xref: /openbmc/qemu/block/mirror.c (revision 46627f41)
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     bool is_commit;
93 } MirrorBDSOpaque;
94 
95 struct MirrorOp {
96     MirrorBlockJob *s;
97     QEMUIOVector qiov;
98     int64_t offset;
99     uint64_t bytes;
100 
101     /* The pointee is set by mirror_co_read(), mirror_co_zero(), and
102      * mirror_co_discard() before yielding for the first time */
103     int64_t *bytes_handled;
104 
105     bool is_pseudo_op;
106     bool is_active_write;
107     bool is_in_flight;
108     CoQueue waiting_requests;
109     Coroutine *co;
110 
111     QTAILQ_ENTRY(MirrorOp) next;
112 };
113 
114 typedef enum MirrorMethod {
115     MIRROR_METHOD_COPY,
116     MIRROR_METHOD_ZERO,
117     MIRROR_METHOD_DISCARD,
118 } MirrorMethod;
119 
120 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
121                                             int error)
122 {
123     s->synced = false;
124     s->actively_synced = false;
125     if (read) {
126         return block_job_error_action(&s->common, s->on_source_error,
127                                       true, error);
128     } else {
129         return block_job_error_action(&s->common, s->on_target_error,
130                                       false, error);
131     }
132 }
133 
134 static void coroutine_fn mirror_wait_on_conflicts(MirrorOp *self,
135                                                   MirrorBlockJob *s,
136                                                   uint64_t offset,
137                                                   uint64_t bytes)
138 {
139     uint64_t self_start_chunk = offset / s->granularity;
140     uint64_t self_end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
141     uint64_t self_nb_chunks = self_end_chunk - self_start_chunk;
142 
143     while (find_next_bit(s->in_flight_bitmap, self_end_chunk,
144                          self_start_chunk) < self_end_chunk &&
145            s->ret >= 0)
146     {
147         MirrorOp *op;
148 
149         QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
150             uint64_t op_start_chunk = op->offset / s->granularity;
151             uint64_t op_nb_chunks = DIV_ROUND_UP(op->offset + op->bytes,
152                                                  s->granularity) -
153                                     op_start_chunk;
154 
155             if (op == self) {
156                 continue;
157             }
158 
159             if (ranges_overlap(self_start_chunk, self_nb_chunks,
160                                op_start_chunk, op_nb_chunks))
161             {
162                 qemu_co_queue_wait(&op->waiting_requests, NULL);
163                 break;
164             }
165         }
166     }
167 }
168 
169 static void coroutine_fn mirror_iteration_done(MirrorOp *op, int ret)
170 {
171     MirrorBlockJob *s = op->s;
172     struct iovec *iov;
173     int64_t chunk_num;
174     int i, nb_chunks;
175 
176     trace_mirror_iteration_done(s, op->offset, op->bytes, ret);
177 
178     s->in_flight--;
179     s->bytes_in_flight -= op->bytes;
180     iov = op->qiov.iov;
181     for (i = 0; i < op->qiov.niov; i++) {
182         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
183         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
184         s->buf_free_count++;
185     }
186 
187     chunk_num = op->offset / s->granularity;
188     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
189 
190     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
191     QTAILQ_REMOVE(&s->ops_in_flight, op, next);
192     if (ret >= 0) {
193         if (s->cow_bitmap) {
194             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
195         }
196         if (!s->initial_zeroing_ongoing) {
197             job_progress_update(&s->common.job, op->bytes);
198         }
199     }
200     qemu_iovec_destroy(&op->qiov);
201 
202     qemu_co_queue_restart_all(&op->waiting_requests);
203     g_free(op);
204 }
205 
206 static void coroutine_fn mirror_write_complete(MirrorOp *op, int ret)
207 {
208     MirrorBlockJob *s = op->s;
209 
210     if (ret < 0) {
211         BlockErrorAction action;
212 
213         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
214         action = mirror_error_action(s, false, -ret);
215         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
216             s->ret = ret;
217         }
218     }
219 
220     mirror_iteration_done(op, ret);
221 }
222 
223 static void coroutine_fn mirror_read_complete(MirrorOp *op, int ret)
224 {
225     MirrorBlockJob *s = op->s;
226 
227     if (ret < 0) {
228         BlockErrorAction action;
229 
230         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
