xref: /openbmc/qemu/block/mirror.c (revision 284d697c)
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     BlockDriverState *base_overlay;
46 
47     /* The name of the graph node to replace */
48     char *replaces;
49     /* The BDS to replace */
50     BlockDriverState *to_replace;
51     /* Used to block operations on the drive-mirror-replace target */
52     Error *replace_blocker;
53     bool is_none_mode;
54     BlockMirrorBackingMode backing_mode;
55     /* Whether the target image requires explicit zero-initialization */
56     bool zero_target;
57     MirrorCopyMode copy_mode;
58     BlockdevOnError on_source_error, on_target_error;
59     bool synced;
60     /* Set when the target is synced (dirty bitmap is clean, nothing
61      * in flight) and the job is running in active mode */
62     bool actively_synced;
63     bool should_complete;
64     int64_t granularity;
65     size_t buf_size;
66     int64_t bdev_length;
67     unsigned long *cow_bitmap;
68     BdrvDirtyBitmap *dirty_bitmap;
69     BdrvDirtyBitmapIter *dbi;
70     uint8_t *buf;
71     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
72     int buf_free_count;
73 
74     uint64_t last_pause_ns;
75     unsigned long *in_flight_bitmap;
76     int in_flight;
77     int64_t bytes_in_flight;
78     QTAILQ_HEAD(, MirrorOp) ops_in_flight;
79     int ret;
80     bool unmap;
81     int target_cluster_size;
82     int max_iov;
83     bool initial_zeroing_ongoing;
84     int in_active_write_counter;
85     bool prepared;
86     bool in_drain;
87 } MirrorBlockJob;
88 
89 typedef struct MirrorBDSOpaque {
90     MirrorBlockJob *job;
91     bool stop;
92 } MirrorBDSOpaque;
93 
94 struct MirrorOp {
95     MirrorBlockJob *s;
96     QEMUIOVector qiov;
97     int64_t offset;
98     uint64_t bytes;
99 
100     /* The pointee is set by mirror_co_read(), mirror_co_zero(), and
101      * mirror_co_discard() before yielding for the first time */
102     int64_t *bytes_handled;
103 
104     bool is_pseudo_op;
105     bool is_active_write;
106     bool is_in_flight;
107     CoQueue waiting_requests;
108     Coroutine *co;
109 
110     QTAILQ_ENTRY(MirrorOp) next;
111 };
112 
113 typedef enum MirrorMethod {
114     MIRROR_METHOD_COPY,
115     MIRROR_METHOD_ZERO,
116     MIRROR_METHOD_DISCARD,
117 } MirrorMethod;
118 
119 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
120                                             int error)
121 {
122     s->synced = false;
123     s->actively_synced = false;
124     if (read) {
125         return block_job_error_action(&s->common, s->on_source_error,
126                                       true, error);
127     } else {
128         return block_job_error_action(&s->common, s->on_target_error,
129                                       false, error);
130     }
131 }
132 
133 static void coroutine_fn mirror_wait_on_conflicts(MirrorOp *self,
134                                                   MirrorBlockJob *s,
135                                                   uint64_t offset,
136                                                   uint64_t bytes)
137 {
138     uint64_t self_start_chunk = offset / s->granularity;
139     uint64_t self_end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
140     uint64_t self_nb_chunks = self_end_chunk - self_start_chunk;
141 
142     while (find_next_bit(s->in_flight_bitmap, self_end_chunk,
143                          self_start_chunk) < self_end_chunk &&
144            s->ret >= 0)
145     {
146         MirrorOp *op;
147 
148         QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
149             uint64_t op_start_chunk = op->offset / s->granularity;
150             uint64_t op_nb_chunks = DIV_ROUND_UP(op->offset + op->bytes,
151                                                  s->granularity) -
152                                     op_start_chunk;
153 
154             if (op == self) {
155                 continue;
156             }
157 
158             if (ranges_overlap(self_start_chunk, self_nb_chunks,
159                                op_start_chunk, op_nb_chunks))
160             {
161                 qemu_co_queue_wait(&op->waiting_requests, NULL);
162                 break;
163             }
164         }
165     }
166 }
167 
168 static void coroutine_fn mirror_iteration_done(MirrorOp *op, int ret)
169 {
170     MirrorBlockJob *s = op->s;
171     struct iovec *iov;
172     int64_t chunk_num;
173     int i, nb_chunks;
174 
175     trace_mirror_iteration_done(s, op->offset, op->bytes, ret);
176 
177     s->in_flight--;
178     s->bytes_in_flight -= op->bytes;
179     iov = op->qiov.iov;
180     for (i = 0; i < op->qiov.niov; i++) {
181         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
182         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
183         s->buf_free_count++;
184     }
185 
186     chunk_num = op->offset / s->granularity;
187     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
188 
189     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
190     QTAILQ_REMOVE(&s->ops_in_flight, op, next);
191     if (ret >= 0) {
192         if (s->cow_bitmap) {
193             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
194         }
195         if (!s->initial_zeroing_ongoing) {
196             job_progress_update(&s->common.job, op->bytes);
197         }
198     }
199     qemu_iovec_destroy(&op->qiov);
200 
201     qemu_co_queue_restart_all(&op->waiting_requests);
202     g_free(op);
203 }
204 
205 static void coroutine_fn mirror_write_complete(MirrorOp *op, int ret)
206 {
207     MirrorBlockJob *s = op->s;
208 
209     if (ret < 0) {
210         BlockErrorAction action;
211 
212         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
213         action = mirror_error_action(s, false, -ret);
214         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
215             s->ret = ret;
216         }
217     }
218 
219     mirror_iteration_done(op, ret);
220 }
221 
222 static void coroutine_fn mirror_read_complete(MirrorOp *op, int ret)
223 {
224     MirrorBlockJob *s = op->s;
225 
226     if (ret < 0) {
227         BlockErrorAction action;
228 
229         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
230         action = mirror_error_action(s, true, -ret);
231         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
232             s->ret = ret;
233         }
234 
235         mirror_iteration_done(op, ret);
236         return;
237     }
238 
239     ret = blk_co_pwritev(s->target, op->offset, op->qiov.size, &op->qiov, 0);
240     mirror_write_complete(op, ret);
241 }
242 
243 /* Clip bytes relative to offset to not exceed end-of-file */
244 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s,
245                                         int64_t offset,
246                                         int64_t bytes)
247 {
248     return MIN(bytes, s->bdev_length - offset);
249 }
250 
251 /* Round offset and/or bytes to target cluster if COW is needed, and
252  * return the offset of the adjusted tail against original. */
253 static int mirror_cow_align(MirrorBlockJob *s, int64_t *offset,
254                             uint64_t *bytes)
255 {
256     bool need_cow;
257     int ret = 0;
258     int64_t align_offset = *offset;
259     int64_t align_bytes = *bytes;
260     int max_bytes = s->granularity * s->max_iov;
261 
262     need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap);
263     need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity,
264                           s->cow_bitmap);
265     if (need_cow) {
266         bdrv_round_to_clusters(blk_bs(s->target), *offset, *bytes,
267                                &align_offset, &align_bytes);
268     }
269 
270     if (align_bytes > max_bytes) {
271         align_bytes = max_bytes;
272         if (need_cow) {
273             align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size);
274         }
275     }
276     /* Clipping may result in align_bytes unaligned to chunk boundary, but
277      * that doesn't matter because it's already the end of source image. */
278     align_bytes = mirror_clip_bytes(s, align_offset, align_bytes);
279 
280     ret = align_offset + align_bytes - (*offset + *bytes);
281     *offset = align_offset;
282     *bytes = align_bytes;
283     assert(ret >= 0);
284     return ret;
285 }
286 
287 static inline void coroutine_fn
288 mirror_wait_for_any_operation(MirrorBlockJob *s, bool active)
289 {
290     MirrorOp *op;
291 
292     QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
293         /* Do not wait on pseudo ops, because it may in turn wait on
294          * some other operation to start, which may in fact be the
295          * caller of this function.  Since there is only one pseudo op
296          * at any given time, we will always find some real operation
297          * to wait on. */
298         if (!op->is_pseudo_op && op->is_in_flight &&
299             op->is_active_write == active)
300         {
301             qemu_co_queue_wait(&op->waiting_requests, NULL);
302             return;
303         }
304     }
305     abort();
306 }
307 
308 static inline void coroutine_fn
309 mirror_wait_for_free_in_flight_slot(MirrorBlockJob *s)
310 {
311     /* Only non-active operations use up in-flight slots */
312     mirror_wait_for_any_operation(s, false);
313 }
314 
315 /* Perform a mirror copy operation.
