203 {
204     MirrorBlockJob *s = op->s;
205 
206     if (ret < 0) {
207         BlockErrorAction action;
208 
209         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
210         action = mirror_error_action(s, false, -ret);
211         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
212             s->ret = ret;
213         }
214     }
215 
216     mirror_iteration_done(op, ret);
217 }
218 
219 static void coroutine_fn mirror_read_complete(MirrorOp *op, int ret)
220 {
221     MirrorBlockJob *s = op->s;
222 
223     if (ret < 0) {
224         BlockErrorAction action;
225 
226         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
227         action = mirror_error_action(s, true, -ret);
228         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
229             s->ret = ret;
230         }
231 
232         mirror_iteration_done(op, ret);
233         return;
234     }
235 
236     ret = blk_co_pwritev(s->target, op->offset, op->qiov.size, &op->qiov, 0);
237     mirror_write_complete(op, ret);
238 }
239 
240 /* Clip bytes relative to offset to not exceed end-of-file */
241 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s,
242                                         int64_t offset,
243                                         int64_t bytes)
244 {
245     return MIN(bytes, s->bdev_length - offset);
246 }
247 
248 /* Round offset and/or bytes to target cluster if COW is needed, and
249  * return the offset of the adjusted tail against original. */
250 static int mirror_cow_align(MirrorBlockJob *s, int64_t *offset,
251                             uint64_t *bytes)
252 {
253     bool need_cow;
254     int ret = 0;
255     int64_t align_offset = *offset;
256     int64_t align_bytes = *bytes;
257     int max_bytes = s->granularity * s->max_iov;
258 
259     need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap);
260     need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity,
261                           s->cow_bitmap);
262     if (need_cow) {
263         bdrv_round_to_clusters(blk_bs(s->target), *offset, *bytes,
264                                &align_offset, &align_bytes);
265     }
266 
267     if (align_bytes > max_bytes) {
268         align_bytes = max_bytes;
269         if (need_cow) {
270             align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size);
271         }
272     }
273     /* Clipping may result in align_bytes unaligned to chunk boundary, but
274      * that doesn't matter because it's already the end of source image. */
275     align_bytes = mirror_clip_bytes(s, align_offset, align_bytes);
276 
277     ret = align_offset + align_bytes - (*offset + *bytes);
278     *offset = align_offset;
279     *bytes = align_bytes;
280     assert(ret >= 0);
281     return ret;
282 }
283 
284 static inline void coroutine_fn
285 mirror_wait_for_any_operation(MirrorBlockJob *s, bool active)
286 {
287     MirrorOp *op;
288 
289     QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
290         /* Do not wait on pseudo ops, because it may in turn wait on
291          * some other operation to start, which may in fact be the
292          * caller of this function.  Since there is only one pseudo op
293          * at any given time, we will always find some real operation
294          * to wait on. */
295         if (!op->is_pseudo_op && op->is_active_write == active) {
296             qemu_co_queue_wait(&op->waiting_requests, NULL);
297             return;
298         }
299     }
300     abort();
301 }
302 
303 static inline void coroutine_fn
304 mirror_wait_for_free_in_flight_slot(MirrorBlockJob *s)
305 {
306     /* Only non-active operations use up in-flight slots */
307     mirror_wait_for_any_operation(s, false);
308 }
309 
310 /* Perform a mirror copy operation.
311  *
312  * *op->bytes_handled is set to the number of bytes copied after and
313  * including offset, excluding any bytes copied prior to offset due
314  * to alignment.  This will be op->bytes if no alignment is necessary,
315  * or (new_end - op->offset) if the tail is rounded up or down due to
316  * alignment or buffer limit.
317  */
318 static void coroutine_fn mirror_co_read(void *opaque)
319 {
320     MirrorOp *op = opaque;
321     MirrorBlockJob *s = op->s;
322     int nb_chunks;
323     uint64_t ret;
324     uint64_t max_bytes;
325 
326     max_bytes = s->granularity * s->max_iov;
327 
328     /* We can only handle as much as buf_size at a time. */
329     op->bytes = MIN(s->buf_size, MIN(max_bytes, op->bytes));
330     assert(op->bytes);
331     assert(op->bytes < BDRV_REQUEST_MAX_BYTES);
332     *op->bytes_handled = op->bytes;
333 
334     if (s->cow_bitmap) {
335         *op->bytes_handled += mirror_cow_align(s, &op->offset, &op->bytes);
336     }
337     /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */
338     assert(*op->bytes_handled <= UINT_MAX);
339     assert(op->bytes <= s->buf_size);
340     /* The offset is granularity-aligned because:
341      * 1) Caller passes in aligned values;
342      * 2) mirror_cow_align is used only when target cluster is larger. */
343     assert(QEMU_IS_ALIGNED(op->offset, s->granularity));
344     /* The range is sector-aligned, since bdrv_getlength() rounds up. */
345     assert(QEMU_IS_ALIGNED(op->bytes, BDRV_SECTOR_SIZE));
346     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
347 
348     while (s->buf_free_count < nb_chunks) {
349         trace_mirror_yield_in_flight(s, op->offset, s->in_flight);
350         mirror_wait_for_free_in_flight_slot(s);
351     }
352 
353     /* Now make a QEMUIOVector taking enough granularity-sized chunks
354      * from s->buf_free.
