xref: /openbmc/qemu/block/mirror.c (revision 88cd34ee)
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     Coroutine *co;
107 
108     QTAILQ_ENTRY(MirrorOp) next;
109 };
110 
111 typedef enum MirrorMethod {
112     MIRROR_METHOD_COPY,
113     MIRROR_METHOD_ZERO,
114     MIRROR_METHOD_DISCARD,
115 } MirrorMethod;
116 
117 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
118                                             int error)
119 {
120     s->synced = false;
121     s->actively_synced = false;
122     if (read) {
123         return block_job_error_action(&s->common, s->on_source_error,
124                                       true, error);
125     } else {
126         return block_job_error_action(&s->common, s->on_target_error,
127                                       false, error);
128     }
129 }
130 
131 static void coroutine_fn mirror_wait_on_conflicts(MirrorOp *self,
132                                                   MirrorBlockJob *s,
133                                                   uint64_t offset,
134                                                   uint64_t bytes)
135 {
136     uint64_t self_start_chunk = offset / s->granularity;
137     uint64_t self_end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
138     uint64_t self_nb_chunks = self_end_chunk - self_start_chunk;
139 
140     while (find_next_bit(s->in_flight_bitmap, self_end_chunk,
141                          self_start_chunk) < self_end_chunk &&
142            s->ret >= 0)
143     {
144         MirrorOp *op;
145 
146         QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
147             uint64_t op_start_chunk = op->offset / s->granularity;
148             uint64_t op_nb_chunks = DIV_ROUND_UP(op->offset + op->bytes,
149                                                  s->granularity) -
150                                     op_start_chunk;
151 
152             if (op == self) {
153                 continue;
154             }
155 
156             if (ranges_overlap(self_start_chunk, self_nb_chunks,
157                                op_start_chunk, op_nb_chunks))
158             {
159                 qemu_co_queue_wait(&op->waiting_requests, NULL);
160                 break;
161             }
162         }
163     }
164 }
165 
166 static void coroutine_fn mirror_iteration_done(MirrorOp *op, int ret)
167 {
168     MirrorBlockJob *s = op->s;
169     struct iovec *iov;
170     int64_t chunk_num;
171     int i, nb_chunks;
172 
173     trace_mirror_iteration_done(s, op->offset, op->bytes, ret);
174 
175     s->in_flight--;
176     s->bytes_in_flight -= op->bytes;
177     iov = op->qiov.iov;
178     for (i = 0; i < op->qiov.niov; i++) {
179         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
180         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
181         s->buf_free_count++;
182     }
183 
184     chunk_num = op->offset / s->granularity;
185     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
186 
187     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
188     QTAILQ_REMOVE(&s->ops_in_flight, op, next);
189     if (ret >= 0) {
190         if (s->cow_bitmap) {
191             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
192         }
193         if (!s->initial_zeroing_ongoing) {
194             job_progress_update(&s->common.job, op->bytes);
195         }
196     }
197     qemu_iovec_destroy(&op->qiov);
198 
199     qemu_co_queue_restart_all(&op->waiting_requests);
200     g_free(op);
201 }
202 
203 static void coroutine_fn mirror_write_complete(MirrorOp *op, int ret)
204 {
205     MirrorBlockJob *s = op->s;
206 
207     if (ret < 0) {
208         BlockErrorAction action;
209 
210         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
211         action = mirror_error_action(s, false, -ret);
212         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
213             s->ret = ret;
214         }
215     }
216 
217     mirror_iteration_done(op, ret);
218 }
219 
220 static void coroutine_fn mirror_read_complete(MirrorOp *op, int ret)
221 {
222     MirrorBlockJob *s = op->s;
223 
224     if (ret < 0) {
225         BlockErrorAction action;
226 
227         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
228         action = mirror_error_action(s, true, -ret);
229         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
230             s->ret = ret;
231         }
232 
233         mirror_iteration_done(op, ret);
234         return;
235     }
236 
237     ret = blk_co_pwritev(s->target, op->offset, op->qiov.size, &op->qiov, 0);
238     mirror_write_complete(op, ret);
239 }
240 
241 /* Clip bytes relative to offset to not exceed end-of-file */
242 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s,
243                                         int64_t offset,
244                                         int64_t bytes)
245 {
246     return MIN(bytes, s->bdev_length - offset);
247 }
248 
249 /* Round offset and/or bytes to target cluster if COW is needed, and
250  * return the offset of the adjusted tail against original. */
251 static int mirror_cow_align(MirrorBlockJob *s, int64_t *offset,
252                             uint64_t *bytes)
253 {
254     bool need_cow;
255     int ret = 0;
256     int64_t align_offset = *offset;
257     int64_t align_bytes = *bytes;
258     int max_bytes = s->granularity * s->max_iov;
259 
260     need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap);
261     need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity,
262                           s->cow_bitmap);
263     if (need_cow) {
264         bdrv_round_to_clusters(blk_bs(s->target), *offset, *bytes,
265                                &align_offset, &align_bytes);
266     }
267 
268     if (align_bytes > max_bytes) {
269         align_bytes = max_bytes;
270         if (need_cow) {
271             align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size);
272         }
273     }
274     /* Clipping may result in align_bytes unaligned to chunk boundary, but
275      * that doesn't matter because it's already the end of source image. */
276     align_bytes = mirror_clip_bytes(s, align_offset, align_bytes);
277 
278     ret = align_offset + align_bytes - (*offset + *bytes);
279     *offset = align_offset;
280     *bytes = align_bytes;
281     assert(ret >= 0);
282     return ret;
283 }
284 
285 static inline void coroutine_fn
286 mirror_wait_for_any_operation(MirrorBlockJob *s, MirrorOp *self, bool active)
287 {
288     MirrorOp *op;
289 
290     QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
291         if (self == op) {
292             continue;
293         }
294         /* Do not wait on pseudo ops, because it may in turn wait on
295          * some other operation to start, which may in fact be the
296          * caller of this function.  Since there is only one pseudo op
297          * at any given time, we will always find some real operation
298          * to wait on. */
299         if (!op->is_pseudo_op && op->is_active_write == active) {
300             qemu_co_queue_wait(&op->waiting_requests, NULL);
301             return;
302         }
303     }
304     abort();
305 }
306 
307 static inline void coroutine_fn
308 mirror_wait_for_free_in_flight_slot(MirrorBlockJob *s, MirrorOp *self)
309 {
310     /* Only non-active operations use up in-flight slots */
311     mirror_wait_for_any_operation(s, self, false);
312 }
313 
314 /* Perform a mirror copy operation.
315  *
316  * *op->bytes_handled is set to the number of bytes copied after and
317  * including offset, excluding any bytes copied prior to offset due
318  * to alignment.  This will be op->bytes if no alignment is necessary,
319  * or (new_end - op->offset) if the tail is rounded up or down due to
320  * alignment or buffer limit.