316  *
317  * *op->bytes_handled is set to the number of bytes copied after and
318  * including offset, excluding any bytes copied prior to offset due
319  * to alignment.  This will be op->bytes if no alignment is necessary,
320  * or (new_end - op->offset) if the tail is rounded up or down due to
321  * alignment or buffer limit.
322  */
323 static void coroutine_fn mirror_co_read(void *opaque)
324 {
325     MirrorOp *op = opaque;
326     MirrorBlockJob *s = op->s;
327     int nb_chunks;
328     uint64_t ret;
329     uint64_t max_bytes;
330 
331     max_bytes = s->granularity * s->max_iov;
332 
333     /* We can only handle as much as buf_size at a time. */
334     op->bytes = MIN(s->buf_size, MIN(max_bytes, op->bytes));
335     assert(op->bytes);
336     assert(op->bytes < BDRV_REQUEST_MAX_BYTES);
337     *op->bytes_handled = op->bytes;
338 
339     if (s->cow_bitmap) {
340         *op->bytes_handled += mirror_cow_align(s, &op->offset, &op->bytes);
341     }
342     /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */
343     assert(*op->bytes_handled <= UINT_MAX);
344     assert(op->bytes <= s->buf_size);
345     /* The offset is granularity-aligned because:
346      * 1) Caller passes in aligned values;
347      * 2) mirror_cow_align is used only when target cluster is larger. */
348     assert(QEMU_IS_ALIGNED(op->offset, s->granularity));
349     /* The range is sector-aligned, since bdrv_getlength() rounds up. */
350     assert(QEMU_IS_ALIGNED(op->bytes, BDRV_SECTOR_SIZE));
351     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
352 
353     while (s->buf_free_count < nb_chunks) {
354         trace_mirror_yield_in_flight(s, op->offset, s->in_flight);
355         mirror_wait_for_free_in_flight_slot(s);
356     }
357 
358     /* Now make a QEMUIOVector taking enough granularity-sized chunks
359      * from s->buf_free.
360      */
361     qemu_iovec_init(&op->qiov, nb_chunks);
362     while (nb_chunks-- > 0) {
363         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
364         size_t remaining = op->bytes - op->qiov.size;
365 
366         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
367         s->buf_free_count--;
368         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
369     }
370 
371     /* Copy the dirty cluster.  */
372     s->in_flight++;
373     s->bytes_in_flight += op->bytes;
374     op->is_in_flight = true;
375     trace_mirror_one_iteration(s, op->offset, op->bytes);
376 
377     ret = bdrv_co_preadv(s->mirror_top_bs->backing, op->offset, op->bytes,
378                          &op->qiov, 0);
379     mirror_read_complete(op, ret);
380 }
381 
382 static void coroutine_fn mirror_co_zero(void *opaque)
383 {
384     MirrorOp *op = opaque;
385     int ret;
386 
387     op->s->in_flight++;
388     op->s->bytes_in_flight += op->bytes;
389     *op->bytes_handled = op->bytes;
390     op->is_in_flight = true;
391 
392     ret = blk_co_pwrite_zeroes(op->s->target, op->offset, op->bytes,
393                                op->s->unmap ? BDRV_REQ_MAY_UNMAP : 0);
394     mirror_write_complete(op, ret);
395 }
396 
397 static void coroutine_fn mirror_co_discard(void *opaque)
398 {
399     MirrorOp *op = opaque;
400     int ret;
401 
402     op->s->in_flight++;
403     op->s->bytes_in_flight += op->bytes;
404     *op->bytes_handled = op->bytes;
405     op->is_in_flight = true;
406 
407     ret = blk_co_pdiscard(op->s->target, op->offset, op->bytes);
408     mirror_write_complete(op, ret);
409 }
410 
411 static unsigned mirror_perform(MirrorBlockJob *s, int64_t offset,
412                                unsigned bytes, MirrorMethod mirror_method)
413 {
414     MirrorOp *op;
415     Coroutine *co;
416     int64_t bytes_handled = -1;
417 
418     op = g_new(MirrorOp, 1);
419     *op = (MirrorOp){
420         .s              = s,
421         .offset         = offset,
422         .bytes          = bytes,
423         .bytes_handled  = &bytes_handled,
424     };
425     qemu_co_queue_init(&op->waiting_requests);
426 
427     switch (mirror_method) {
428     case MIRROR_METHOD_COPY:
429         co = qemu_coroutine_create(mirror_co_read, op);
430         break;
431     case MIRROR_METHOD_ZERO:
432         co = qemu_coroutine_create(mirror_co_zero, op);
433         break;
434     case MIRROR_METHOD_DISCARD:
435         co = qemu_coroutine_create(mirror_co_discard, op);
436         break;
437     default:
438         abort();
439     }
440     op->co = co;
441 
442     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
443     qemu_coroutine_enter(co);
444     /* At this point, ownership of op has been moved to the coroutine
445      * and the object may already be freed */
446 
447     /* Assert that this value has been set */
448     assert(bytes_handled >= 0);
449 
450     /* Same assertion as in mirror_co_read() (and for mirror_co_read()
451      * and mirror_co_discard(), bytes_handled == op->bytes, which
452      * is the @bytes parameter given to this function) */
453     assert(bytes_handled <= UINT_MAX);
454     return bytes_handled;
455 }
456 
457 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
458 {
459     BlockDriverState *source = s->mirror_top_bs->backing->bs;
460     MirrorOp *pseudo_op;
461     int64_t offset;
462     uint64_t delay_ns = 0, ret = 0;
463     /* At least the first dirty chunk is mirrored in one iteration. */
464     int nb_chunks = 1;
465     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
466     int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES);
467 
468     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
469     offset = bdrv_dirty_iter_next(s->dbi);
470     if (offset < 0) {
471         bdrv_set_dirty_iter(s->dbi, 0);
472         offset = bdrv_dirty_iter_next(s->dbi);
473         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
474         assert(offset >= 0);
475     }
476     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
477 
478     mirror_wait_on_conflicts(NULL, s, offset, 1);
479 
480     job_pause_point(&s->common.job);
481 
482     /* Find the number of consective dirty chunks following the first dirty
483      * one, and wait for in flight requests in them. */
484     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
485     while (nb_chunks * s->granularity < s->buf_size) {
486         int64_t next_dirty;
487         int64_t next_offset = offset + nb_chunks * s->granularity;
488         int64_t next_chunk = next_offset / s->granularity;
489         if (next_offset >= s->bdev_length ||
490             !bdrv_dirty_bitmap_get_locked(s->dirty_bitmap, next_offset)) {
491             break;
492         }
493         if (test_bit(next_chunk, s->in_flight_bitmap)) {
494             break;
495         }
496 
497         next_dirty = bdrv_dirty_iter_next(s->dbi);
498         if (next_dirty > next_offset || next_dirty < 0) {
499             /* The bitmap iterator's cache is stale, refresh it */
500             bdrv_set_dirty_iter(s->dbi, next_offset);
501             next_dirty = bdrv_dirty_iter_next(s->dbi);
502         }
503         assert(next_dirty == next_offset);
504         nb_chunks++;
505     }
506 
507     /* Clear dirty bits before querying the block status, because
508      * calling bdrv_block_status_above could yield - if some blocks are
509      * marked dirty in this window, we need to know.