355      */
356     qemu_iovec_init(&op->qiov, nb_chunks);
357     while (nb_chunks-- > 0) {
358         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
359         size_t remaining = op->bytes - op->qiov.size;
360 
361         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
362         s->buf_free_count--;
363         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
364     }
365 
366     /* Copy the dirty cluster.  */
367     s->in_flight++;
368     s->bytes_in_flight += op->bytes;
369     trace_mirror_one_iteration(s, op->offset, op->bytes);
370 
371     ret = bdrv_co_preadv(s->mirror_top_bs->backing, op->offset, op->bytes,
372                          &op->qiov, 0);
373     mirror_read_complete(op, ret);
374 }
375 
376 static void coroutine_fn mirror_co_zero(void *opaque)
377 {
378     MirrorOp *op = opaque;
379     int ret;
380 
381     op->s->in_flight++;
382     op->s->bytes_in_flight += op->bytes;
383     *op->bytes_handled = op->bytes;
384 
385     ret = blk_co_pwrite_zeroes(op->s->target, op->offset, op->bytes,
386                                op->s->unmap ? BDRV_REQ_MAY_UNMAP : 0);
387     mirror_write_complete(op, ret);
388 }
389 
390 static void coroutine_fn mirror_co_discard(void *opaque)
391 {
392     MirrorOp *op = opaque;
393     int ret;
394 
395     op->s->in_flight++;
396     op->s->bytes_in_flight += op->bytes;
397     *op->bytes_handled = op->bytes;
398 
399     ret = blk_co_pdiscard(op->s->target, op->offset, op->bytes);
400     mirror_write_complete(op, ret);
401 }
402 
403 static unsigned mirror_perform(MirrorBlockJob *s, int64_t offset,
404                                unsigned bytes, MirrorMethod mirror_method)
405 {
406     MirrorOp *op;
407     Coroutine *co;
408     int64_t bytes_handled = -1;
409 
410     op = g_new(MirrorOp, 1);
411     *op = (MirrorOp){
412         .s              = s,
413         .offset         = offset,
414         .bytes          = bytes,
415         .bytes_handled  = &bytes_handled,
416     };
417     qemu_co_queue_init(&op->waiting_requests);
418 
419     switch (mirror_method) {
420     case MIRROR_METHOD_COPY:
421         co = qemu_coroutine_create(mirror_co_read, op);
422         break;
423     case MIRROR_METHOD_ZERO:
424         co = qemu_coroutine_create(mirror_co_zero, op);
425         break;
426     case MIRROR_METHOD_DISCARD:
427         co = qemu_coroutine_create(mirror_co_discard, op);
428         break;
429     default:
430         abort();
431     }
432 
433     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
434     qemu_coroutine_enter(co);
435     /* At this point, ownership of op has been moved to the coroutine
436      * and the object may already be freed */
437 
438     /* Assert that this value has been set */
439     assert(bytes_handled >= 0);
440 
441     /* Same assertion as in mirror_co_read() (and for mirror_co_read()
442      * and mirror_co_discard(), bytes_handled == op->bytes, which
443      * is the @bytes parameter given to this function) */
444     assert(bytes_handled <= UINT_MAX);
445     return bytes_handled;
446 }
447 
448 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
449 {
450     BlockDriverState *source = s->mirror_top_bs->backing->bs;
451     MirrorOp *pseudo_op;
452     int64_t offset;
453     uint64_t delay_ns = 0, ret = 0;
454     /* At least the first dirty chunk is mirrored in one iteration. */
455     int nb_chunks = 1;
456     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
457     int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES);
458 
459     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
460     offset = bdrv_dirty_iter_next(s->dbi);
461     if (offset < 0) {
462         bdrv_set_dirty_iter(s->dbi, 0);
463         offset = bdrv_dirty_iter_next(s->dbi);
464         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
465         assert(offset >= 0);
466     }
467     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
468 
469     mirror_wait_on_conflicts(NULL, s, offset, 1);
470 
471     job_pause_point(&s->common.job);
472 
473     /* Find the number of consective dirty chunks following the first dirty
474      * one, and wait for in flight requests in them. */
475     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
476     while (nb_chunks * s->granularity < s->buf_size) {
477         int64_t next_dirty;
478         int64_t next_offset = offset + nb_chunks * s->granularity;
479         int64_t next_chunk = next_offset / s->granularity;
480         if (next_offset >= s->bdev_length ||
481             !bdrv_dirty_bitmap_get_locked(s->dirty_bitmap, next_offset)) {
482             break;
483         }
484         if (test_bit(next_chunk, s->in_flight_bitmap)) {
485             break;
486         }
487 
488         next_dirty = bdrv_dirty_iter_next(s->dbi);
489         if (next_dirty > next_offset || next_dirty < 0) {
490             /* The bitmap iterator's cache is stale, refresh it */
491             bdrv_set_dirty_iter(s->dbi, next_offset);
492             next_dirty = bdrv_dirty_iter_next(s->dbi);
493         }
494         assert(next_dirty == next_offset);
495         nb_chunks++;
496     }
497 
498     /* Clear dirty bits before querying the block status, because
499      * calling bdrv_block_status_above could yield - if some blocks are
500      * marked dirty in this window, we need to know.
501      */
502     bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset,
503                                    nb_chunks * s->granularity);
504     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
505 
506     /* Before claiming an area in the in-flight bitmap, we have to
507      * create a MirrorOp for it so that conflicting requests can wait
508      * for it.  mirror_perform() will create the real MirrorOps later,
509      * for now we just create a pseudo operation that will wake up all
510      * conflicting requests once all real operations have been
511      * launched. */
512     pseudo_op = g_new(MirrorOp, 1);
513     *pseudo_op = (MirrorOp){
514         .offset         = offset,
515         .bytes          = nb_chunks * s->granularity,
516         .is_pseudo_op   = true,
517     };
518     qemu_co_queue_init(&pseudo_op->waiting_requests);
519     QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next);
520 
521     bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks);
522     while (nb_chunks > 0 && offset < s->bdev_length) {
523         int ret;
524         int64_t io_bytes;
525         int64_t io_bytes_acct;
526         MirrorMethod mirror_method = MIRROR_METHOD_COPY;
527 
528         assert(!(offset % s->granularity));
529         ret = bdrv_block_status_above(source, NULL, offset,
530                                       nb_chunks * s->granularity,
531                                       &io_bytes, NULL, NULL);
532         if (ret < 0) {
533             io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes);
534         } else if (ret & BDRV_BLOCK_DATA) {
535             io_bytes = MIN(io_bytes, max_io_bytes);
536         }
537 
538         io_bytes -= io_bytes % s->granularity;
539         if (io_bytes < s->granularity) {
540             io_bytes = s->granularity;
541         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
542             int64_t target_offset;
543             int64_t target_bytes;
544             bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes,
545                                    &target_offset, &target_bytes);
546             if (target_offset == offset &&
547                 target_bytes == io_bytes) {
548                 mirror_method = ret & BDRV_BLOCK_ZERO ?
549                                     MIRROR_METHOD_ZERO :
550                                     MIRROR_METHOD_DISCARD;
551             }
552         }
553 
554         while (s->in_flight >= MAX_IN_FLIGHT) {
555             trace_mirror_yield_in_flight(s, offset, s->in_flight);
556             mirror_wait_for_free_in_flight_slot(s);
557         }
558 
559         if (s->ret < 0) {
560             ret = 0;
561             goto fail;
562         }
563 
564         io_bytes = mirror_clip_bytes(s, offset, io_bytes);
565         io_bytes = mirror_perform(s, offset, io_bytes, mirror_method);
566         if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) {
567             io_bytes_acct = 0;
568         } else {
569             io_bytes_acct = io_bytes;
570         }
571         assert(io_bytes);
572         offset += io_bytes;
573         nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity);
574         delay_ns = block_job_ratelimit_get_delay(&s->common, io_bytes_acct);
575     }
576 
577     ret = delay_ns;
578 fail:
579     QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next);
580     qemu_co_queue_restart_all(&pseudo_op->waiting_requests);
581     g_free(pseudo_op);
582 
583     return ret;
584 }
585 
586 static void mirror_free_init(MirrorBlockJob *s)
587 {
588     int granularity = s->granularity;
589     size_t buf_size = s->buf_size;
590     uint8_t *buf = s->buf;
591 
592     assert(s->buf_free_count == 0);
593     QSIMPLEQ_INIT(&s->buf_free);
594     while (buf_size != 0) {
595         MirrorBuffer *cur = (MirrorBuffer *)buf;
596         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
597         s->buf_free_count++;
598         buf_size -= granularity;
599         buf += granularity;
600     }
601 }
602 
603 /* This is also used for the .pause callback. There is no matching
604  * mirror_resume() because mirror_run() will begin iterating again
605  * when the job is resumed.