321  */
322 static void coroutine_fn mirror_co_read(void *opaque)
323 {
324     MirrorOp *op = opaque;
325     MirrorBlockJob *s = op->s;
326     int nb_chunks;
327     uint64_t ret;
328     uint64_t max_bytes;
329 
330     max_bytes = s->granularity * s->max_iov;
331 
332     /* We can only handle as much as buf_size at a time. */
333     op->bytes = MIN(s->buf_size, MIN(max_bytes, op->bytes));
334     assert(op->bytes);
335     assert(op->bytes < BDRV_REQUEST_MAX_BYTES);
336     *op->bytes_handled = op->bytes;
337 
338     if (s->cow_bitmap) {
339         *op->bytes_handled += mirror_cow_align(s, &op->offset, &op->bytes);
340     }
341     /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */
342     assert(*op->bytes_handled <= UINT_MAX);
343     assert(op->bytes <= s->buf_size);
344     /* The offset is granularity-aligned because:
345      * 1) Caller passes in aligned values;
346      * 2) mirror_cow_align is used only when target cluster is larger. */
347     assert(QEMU_IS_ALIGNED(op->offset, s->granularity));
348     /* The range is sector-aligned, since bdrv_getlength() rounds up. */
349     assert(QEMU_IS_ALIGNED(op->bytes, BDRV_SECTOR_SIZE));
350     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
351 
352     while (s->buf_free_count < nb_chunks) {
353         trace_mirror_yield_in_flight(s, op->offset, s->in_flight);
354         mirror_wait_for_free_in_flight_slot(s, op);
355     }
356 
357     /* Now make a QEMUIOVector taking enough granularity-sized chunks
358      * from s->buf_free.
359      */
360     qemu_iovec_init(&op->qiov, nb_chunks);
361     while (nb_chunks-- > 0) {
362         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
363         size_t remaining = op->bytes - op->qiov.size;
364 
365         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
366         s->buf_free_count--;
367         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
368     }
369 
370     /* Copy the dirty cluster.  */
371     s->in_flight++;
372     s->bytes_in_flight += op->bytes;
373     trace_mirror_one_iteration(s, op->offset, op->bytes);
374 
375     ret = bdrv_co_preadv(s->mirror_top_bs->backing, op->offset, op->bytes,
376                          &op->qiov, 0);
377     mirror_read_complete(op, ret);
378 }
379 
380 static void coroutine_fn mirror_co_zero(void *opaque)
381 {
382     MirrorOp *op = opaque;
383     int ret;
384 
385     op->s->in_flight++;
386     op->s->bytes_in_flight += op->bytes;
387     *op->bytes_handled = op->bytes;
388 
389     ret = blk_co_pwrite_zeroes(op->s->target, op->offset, op->bytes,
390                                op->s->unmap ? BDRV_REQ_MAY_UNMAP : 0);
391     mirror_write_complete(op, ret);
392 }
393 
394 static void coroutine_fn mirror_co_discard(void *opaque)
395 {
396     MirrorOp *op = opaque;
397     int ret;
398 
399     op->s->in_flight++;
400     op->s->bytes_in_flight += op->bytes;
401     *op->bytes_handled = op->bytes;
402 
403     ret = blk_co_pdiscard(op->s->target, op->offset, op->bytes);
404     mirror_write_complete(op, ret);
405 }
406 
407 static unsigned mirror_perform(MirrorBlockJob *s, int64_t offset,
408                                unsigned bytes, MirrorMethod mirror_method)
409 {
410     MirrorOp *op;
411     Coroutine *co;
412     int64_t bytes_handled = -1;
413 
414     op = g_new(MirrorOp, 1);
415     *op = (MirrorOp){
416         .s              = s,
417         .offset         = offset,
418         .bytes          = bytes,
419         .bytes_handled  = &bytes_handled,
420     };
421     qemu_co_queue_init(&op->waiting_requests);
422 
423     switch (mirror_method) {
424     case MIRROR_METHOD_COPY:
425         co = qemu_coroutine_create(mirror_co_read, op);
426         break;
427     case MIRROR_METHOD_ZERO:
428         co = qemu_coroutine_create(mirror_co_zero, op);
429         break;
430     case MIRROR_METHOD_DISCARD:
431         co = qemu_coroutine_create(mirror_co_discard, op);
432         break;
433     default:
434         abort();
435     }
436     op->co = co;
437 
438     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
439     qemu_coroutine_enter(co);
440     /* At this point, ownership of op has been moved to the coroutine
441      * and the object may already be freed */
442 
443     /* Assert that this value has been set */
444     assert(bytes_handled >= 0);
445 
446     /* Same assertion as in mirror_co_read() (and for mirror_co_read()
447      * and mirror_co_discard(), bytes_handled == op->bytes, which
448      * is the @bytes parameter given to this function) */
449     assert(bytes_handled <= UINT_MAX);
450     return bytes_handled;
451 }
452 
453 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
454 {
455     BlockDriverState *source = s->mirror_top_bs->backing->bs;
456     MirrorOp *pseudo_op;
457     int64_t offset;
458     uint64_t delay_ns = 0, ret = 0;
459     /* At least the first dirty chunk is mirrored in one iteration. */
460     int nb_chunks = 1;
461     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
462     int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES);
463 
464     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
465     offset = bdrv_dirty_iter_next(s->dbi);
466     if (offset < 0) {
467         bdrv_set_dirty_iter(s->dbi, 0);
468         offset = bdrv_dirty_iter_next(s->dbi);
469         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
470         assert(offset >= 0);
471     }
472     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
473 
474     mirror_wait_on_conflicts(NULL, s, offset, 1);
475 
476     job_pause_point(&s->common.job);
477 
478     /* Find the number of consective dirty chunks following the first dirty
479      * one, and wait for in flight requests in them. */
480     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
481     while (nb_chunks * s->granularity < s->buf_size) {
482         int64_t next_dirty;
483         int64_t next_offset = offset + nb_chunks * s->granularity;
484         int64_t next_chunk = next_offset / s->granularity;
485         if (next_offset >= s->bdev_length ||
486             !bdrv_dirty_bitmap_get_locked(s->dirty_bitmap, next_offset)) {
487             break;
488         }
489         if (test_bit(next_chunk, s->in_flight_bitmap)) {
490             break;
491         }
492 
493         next_dirty = bdrv_dirty_iter_next(s->dbi);
494         if (next_dirty > next_offset || next_dirty < 0) {
495             /* The bitmap iterator's cache is stale, refresh it */
496             bdrv_set_dirty_iter(s->dbi, next_offset);
497             next_dirty = bdrv_dirty_iter_next(s->dbi);
498         }
499         assert(next_dirty == next_offset);
500         nb_chunks++;
501     }
502 
503     /* Clear dirty bits before querying the block status, because
504      * calling bdrv_block_status_above could yield - if some blocks are
505      * marked dirty in this window, we need to know.