510      */
511     bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset,
512                                    nb_chunks * s->granularity);
513     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
514 
515     /* Before claiming an area in the in-flight bitmap, we have to
516      * create a MirrorOp for it so that conflicting requests can wait
517      * for it.  mirror_perform() will create the real MirrorOps later,
518      * for now we just create a pseudo operation that will wake up all
519      * conflicting requests once all real operations have been
520      * launched. */
521     pseudo_op = g_new(MirrorOp, 1);
522     *pseudo_op = (MirrorOp){
523         .offset         = offset,
524         .bytes          = nb_chunks * s->granularity,
525         .is_pseudo_op   = true,
526     };
527     qemu_co_queue_init(&pseudo_op->waiting_requests);
528     QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next);
529 
530     bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks);
531     while (nb_chunks > 0 && offset < s->bdev_length) {
532         int ret;
533         int64_t io_bytes;
534         int64_t io_bytes_acct;
535         MirrorMethod mirror_method = MIRROR_METHOD_COPY;
536 
537         assert(!(offset % s->granularity));
538         ret = bdrv_block_status_above(source, NULL, offset,
539                                       nb_chunks * s->granularity,
540                                       &io_bytes, NULL, NULL);
541         if (ret < 0) {
542             io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes);
543         } else if (ret & BDRV_BLOCK_DATA) {
544             io_bytes = MIN(io_bytes, max_io_bytes);
545         }
546 
547         io_bytes -= io_bytes % s->granularity;
548         if (io_bytes < s->granularity) {
549             io_bytes = s->granularity;
550         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
551             int64_t target_offset;
552             int64_t target_bytes;
553             bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes,
554                                    &target_offset, &target_bytes);
555             if (target_offset == offset &&
556                 target_bytes == io_bytes) {
557                 mirror_method = ret & BDRV_BLOCK_ZERO ?
558                                     MIRROR_METHOD_ZERO :
559                                     MIRROR_METHOD_DISCARD;
560             }
561         }
562 
563         while (s->in_flight >= MAX_IN_FLIGHT) {
564             trace_mirror_yield_in_flight(s, offset, s->in_flight);
565             mirror_wait_for_free_in_flight_slot(s);
566         }
567 
568         if (s->ret < 0) {
569             ret = 0;
570             goto fail;
571         }
572 
573         io_bytes = mirror_clip_bytes(s, offset, io_bytes);
574         io_bytes = mirror_perform(s, offset, io_bytes, mirror_method);
575         if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) {
576             io_bytes_acct = 0;
577         } else {
578             io_bytes_acct = io_bytes;
579         }
580         assert(io_bytes);
581         offset += io_bytes;
582         nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity);
583         delay_ns = block_job_ratelimit_get_delay(&s->common, io_bytes_acct);
584     }
585 
586     ret = delay_ns;
587 fail:
588     QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next);
589     qemu_co_queue_restart_all(&pseudo_op->waiting_requests);
590     g_free(pseudo_op);
591 
592     return ret;
593 }
594 
595 static void mirror_free_init(MirrorBlockJob *s)
596 {
597     int granularity = s->granularity;
598     size_t buf_size = s->buf_size;
599     uint8_t *buf = s->buf;
600 
601     assert(s->buf_free_count == 0);
602     QSIMPLEQ_INIT(&s->buf_free);
603     while (buf_size != 0) {
604         MirrorBuffer *cur = (MirrorBuffer *)buf;
605         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
606         s->buf_free_count++;
607         buf_size -= granularity;
608         buf += granularity;
609     }
610 }
611 
612 /* This is also used for the .pause callback. There is no matching
613  * mirror_resume() because mirror_run() will begin iterating again
614  * when the job is resumed.
615  */
616 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s)
617 {
618     while (s->in_flight > 0) {
619         mirror_wait_for_free_in_flight_slot(s);
620     }
621 }
622 
623 /**
624  * mirror_exit_common: handle both abort() and prepare() cases.
625  * for .prepare, returns 0 on success and -errno on failure.
626  * for .abort cases, denoted by abort = true, MUST return 0.
627  */
628 static int mirror_exit_common(Job *job)
629 {
630     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
631     BlockJob *bjob = &s->common;
632     MirrorBDSOpaque *bs_opaque;
633     AioContext *replace_aio_context = NULL;
634     BlockDriverState *src;
635     BlockDriverState *target_bs;
636     BlockDriverState *mirror_top_bs;
637     Error *local_err = NULL;
638     bool abort = job->ret < 0;
639     int ret = 0;
640 
641     if (s->prepared) {
642         return 0;
643     }
644     s->prepared = true;
645 
646     mirror_top_bs = s->mirror_top_bs;
647     bs_opaque = mirror_top_bs->opaque;
648     src = mirror_top_bs->backing->bs;
649     target_bs = blk_bs(s->target);
650 
651     if (bdrv_chain_contains(src, target_bs)) {
652         bdrv_unfreeze_backing_chain(mirror_top_bs, target_bs);
653     }
654 
655     bdrv_release_dirty_bitmap(s->dirty_bitmap);
656 
657     /* Make sure that the source BDS doesn't go away during bdrv_replace_node,
658      * before we can call bdrv_drained_end */
659     bdrv_ref(src);
660     bdrv_ref(mirror_top_bs);
661     bdrv_ref(target_bs);
662 
663     /*
664      * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
665      * inserting target_bs at s->to_replace, where we might not be able to get
666      * these permissions.
667      */
668     blk_unref(s->target);
669     s->target = NULL;
670 
671     /* We don't access the source any more. Dropping any WRITE/RESIZE is
672      * required before it could become a backing file of target_bs. Not having
673      * these permissions any more means that we can't allow any new requests on
674      * mirror_top_bs from now on, so keep it drained. */
675     bdrv_drained_begin(mirror_top_bs);
676     bs_opaque->stop = true;
677     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
678                              &error_abort);
679     if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
680         BlockDriverState *backing = s->is_none_mode ? src : s->base;
681         BlockDriverState *unfiltered_target = bdrv_skip_filters(target_bs);
682 
683         if (bdrv_cow_bs(unfiltered_target) != backing) {
684             bdrv_set_backing_hd(unfiltered_target, backing, &local_err);
685             if (local_err) {
686                 error_report_err(local_err);
687                 local_err = NULL;
688                 ret = -EPERM;
689             }
690         }
691     }
692 
693     if (s->to_replace) {
694         replace_aio_context = bdrv_get_aio_context(s->to_replace);
695         aio_context_acquire(replace_aio_context);
696     }
697 
698     if (s->should_complete && !abort) {
699         BlockDriverState *to_replace = s->to_replace ?: src;
700         bool ro = bdrv_is_read_only(to_replace);
701 
702         if (ro != bdrv_is_read_only(target_bs)) {
703             bdrv_reopen_set_read_only(target_bs, ro, NULL);
704         }
705 
706         /* The mirror job has no requests in flight any more, but we need to
707          * drain potential other users of the BDS before changing the graph. */
708         assert(s->in_drain);
709         bdrv_drained_begin(target_bs);
710         /*
711          * Cannot use check_to_replace_node() here, because that would
712          * check for an op blocker on @to_replace, and we have our own
713          * there.