231         action = mirror_error_action(s, true, -ret);
232         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
233             s->ret = ret;
234         }
235 
236         mirror_iteration_done(op, ret);
237         return;
238     }
239 
240     ret = blk_co_pwritev(s->target, op->offset, op->qiov.size, &op->qiov, 0);
241     mirror_write_complete(op, ret);
242 }
243 
244 /* Clip bytes relative to offset to not exceed end-of-file */
245 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s,
246                                         int64_t offset,
247                                         int64_t bytes)
248 {
249     return MIN(bytes, s->bdev_length - offset);
250 }
251 
252 /* Round offset and/or bytes to target cluster if COW is needed, and
253  * return the offset of the adjusted tail against original. */
254 static int mirror_cow_align(MirrorBlockJob *s, int64_t *offset,
255                             uint64_t *bytes)
256 {
257     bool need_cow;
258     int ret = 0;
259     int64_t align_offset = *offset;
260     int64_t align_bytes = *bytes;
261     int max_bytes = s->granularity * s->max_iov;
262 
263     need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap);
264     need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity,
265                           s->cow_bitmap);
266     if (need_cow) {
267         bdrv_round_to_clusters(blk_bs(s->target), *offset, *bytes,
268                                &align_offset, &align_bytes);
269     }
270 
271     if (align_bytes > max_bytes) {
272         align_bytes = max_bytes;
273         if (need_cow) {
274             align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size);
275         }
276     }
277     /* Clipping may result in align_bytes unaligned to chunk boundary, but
278      * that doesn't matter because it's already the end of source image. */
279     align_bytes = mirror_clip_bytes(s, align_offset, align_bytes);
280 
281     ret = align_offset + align_bytes - (*offset + *bytes);
282     *offset = align_offset;
283     *bytes = align_bytes;
284     assert(ret >= 0);
285     return ret;
286 }
287 
288 static inline void coroutine_fn
289 mirror_wait_for_any_operation(MirrorBlockJob *s, bool active)
290 {
291     MirrorOp *op;
292 
293     QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
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_in_flight &&
300             op->is_active_write == active)
301         {
302             qemu_co_queue_wait(&op->waiting_requests, NULL);
303             return;
304         }
305     }
306     abort();
307 }
308 
309 static inline void coroutine_fn
310 mirror_wait_for_free_in_flight_slot(MirrorBlockJob *s)
311 {
312     /* Only non-active operations use up in-flight slots */
313     mirror_wait_for_any_operation(s, false);
314 }
315 
316 /* Perform a mirror copy operation.
317  *
318  * *op->bytes_handled is set to the number of bytes copied after and
319  * including offset, excluding any bytes copied prior to offset due
320  * to alignment.  This will be op->bytes if no alignment is necessary,
321  * or (new_end - op->offset) if the tail is rounded up or down due to
322  * alignment or buffer limit.
323  */
324 static void coroutine_fn mirror_co_read(void *opaque)
325 {
326     MirrorOp *op = opaque;
327     MirrorBlockJob *s = op->s;
328     int nb_chunks;
329     uint64_t ret;
330     uint64_t max_bytes;
331 
332     max_bytes = s->granularity * s->max_iov;
333 
334     /* We can only handle as much as buf_size at a time. */
335     op->bytes = MIN(s->buf_size, MIN(max_bytes, op->bytes));
336     assert(op->bytes);
337     assert(op->bytes < BDRV_REQUEST_MAX_BYTES);
338     *op->bytes_handled = op->bytes;
339 
340     if (s->cow_bitmap) {
341         *op->bytes_handled += mirror_cow_align(s, &op->offset, &op->bytes);
342     }
343     /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */
344     assert(*op->bytes_handled <= UINT_MAX);
345     assert(op->bytes <= s->buf_size);
346     /* The offset is granularity-aligned because:
347      * 1) Caller passes in aligned values;
348      * 2) mirror_cow_align is used only when target cluster is larger. */
349     assert(QEMU_IS_ALIGNED(op->offset, s->granularity));
350     /* The range is sector-aligned, since bdrv_getlength() rounds up. */
351     assert(QEMU_IS_ALIGNED(op->bytes, BDRV_SECTOR_SIZE));
352     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
353 
354     while (s->buf_free_count < nb_chunks) {
355         trace_mirror_yield_in_flight(s, op->offset, s->in_flight);
356         mirror_wait_for_free_in_flight_slot(s);
357     }
358 
359     /* Now make a QEMUIOVector taking enough granularity-sized chunks
360      * from s->buf_free.