506      */
507     bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset,
508                                    nb_chunks * s->granularity);
509     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
510 
511     /* Before claiming an area in the in-flight bitmap, we have to
512      * create a MirrorOp for it so that conflicting requests can wait
513      * for it.  mirror_perform() will create the real MirrorOps later,
514      * for now we just create a pseudo operation that will wake up all
515      * conflicting requests once all real operations have been
516      * launched. */
517     pseudo_op = g_new(MirrorOp, 1);
518     *pseudo_op = (MirrorOp){
519         .offset         = offset,
520         .bytes          = nb_chunks * s->granularity,
521         .is_pseudo_op   = true,
522     };
523     qemu_co_queue_init(&pseudo_op->waiting_requests);
524     QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next);
525 
526     bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks);
527     while (nb_chunks > 0 && offset < s->bdev_length) {
528         int ret;
529         int64_t io_bytes;
530         int64_t io_bytes_acct;
531         MirrorMethod mirror_method = MIRROR_METHOD_COPY;
532 
533         assert(!(offset % s->granularity));
534         ret = bdrv_block_status_above(source, NULL, offset,
535                                       nb_chunks * s->granularity,
536                                       &io_bytes, NULL, NULL);
537         if (ret < 0) {
538             io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes);
539         } else if (ret & BDRV_BLOCK_DATA) {
540             io_bytes = MIN(io_bytes, max_io_bytes);
541         }
542 
543         io_bytes -= io_bytes % s->granularity;
544         if (io_bytes < s->granularity) {
545             io_bytes = s->granularity;
546         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
547             int64_t target_offset;
548             int64_t target_bytes;
549             bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes,
550                                    &target_offset, &target_bytes);
551             if (target_offset == offset &&
552                 target_bytes == io_bytes) {
553                 mirror_method = ret & BDRV_BLOCK_ZERO ?
554                                     MIRROR_METHOD_ZERO :
555                                     MIRROR_METHOD_DISCARD;
556             }
557         }
558 
559         while (s->in_flight >= MAX_IN_FLIGHT) {
560             trace_mirror_yield_in_flight(s, offset, s->in_flight);
561             mirror_wait_for_free_in_flight_slot(s, pseudo_op);
562         }
563 
564         if (s->ret < 0) {
565             ret = 0;
566             goto fail;
567         }
568 
569         io_bytes = mirror_clip_bytes(s, offset, io_bytes);
570         io_bytes = mirror_perform(s, offset, io_bytes, mirror_method);
571         if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) {
572             io_bytes_acct = 0;
573         } else {
574             io_bytes_acct = io_bytes;
575         }
576         assert(io_bytes);
577         offset += io_bytes;
578         nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity);
579         delay_ns = block_job_ratelimit_get_delay(&s->common, io_bytes_acct);
580     }
581 
582     ret = delay_ns;
583 fail:
584     QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next);
585     qemu_co_queue_restart_all(&pseudo_op->waiting_requests);
586     g_free(pseudo_op);
587 
588     return ret;
589 }
590 
591 static void mirror_free_init(MirrorBlockJob *s)
592 {
593     int granularity = s->granularity;
594     size_t buf_size = s->buf_size;
595     uint8_t *buf = s->buf;
596 
597     assert(s->buf_free_count == 0);
598     QSIMPLEQ_INIT(&s->buf_free);
599     while (buf_size != 0) {
600         MirrorBuffer *cur = (MirrorBuffer *)buf;
601         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
602         s->buf_free_count++;
603         buf_size -= granularity;
604         buf += granularity;
605     }
606 }
607 
608 /* This is also used for the .pause callback. There is no matching
609  * mirror_resume() because mirror_run() will begin iterating again
610  * when the job is resumed.
611  */
612 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s)
613 {
614     while (s->in_flight > 0) {
615         mirror_wait_for_free_in_flight_slot(s, NULL);
616     }
617 }
618 
619 /**
620  * mirror_exit_common: handle both abort() and prepare() cases.
621  * for .prepare, returns 0 on success and -errno on failure.
622  * for .abort cases, denoted by abort = true, MUST return 0.
623  */
624 static int mirror_exit_common(Job *job)
625 {
626     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
627     BlockJob *bjob = &s->common;
628     MirrorBDSOpaque *bs_opaque;
629     AioContext *replace_aio_context = NULL;
630     BlockDriverState *src;
631     BlockDriverState *target_bs;
632     BlockDriverState *mirror_top_bs;
633     Error *local_err = NULL;
634     bool abort = job->ret < 0;
635     int ret = 0;
636 
637     if (s->prepared) {
638         return 0;
639     }
640     s->prepared = true;
641 
642     mirror_top_bs = s->mirror_top_bs;
643     bs_opaque = mirror_top_bs->opaque;
644     src = mirror_top_bs->backing->bs;
645     target_bs = blk_bs(s->target);
646 
647     if (bdrv_chain_contains(src, target_bs)) {
648         bdrv_unfreeze_backing_chain(mirror_top_bs, target_bs);
649     }
650 
651     bdrv_release_dirty_bitmap(s->dirty_bitmap);
652 
653     /* Make sure that the source BDS doesn't go away during bdrv_replace_node,
654      * before we can call bdrv_drained_end */
655     bdrv_ref(src);
656     bdrv_ref(mirror_top_bs);
657     bdrv_ref(target_bs);
658 
659     /*
660      * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
661      * inserting target_bs at s->to_replace, where we might not be able to get
662      * these permissions.
663      */
664     blk_unref(s->target);
665     s->target = NULL;
666 
667     /* We don't access the source any more. Dropping any WRITE/RESIZE is
668      * required before it could become a backing file of target_bs. Not having
669      * these permissions any more means that we can't allow any new requests on
670      * mirror_top_bs from now on, so keep it drained. */
671     bdrv_drained_begin(mirror_top_bs);
672     bs_opaque->stop = true;
673     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
674                              &error_abort);
675     if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
676         BlockDriverState *backing = s->is_none_mode ? src : s->base;
677         if (backing_bs(target_bs) != backing) {
678             bdrv_set_backing_hd(target_bs, backing, &local_err);
679             if (local_err) {
680                 error_report_err(local_err);
681                 ret = -EPERM;
682             }
683         }
684     }
685 
686     if (s->to_replace) {
687         replace_aio_context = bdrv_get_aio_context(s->to_replace);
688         aio_context_acquire(replace_aio_context);
689     }
690 
691     if (s->should_complete && !abort) {
692         BlockDriverState *to_replace = s->to_replace ?: src;
693         bool ro = bdrv_is_read_only(to_replace);
694 
695         if (ro != bdrv_is_read_only(target_bs)) {
696             bdrv_reopen_set_read_only(target_bs, ro, NULL);
697         }
698 
699         /* The mirror job has no requests in flight any more, but we need to
700          * drain potential other users of the BDS before changing the graph. */
701         assert(s->in_drain);
702         bdrv_drained_begin(target_bs);
703         /*
704          * Cannot use check_to_replace_node() here, because that would
705          * check for an op blocker on @to_replace, and we have our own
706          * there.
707          */
708         if (bdrv_recurse_can_replace(src, to_replace)) {
709             bdrv_replace_node(to_replace, target_bs, &local_err);
710         } else {
711             error_setg(&local_err, "Can no longer replace '%s' by '%s', "
712                        "because it can no longer be guaranteed that doing so "
713                        "would not lead to an abrupt change of visible data",
714                        to_replace->node_name, target_bs->node_name);
715         }
716         bdrv_drained_end(target_bs);
717         if (local_err) {
718             error_report_err(local_err);
719             ret = -EPERM;
720         }
721     }
722     if (s->to_replace) {
723         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
724         error_free(s->replace_blocker);
725         bdrv_unref(s->to_replace);
726     }
727     if (replace_aio_context) {
728         aio_context_release(replace_aio_context);
729     }
730     g_free(s->replaces);
731     bdrv_unref(target_bs);
732 
733     /*
734      * Remove the mirror filter driver from the graph. Before this, get rid of
735      * the blockers on the intermediate nodes so that the resulting state is
736      * valid.