714          */
715         if (bdrv_recurse_can_replace(src, to_replace)) {
716             bdrv_replace_node(to_replace, target_bs, &local_err);
717         } else {
718             error_setg(&local_err, "Can no longer replace '%s' by '%s', "
719                        "because it can no longer be guaranteed that doing so "
720                        "would not lead to an abrupt change of visible data",
721                        to_replace->node_name, target_bs->node_name);
722         }
723         bdrv_drained_end(target_bs);
724         if (local_err) {
725             error_report_err(local_err);
726             ret = -EPERM;
727         }
728     }
729     if (s->to_replace) {
730         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
731         error_free(s->replace_blocker);
732         bdrv_unref(s->to_replace);
733     }
734     if (replace_aio_context) {
735         aio_context_release(replace_aio_context);
736     }
737     g_free(s->replaces);
738     bdrv_unref(target_bs);
739 
740     /*
741      * Remove the mirror filter driver from the graph. Before this, get rid of
742      * the blockers on the intermediate nodes so that the resulting state is
743      * valid.
744      */
745     block_job_remove_all_bdrv(bjob);
746     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
747 
748     /* We just changed the BDS the job BB refers to (with either or both of the
749      * bdrv_replace_node() calls), so switch the BB back so the cleanup does
750      * the right thing. We don't need any permissions any more now. */
751     blk_remove_bs(bjob->blk);
752     blk_set_perm(bjob->blk, 0, BLK_PERM_ALL, &error_abort);
753     blk_insert_bs(bjob->blk, mirror_top_bs, &error_abort);
754 
755     bs_opaque->job = NULL;
756 
757     bdrv_drained_end(src);
758     bdrv_drained_end(mirror_top_bs);
759     s->in_drain = false;
760     bdrv_unref(mirror_top_bs);
761     bdrv_unref(src);
762 
763     return ret;
764 }
765 
766 static int mirror_prepare(Job *job)
767 {
768     return mirror_exit_common(job);
769 }
770 
771 static void mirror_abort(Job *job)
772 {
773     int ret = mirror_exit_common(job);
774     assert(ret == 0);
775 }
776 
777 static void coroutine_fn mirror_throttle(MirrorBlockJob *s)
778 {
779     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
780 
781     if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) {
782         s->last_pause_ns = now;
783         job_sleep_ns(&s->common.job, 0);
784     } else {
785         job_pause_point(&s->common.job);
786     }
787 }
788 
789 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
790 {
791     int64_t offset;
792     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
793     BlockDriverState *target_bs = blk_bs(s->target);
794     int ret;
795     int64_t count;
796 
797     if (s->zero_target) {
798         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
799             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
800             return 0;
801         }
802 
803         s->initial_zeroing_ongoing = true;
804         for (offset = 0; offset < s->bdev_length; ) {
805             int bytes = MIN(s->bdev_length - offset,
806                             QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
807 
808             mirror_throttle(s);
809 
810             if (job_is_cancelled(&s->common.job)) {
811                 s->initial_zeroing_ongoing = false;
812                 return 0;
813             }
814 
815             if (s->in_flight >= MAX_IN_FLIGHT) {
816                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
817                                    s->in_flight);
818                 mirror_wait_for_free_in_flight_slot(s);
819                 continue;
820             }
821 
822             mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO);
823             offset += bytes;
824         }
825 
826         mirror_wait_for_all_io(s);
827         s->initial_zeroing_ongoing = false;
828     }
829 
830     /* First part, loop on the sectors and initialize the dirty bitmap.  */
831     for (offset = 0; offset < s->bdev_length; ) {
832         /* Just to make sure we are not exceeding int limit. */
833         int bytes = MIN(s->bdev_length - offset,
834                         QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
835 
836         mirror_throttle(s);
837 
838         if (job_is_cancelled(&s->common.job)) {
839             return 0;
840         }
841 
842         ret = bdrv_is_allocated_above(bs, s->base_overlay, true, offset, bytes,
843                                       &count);
844         if (ret < 0) {
845             return ret;
846         }
847 
848         assert(count);
849         if (ret > 0) {
850             bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
851         }
852         offset += count;
853     }
854     return 0;
855 }
856 
857 /* Called when going out of the streaming phase to flush the bulk of the
858  * data to the medium, or just before completing.
859  */
860 static int mirror_flush(MirrorBlockJob *s)
861 {
862     int ret = blk_flush(s->target);
863     if (ret < 0) {
864         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
865             s->ret = ret;
866         }
867     }
868     return ret;
869 }
870 
871 static int coroutine_fn mirror_run(Job *job, Error **errp)
872 {
873     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
874     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
875     BlockDriverState *target_bs = blk_bs(s->target);
876     bool need_drain = true;
877     int64_t length;
878     int64_t target_length;
879     BlockDriverInfo bdi;
880     char backing_filename[2]; /* we only need 2 characters because we are only
881                                  checking for a NULL string */
882     int ret = 0;
883 
884     if (job_is_cancelled(&s->common.job)) {
885         goto immediate_exit;
886     }
887 
888     s->bdev_length = bdrv_getlength(bs);
889     if (s->bdev_length < 0) {
890         ret = s->bdev_length;
891         goto immediate_exit;
892     }
893 
894     target_length = blk_getlength(s->target);
895     if (target_length < 0) {
896         ret = target_length;
897         goto immediate_exit;
898     }
899 
900     /* Active commit must resize the base image if its size differs from the
901      * active layer. */
902     if (s->base == blk_bs(s->target)) {
903         if (s->bdev_length > target_length) {
904             ret = blk_truncate(s->target, s->bdev_length, false,
905                                PREALLOC_MODE_OFF, 0, NULL);
906             if (ret < 0) {
907                 goto immediate_exit;
908             }
909         }
910     } else if (s->bdev_length != target_length) {
911         error_setg(errp, "Source and target image have different sizes");
912         ret = -EINVAL;
913         goto immediate_exit;
914     }
915 
916     if (s->bdev_length == 0) {
917         /* Transition to the READY state and wait for complete. */
918         job_transition_to_ready(&s->common.job);
919         s->synced = true;
920         s->actively_synced = true;
921         while (!job_is_cancelled(&s->common.job) && !s->should_complete) {
922             job_yield(&s->common.job);
923         }
924         s->common.job.cancelled = false;
925         goto immediate_exit;
926     }
927 
928     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
929     s->in_flight_bitmap = bitmap_new(length);
930 
931     /* If we have no backing file yet in the destination, we cannot let
932      * the destination do COW.  Instead, we copy sectors around the
933      * dirty data if needed.  We need a bitmap to do that.
934      */
935     bdrv_get_backing_filename(target_bs, backing_filename,
936                               sizeof(backing_filename));
937     if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) {
938         s->target_cluster_size = bdi.cluster_size;
939     } else {
940         s->target_cluster_size = BDRV_SECTOR_SIZE;
941     }
942     if (backing_filename[0] && !bdrv_backing_chain_next(target_bs) &&
943         s->granularity < s->target_cluster_size) {
944         s->buf_size = MAX(s->buf_size, s->target_cluster_size);
945         s->cow_bitmap = bitmap_new(length);
946     }
947     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
948 
949     s->buf = qemu_try_blockalign(bs, s->buf_size);
950     if (s->buf == NULL) {
951         ret = -ENOMEM;
952         goto immediate_exit;
953     }
954 
955     mirror_free_init(s);
956 
957     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
958     if (!s->is_none_mode) {
959         ret = mirror_dirty_init(s);
960         if (ret < 0 || job_is_cancelled(&s->common.job)) {
961             goto immediate_exit;
962         }
963     }
964 
965     assert(!s->dbi);
966     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
967     for (;;) {
968         uint64_t delay_ns = 0;
969         int64_t cnt, delta;
970         bool should_complete;
971 
972         /* Do not start passive operations while there are active
973          * writes in progress */
974         while (s->in_active_write_counter) {
975             mirror_wait_for_any_operation(s, true);
976         }
977 
978         if (s->ret < 0) {
979             ret = s->ret;
980             goto immediate_exit;
981         }
982 
983         job_pause_point(&s->common.job);
984 
985         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
986         /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
987          * the number of bytes currently being processed; together those are
988          * the current remaining operation length */
989         job_progress_set_remaining(&s->common.job, s->bytes_in_flight + cnt);
990 
991         /* Note that even when no rate limit is applied we need to yield
992          * periodically with no pending I/O so that bdrv_drain_all() returns.