606  */
607 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s)
608 {
609     while (s->in_flight > 0) {
610         mirror_wait_for_free_in_flight_slot(s);
611     }
612 }
613 
614 /**
615  * mirror_exit_common: handle both abort() and prepare() cases.
616  * for .prepare, returns 0 on success and -errno on failure.
617  * for .abort cases, denoted by abort = true, MUST return 0.
618  */
619 static int mirror_exit_common(Job *job)
620 {
621     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
622     BlockJob *bjob = &s->common;
623     MirrorBDSOpaque *bs_opaque = s->mirror_top_bs->opaque;
624     AioContext *replace_aio_context = NULL;
625     BlockDriverState *src = s->mirror_top_bs->backing->bs;
626     BlockDriverState *target_bs = blk_bs(s->target);
627     BlockDriverState *mirror_top_bs = s->mirror_top_bs;
628     Error *local_err = NULL;
629     bool abort = job->ret < 0;
630     int ret = 0;
631 
632     if (s->prepared) {
633         return 0;
634     }
635     s->prepared = true;
636 
637     if (bdrv_chain_contains(src, target_bs)) {
638         bdrv_unfreeze_backing_chain(mirror_top_bs, target_bs);
639     }
640 
641     bdrv_release_dirty_bitmap(src, s->dirty_bitmap);
642 
643     /* Make sure that the source BDS doesn't go away during bdrv_replace_node,
644      * before we can call bdrv_drained_end */
645     bdrv_ref(src);
646     bdrv_ref(mirror_top_bs);
647     bdrv_ref(target_bs);
648 
649     /* Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
650      * inserting target_bs at s->to_replace, where we might not be able to get
651      * these permissions.
652      *
653      * Note that blk_unref() alone doesn't necessarily drop permissions because
654      * we might be running nested inside mirror_drain(), which takes an extra
655      * reference, so use an explicit blk_set_perm() first. */
656     blk_set_perm(s->target, 0, BLK_PERM_ALL, &error_abort);
657     blk_unref(s->target);
658     s->target = NULL;
659 
660     /* We don't access the source any more. Dropping any WRITE/RESIZE is
661      * required before it could become a backing file of target_bs. Not having
662      * these permissions any more means that we can't allow any new requests on
663      * mirror_top_bs from now on, so keep it drained. */
664     bdrv_drained_begin(mirror_top_bs);
665     bs_opaque->stop = true;
666     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
667                              &error_abort);
668     if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
669         BlockDriverState *backing = s->is_none_mode ? src : s->base;
670         if (backing_bs(target_bs) != backing) {
671             bdrv_set_backing_hd(target_bs, backing, &local_err);
672             if (local_err) {
673                 error_report_err(local_err);
674                 ret = -EPERM;
675             }
676         }
677     }
678 
679     if (s->to_replace) {
680         replace_aio_context = bdrv_get_aio_context(s->to_replace);
681         aio_context_acquire(replace_aio_context);
682     }
683 
684     if (s->should_complete && !abort) {
685         BlockDriverState *to_replace = s->to_replace ?: src;
686         bool ro = bdrv_is_read_only(to_replace);
687 
688         if (ro != bdrv_is_read_only(target_bs)) {
689             bdrv_reopen_set_read_only(target_bs, ro, NULL);
690         }
691 
692         /* The mirror job has no requests in flight any more, but we need to
693          * drain potential other users of the BDS before changing the graph. */
694         assert(s->in_drain);
695         bdrv_drained_begin(target_bs);
696         bdrv_replace_node(to_replace, target_bs, &local_err);
697         bdrv_drained_end(target_bs);
698         if (local_err) {
699             error_report_err(local_err);
700             ret = -EPERM;
701         }
702     }
703     if (s->to_replace) {
704         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
705         error_free(s->replace_blocker);
706         bdrv_unref(s->to_replace);
707     }
708     if (replace_aio_context) {
709         aio_context_release(replace_aio_context);
710     }
711     g_free(s->replaces);
712     bdrv_unref(target_bs);
713 
714     /*
715      * Remove the mirror filter driver from the graph. Before this, get rid of
716      * the blockers on the intermediate nodes so that the resulting state is
717      * valid.
718      */
719     block_job_remove_all_bdrv(bjob);
720     bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
721 
722     /* We just changed the BDS the job BB refers to (with either or both of the
723      * bdrv_replace_node() calls), so switch the BB back so the cleanup does
724      * the right thing. We don't need any permissions any more now. */
725     blk_remove_bs(bjob->blk);
726     blk_set_perm(bjob->blk, 0, BLK_PERM_ALL, &error_abort);
727     blk_insert_bs(bjob->blk, mirror_top_bs, &error_abort);
728 
729     bs_opaque->job = NULL;
730 
731     bdrv_drained_end(src);
732     bdrv_drained_end(mirror_top_bs);
733     s->in_drain = false;
734     bdrv_unref(mirror_top_bs);
735     bdrv_unref(src);
736 
737     return ret;
738 }
739 
740 static int mirror_prepare(Job *job)
741 {
742     return mirror_exit_common(job);
743 }
744 
745 static void mirror_abort(Job *job)
746 {
747     int ret = mirror_exit_common(job);
748     assert(ret == 0);
749 }
750 
751 static void coroutine_fn mirror_throttle(MirrorBlockJob *s)
752 {
753     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
754 
755     if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) {
756         s->last_pause_ns = now;
757         job_sleep_ns(&s->common.job, 0);
758     } else {
759         job_pause_point(&s->common.job);
760     }
761 }
762 
763 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
764 {
765     int64_t offset;
766     BlockDriverState *base = s->base;
767     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
768     BlockDriverState *target_bs = blk_bs(s->target);
769     int ret;
770     int64_t count;
771 
772     if (s->zero_target) {
773         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
774             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
775             return 0;
776         }
777 
778         s->initial_zeroing_ongoing = true;
779         for (offset = 0; offset < s->bdev_length; ) {
780             int bytes = MIN(s->bdev_length - offset,
781                             QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
782 
783             mirror_throttle(s);
784 
785             if (job_is_cancelled(&s->common.job)) {
786                 s->initial_zeroing_ongoing = false;
787                 return 0;
788             }
789 
790             if (s->in_flight >= MAX_IN_FLIGHT) {
791                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
792                                    s->in_flight);
793                 mirror_wait_for_free_in_flight_slot(s);
794                 continue;
795             }
796 
797             mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO);
798             offset += bytes;
799         }
800 
801         mirror_wait_for_all_io(s);
802         s->initial_zeroing_ongoing = false;
803     }
804 
805     /* First part, loop on the sectors and initialize the dirty bitmap.  */
806     for (offset = 0; offset < s->bdev_length; ) {
807         /* Just to make sure we are not exceeding int limit. */
808         int bytes = MIN(s->bdev_length - offset,
809                         QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
810 
811         mirror_throttle(s);
812 
813         if (job_is_cancelled(&s->common.job)) {
814             return 0;
815         }
816 
817         ret = bdrv_is_allocated_above(bs, base, false, offset, bytes, &count);
818         if (ret < 0) {
819             return ret;
820         }
821 
822         assert(count);
823         if (ret == 1) {
824             bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
825         }
826         offset += count;
827     }
828     return 0;
829 }
830 
831 /* Called when going out of the streaming phase to flush the bulk of the
832  * data to the medium, or just before completing.