737      */
738     block_job_remove_all_bdrv(bjob);
739     bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
740 
741     /* We just changed the BDS the job BB refers to (with either or both of the
742      * bdrv_replace_node() calls), so switch the BB back so the cleanup does
743      * the right thing. We don't need any permissions any more now. */
744     blk_remove_bs(bjob->blk);
745     blk_set_perm(bjob->blk, 0, BLK_PERM_ALL, &error_abort);
746     blk_insert_bs(bjob->blk, mirror_top_bs, &error_abort);
747 
748     bs_opaque->job = NULL;
749 
750     bdrv_drained_end(src);
751     bdrv_drained_end(mirror_top_bs);
752     s->in_drain = false;
753     bdrv_unref(mirror_top_bs);
754     bdrv_unref(src);
755 
756     return ret;
757 }
758 
759 static int mirror_prepare(Job *job)
760 {
761     return mirror_exit_common(job);
762 }
763 
764 static void mirror_abort(Job *job)
765 {
766     int ret = mirror_exit_common(job);
767     assert(ret == 0);
768 }
769 
770 static void coroutine_fn mirror_throttle(MirrorBlockJob *s)
771 {
772     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
773 
774     if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) {
775         s->last_pause_ns = now;
776         job_sleep_ns(&s->common.job, 0);
777     } else {
778         job_pause_point(&s->common.job);
779     }
780 }
781 
782 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
783 {
784     int64_t offset;
785     BlockDriverState *base = s->base;
786     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
787     BlockDriverState *target_bs = blk_bs(s->target);
788     int ret;
789     int64_t count;
790 
791     if (s->zero_target) {
792         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
793             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
794             return 0;
795         }
796 
797         s->initial_zeroing_ongoing = true;
798         for (offset = 0; offset < s->bdev_length; ) {
799             int bytes = MIN(s->bdev_length - offset,
800                             QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
801 
802             mirror_throttle(s);
803 
804             if (job_is_cancelled(&s->common.job)) {
805                 s->initial_zeroing_ongoing = false;
806                 return 0;
807             }
808 
809             if (s->in_flight >= MAX_IN_FLIGHT) {
810                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
811                                    s->in_flight);
812                 mirror_wait_for_free_in_flight_slot(s, NULL);
813                 continue;
814             }
815 
816             mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO);
817             offset += bytes;
818         }
819 
820         mirror_wait_for_all_io(s);
821         s->initial_zeroing_ongoing = false;
822     }
823 
824     /* First part, loop on the sectors and initialize the dirty bitmap.  */
825     for (offset = 0; offset < s->bdev_length; ) {
826         /* Just to make sure we are not exceeding int limit. */
827         int bytes = MIN(s->bdev_length - offset,
828                         QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
829 
830         mirror_throttle(s);
831 
832         if (job_is_cancelled(&s->common.job)) {
833             return 0;
834         }
835 
836         ret = bdrv_is_allocated_above(bs, base, false, offset, bytes, &count);
837         if (ret < 0) {
838             return ret;
839         }
840 
841         assert(count);
842         if (ret == 1) {
843             bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
844         }
845         offset += count;
846     }
847     return 0;
848 }
849 
850 /* Called when going out of the streaming phase to flush the bulk of the
851  * data to the medium, or just before completing.
852  */
853 static int mirror_flush(MirrorBlockJob *s)
854 {
855     int ret = blk_flush(s->target);
856     if (ret < 0) {
857         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
858             s->ret = ret;
859         }
860     }
861     return ret;
862 }
863 
864 static int coroutine_fn mirror_run(Job *job, Error **errp)
865 {
866     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
867     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
868     BlockDriverState *target_bs = blk_bs(s->target);
869     bool need_drain = true;
870     int64_t length;
871     BlockDriverInfo bdi;
872     char backing_filename[2]; /* we only need 2 characters because we are only
873                                  checking for a NULL string */
874     int ret = 0;
875 
876     if (job_is_cancelled(&s->common.job)) {
877         goto immediate_exit;
878     }
879 
880     s->bdev_length = bdrv_getlength(bs);
881     if (s->bdev_length < 0) {
882         ret = s->bdev_length;
883         goto immediate_exit;
884     }
885 
886     /* Active commit must resize the base image if its size differs from the
887      * active layer. */
888     if (s->base == blk_bs(s->target)) {
889         int64_t base_length;
890 
891         base_length = blk_getlength(s->target);
892         if (base_length < 0) {
893             ret = base_length;
894             goto immediate_exit;
895         }
896 
897         if (s->bdev_length > base_length) {
898             ret = blk_truncate(s->target, s->bdev_length, false,
899                                PREALLOC_MODE_OFF, NULL);
900             if (ret < 0) {
901                 goto immediate_exit;
902             }
903         }
904     }
905 
906     if (s->bdev_length == 0) {
907         /* Transition to the READY state and wait for complete. */
908         job_transition_to_ready(&s->common.job);
909         s->synced = true;
910         s->actively_synced = true;
911         while (!job_is_cancelled(&s->common.job) && !s->should_complete) {
912             job_yield(&s->common.job);
913         }
914         s->common.job.cancelled = false;
915         goto immediate_exit;
916     }
917 
918     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
919     s->in_flight_bitmap = bitmap_new(length);
920 
921     /* If we have no backing file yet in the destination, we cannot let
922      * the destination do COW.  Instead, we copy sectors around the
923      * dirty data if needed.  We need a bitmap to do that.
924      */
925     bdrv_get_backing_filename(target_bs, backing_filename,
926                               sizeof(backing_filename));
927     if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) {
928         s->target_cluster_size = bdi.cluster_size;
929     } else {
930         s->target_cluster_size = BDRV_SECTOR_SIZE;
931     }
932     if (backing_filename[0] && !target_bs->backing &&
933         s->granularity < s->target_cluster_size) {
934         s->buf_size = MAX(s->buf_size, s->target_cluster_size);
935         s->cow_bitmap = bitmap_new(length);
936     }
937     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
938 
939     s->buf = qemu_try_blockalign(bs, s->buf_size);
940     if (s->buf == NULL) {
941         ret = -ENOMEM;
942         goto immediate_exit;
943     }
944 
945     mirror_free_init(s);
946 
947     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
948     if (!s->is_none_mode) {
949         ret = mirror_dirty_init(s);
950         if (ret < 0 || job_is_cancelled(&s->common.job)) {
951             goto immediate_exit;
952         }
953     }
954 
955     assert(!s->dbi);
956     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
957     for (;;) {
958         uint64_t delay_ns = 0;
959         int64_t cnt, delta;
960         bool should_complete;
961 
962         /* Do not start passive operations while there are active
963          * writes in progress */
964         while (s->in_active_write_counter) {
965             mirror_wait_for_any_operation(s, NULL, true);
966         }
967 
968         if (s->ret < 0) {
969             ret = s->ret;
970             goto immediate_exit;
971         }
972 
973         job_pause_point(&s->common.job);
974 
975         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
976         /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
977          * the number of bytes currently being processed; together those are
978          * the current remaining operation length */
979         job_progress_set_remaining(&s->common.job, s->bytes_in_flight + cnt);
980 
981         /* Note that even when no rate limit is applied we need to yield
982          * periodically with no pending I/O so that bdrv_drain_all() returns.