993          * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
994          * an error, or when the source is clean, whichever comes first. */
995         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
996         if (delta < BLOCK_JOB_SLICE_TIME &&
997             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
998             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
999                 (cnt == 0 && s->in_flight > 0)) {
1000                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
1001                 mirror_wait_for_free_in_flight_slot(s);
1002                 continue;
1003             } else if (cnt != 0) {
1004                 delay_ns = mirror_iteration(s);
1005             }
1006         }
1007 
1008         should_complete = false;
1009         if (s->in_flight == 0 && cnt == 0) {
1010             trace_mirror_before_flush(s);
1011             if (!s->synced) {
1012                 if (mirror_flush(s) < 0) {
1013                     /* Go check s->ret.  */
1014                     continue;
1015                 }
1016                 /* We're out of the streaming phase.  From now on, if the job
1017                  * is cancelled we will actually complete all pending I/O and
1018                  * report completion.  This way, block-job-cancel will leave
1019                  * the target in a consistent state.
1020                  */
1021                 job_transition_to_ready(&s->common.job);
1022                 s->synced = true;
1023                 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) {
1024                     s->actively_synced = true;
1025                 }
1026             }
1027 
1028             should_complete = s->should_complete ||
1029                 job_is_cancelled(&s->common.job);
1030             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1031         }
1032 
1033         if (cnt == 0 && should_complete) {
1034             /* The dirty bitmap is not updated while operations are pending.
1035              * If we're about to exit, wait for pending operations before
1036              * calling bdrv_get_dirty_count(bs), or we may exit while the
1037              * source has dirty data to copy!
1038              *
1039              * Note that I/O can be submitted by the guest while
1040              * mirror_populate runs, so pause it now.  Before deciding
1041              * whether to switch to target check one last time if I/O has
1042              * come in the meanwhile, and if not flush the data to disk.
1043              */
1044             trace_mirror_before_drain(s, cnt);
1045 
1046             s->in_drain = true;
1047             bdrv_drained_begin(bs);
1048             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1049             if (cnt > 0 || mirror_flush(s) < 0) {
1050                 bdrv_drained_end(bs);
1051                 s->in_drain = false;
1052                 continue;
1053             }
1054 
1055             /* The two disks are in sync.  Exit and report successful
1056              * completion.
1057              */
1058             assert(QLIST_EMPTY(&bs->tracked_requests));
1059             s->common.job.cancelled = false;
1060             need_drain = false;
1061             break;
1062         }
1063 
1064         ret = 0;
1065 
1066         if (s->synced && !should_complete) {
1067             delay_ns = (s->in_flight == 0 &&
1068                         cnt == 0 ? BLOCK_JOB_SLICE_TIME : 0);
1069         }
1070         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
1071         job_sleep_ns(&s->common.job, delay_ns);
1072         if (job_is_cancelled(&s->common.job) &&
1073             (!s->synced || s->common.job.force_cancel))
1074         {
1075             break;
1076         }
1077         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1078     }
1079 
1080 immediate_exit:
1081     if (s->in_flight > 0) {
1082         /* We get here only if something went wrong.  Either the job failed,
1083          * or it was cancelled prematurely so that we do not guarantee that
1084          * the target is a copy of the source.
1085          */
1086         assert(ret < 0 || ((s->common.job.force_cancel || !s->synced) &&
1087                job_is_cancelled(&s->common.job)));
1088         assert(need_drain);
1089         mirror_wait_for_all_io(s);
1090     }
1091 
1092     assert(s->in_flight == 0);
1093     qemu_vfree(s->buf);
1094     g_free(s->cow_bitmap);
1095     g_free(s->in_flight_bitmap);
1096     bdrv_dirty_iter_free(s->dbi);
1097 
1098     if (need_drain) {
1099         s->in_drain = true;
1100         bdrv_drained_begin(bs);
1101     }
1102 
1103     return ret;
1104 }
1105 
1106 static void mirror_complete(Job *job, Error **errp)
1107 {
1108     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1109     BlockDriverState *target;
1110 
1111     target = blk_bs(s->target);
1112 
1113     if (!s->synced) {
1114         error_setg(errp, "The active block job '%s' cannot be completed",
1115                    job->id);
1116         return;
1117     }
1118 
1119     if (s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
1120         int ret;
1121 
1122         assert(!bdrv_backing_chain_next(target));
1123         ret = bdrv_open_backing_file(bdrv_skip_filters(target), NULL,
1124                                      "backing", errp);
1125         if (ret < 0) {
1126             return;
1127         }
1128     }
1129 
1130     /* block all operations on to_replace bs */
1131     if (s->replaces) {
1132         AioContext *replace_aio_context;
1133 
1134         s->to_replace = bdrv_find_node(s->replaces);
1135         if (!s->to_replace) {
1136             error_setg(errp, "Node name '%s' not found", s->replaces);
1137             return;
1138         }
1139 
1140         replace_aio_context = bdrv_get_aio_context(s->to_replace);
1141         aio_context_acquire(replace_aio_context);
1142 
1143         /* TODO Translate this into permission system. Current definition of
1144          * GRAPH_MOD would require to request it for the parents; they might
1145          * not even be BlockDriverStates, however, so a BdrvChild can't address
1146          * them. May need redefinition of GRAPH_MOD. */
1147         error_setg(&s->replace_blocker,
1148                    "block device is in use by block-job-complete");
1149         bdrv_op_block_all(s->to_replace, s->replace_blocker);
1150         bdrv_ref(s->to_replace);
1151 
1152         aio_context_release(replace_aio_context);
1153     }
1154 
1155     s->should_complete = true;
1156     job_enter(job);
1157 }
1158 
1159 static void coroutine_fn mirror_pause(Job *job)
1160 {
1161     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1162 
1163     mirror_wait_for_all_io(s);
1164 }
1165 
1166 static bool mirror_drained_poll(BlockJob *job)
1167 {
1168     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1169 
1170     /* If the job isn't paused nor cancelled, we can't be sure that it won't
1171      * issue more requests. We make an exception if we've reached this point
1172      * from one of our own drain sections, to avoid a deadlock waiting for
1173      * ourselves.
1174      */
1175     if (!s->common.job.paused && !s->common.job.cancelled && !s->in_drain) {
1176         return true;
1177     }
1178 
1179     return !!s->in_flight;
1180 }
1181 
1182 static const BlockJobDriver mirror_job_driver = {
1183     .job_driver = {
1184         .instance_size          = sizeof(MirrorBlockJob),
1185         .job_type               = JOB_TYPE_MIRROR,
1186         .free                   = block_job_free,
1187         .user_resume            = block_job_user_resume,
1188         .run                    = mirror_run,
1189         .prepare                = mirror_prepare,
1190         .abort                  = mirror_abort,
1191         .pause                  = mirror_pause,
1192         .complete               = mirror_complete,
1193     },
1194     .drained_poll           = mirror_drained_poll,
1195 };
1196 
1197 static const BlockJobDriver commit_active_job_driver = {
1198     .job_driver = {
1199         .instance_size          = sizeof(MirrorBlockJob),
1200         .job_type               = JOB_TYPE_COMMIT,
1201         .free                   = block_job_free,
1202         .user_resume            = block_job_user_resume,
1203         .run                    = mirror_run,
1204         .prepare                = mirror_prepare,
1205         .abort                  = mirror_abort,
1206         .pause                  = mirror_pause,
1207         .complete               = mirror_complete,
1208     },
1209     .drained_poll           = mirror_drained_poll,
1210 };
1211 
1212 static void coroutine_fn
1213 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method,
1214                      uint64_t offset, uint64_t bytes,
1215                      QEMUIOVector *qiov, int flags)
1216 {
1217     int ret;
1218     size_t qiov_offset = 0;
1219     int64_t bitmap_offset, bitmap_end;
1220 
1221     if (!QEMU_IS_ALIGNED(offset, job->granularity) &&
1222         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset))
1223     {
1224             /*
1225              * Dirty unaligned padding: ignore it.