833  */
834 static int mirror_flush(MirrorBlockJob *s)
835 {
836     int ret = blk_flush(s->target);
837     if (ret < 0) {
838         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
839             s->ret = ret;
840         }
841     }
842     return ret;
843 }
844 
845 static int coroutine_fn mirror_run(Job *job, Error **errp)
846 {
847     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
848     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
849     BlockDriverState *target_bs = blk_bs(s->target);
850     bool need_drain = true;
851     int64_t length;
852     BlockDriverInfo bdi;
853     char backing_filename[2]; /* we only need 2 characters because we are only
854                                  checking for a NULL string */
855     int ret = 0;
856 
857     if (job_is_cancelled(&s->common.job)) {
858         goto immediate_exit;
859     }
860 
861     s->bdev_length = bdrv_getlength(bs);
862     if (s->bdev_length < 0) {
863         ret = s->bdev_length;
864         goto immediate_exit;
865     }
866 
867     /* Active commit must resize the base image if its size differs from the
868      * active layer. */
869     if (s->base == blk_bs(s->target)) {
870         int64_t base_length;
871 
872         base_length = blk_getlength(s->target);
873         if (base_length < 0) {
874             ret = base_length;
875             goto immediate_exit;
876         }
877 
878         if (s->bdev_length > base_length) {
879             ret = blk_truncate(s->target, s->bdev_length, PREALLOC_MODE_OFF,
880                                NULL);
881             if (ret < 0) {
882                 goto immediate_exit;
883             }
884         }
885     }
886 
887     if (s->bdev_length == 0) {
888         /* Transition to the READY state and wait for complete. */
889         job_transition_to_ready(&s->common.job);
890         s->synced = true;
891         s->actively_synced = true;
892         while (!job_is_cancelled(&s->common.job) && !s->should_complete) {
893             job_yield(&s->common.job);
894         }
895         s->common.job.cancelled = false;
896         goto immediate_exit;
897     }
898 
899     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
900     s->in_flight_bitmap = bitmap_new(length);
901 
902     /* If we have no backing file yet in the destination, we cannot let
903      * the destination do COW.  Instead, we copy sectors around the
904      * dirty data if needed.  We need a bitmap to do that.
905      */
906     bdrv_get_backing_filename(target_bs, backing_filename,
907                               sizeof(backing_filename));
908     if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) {
909         s->target_cluster_size = bdi.cluster_size;
910     } else {
911         s->target_cluster_size = BDRV_SECTOR_SIZE;
912     }
913     if (backing_filename[0] && !target_bs->backing &&
914         s->granularity < s->target_cluster_size) {
915         s->buf_size = MAX(s->buf_size, s->target_cluster_size);
916         s->cow_bitmap = bitmap_new(length);
917     }
918     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
919 
920     s->buf = qemu_try_blockalign(bs, s->buf_size);
921     if (s->buf == NULL) {
922         ret = -ENOMEM;
923         goto immediate_exit;
924     }
925 
926     mirror_free_init(s);
927 
928     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
929     if (!s->is_none_mode) {
930         ret = mirror_dirty_init(s);
931         if (ret < 0 || job_is_cancelled(&s->common.job)) {
932             goto immediate_exit;
933         }
934     }
935 
936     assert(!s->dbi);
937     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
938     for (;;) {
939         uint64_t delay_ns = 0;
940         int64_t cnt, delta;
941         bool should_complete;
942 
943         /* Do not start passive operations while there are active
944          * writes in progress */
945         while (s->in_active_write_counter) {
946             mirror_wait_for_any_operation(s, true);
947         }
948 
949         if (s->ret < 0) {
950             ret = s->ret;
951             goto immediate_exit;
952         }
953 
954         job_pause_point(&s->common.job);
955 
956         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
957         /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
958          * the number of bytes currently being processed; together those are
959          * the current remaining operation length */
960         job_progress_set_remaining(&s->common.job, s->bytes_in_flight + cnt);
961 
962         /* Note that even when no rate limit is applied we need to yield
963          * periodically with no pending I/O so that bdrv_drain_all() returns.
964          * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
965          * an error, or when the source is clean, whichever comes first. */
966         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
967         if (delta < BLOCK_JOB_SLICE_TIME &&
968             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
969             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
970                 (cnt == 0 && s->in_flight > 0)) {
971                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
972                 mirror_wait_for_free_in_flight_slot(s);
973                 continue;
974             } else if (cnt != 0) {
975                 delay_ns = mirror_iteration(s);
976             }
977         }
978 
979         should_complete = false;
980         if (s->in_flight == 0 && cnt == 0) {
981             trace_mirror_before_flush(s);
982             if (!s->synced) {
983                 if (mirror_flush(s) < 0) {
984                     /* Go check s->ret.  */
985                     continue;
986                 }
987                 /* We're out of the streaming phase.  From now on, if the job
988                  * is cancelled we will actually complete all pending I/O and
989                  * report completion.  This way, block-job-cancel will leave
990                  * the target in a consistent state.
991                  */
992                 job_transition_to_ready(&s->common.job);
993                 s->synced = true;
994                 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) {
995                     s->actively_synced = true;
996                 }
997             }
998 
999             should_complete = s->should_complete ||
1000                 job_is_cancelled(&s->common.job);
1001             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1002         }
1003 
1004         if (cnt == 0 && should_complete) {
1005             /* The dirty bitmap is not updated while operations are pending.
1006              * If we're about to exit, wait for pending operations before
1007              * calling bdrv_get_dirty_count(bs), or we may exit while the
1008              * source has dirty data to copy!
1009              *
1010              * Note that I/O can be submitted by the guest while
1011              * mirror_populate runs, so pause it now.  Before deciding
1012              * whether to switch to target check one last time if I/O has
1013              * come in the meanwhile, and if not flush the data to disk.
1014              */
1015             trace_mirror_before_drain(s, cnt);
1016 
1017             s->in_drain = true;
1018             bdrv_drained_begin(bs);
1019             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1020             if (cnt > 0 || mirror_flush(s) < 0) {
1021                 bdrv_drained_end(bs);
1022                 s->in_drain = false;
1023                 continue;
1024             }
1025 
1026             /* The two disks are in sync.  Exit and report successful
1027              * completion.