983          * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
984          * an error, or when the source is clean, whichever comes first. */
985         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
986         if (delta < BLOCK_JOB_SLICE_TIME &&
987             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
988             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
989                 (cnt == 0 && s->in_flight > 0)) {
990                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
991                 mirror_wait_for_free_in_flight_slot(s, NULL);
992                 continue;
993             } else if (cnt != 0) {
994                 delay_ns = mirror_iteration(s);
995             }
996         }
997 
998         should_complete = false;
999         if (s->in_flight == 0 && cnt == 0) {
1000             trace_mirror_before_flush(s);
1001             if (!s->synced) {
1002                 if (mirror_flush(s) < 0) {
1003                     /* Go check s->ret.  */
1004                     continue;
1005                 }
1006                 /* We're out of the streaming phase.  From now on, if the job
1007                  * is cancelled we will actually complete all pending I/O and
1008                  * report completion.  This way, block-job-cancel will leave
1009                  * the target in a consistent state.
1010                  */
1011                 job_transition_to_ready(&s->common.job);
1012                 s->synced = true;
1013                 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) {
1014                     s->actively_synced = true;
1015                 }
1016             }
1017 
1018             should_complete = s->should_complete ||
1019                 job_is_cancelled(&s->common.job);
1020             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1021         }
1022 
1023         if (cnt == 0 && should_complete) {
1024             /* The dirty bitmap is not updated while operations are pending.
1025              * If we're about to exit, wait for pending operations before
1026              * calling bdrv_get_dirty_count(bs), or we may exit while the
1027              * source has dirty data to copy!
1028              *
1029              * Note that I/O can be submitted by the guest while
1030              * mirror_populate runs, so pause it now.  Before deciding
1031              * whether to switch to target check one last time if I/O has
1032              * come in the meanwhile, and if not flush the data to disk.
1033              */
1034             trace_mirror_before_drain(s, cnt);
1035 
1036             s->in_drain = true;
1037             bdrv_drained_begin(bs);
1038             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1039             if (cnt > 0 || mirror_flush(s) < 0) {
1040                 bdrv_drained_end(bs);
1041                 s->in_drain = false;
1042                 continue;
1043             }
1044 
1045             /* The two disks are in sync.  Exit and report successful
1046              * completion.
1047              */
1048             assert(QLIST_EMPTY(&bs->tracked_requests));
1049             s->common.job.cancelled = false;
1050             need_drain = false;
1051             break;
1052         }
1053 
1054         ret = 0;
1055 
1056         if (s->synced && !should_complete) {
1057             delay_ns = (s->in_flight == 0 &&
1058                         cnt == 0 ? BLOCK_JOB_SLICE_TIME : 0);
1059         }
1060         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
1061         job_sleep_ns(&s->common.job, delay_ns);
1062         if (job_is_cancelled(&s->common.job) &&
1063             (!s->synced || s->common.job.force_cancel))
1064         {
1065             break;
1066         }
1067         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1068     }
1069 
1070 immediate_exit:
1071     if (s->in_flight > 0) {
1072         /* We get here only if something went wrong.  Either the job failed,
1073          * or it was cancelled prematurely so that we do not guarantee that
1074          * the target is a copy of the source.
1075          */
1076         assert(ret < 0 || ((s->common.job.force_cancel || !s->synced) &&
1077                job_is_cancelled(&s->common.job)));
1078         assert(need_drain);
1079         mirror_wait_for_all_io(s);
1080     }
1081 
1082     assert(s->in_flight == 0);
1083     qemu_vfree(s->buf);
1084     g_free(s->cow_bitmap);
1085     g_free(s->in_flight_bitmap);
1086     bdrv_dirty_iter_free(s->dbi);
1087 
1088     if (need_drain) {
1089         s->in_drain = true;
1090         bdrv_drained_begin(bs);
1091     }
1092 
1093     return ret;
1094 }
1095 
1096 static void mirror_complete(Job *job, Error **errp)
1097 {
1098     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1099     BlockDriverState *target;
1100 
1101     target = blk_bs(s->target);
1102 
1103     if (!s->synced) {
1104         error_setg(errp, "The active block job '%s' cannot be completed",
1105                    job->id);
1106         return;
1107     }
1108 
1109     if (s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
1110         int ret;
1111 
1112         assert(!target->backing);
1113         ret = bdrv_open_backing_file(target, NULL, "backing", errp);
1114         if (ret < 0) {
1115             return;
1116         }
1117     }
1118 
1119     /* block all operations on to_replace bs */
1120     if (s->replaces) {
1121         AioContext *replace_aio_context;
1122 
1123         s->to_replace = bdrv_find_node(s->replaces);
1124         if (!s->to_replace) {
1125             error_setg(errp, "Node name '%s' not found", s->replaces);
1126             return;
1127         }
1128 
1129         replace_aio_context = bdrv_get_aio_context(s->to_replace);
1130         aio_context_acquire(replace_aio_context);
1131 
1132         /* TODO Translate this into permission system. Current definition of
1133          * GRAPH_MOD would require to request it for the parents; they might
1134          * not even be BlockDriverStates, however, so a BdrvChild can't address
1135          * them. May need redefinition of GRAPH_MOD. */
1136         error_setg(&s->replace_blocker,
1137                    "block device is in use by block-job-complete");
1138         bdrv_op_block_all(s->to_replace, s->replace_blocker);
1139         bdrv_ref(s->to_replace);
1140 
1141         aio_context_release(replace_aio_context);
1142     }
1143 
1144     s->should_complete = true;
1145     job_enter(job);
1146 }
1147 
1148 static void coroutine_fn mirror_pause(Job *job)
1149 {
1150     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1151 
1152     mirror_wait_for_all_io(s);
1153 }
1154 
1155 static bool mirror_drained_poll(BlockJob *job)
1156 {
1157     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1158 
1159     /* If the job isn't paused nor cancelled, we can't be sure that it won't
1160      * issue more requests. We make an exception if we've reached this point
1161      * from one of our own drain sections, to avoid a deadlock waiting for
1162      * ourselves.
1163      */
1164     if (!s->common.job.paused && !s->common.job.cancelled && !s->in_drain) {
1165         return true;
1166     }
1167 
1168     return !!s->in_flight;
1169 }
1170 
1171 static const BlockJobDriver mirror_job_driver = {
1172     .job_driver = {
1173         .instance_size          = sizeof(MirrorBlockJob),
1174         .job_type               = JOB_TYPE_MIRROR,
1175         .free                   = block_job_free,
1176         .user_resume            = block_job_user_resume,
1177         .run                    = mirror_run,
1178         .prepare                = mirror_prepare,
1179         .abort                  = mirror_abort,
1180         .pause                  = mirror_pause,
1181         .complete               = mirror_complete,
1182     },
1183     .drained_poll           = mirror_drained_poll,
1184 };
1185 
1186 static const BlockJobDriver commit_active_job_driver = {
1187     .job_driver = {
1188         .instance_size          = sizeof(MirrorBlockJob),
1189         .job_type               = JOB_TYPE_COMMIT,
1190         .free                   = block_job_free,
1191         .user_resume            = block_job_user_resume,
1192         .run                    = mirror_run,
1193         .prepare                = mirror_prepare,
1194         .abort                  = mirror_abort,
1195         .pause                  = mirror_pause,
1196         .complete               = mirror_complete,
1197     },
1198     .drained_poll           = mirror_drained_poll,
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     int ret;
1207     size_t qiov_offset = 0;
1208     int64_t bitmap_offset, bitmap_end;
1209 
1210     if (!QEMU_IS_ALIGNED(offset, job->granularity) &&
1211         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset))
1212     {
1213             /*
1214              * Dirty unaligned padding: ignore it.