1226              *
1227              * Reasoning:
1228              * 1. If we copy it, we can't reset corresponding bit in
1229              *    dirty_bitmap as there may be some "dirty" bytes still not
1230              *    copied.
1231              * 2. It's already dirty, so skipping it we don't diverge mirror
1232              *    progress.
1233              *
1234              * Note, that because of this, guest write may have no contribution
1235              * into mirror converge, but that's not bad, as we have background
1236              * process of mirroring. If under some bad circumstances (high guest
1237              * IO load) background process starve, we will not converge anyway,
1238              * even if each write will contribute, as guest is not guaranteed to
1239              * rewrite the whole disk.
1240              */
1241             qiov_offset = QEMU_ALIGN_UP(offset, job->granularity) - offset;
1242             if (bytes <= qiov_offset) {
1243                 /* nothing to do after shrink */
1244                 return;
1245             }
1246             offset += qiov_offset;
1247             bytes -= qiov_offset;
1248     }
1249 
1250     if (!QEMU_IS_ALIGNED(offset + bytes, job->granularity) &&
1251         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset + bytes - 1))
1252     {
1253         uint64_t tail = (offset + bytes) % job->granularity;
1254 
1255         if (bytes <= tail) {
1256             /* nothing to do after shrink */
1257             return;
1258         }
1259         bytes -= tail;
1260     }
1261 
1262     /*
1263      * Tails are either clean or shrunk, so for bitmap resetting
1264      * we safely align the range down.
1265      */
1266     bitmap_offset = QEMU_ALIGN_UP(offset, job->granularity);
1267     bitmap_end = QEMU_ALIGN_DOWN(offset + bytes, job->granularity);
1268     if (bitmap_offset < bitmap_end) {
1269         bdrv_reset_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1270                                 bitmap_end - bitmap_offset);
1271     }
1272 
1273     job_progress_increase_remaining(&job->common.job, bytes);
1274 
1275     switch (method) {
1276     case MIRROR_METHOD_COPY:
1277         ret = blk_co_pwritev_part(job->target, offset, bytes,
1278                                   qiov, qiov_offset, flags);
1279         break;
1280 
1281     case MIRROR_METHOD_ZERO:
1282         assert(!qiov);
1283         ret = blk_co_pwrite_zeroes(job->target, offset, bytes, flags);
1284         break;
1285 
1286     case MIRROR_METHOD_DISCARD:
1287         assert(!qiov);
1288         ret = blk_co_pdiscard(job->target, offset, bytes);
1289         break;
1290 
1291     default:
1292         abort();
1293     }
1294 
1295     if (ret >= 0) {
1296         job_progress_update(&job->common.job, bytes);
1297     } else {
1298         BlockErrorAction action;
1299 
1300         /*
1301          * We failed, so we should mark dirty the whole area, aligned up.
1302          * Note that we don't care about shrunk tails if any: they were dirty
1303          * at function start, and they must be still dirty, as we've locked
1304          * the region for in-flight op.
1305          */
1306         bitmap_offset = QEMU_ALIGN_DOWN(offset, job->granularity);
1307         bitmap_end = QEMU_ALIGN_UP(offset + bytes, job->granularity);
1308         bdrv_set_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1309                               bitmap_end - bitmap_offset);
1310         job->actively_synced = false;
1311 
1312         action = mirror_error_action(job, false, -ret);
1313         if (action == BLOCK_ERROR_ACTION_REPORT) {
1314             if (!job->ret) {
1315                 job->ret = ret;
1316             }
1317         }
1318     }
1319 }
1320 
1321 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s,
1322                                                    uint64_t offset,
1323                                                    uint64_t bytes)
1324 {
1325     MirrorOp *op;
1326     uint64_t start_chunk = offset / s->granularity;
1327     uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
1328 
1329     op = g_new(MirrorOp, 1);
1330     *op = (MirrorOp){
1331         .s                  = s,
1332         .offset             = offset,
1333         .bytes              = bytes,
1334         .is_active_write    = true,
1335         .is_in_flight       = true,
1336     };
1337     qemu_co_queue_init(&op->waiting_requests);
1338     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
1339 
1340     s->in_active_write_counter++;
1341 
1342     mirror_wait_on_conflicts(op, s, offset, bytes);
1343 
1344     bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1345 
1346     return op;
1347 }
1348 
1349 static void coroutine_fn active_write_settle(MirrorOp *op)
1350 {
1351     uint64_t start_chunk = op->offset / op->s->granularity;
1352     uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes,
1353                                       op->s->granularity);
1354 
1355     if (!--op->s->in_active_write_counter && op->s->actively_synced) {
1356         BdrvChild *source = op->s->mirror_top_bs->backing;
1357 
1358         if (QLIST_FIRST(&source->bs->parents) == source &&
1359             QLIST_NEXT(source, next_parent) == NULL)
1360         {
1361             /* Assert that we are back in sync once all active write
1362              * operations are settled.
1363              * Note that we can only assert this if the mirror node
1364              * is the source node's only parent. */
1365             assert(!bdrv_get_dirty_count(op->s->dirty_bitmap));
1366         }
1367     }
1368     bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1369     QTAILQ_REMOVE(&op->s->ops_in_flight, op, next);
1370     qemu_co_queue_restart_all(&op->waiting_requests);
1371     g_free(op);
1372 }
1373 
1374 static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs,
1375     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1376 {
1377     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1378 }
1379 
1380 static int coroutine_fn bdrv_mirror_top_do_write(BlockDriverState *bs,
1381     MirrorMethod method, uint64_t offset, uint64_t bytes, QEMUIOVector *qiov,
1382     int flags)
1383 {
1384     MirrorOp *op = NULL;
1385     MirrorBDSOpaque *s = bs->opaque;
1386     int ret = 0;
1387     bool copy_to_target;
1388 
1389     copy_to_target = s->job->ret >= 0 &&
1390                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1391 
1392     if (copy_to_target) {
1393         op = active_write_prepare(s->job, offset, bytes);
1394     }
1395 
1396     switch (method) {
1397     case MIRROR_METHOD_COPY:
1398         ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1399         break;
1400 
1401     case MIRROR_METHOD_ZERO:
1402         ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1403         break;
1404 
1405     case MIRROR_METHOD_DISCARD:
1406         ret = bdrv_co_pdiscard(bs->backing, offset, bytes);
1407         break;
1408 
1409     default:
1410         abort();
1411     }
1412 
1413     if (ret < 0) {
1414         goto out;
1415     }
1416 
1417     if (copy_to_target) {
1418         do_sync_target_write(s->job, method, offset, bytes, qiov, flags);
1419     }
1420 
1421 out:
1422     if (copy_to_target) {
1423         active_write_settle(op);
1424     }
1425     return ret;
1426 }
1427 
1428 static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs,
1429     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1430 {
1431     MirrorBDSOpaque *s = bs->opaque;
1432     QEMUIOVector bounce_qiov;
1433     void *bounce_buf;
1434     int ret = 0;
1435     bool copy_to_target;
1436 
1437     copy_to_target = s->job->ret >= 0 &&
1438                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1439 
1440     if (copy_to_target) {
1441         /* The guest might concurrently modify the data to write; but
1442          * the data on source and destination must match, so we have
1443          * to use a bounce buffer if we are going to write to the
1444          * target now. */
1445         bounce_buf = qemu_blockalign(bs, bytes);
1446         iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes);
1447 
1448         qemu_iovec_init(&bounce_qiov, 1);
1449         qemu_iovec_add(&bounce_qiov, bounce_buf, bytes);
1450         qiov = &bounce_qiov;
1451     }
1452 
1453     ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov,
1454                                    flags);
1455 
1456     if (copy_to_target) {
1457         qemu_iovec_destroy(&bounce_qiov);
1458         qemu_vfree(bounce_buf);
1459     }
1460 
1461     return ret;
1462 }
1463 
1464 static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs)
1465 {
1466     if (bs->backing == NULL) {
1467         /* we can be here after failed bdrv_append in mirror_start_job */
1468         return 0;
1469     }
1470     return bdrv_co_flush(bs->backing->bs);
1471 }
1472 
1473 static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs,
1474     int64_t offset, int bytes, BdrvRequestFlags flags)
1475 {
1476     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL,
1477                                     flags);
1478 }
1479 
1480 static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs,
1481     int64_t offset, int bytes)
1482 {
1483     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes,
1484                                     NULL, 0);
1485 }
1486 
1487 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs)
1488 {
1489     if (bs->backing == NULL) {
1490         /* we can be here after failed bdrv_attach_child in
1491          * bdrv_set_backing_hd */
1492         return;
1493     }
1494     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1495             bs->backing->bs->filename);
1496 }
1497 
1498 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1499                                        BdrvChildRole role,
1500                                        BlockReopenQueue *reopen_queue,
1501                                        uint64_t perm, uint64_t shared,
1502                                        uint64_t *nperm, uint64_t *nshared)
1503 {
1504     MirrorBDSOpaque *s = bs->opaque;
1505 
1506     if (s->stop) {
1507         /*
1508          * If the job is to be stopped, we do not need to forward
1509          * anything to the real image.