1028              */
1029             assert(QLIST_EMPTY(&bs->tracked_requests));
1030             s->common.job.cancelled = false;
1031             need_drain = false;
1032             break;
1033         }
1034 
1035         ret = 0;
1036 
1037         if (s->synced && !should_complete) {
1038             delay_ns = (s->in_flight == 0 &&
1039                         cnt == 0 ? BLOCK_JOB_SLICE_TIME : 0);
1040         }
1041         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
1042         job_sleep_ns(&s->common.job, delay_ns);
1043         if (job_is_cancelled(&s->common.job) &&
1044             (!s->synced || s->common.job.force_cancel))
1045         {
1046             break;
1047         }
1048         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1049     }
1050 
1051 immediate_exit:
1052     if (s->in_flight > 0) {
1053         /* We get here only if something went wrong.  Either the job failed,
1054          * or it was cancelled prematurely so that we do not guarantee that
1055          * the target is a copy of the source.
1056          */
1057         assert(ret < 0 || ((s->common.job.force_cancel || !s->synced) &&
1058                job_is_cancelled(&s->common.job)));
1059         assert(need_drain);
1060         mirror_wait_for_all_io(s);
1061     }
1062 
1063     assert(s->in_flight == 0);
1064     qemu_vfree(s->buf);
1065     g_free(s->cow_bitmap);
1066     g_free(s->in_flight_bitmap);
1067     bdrv_dirty_iter_free(s->dbi);
1068 
1069     if (need_drain) {
1070         s->in_drain = true;
1071         bdrv_drained_begin(bs);
1072     }
1073 
1074     return ret;
1075 }
1076 
1077 static void mirror_complete(Job *job, Error **errp)
1078 {
1079     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1080     BlockDriverState *target;
1081 
1082     target = blk_bs(s->target);
1083 
1084     if (!s->synced) {
1085         error_setg(errp, "The active block job '%s' cannot be completed",
1086                    job->id);
1087         return;
1088     }
1089 
1090     if (s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
1091         int ret;
1092 
1093         assert(!target->backing);
1094         ret = bdrv_open_backing_file(target, NULL, "backing", errp);
1095         if (ret < 0) {
1096             return;
1097         }
1098     }
1099 
1100     /* block all operations on to_replace bs */
1101     if (s->replaces) {
1102         AioContext *replace_aio_context;
1103 
1104         s->to_replace = bdrv_find_node(s->replaces);
1105         if (!s->to_replace) {
1106             error_setg(errp, "Node name '%s' not found", s->replaces);
1107             return;
1108         }
1109 
1110         replace_aio_context = bdrv_get_aio_context(s->to_replace);
1111         aio_context_acquire(replace_aio_context);
1112 
1113         /* TODO Translate this into permission system. Current definition of
1114          * GRAPH_MOD would require to request it for the parents; they might
1115          * not even be BlockDriverStates, however, so a BdrvChild can't address
1116          * them. May need redefinition of GRAPH_MOD. */
1117         error_setg(&s->replace_blocker,
1118                    "block device is in use by block-job-complete");
1119         bdrv_op_block_all(s->to_replace, s->replace_blocker);
1120         bdrv_ref(s->to_replace);
1121 
1122         aio_context_release(replace_aio_context);
1123     }
1124 
1125     s->should_complete = true;
1126     job_enter(job);
1127 }
1128 
1129 static void coroutine_fn mirror_pause(Job *job)
1130 {
1131     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1132 
1133     mirror_wait_for_all_io(s);
1134 }
1135 
1136 static bool mirror_drained_poll(BlockJob *job)
1137 {
1138     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1139 
1140     /* If the job isn't paused nor cancelled, we can't be sure that it won't
1141      * issue more requests. We make an exception if we've reached this point
1142      * from one of our own drain sections, to avoid a deadlock waiting for
1143      * ourselves.
1144      */
1145     if (!s->common.job.paused && !s->common.job.cancelled && !s->in_drain) {
1146         return true;
1147     }
1148 
1149     return !!s->in_flight;
1150 }
1151 
1152 static void mirror_drain(BlockJob *job)
1153 {
1154     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1155 
1156     /* Need to keep a reference in case blk_drain triggers execution
1157      * of mirror_complete...
1158      */
1159     if (s->target) {
1160         BlockBackend *target = s->target;
1161         blk_ref(target);
1162         blk_drain(target);
1163         blk_unref(target);
1164     }
1165 }
1166 
1167 static const BlockJobDriver mirror_job_driver = {
1168     .job_driver = {
1169         .instance_size          = sizeof(MirrorBlockJob),
1170         .job_type               = JOB_TYPE_MIRROR,
1171         .free                   = block_job_free,
1172         .user_resume            = block_job_user_resume,
1173         .drain                  = block_job_drain,
1174         .run                    = mirror_run,
1175         .prepare                = mirror_prepare,
1176         .abort                  = mirror_abort,
1177         .pause                  = mirror_pause,
1178         .complete               = mirror_complete,
1179     },
1180     .drained_poll           = mirror_drained_poll,
1181     .drain                  = mirror_drain,
1182 };
1183 
1184 static const BlockJobDriver commit_active_job_driver = {
1185     .job_driver = {
1186         .instance_size          = sizeof(MirrorBlockJob),
1187         .job_type               = JOB_TYPE_COMMIT,
1188         .free                   = block_job_free,
1189         .user_resume            = block_job_user_resume,
1190         .drain                  = block_job_drain,
1191         .run                    = mirror_run,
1192         .prepare                = mirror_prepare,
1193         .abort                  = mirror_abort,
1194         .pause                  = mirror_pause,
1195         .complete               = mirror_complete,
1196     },
1197     .drained_poll           = mirror_drained_poll,
1198     .drain                  = mirror_drain,
1199 };
1200 
1201 static void coroutine_fn
1202 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method,
1203                      uint64_t offset, uint64_t bytes,
1204                      QEMUIOVector *qiov, int flags)
1205 {
1206     QEMUIOVector target_qiov;
1207     uint64_t dirty_offset = offset;
1208     uint64_t dirty_bytes;
1209 
1210     if (qiov) {
1211         qemu_iovec_init(&target_qiov, qiov->niov);
1212     }
1213 
1214     while (true) {
1215         bool valid_area;
1216         int ret;
1217 
1218         bdrv_dirty_bitmap_lock(job->dirty_bitmap);
1219         dirty_bytes = MIN(offset + bytes - dirty_offset, INT_MAX);
1220         valid_area = bdrv_dirty_bitmap_next_dirty_area(job->dirty_bitmap,
1221                                                        &dirty_offset,
1222                                                        &dirty_bytes);
1223         if (!valid_area) {
1224             bdrv_dirty_bitmap_unlock(job->dirty_bitmap);
1225             break;
1226         }
1227 
1228         bdrv_reset_dirty_bitmap_locked(job->dirty_bitmap,
1229                                        dirty_offset, dirty_bytes);
1230         bdrv_dirty_bitmap_unlock(job->dirty_bitmap);
1231 
1232         job_progress_increase_remaining(&job->common.job, dirty_bytes);
1233 
1234         assert(dirty_offset - offset <= SIZE_MAX);
1235         if (qiov) {
1236             qemu_iovec_reset(&target_qiov);
1237             qemu_iovec_concat(&target_qiov, qiov,
1238                               dirty_offset - offset, dirty_bytes);
1239         }
1240 
1241         switch (method) {
1242         case MIRROR_METHOD_COPY:
1243             ret = blk_co_pwritev(job->target, dirty_offset, dirty_bytes,