1215              *
1216              * Reasoning:
1217              * 1. If we copy it, we can't reset corresponding bit in
1218              *    dirty_bitmap as there may be some "dirty" bytes still not
1219              *    copied.
1220              * 2. It's already dirty, so skipping it we don't diverge mirror
1221              *    progress.
1222              *
1223              * Note, that because of this, guest write may have no contribution
1224              * into mirror converge, but that's not bad, as we have background
1225              * process of mirroring. If under some bad circumstances (high guest
1226              * IO load) background process starve, we will not converge anyway,
1227              * even if each write will contribute, as guest is not guaranteed to
1228              * rewrite the whole disk.
1229              */
1230             qiov_offset = QEMU_ALIGN_UP(offset, job->granularity) - offset;
1231             if (bytes <= qiov_offset) {
1232                 /* nothing to do after shrink */
1233                 return;
1234             }
1235             offset += qiov_offset;
1236             bytes -= qiov_offset;
1237     }
1238 
1239     if (!QEMU_IS_ALIGNED(offset + bytes, job->granularity) &&
1240         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset + bytes - 1))
1241     {
1242         uint64_t tail = (offset + bytes) % job->granularity;
1243 
1244         if (bytes <= tail) {
1245             /* nothing to do after shrink */
1246             return;
1247         }
1248         bytes -= tail;
1249     }
1250 
1251     /*
1252      * Tails are either clean or shrunk, so for bitmap resetting
1253      * we safely align the range down.
1254      */
1255     bitmap_offset = QEMU_ALIGN_UP(offset, job->granularity);
1256     bitmap_end = QEMU_ALIGN_DOWN(offset + bytes, job->granularity);
1257     if (bitmap_offset < bitmap_end) {
1258         bdrv_reset_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1259                                 bitmap_end - bitmap_offset);
1260     }
1261 
1262     job_progress_increase_remaining(&job->common.job, bytes);
1263 
1264     switch (method) {
1265     case MIRROR_METHOD_COPY:
1266         ret = blk_co_pwritev_part(job->target, offset, bytes,
1267                                   qiov, qiov_offset, flags);
1268         break;
1269 
1270     case MIRROR_METHOD_ZERO:
1271         assert(!qiov);
1272         ret = blk_co_pwrite_zeroes(job->target, offset, bytes, flags);
1273         break;
1274 
1275     case MIRROR_METHOD_DISCARD:
1276         assert(!qiov);
1277         ret = blk_co_pdiscard(job->target, offset, bytes);
1278         break;
1279 
1280     default:
1281         abort();
1282     }
1283 
1284     if (ret >= 0) {
1285         job_progress_update(&job->common.job, bytes);
1286     } else {
1287         BlockErrorAction action;
1288 
1289         /*
1290          * We failed, so we should mark dirty the whole area, aligned up.
1291          * Note that we don't care about shrunk tails if any: they were dirty
1292          * at function start, and they must be still dirty, as we've locked
1293          * the region for in-flight op.
1294          */
1295         bitmap_offset = QEMU_ALIGN_DOWN(offset, job->granularity);
1296         bitmap_end = QEMU_ALIGN_UP(offset + bytes, job->granularity);
1297         bdrv_set_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1298                               bitmap_end - bitmap_offset);
1299         job->actively_synced = false;
1300 
1301         action = mirror_error_action(job, false, -ret);
1302         if (action == BLOCK_ERROR_ACTION_REPORT) {
1303             if (!job->ret) {
1304                 job->ret = ret;
1305             }
1306         }
1307     }
1308 }
1309 
1310 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s,
1311                                                    uint64_t offset,
1312                                                    uint64_t bytes)
1313 {
1314     MirrorOp *op;
1315     uint64_t start_chunk = offset / s->granularity;
1316     uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
1317 
1318     op = g_new(MirrorOp, 1);
1319     *op = (MirrorOp){
1320         .s                  = s,
1321         .offset             = offset,
1322         .bytes              = bytes,
1323         .is_active_write    = true,
1324     };
1325     qemu_co_queue_init(&op->waiting_requests);
1326     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
1327 
1328     s->in_active_write_counter++;
1329 
1330     mirror_wait_on_conflicts(op, s, offset, bytes);
1331 
1332     bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1333 
1334     return op;
1335 }
1336 
1337 static void coroutine_fn active_write_settle(MirrorOp *op)
1338 {
1339     uint64_t start_chunk = op->offset / op->s->granularity;
1340     uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes,
1341                                       op->s->granularity);
1342 
1343     if (!--op->s->in_active_write_counter && op->s->actively_synced) {
1344         BdrvChild *source = op->s->mirror_top_bs->backing;
1345 
1346         if (QLIST_FIRST(&source->bs->parents) == source &&
1347             QLIST_NEXT(source, next_parent) == NULL)
1348         {
1349             /* Assert that we are back in sync once all active write
1350              * operations are settled.