1510          */
1511         *nperm = 0;
1512         *nshared = BLK_PERM_ALL;
1513         return;
1514     }
1515 
1516     /* Must be able to forward guest writes to the real image */
1517     *nperm = 0;
1518     if (perm & BLK_PERM_WRITE) {
1519         *nperm |= BLK_PERM_WRITE;
1520     }
1521 
1522     *nshared = BLK_PERM_ALL;
1523 }
1524 
1525 /* Dummy node that provides consistent read to its users without requiring it
1526  * from its backing file and that allows writes on the backing file chain. */
1527 static BlockDriver bdrv_mirror_top = {
1528     .format_name                = "mirror_top",
1529     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1530     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1531     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1532     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1533     .bdrv_co_flush              = bdrv_mirror_top_flush,
1534     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1535     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1536 
1537     .is_filter                  = true,
1538 };
1539 
1540 static BlockJob *mirror_start_job(
1541                              const char *job_id, BlockDriverState *bs,
1542                              int creation_flags, BlockDriverState *target,
1543                              const char *replaces, int64_t speed,
1544                              uint32_t granularity, int64_t buf_size,
1545                              BlockMirrorBackingMode backing_mode,
1546                              bool zero_target,
1547                              BlockdevOnError on_source_error,
1548                              BlockdevOnError on_target_error,
1549                              bool unmap,
1550                              BlockCompletionFunc *cb,
1551                              void *opaque,
1552                              const BlockJobDriver *driver,
1553                              bool is_none_mode, BlockDriverState *base,
1554                              bool auto_complete, const char *filter_node_name,
1555                              bool is_mirror, MirrorCopyMode copy_mode,
1556                              Error **errp)
1557 {
1558     MirrorBlockJob *s;
1559     MirrorBDSOpaque *bs_opaque;
1560     BlockDriverState *mirror_top_bs;
1561     bool target_is_backing;
1562     uint64_t target_perms, target_shared_perms;
1563     Error *local_err = NULL;
1564     int ret;
1565 
1566     if (granularity == 0) {
1567         granularity = bdrv_get_default_bitmap_granularity(target);
1568     }
1569 
1570     assert(is_power_of_2(granularity));
1571 
1572     if (buf_size < 0) {
1573         error_setg(errp, "Invalid parameter 'buf-size'");
1574         return NULL;
1575     }
1576 
1577     if (buf_size == 0) {
1578         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1579     }
1580 
1581     if (bdrv_skip_filters(bs) == bdrv_skip_filters(target)) {
1582         error_setg(errp, "Can't mirror node into itself");
1583         return NULL;
1584     }
1585 
1586     /* In the case of active commit, add dummy driver to provide consistent
1587      * reads on the top, while disabling it in the intermediate nodes, and make
1588      * the backing chain writable. */
1589     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1590                                          BDRV_O_RDWR, errp);
1591     if (mirror_top_bs == NULL) {
1592         return NULL;
1593     }
1594     if (!filter_node_name) {
1595         mirror_top_bs->implicit = true;
1596     }
1597 
1598     /* So that we can always drop this node */
1599     mirror_top_bs->never_freeze = true;
1600 
1601     mirror_top_bs->total_sectors = bs->total_sectors;
1602     mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
1603     mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
1604                                           BDRV_REQ_NO_FALLBACK;
1605     bs_opaque = g_new0(MirrorBDSOpaque, 1);
1606     mirror_top_bs->opaque = bs_opaque;
1607 
1608     /* bdrv_append takes ownership of the mirror_top_bs reference, need to keep
1609      * it alive until block_job_create() succeeds even if bs has no parent. */
1610     bdrv_ref(mirror_top_bs);
1611     bdrv_drained_begin(bs);
1612     bdrv_append(mirror_top_bs, bs, &local_err);
1613     bdrv_drained_end(bs);
1614 
1615     if (local_err) {
1616         bdrv_unref(mirror_top_bs);
1617         error_propagate(errp, local_err);
1618         return NULL;
1619     }
1620 
1621     /* Make sure that the source is not resized while the job is running */
1622     s = block_job_create(job_id, driver, NULL, mirror_top_bs,
1623                          BLK_PERM_CONSISTENT_READ,
1624                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1625                          BLK_PERM_WRITE | BLK_PERM_GRAPH_MOD, speed,
1626                          creation_flags, cb, opaque, errp);
1627     if (!s) {
1628         goto fail;
1629     }
1630     bs_opaque->job = s;
1631 
1632     /* The block job now has a reference to this node */
1633     bdrv_unref(mirror_top_bs);
1634 
1635     s->mirror_top_bs = mirror_top_bs;
1636 
1637     /* No resize for the target either; while the mirror is still running, a
1638      * consistent read isn't necessarily possible. We could possibly allow
1639      * writes and graph modifications, though it would likely defeat the
1640      * purpose of a mirror, so leave them blocked for now.
1641      *
1642      * In the case of active commit, things look a bit different, though,
1643      * because the target is an already populated backing file in active use.
1644      * We can allow anything except resize there.*/
1645 
1646     target_perms = BLK_PERM_WRITE;
1647     target_shared_perms = BLK_PERM_WRITE_UNCHANGED;
1648 
1649     target_is_backing = bdrv_chain_contains(bs, target);
1650     if (target_is_backing) {
1651         int64_t bs_size, target_size;
1652         bs_size = bdrv_getlength(bs);
1653         if (bs_size < 0) {
1654             error_setg_errno(errp, -bs_size,
1655                              "Could not inquire top image size");
1656             goto fail;
1657         }
1658 
1659         target_size = bdrv_getlength(target);
1660         if (target_size < 0) {
1661             error_setg_errno(errp, -target_size,
1662                              "Could not inquire base image size");
1663             goto fail;
1664         }
1665 
1666         if (target_size < bs_size) {
1667             target_perms |= BLK_PERM_RESIZE;
1668         }
1669 
1670         target_shared_perms |= BLK_PERM_CONSISTENT_READ
1671                             |  BLK_PERM_WRITE
1672                             |  BLK_PERM_GRAPH_MOD;
1673     } else if (bdrv_chain_contains(bs, bdrv_skip_filters(target))) {
1674         /*
1675          * We may want to allow this in the future, but it would
1676          * require taking some extra care.