1244                                  qiov ? &target_qiov : NULL, flags);
1245             break;
1246 
1247         case MIRROR_METHOD_ZERO:
1248             assert(!qiov);
1249             ret = blk_co_pwrite_zeroes(job->target, dirty_offset, dirty_bytes,
1250                                        flags);
1251             break;
1252 
1253         case MIRROR_METHOD_DISCARD:
1254             assert(!qiov);
1255             ret = blk_co_pdiscard(job->target, dirty_offset, dirty_bytes);
1256             break;
1257 
1258         default:
1259             abort();
1260         }
1261 
1262         if (ret >= 0) {
1263             job_progress_update(&job->common.job, dirty_bytes);
1264         } else {
1265             BlockErrorAction action;
1266 
1267             bdrv_set_dirty_bitmap(job->dirty_bitmap, dirty_offset, dirty_bytes);
1268             job->actively_synced = false;
1269 
1270             action = mirror_error_action(job, false, -ret);
1271             if (action == BLOCK_ERROR_ACTION_REPORT) {
1272                 if (!job->ret) {
1273                     job->ret = ret;
1274                 }
1275                 break;
1276             }
1277         }
1278 
1279         dirty_offset += dirty_bytes;
1280     }
1281 
1282     if (qiov) {
1283         qemu_iovec_destroy(&target_qiov);
1284     }
1285 }
1286 
1287 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s,
1288                                                    uint64_t offset,
1289                                                    uint64_t bytes)
1290 {
1291     MirrorOp *op;
1292     uint64_t start_chunk = offset / s->granularity;
1293     uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
1294 
1295     op = g_new(MirrorOp, 1);
1296     *op = (MirrorOp){
1297         .s                  = s,
1298         .offset             = offset,
1299         .bytes              = bytes,
1300         .is_active_write    = true,
1301     };
1302     qemu_co_queue_init(&op->waiting_requests);
1303     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
1304 
1305     s->in_active_write_counter++;
1306 
1307     mirror_wait_on_conflicts(op, s, offset, bytes);
1308 
1309     bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1310 
1311     return op;
1312 }
1313 
1314 static void coroutine_fn active_write_settle(MirrorOp *op)
1315 {
1316     uint64_t start_chunk = op->offset / op->s->granularity;
1317     uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes,
1318                                       op->s->granularity);
1319 
1320     if (!--op->s->in_active_write_counter && op->s->actively_synced) {
1321         BdrvChild *source = op->s->mirror_top_bs->backing;
1322 
1323         if (QLIST_FIRST(&source->bs->parents) == source &&
1324             QLIST_NEXT(source, next_parent) == NULL)
1325         {
1326             /* Assert that we are back in sync once all active write
1327              * operations are settled.
1328              * Note that we can only assert this if the mirror node
1329              * is the source node's only parent. */
1330             assert(!bdrv_get_dirty_count(op->s->dirty_bitmap));
1331         }
1332     }
1333     bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1334     QTAILQ_REMOVE(&op->s->ops_in_flight, op, next);
1335     qemu_co_queue_restart_all(&op->waiting_requests);
1336     g_free(op);
1337 }
1338 
1339 static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs,
1340     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1341 {
1342     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1343 }
1344 
1345 static int coroutine_fn bdrv_mirror_top_do_write(BlockDriverState *bs,
1346     MirrorMethod method, uint64_t offset, uint64_t bytes, QEMUIOVector *qiov,
1347     int flags)
1348 {
1349     MirrorOp *op = NULL;
1350     MirrorBDSOpaque *s = bs->opaque;
1351     int ret = 0;
1352     bool copy_to_target;
1353 
1354     copy_to_target = s->job->ret >= 0 &&
1355                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1356 
1357     if (copy_to_target) {
1358         op = active_write_prepare(s->job, offset, bytes);
1359     }
1360 
1361     switch (method) {
1362     case MIRROR_METHOD_COPY:
1363         ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1364         break;
1365 
1366     case MIRROR_METHOD_ZERO:
1367         ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1368         break;
1369 
1370     case MIRROR_METHOD_DISCARD:
1371         ret = bdrv_co_pdiscard(bs->backing, offset, bytes);
1372         break;
1373 
1374     default:
1375         abort();
1376     }
1377 
1378     if (ret < 0) {
1379         goto out;
1380     }
1381 
1382     if (copy_to_target) {
1383         do_sync_target_write(s->job, method, offset, bytes, qiov, flags);
1384     }
1385 
1386 out:
1387     if (copy_to_target) {
1388         active_write_settle(op);
1389     }
1390     return ret;
1391 }
1392 
1393 static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs,
1394     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1395 {
1396     MirrorBDSOpaque *s = bs->opaque;
1397     QEMUIOVector bounce_qiov;
1398     void *bounce_buf;
1399     int ret = 0;
1400     bool copy_to_target;
1401 
1402     copy_to_target = s->job->ret >= 0 &&
1403                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1404 
1405     if (copy_to_target) {
1406         /* The guest might concurrently modify the data to write; but
1407          * the data on source and destination must match, so we have
1408          * to use a bounce buffer if we are going to write to the
1409          * target now. */
1410         bounce_buf = qemu_blockalign(bs, bytes);
1411         iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes);
1412 
1413         qemu_iovec_init(&bounce_qiov, 1);
1414         qemu_iovec_add(&bounce_qiov, bounce_buf, bytes);
1415         qiov = &bounce_qiov;
1416     }
1417 
1418     ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov,
1419                                    flags);
1420 
1421     if (copy_to_target) {
1422         qemu_iovec_destroy(&bounce_qiov);
1423         qemu_vfree(bounce_buf);
1424     }
1425 
1426     return ret;
1427 }
1428 
1429 static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs)
1430 {
1431     if (bs->backing == NULL) {
1432         /* we can be here after failed bdrv_append in mirror_start_job */
1433         return 0;
1434     }
1435     return bdrv_co_flush(bs->backing->bs);
1436 }
1437 
1438 static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs,
1439     int64_t offset, int bytes, BdrvRequestFlags flags)
1440 {
1441     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL,
1442                                     flags);
1443 }
1444 
1445 static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs,
1446     int64_t offset, int bytes)
1447 {
1448     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes,
1449                                     NULL, 0);
1450 }
1451 
1452 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs)
1453 {
1454     if (bs->backing == NULL) {
1455         /* we can be here after failed bdrv_attach_child in
1456          * bdrv_set_backing_hd */
1457         return;
1458     }
1459     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1460             bs->backing->bs->filename);
1461 }
1462 
1463 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1464                                        const BdrvChildRole *role,
1465                                        BlockReopenQueue *reopen_queue,
1466                                        uint64_t perm, uint64_t shared,
1467                                        uint64_t *nperm, uint64_t *nshared)
1468 {
1469     MirrorBDSOpaque *s = bs->opaque;
1470 
1471     if (s->stop) {
1472         /*
1473          * If the job is to be stopped, we do not need to forward
1474          * anything to the real image.