1351              * Note that we can only assert this if the mirror node
1352              * is the source node's only parent. */
1353             assert(!bdrv_get_dirty_count(op->s->dirty_bitmap));
1354         }
1355     }
1356     bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1357     QTAILQ_REMOVE(&op->s->ops_in_flight, op, next);
1358     qemu_co_queue_restart_all(&op->waiting_requests);
1359     g_free(op);
1360 }
1361 
1362 static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs,
1363     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1364 {
1365     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1366 }
1367 
1368 static int coroutine_fn bdrv_mirror_top_do_write(BlockDriverState *bs,
1369     MirrorMethod method, uint64_t offset, uint64_t bytes, QEMUIOVector *qiov,
1370     int flags)
1371 {
1372     MirrorOp *op = NULL;
1373     MirrorBDSOpaque *s = bs->opaque;
1374     int ret = 0;
1375     bool copy_to_target;
1376 
1377     copy_to_target = s->job->ret >= 0 &&
1378                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1379 
1380     if (copy_to_target) {
1381         op = active_write_prepare(s->job, offset, bytes);
1382     }
1383 
1384     switch (method) {
1385     case MIRROR_METHOD_COPY:
1386         ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1387         break;
1388 
1389     case MIRROR_METHOD_ZERO:
1390         ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1391         break;
1392 
1393     case MIRROR_METHOD_DISCARD:
1394         ret = bdrv_co_pdiscard(bs->backing, offset, bytes);
1395         break;
1396 
1397     default:
1398         abort();
1399     }
1400 
1401     if (ret < 0) {
1402         goto out;
1403     }
1404 
1405     if (copy_to_target) {
1406         do_sync_target_write(s->job, method, offset, bytes, qiov, flags);
1407     }
1408 
1409 out:
1410     if (copy_to_target) {
1411         active_write_settle(op);
1412     }
1413     return ret;
1414 }
1415 
1416 static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs,
1417     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1418 {
1419     MirrorBDSOpaque *s = bs->opaque;
1420     QEMUIOVector bounce_qiov;
1421     void *bounce_buf;
1422     int ret = 0;
1423     bool copy_to_target;
1424 
1425     copy_to_target = s->job->ret >= 0 &&
1426                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1427 
1428     if (copy_to_target) {
1429         /* The guest might concurrently modify the data to write; but
1430          * the data on source and destination must match, so we have
1431          * to use a bounce buffer if we are going to write to the
1432          * target now. */
1433         bounce_buf = qemu_blockalign(bs, bytes);
1434         iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes);
1435 
1436         qemu_iovec_init(&bounce_qiov, 1);
1437         qemu_iovec_add(&bounce_qiov, bounce_buf, bytes);
1438         qiov = &bounce_qiov;
1439     }
1440 
1441     ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov,
1442                                    flags);
1443 
1444     if (copy_to_target) {
1445         qemu_iovec_destroy(&bounce_qiov);
1446         qemu_vfree(bounce_buf);
1447     }
1448 
1449     return ret;
1450 }
1451 
1452 static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs)
1453 {
1454     if (bs->backing == NULL) {
1455         /* we can be here after failed bdrv_append in mirror_start_job */
1456         return 0;
1457     }
1458     return bdrv_co_flush(bs->backing->bs);
1459 }
1460 
1461 static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs,
1462     int64_t offset, int bytes, BdrvRequestFlags flags)
1463 {
1464     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL,
1465                                     flags);
1466 }
1467 
1468 static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs,
1469     int64_t offset, int bytes)
1470 {
1471     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes,
1472                                     NULL, 0);
1473 }
1474 
1475 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs)
1476 {
1477     if (bs->backing == NULL) {
1478         /* we can be here after failed bdrv_attach_child in
1479          * bdrv_set_backing_hd */
1480         return;
1481     }
1482     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1483             bs->backing->bs->filename);
1484 }
1485 
1486 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1487                                        const BdrvChildRole *role,
1488                                        BlockReopenQueue *reopen_queue,
1489                                        uint64_t perm, uint64_t shared,
1490                                        uint64_t *nperm, uint64_t *nshared)
1491 {
1492     MirrorBDSOpaque *s = bs->opaque;
1493 
1494     if (s->stop) {
1495         /*
1496          * If the job is to be stopped, we do not need to forward
1497          * anything to the real image.
1498          */
1499         *nperm = 0;
1500         *nshared = BLK_PERM_ALL;
1501         return;
1502     }
1503 
1504     /* Must be able to forward guest writes to the real image */
1505     *nperm = 0;
1506     if (perm & BLK_PERM_WRITE) {
1507         *nperm |= BLK_PERM_WRITE;
1508     }
1509 
1510     *nshared = BLK_PERM_ALL;
1511 }
1512 
1513 /* Dummy node that provides consistent read to its users without requiring it
1514  * from its backing file and that allows writes on the backing file chain. */
1515 static BlockDriver bdrv_mirror_top = {
1516     .format_name                = "mirror_top",
1517     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1518     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1519     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1520     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1521     .bdrv_co_flush              = bdrv_mirror_top_flush,
1522     .bdrv_co_block_status       = bdrv_co_block_status_from_backing,
1523     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1524     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1525 };
1526 
1527 static BlockJob *mirror_start_job(
1528                              const char *job_id, BlockDriverState *bs,
1529                              int creation_flags, BlockDriverState *target,
1530                              const char *replaces, int64_t speed,
1531                              uint32_t granularity, int64_t buf_size,
1532                              BlockMirrorBackingMode backing_mode,
1533                              bool zero_target,
1534                              BlockdevOnError on_source_error,
1535                              BlockdevOnError on_target_error,
1536                              bool unmap,
1537                              BlockCompletionFunc *cb,
1538                              void *opaque,
1539                              const BlockJobDriver *driver,
1540                              bool is_none_mode, BlockDriverState *base,
1541                              bool auto_complete, const char *filter_node_name,
1542                              bool is_mirror, MirrorCopyMode copy_mode,
1543                              Error **errp)
1544 {
1545     MirrorBlockJob *s;
1546     MirrorBDSOpaque *bs_opaque;
1547     BlockDriverState *mirror_top_bs;
1548     bool target_graph_mod;
1549     bool target_is_backing;
1550     Error *local_err = NULL;
1551     int ret;
1552 
1553     if (granularity == 0) {
1554         granularity = bdrv_get_default_bitmap_granularity(target);
1555     }
1556 
1557     assert(is_power_of_2(granularity));
1558 
1559     if (buf_size < 0) {
1560         error_setg(errp, "Invalid parameter 'buf-size'");
1561         return NULL;
1562     }
1563 
1564     if (buf_size == 0) {
1565         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1566     }
1567 
1568     if (bs == target) {
1569         error_setg(errp, "Can't mirror node into itself");
1570         return NULL;
1571     }
1572 
1573     /* In the case of active commit, add dummy driver to provide consistent
1574      * reads on the top, while disabling it in the intermediate nodes, and make
1575      * the backing chain writable. */
1576     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1577                                          BDRV_O_RDWR, errp);
1578     if (mirror_top_bs == NULL) {
1579         return NULL;
1580     }
1581     if (!filter_node_name) {
1582         mirror_top_bs->implicit = true;
1583     }
1584 
1585     /* So that we can always drop this node */
1586     mirror_top_bs->never_freeze = true;
1587 
1588     mirror_top_bs->total_sectors = bs->total_sectors;
1589     mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
1590     mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
1591                                           BDRV_REQ_NO_FALLBACK;
1592     bs_opaque = g_new0(MirrorBDSOpaque, 1);
1593     mirror_top_bs->opaque = bs_opaque;
1594 
1595     /* bdrv_append takes ownership of the mirror_top_bs reference, need to keep
1596      * it alive until block_job_create() succeeds even if bs has no parent. */
1597     bdrv_ref(mirror_top_bs);
1598     bdrv_drained_begin(bs);
1599     bdrv_append(mirror_top_bs, bs, &local_err);
1600     bdrv_drained_end(bs);
1601 
1602     if (local_err) {
1603         bdrv_unref(mirror_top_bs);
1604         error_propagate(errp, local_err);
1605         return NULL;
1606     }
1607 
1608     /* Make sure that the source is not resized while the job is running */
1609     s = block_job_create(job_id, driver, NULL, mirror_top_bs,
1610                          BLK_PERM_CONSISTENT_READ,
1611                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1612                          BLK_PERM_WRITE | BLK_PERM_GRAPH_MOD, speed,
1613                          creation_flags, cb, opaque, errp);
1614     if (!s) {
1615         goto fail;
1616     }
1617     bs_opaque->job = s;
1618 
1619     /* The block job now has a reference to this node */
1620     bdrv_unref(mirror_top_bs);
1621 
1622     s->mirror_top_bs = mirror_top_bs;
1623 
1624     /* No resize for the target either; while the mirror is still running, a
1625      * consistent read isn't necessarily possible. We could possibly allow
1626      * writes and graph modifications, though it would likely defeat the
1627      * purpose of a mirror, so leave them blocked for now.