1677          */
1678         error_setg(errp, "Cannot mirror to a filter on top of a node in the "
1679                    "source's backing chain");
1680         goto fail;
1681     }
1682 
1683     if (backing_mode != MIRROR_LEAVE_BACKING_CHAIN) {
1684         target_perms |= BLK_PERM_GRAPH_MOD;
1685     }
1686 
1687     s->target = blk_new(s->common.job.aio_context,
1688                         target_perms, target_shared_perms);
1689     ret = blk_insert_bs(s->target, target, errp);
1690     if (ret < 0) {
1691         goto fail;
1692     }
1693     if (is_mirror) {
1694         /* XXX: Mirror target could be a NBD server of target QEMU in the case
1695          * of non-shared block migration. To allow migration completion, we
1696          * have to allow "inactivate" of the target BB.  When that happens, we
1697          * know the job is drained, and the vcpus are stopped, so no write
1698          * operation will be performed. Block layer already has assertions to
1699          * ensure that. */
1700         blk_set_force_allow_inactivate(s->target);
1701     }
1702     blk_set_allow_aio_context_change(s->target, true);
1703     blk_set_disable_request_queuing(s->target, true);
1704 
1705     s->replaces = g_strdup(replaces);
1706     s->on_source_error = on_source_error;
1707     s->on_target_error = on_target_error;
1708     s->is_none_mode = is_none_mode;
1709     s->backing_mode = backing_mode;
1710     s->zero_target = zero_target;
1711     s->copy_mode = copy_mode;
1712     s->base = base;
1713     s->base_overlay = bdrv_find_overlay(bs, base);
1714     s->granularity = granularity;
1715     s->buf_size = ROUND_UP(buf_size, granularity);
1716     s->unmap = unmap;
1717     if (auto_complete) {
1718         s->should_complete = true;
1719     }
1720 
1721     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1722     if (!s->dirty_bitmap) {
1723         goto fail;
1724     }
1725     if (s->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING) {
1726         bdrv_disable_dirty_bitmap(s->dirty_bitmap);
1727     }
1728 
1729     ret = block_job_add_bdrv(&s->common, "source", bs, 0,
1730                              BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE |
1731                              BLK_PERM_CONSISTENT_READ,
1732                              errp);
1733     if (ret < 0) {
1734         goto fail;
1735     }
1736 
1737     /* Required permissions are already taken with blk_new() */
1738     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1739                        &error_abort);
1740 
1741     /* In commit_active_start() all intermediate nodes disappear, so
1742      * any jobs in them must be blocked */
1743     if (target_is_backing) {
1744         BlockDriverState *iter, *filtered_target;
1745         uint64_t iter_shared_perms;
1746 
1747         /*
1748          * The topmost node with
1749          * bdrv_skip_filters(filtered_target) == bdrv_skip_filters(target)
1750          */
1751         filtered_target = bdrv_cow_bs(bdrv_find_overlay(bs, target));
1752 
1753         assert(bdrv_skip_filters(filtered_target) ==
1754                bdrv_skip_filters(target));
1755 
1756         /*
1757          * XXX BLK_PERM_WRITE needs to be allowed so we don't block
1758          * ourselves at s->base (if writes are blocked for a node, they are
1759          * also blocked for its backing file). The other options would be a
1760          * second filter driver above s->base (== target).
1761          */
1762         iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE;
1763 
1764         for (iter = bdrv_filter_or_cow_bs(bs); iter != target;
1765              iter = bdrv_filter_or_cow_bs(iter))
1766         {
1767             if (iter == filtered_target) {
1768                 /*
1769                  * From here on, all nodes are filters on the base.
1770                  * This allows us to share BLK_PERM_CONSISTENT_READ.
1771                  */
1772                 iter_shared_perms |= BLK_PERM_CONSISTENT_READ;
1773             }
1774 
1775             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1776                                      iter_shared_perms, errp);
1777             if (ret < 0) {
1778                 goto fail;
1779             }
1780         }
1781 
1782         if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) {
1783             goto fail;
1784         }
1785     }
1786 
1787     QTAILQ_INIT(&s->ops_in_flight);
1788 
1789     trace_mirror_start(bs, s, opaque);
1790     job_start(&s->common.job);
1791 
1792     return &s->common;
1793 
1794 fail:
1795     if (s) {
1796         /* Make sure this BDS does not go away until we have completed the graph
1797          * changes below */
1798         bdrv_ref(mirror_top_bs);
1799 
1800         g_free(s->replaces);
1801         blk_unref(s->target);
1802         bs_opaque->job = NULL;
1803         if (s->dirty_bitmap) {
1804             bdrv_release_dirty_bitmap(s->dirty_bitmap);
1805         }
1806         job_early_fail(&s->common.job);
1807     }
1808 
1809     bs_opaque->stop = true;
1810     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
1811                              &error_abort);
1812     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
1813 
1814     bdrv_unref(mirror_top_bs);
1815 
1816     return NULL;
1817 }
1818 
1819 void mirror_start(const char *job_id, BlockDriverState *bs,
1820                   BlockDriverState *target, const char *replaces,
1821                   int creation_flags, int64_t speed,
1822                   uint32_t granularity, int64_t buf_size,
1823                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1824                   bool zero_target,
1825                   BlockdevOnError on_source_error,
1826                   BlockdevOnError on_target_error,
1827                   bool unmap, const char *filter_node_name,
1828                   MirrorCopyMode copy_mode, Error **errp)
1829 {
1830     bool is_none_mode;
1831     BlockDriverState *base;
1832 
1833     if ((mode == MIRROR_SYNC_MODE_INCREMENTAL) ||
1834         (mode == MIRROR_SYNC_MODE_BITMAP)) {
1835         error_setg(errp, "Sync mode '%s' not supported",
1836                    MirrorSyncMode_str(mode));
1837         return;
1838     }
1839     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1840     base = mode == MIRROR_SYNC_MODE_TOP ? bdrv_backing_chain_next(bs) : NULL;
1841     mirror_start_job(job_id, bs, creation_flags, target, replaces,
1842                      speed, granularity, buf_size, backing_mode, zero_target,
1843                      on_source_error, on_target_error, unmap, NULL, NULL,
1844                      &mirror_job_driver, is_none_mode, base, false,
1845                      filter_node_name, true, copy_mode, errp);
1846 }
1847 
1848 BlockJob *commit_active_start(const char *job_id, BlockDriverState *bs,
1849                               BlockDriverState *base, int creation_flags,
1850                               int64_t speed, BlockdevOnError on_error,
1851                               const char *filter_node_name,
1852                               BlockCompletionFunc *cb, void *opaque,
1853                               bool auto_complete, Error **errp)
1854 {
1855     bool base_read_only;
1856     Error *local_err = NULL;
1857     BlockJob *ret;
1858 
1859     base_read_only = bdrv_is_read_only(base);
1860 
1861     if (base_read_only) {
1862         if (bdrv_reopen_set_read_only(base, false, errp) < 0) {
1863             return NULL;
1864         }
1865     }
1866 
1867     ret = mirror_start_job(
1868                      job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1869                      MIRROR_LEAVE_BACKING_CHAIN, false,
1870                      on_error, on_error, true, cb, opaque,
1871                      &commit_active_job_driver, false, base, auto_complete,
1872                      filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND,
1873                      &local_err);
1874     if (local_err) {
1875         error_propagate(errp, local_err);
1876         goto error_restore_flags;
1877     }
1878 
1879     return ret;
1880 
1881 error_restore_flags:
1882     /* ignore error and errp for bdrv_reopen, because we want to propagate
1883      * the original error */
1884     if (base_read_only) {
1885         bdrv_reopen_set_read_only(base, true, NULL);
1886     }
1887     return NULL;
1888 }
1889