1475          */
1476         *nperm = 0;
1477         *nshared = BLK_PERM_ALL;
1478         return;
1479     }
1480 
1481     /* Must be able to forward guest writes to the real image */
1482     *nperm = 0;
1483     if (perm & BLK_PERM_WRITE) {
1484         *nperm |= BLK_PERM_WRITE;
1485     }
1486 
1487     *nshared = BLK_PERM_ALL;
1488 }
1489 
1490 static void bdrv_mirror_top_refresh_limits(BlockDriverState *bs, Error **errp)
1491 {
1492     MirrorBDSOpaque *s = bs->opaque;
1493 
1494     if (s && s->job && s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING) {
1495         bs->bl.request_alignment = s->job->granularity;
1496     }
1497 }
1498 
1499 /* Dummy node that provides consistent read to its users without requiring it
1500  * from its backing file and that allows writes on the backing file chain. */
1501 static BlockDriver bdrv_mirror_top = {
1502     .format_name                = "mirror_top",
1503     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1504     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1505     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1506     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1507     .bdrv_co_flush              = bdrv_mirror_top_flush,
1508     .bdrv_co_block_status       = bdrv_co_block_status_from_backing,
1509     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1510     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1511     .bdrv_refresh_limits        = bdrv_mirror_top_refresh_limits,
1512 };
1513 
1514 static BlockJob *mirror_start_job(
1515                              const char *job_id, BlockDriverState *bs,
1516                              int creation_flags, BlockDriverState *target,
1517                              const char *replaces, int64_t speed,
1518                              uint32_t granularity, int64_t buf_size,
1519                              BlockMirrorBackingMode backing_mode,
1520                              bool zero_target,
1521                              BlockdevOnError on_source_error,
1522                              BlockdevOnError on_target_error,
1523                              bool unmap,
1524                              BlockCompletionFunc *cb,
1525                              void *opaque,
1526                              const BlockJobDriver *driver,
1527                              bool is_none_mode, BlockDriverState *base,
1528                              bool auto_complete, const char *filter_node_name,
1529                              bool is_mirror, MirrorCopyMode copy_mode,
1530                              Error **errp)
1531 {
1532     MirrorBlockJob *s;
1533     MirrorBDSOpaque *bs_opaque;
1534     BlockDriverState *mirror_top_bs;
1535     bool target_graph_mod;
1536     bool target_is_backing;
1537     Error *local_err = NULL;
1538     int ret;
1539 
1540     if (granularity == 0) {
1541         granularity = bdrv_get_default_bitmap_granularity(target);
1542     }
1543 
1544     assert(is_power_of_2(granularity));
1545 
1546     if (buf_size < 0) {
1547         error_setg(errp, "Invalid parameter 'buf-size'");
1548         return NULL;
1549     }
1550 
1551     if (buf_size == 0) {
1552         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1553     }
1554 
1555     if (bs == target) {
1556         error_setg(errp, "Can't mirror node into itself");
1557         return NULL;
1558     }
1559 
1560     /* In the case of active commit, add dummy driver to provide consistent
1561      * reads on the top, while disabling it in the intermediate nodes, and make
1562      * the backing chain writable. */
1563     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1564                                          BDRV_O_RDWR, errp);
1565     if (mirror_top_bs == NULL) {
1566         return NULL;
1567     }
1568     if (!filter_node_name) {
1569         mirror_top_bs->implicit = true;
1570     }
1571 
1572     /* So that we can always drop this node */
1573     mirror_top_bs->never_freeze = true;
1574 
1575     mirror_top_bs->total_sectors = bs->total_sectors;
1576     mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
1577     mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
1578                                           BDRV_REQ_NO_FALLBACK;
1579     bs_opaque = g_new0(MirrorBDSOpaque, 1);
1580     mirror_top_bs->opaque = bs_opaque;
1581 
1582     /* bdrv_append takes ownership of the mirror_top_bs reference, need to keep
1583      * it alive until block_job_create() succeeds even if bs has no parent. */
1584     bdrv_ref(mirror_top_bs);
1585     bdrv_drained_begin(bs);
1586     bdrv_append(mirror_top_bs, bs, &local_err);
1587     bdrv_drained_end(bs);
1588 
1589     if (local_err) {
1590         bdrv_unref(mirror_top_bs);
1591         error_propagate(errp, local_err);
1592         return NULL;
1593     }
1594 
1595     /* Make sure that the source is not resized while the job is running */
1596     s = block_job_create(job_id, driver, NULL, mirror_top_bs,
1597                          BLK_PERM_CONSISTENT_READ,
1598                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1599                          BLK_PERM_WRITE | BLK_PERM_GRAPH_MOD, speed,
1600                          creation_flags, cb, opaque, errp);
1601     if (!s) {
1602         goto fail;
1603     }
1604     bs_opaque->job = s;
1605 
1606     /* The block job now has a reference to this node */
1607     bdrv_unref(mirror_top_bs);
1608 
1609     s->mirror_top_bs = mirror_top_bs;
1610 
1611     /* No resize for the target either; while the mirror is still running, a
1612      * consistent read isn't necessarily possible. We could possibly allow
1613      * writes and graph modifications, though it would likely defeat the
1614      * purpose of a mirror, so leave them blocked for now.
1615      *
1616      * In the case of active commit, things look a bit different, though,
1617      * because the target is an already populated backing file in active use.