1628      *
1629      * In the case of active commit, things look a bit different, though,
1630      * because the target is an already populated backing file in active use.
1631      * We can allow anything except resize there.*/
1632     target_is_backing = bdrv_chain_contains(bs, target);
1633     target_graph_mod = (backing_mode != MIRROR_LEAVE_BACKING_CHAIN);
1634     s->target = blk_new(s->common.job.aio_context,
1635                         BLK_PERM_WRITE | BLK_PERM_RESIZE |
1636                         (target_graph_mod ? BLK_PERM_GRAPH_MOD : 0),
1637                         BLK_PERM_WRITE_UNCHANGED |
1638                         (target_is_backing ? BLK_PERM_CONSISTENT_READ |
1639                                              BLK_PERM_WRITE |
1640                                              BLK_PERM_GRAPH_MOD : 0));
1641     ret = blk_insert_bs(s->target, target, errp);
1642     if (ret < 0) {
1643         goto fail;
1644     }
1645     if (is_mirror) {
1646         /* XXX: Mirror target could be a NBD server of target QEMU in the case
1647          * of non-shared block migration. To allow migration completion, we
1648          * have to allow "inactivate" of the target BB.  When that happens, we
1649          * know the job is drained, and the vcpus are stopped, so no write
1650          * operation will be performed. Block layer already has assertions to
1651          * ensure that. */
1652         blk_set_force_allow_inactivate(s->target);
1653     }
1654     blk_set_allow_aio_context_change(s->target, true);
1655     blk_set_disable_request_queuing(s->target, true);
1656 
1657     s->replaces = g_strdup(replaces);
1658     s->on_source_error = on_source_error;
1659     s->on_target_error = on_target_error;
1660     s->is_none_mode = is_none_mode;
1661     s->backing_mode = backing_mode;
1662     s->zero_target = zero_target;
1663     s->copy_mode = copy_mode;
1664     s->base = base;
1665     s->granularity = granularity;
1666     s->buf_size = ROUND_UP(buf_size, granularity);
1667     s->unmap = unmap;
1668     if (auto_complete) {
1669         s->should_complete = true;
1670     }
1671 
1672     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1673     if (!s->dirty_bitmap) {
1674         goto fail;
1675     }
1676     if (s->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING) {
1677         bdrv_disable_dirty_bitmap(s->dirty_bitmap);
1678     }
1679 
1680     ret = block_job_add_bdrv(&s->common, "source", bs, 0,
1681                              BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE |
1682                              BLK_PERM_CONSISTENT_READ,
1683                              errp);
1684     if (ret < 0) {
1685         goto fail;
1686     }
1687 
1688     /* Required permissions are already taken with blk_new() */
1689     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1690                        &error_abort);
1691 
1692     /* In commit_active_start() all intermediate nodes disappear, so
1693      * any jobs in them must be blocked */
1694     if (target_is_backing) {
1695         BlockDriverState *iter;
1696         for (iter = backing_bs(bs); iter != target; iter = backing_bs(iter)) {
1697             /* XXX BLK_PERM_WRITE needs to be allowed so we don't block
1698              * ourselves at s->base (if writes are blocked for a node, they are
1699              * also blocked for its backing file). The other options would be a
1700              * second filter driver above s->base (== target). */
1701             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1702                                      BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
1703                                      errp);
1704             if (ret < 0) {
1705                 goto fail;
1706             }
1707         }
1708 
1709         if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) {
1710             goto fail;
1711         }
1712     }
1713 
1714     QTAILQ_INIT(&s->ops_in_flight);
1715 
1716     trace_mirror_start(bs, s, opaque);
1717     job_start(&s->common.job);
1718 
1719     return &s->common;
1720 
1721 fail:
1722     if (s) {
1723         /* Make sure this BDS does not go away until we have completed the graph
1724          * changes below */
1725         bdrv_ref(mirror_top_bs);
1726 
1727         g_free(s->replaces);
1728         blk_unref(s->target);
1729         bs_opaque->job = NULL;
1730         if (s->dirty_bitmap) {
1731             bdrv_release_dirty_bitmap(s->dirty_bitmap);
1732         }
1733         job_early_fail(&s->common.job);
1734     }
1735 
1736     bs_opaque->stop = true;
1737     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
1738                              &error_abort);
1739     bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
1740 
1741     bdrv_unref(mirror_top_bs);
1742 
1743     return NULL;
1744 }
1745 
1746 void mirror_start(const char *job_id, BlockDriverState *bs,
1747                   BlockDriverState *target, const char *replaces,
1748                   int creation_flags, int64_t speed,
1749                   uint32_t granularity, int64_t buf_size,
1750                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1751                   bool zero_target,
1752                   BlockdevOnError on_source_error,
1753                   BlockdevOnError on_target_error,
1754                   bool unmap, const char *filter_node_name,
1755                   MirrorCopyMode copy_mode, Error **errp)
1756 {
1757     bool is_none_mode;
1758     BlockDriverState *base;
1759 
1760     if ((mode == MIRROR_SYNC_MODE_INCREMENTAL) ||
1761         (mode == MIRROR_SYNC_MODE_BITMAP)) {
1762         error_setg(errp, "Sync mode '%s' not supported",
1763                    MirrorSyncMode_str(mode));
1764         return;
1765     }
1766     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1767     base = mode == MIRROR_SYNC_MODE_TOP ? backing_bs(bs) : NULL;
1768     mirror_start_job(job_id, bs, creation_flags, target, replaces,
1769                      speed, granularity, buf_size, backing_mode, zero_target,
1770                      on_source_error, on_target_error, unmap, NULL, NULL,
1771                      &mirror_job_driver, is_none_mode, base, false,
1772                      filter_node_name, true, copy_mode, errp);
1773 }
1774 
1775 BlockJob *commit_active_start(const char *job_id, BlockDriverState *bs,
1776                               BlockDriverState *base, int creation_flags,
1777                               int64_t speed, BlockdevOnError on_error,
1778                               const char *filter_node_name,
1779                               BlockCompletionFunc *cb, void *opaque,
1780                               bool auto_complete, Error **errp)
1781 {
1782     bool base_read_only;
1783     Error *local_err = NULL;
1784     BlockJob *ret;
1785 
1786     base_read_only = bdrv_is_read_only(base);
1787 
1788     if (base_read_only) {
1789         if (bdrv_reopen_set_read_only(base, false, errp) < 0) {
1790             return NULL;
1791         }
1792     }
1793 
1794     ret = mirror_start_job(
1795                      job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1796                      MIRROR_LEAVE_BACKING_CHAIN, false,
1797                      on_error, on_error, true, cb, opaque,
1798                      &commit_active_job_driver, false, base, auto_complete,
1799                      filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND,
1800                      &local_err);
1801     if (local_err) {
1802         error_propagate(errp, local_err);
1803         goto error_restore_flags;
1804     }
1805 
1806     return ret;
1807 
1808 error_restore_flags:
1809     /* ignore error and errp for bdrv_reopen, because we want to propagate
1810      * the original error */
1811     if (base_read_only) {
1812         bdrv_reopen_set_read_only(base, true, NULL);
1813     }
1814     return NULL;
1815 }
1816