1618      * We can allow anything except resize there.*/
1619     target_is_backing = bdrv_chain_contains(bs, target);
1620     target_graph_mod = (backing_mode != MIRROR_LEAVE_BACKING_CHAIN);
1621     s->target = blk_new(s->common.job.aio_context,
1622                         BLK_PERM_WRITE | BLK_PERM_RESIZE |
1623                         (target_graph_mod ? BLK_PERM_GRAPH_MOD : 0),
1624                         BLK_PERM_WRITE_UNCHANGED |
1625                         (target_is_backing ? BLK_PERM_CONSISTENT_READ |
1626                                              BLK_PERM_WRITE |
1627                                              BLK_PERM_GRAPH_MOD : 0));
1628     ret = blk_insert_bs(s->target, target, errp);
1629     if (ret < 0) {
1630         goto fail;
1631     }
1632     if (is_mirror) {
1633         /* XXX: Mirror target could be a NBD server of target QEMU in the case
1634          * of non-shared block migration. To allow migration completion, we
1635          * have to allow "inactivate" of the target BB.  When that happens, we
1636          * know the job is drained, and the vcpus are stopped, so no write
1637          * operation will be performed. Block layer already has assertions to
1638          * ensure that. */
1639         blk_set_force_allow_inactivate(s->target);
1640     }
1641     blk_set_allow_aio_context_change(s->target, true);
1642     blk_set_disable_request_queuing(s->target, true);
1643 
1644     s->replaces = g_strdup(replaces);
1645     s->on_source_error = on_source_error;
1646     s->on_target_error = on_target_error;
1647     s->is_none_mode = is_none_mode;
1648     s->backing_mode = backing_mode;
1649     s->zero_target = zero_target;
1650     s->copy_mode = copy_mode;
1651     s->base = base;
1652     s->granularity = granularity;
1653     s->buf_size = ROUND_UP(buf_size, granularity);
1654     s->unmap = unmap;
1655     if (auto_complete) {
1656         s->should_complete = true;
1657     }
1658 
1659     /*
1660      * Must be called before we start tracking writes, but after
1661      *
1662      *     ((MirrorBlockJob *)
1663      *         ((MirrorBDSOpaque *)
1664      *             mirror_top_bs->opaque
1665      *         )->job
1666      *     )->copy_mode
1667      *
1668      * has the correct value.
1669      * (We start tracking writes as of the following
1670      * bdrv_create_dirty_bitmap() call.)
1671      */
1672     bdrv_refresh_limits(mirror_top_bs, &local_err);
1673     if (local_err) {
1674         error_propagate(errp, local_err);
1675         goto fail;
1676     }
1677 
1678     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1679     if (!s->dirty_bitmap) {
1680         goto fail;
1681     }
1682 
1683     ret = block_job_add_bdrv(&s->common, "source", bs, 0,
1684                              BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE |
1685                              BLK_PERM_CONSISTENT_READ,
1686                              errp);
1687     if (ret < 0) {
1688         goto fail;
1689     }
1690 
1691     /* Required permissions are already taken with blk_new() */
1692     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1693                        &error_abort);
1694 
1695     /* In commit_active_start() all intermediate nodes disappear, so
1696      * any jobs in them must be blocked */
1697     if (target_is_backing) {
1698         BlockDriverState *iter;
1699         for (iter = backing_bs(bs); iter != target; iter = backing_bs(iter)) {
1700             /* XXX BLK_PERM_WRITE needs to be allowed so we don't block
1701              * ourselves at s->base (if writes are blocked for a node, they are
1702              * also blocked for its backing file). The other options would be a
1703              * second filter driver above s->base (== target). */
1704             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1705                                      BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
1706                                      errp);
1707             if (ret < 0) {
1708                 goto fail;
1709             }
1710         }
1711 
1712         if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) {
1713             goto fail;
1714         }
1715     }
1716 
1717     QTAILQ_INIT(&s->ops_in_flight);
1718 
1719     trace_mirror_start(bs, s, opaque);
1720     job_start(&s->common.job);
1721 
1722     return &s->common;
1723 
1724 fail:
1725     if (s) {
1726         /* Make sure this BDS does not go away until we have completed the graph
1727          * changes below */
1728         bdrv_ref(mirror_top_bs);
1729 
1730         g_free(s->replaces);
1731         blk_unref(s->target);
1732         bs_opaque->job = NULL;
1733         if (s->dirty_bitmap) {
1734             bdrv_release_dirty_bitmap(bs, s->dirty_bitmap);
1735         }
1736         job_early_fail(&s->common.job);
1737     }
1738 
1739     bs_opaque->stop = true;
1740     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
1741                              &error_abort);
1742     bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
1743 
1744     bdrv_unref(mirror_top_bs);
1745 
1746     return NULL;
1747 }
1748 
1749 void mirror_start(const char *job_id, BlockDriverState *bs,
1750                   BlockDriverState *target, const char *replaces,
1751                   int creation_flags, int64_t speed,
1752                   uint32_t granularity, int64_t buf_size,
1753                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1754                   bool zero_target,
1755                   BlockdevOnError on_source_error,
1756                   BlockdevOnError on_target_error,
1757                   bool unmap, const char *filter_node_name,
1758                   MirrorCopyMode copy_mode, Error **errp)
1759 {
1760     bool is_none_mode;
1761     BlockDriverState *base;
1762 
1763     if ((mode == MIRROR_SYNC_MODE_INCREMENTAL) ||
1764         (mode == MIRROR_SYNC_MODE_BITMAP)) {
1765         error_setg(errp, "Sync mode '%s' not supported",
1766                    MirrorSyncMode_str(mode));
1767         return;
1768     }
1769     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1770     base = mode == MIRROR_SYNC_MODE_TOP ? backing_bs(bs) : NULL;
1771     mirror_start_job(job_id, bs, creation_flags, target, replaces,
1772                      speed, granularity, buf_size, backing_mode, zero_target,
1773                      on_source_error, on_target_error, unmap, NULL, NULL,
1774                      &mirror_job_driver, is_none_mode, base, false,
1775                      filter_node_name, true, copy_mode, errp);
1776 }
1777 
1778 BlockJob *commit_active_start(const char *job_id, BlockDriverState *bs,
1779                               BlockDriverState *base, int creation_flags,
1780                               int64_t speed, BlockdevOnError on_error,
1781                               const char *filter_node_name,
1782                               BlockCompletionFunc *cb, void *opaque,
1783                               bool auto_complete, Error **errp)
1784 {
1785     bool base_read_only;
1786     Error *local_err = NULL;
1787     BlockJob *ret;
1788 
1789     base_read_only = bdrv_is_read_only(base);
1790 
1791     if (base_read_only) {
1792         if (bdrv_reopen_set_read_only(base, false, errp) < 0) {
1793             return NULL;
1794         }
1795     }
1796 
1797     ret = mirror_start_job(
1798                      job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1799                      MIRROR_LEAVE_BACKING_CHAIN, false,
1800                      on_error, on_error, true, cb, opaque,
1801                      &commit_active_job_driver, false, base, auto_complete,
1802                      filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND,
1803                      &local_err);
1804     if (local_err) {
1805         error_propagate(errp, local_err);
1806         goto error_restore_flags;
1807     }
1808 
1809     return ret;
1810 
1811 error_restore_flags:
1812     /* ignore error and errp for bdrv_reopen, because we want to propagate
1813      * the original error */
1814     if (base_read_only) {
1815         bdrv_reopen_set_read_only(base, true, NULL);
1816     }
1817     return NULL;
1818 }
1819