361      */
362     qemu_iovec_init(&op->qiov, nb_chunks);
363     while (nb_chunks-- > 0) {
364         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
365         size_t remaining = op->bytes - op->qiov.size;
366 
367         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
368         s->buf_free_count--;
369         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
370     }
371 
372     /* Copy the dirty cluster.  */
373     s->in_flight++;
374     s->bytes_in_flight += op->bytes;
375     op->is_in_flight = true;
376     trace_mirror_one_iteration(s, op->offset, op->bytes);
377 
378     ret = bdrv_co_preadv(s->mirror_top_bs->backing, op->offset, op->bytes,
379                          &op->qiov, 0);
380     mirror_read_complete(op, ret);
381 }
382 
383 static void coroutine_fn mirror_co_zero(void *opaque)
384 {
385     MirrorOp *op = opaque;
386     int ret;
387 
388     op->s->in_flight++;
389     op->s->bytes_in_flight += op->bytes;
390     *op->bytes_handled = op->bytes;
391     op->is_in_flight = true;
392 
393     ret = blk_co_pwrite_zeroes(op->s->target, op->offset, op->bytes,
394                                op->s->unmap ? BDRV_REQ_MAY_UNMAP : 0);
395     mirror_write_complete(op, ret);
396 }
397 
398 static void coroutine_fn mirror_co_discard(void *opaque)
399 {
400     MirrorOp *op = opaque;
401     int ret;
402 
403     op->s->in_flight++;
404     op->s->bytes_in_flight += op->bytes;
405     *op->bytes_handled = op->bytes;
406     op->is_in_flight = true;
407 
408     ret = blk_co_pdiscard(op->s->target, op->offset, op->bytes);
409     mirror_write_complete(op, ret);
410 }
411 
412 static unsigned mirror_perform(MirrorBlockJob *s, int64_t offset,
413                                unsigned bytes, MirrorMethod mirror_method)
414 {
415     MirrorOp *op;
416     Coroutine *co;
417     int64_t bytes_handled = -1;
418 
419     op = g_new(MirrorOp, 1);
420     *op = (MirrorOp){
421         .s              = s,
422         .offset         = offset,
423         .bytes          = bytes,
424         .bytes_handled  = &bytes_handled,
425     };
426     qemu_co_queue_init(&op->waiting_requests);
427 
428     switch (mirror_method) {
429     case MIRROR_METHOD_COPY:
430         co = qemu_coroutine_create(mirror_co_read, op);
431         break;
432     case MIRROR_METHOD_ZERO:
433         co = qemu_coroutine_create(mirror_co_zero, op);
434         break;
435     case MIRROR_METHOD_DISCARD:
436         co = qemu_coroutine_create(mirror_co_discard, op);
437         break;
438     default:
439         abort();
440     }
441     op->co = co;
442 
443     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
444     qemu_coroutine_enter(co);
445     /* At this point, ownership of op has been moved to the coroutine
446      * and the object may already be freed */
447 
448     /* Assert that this value has been set */
449     assert(bytes_handled >= 0);
450 
451     /* Same assertion as in mirror_co_read() (and for mirror_co_read()
452      * and mirror_co_discard(), bytes_handled == op->bytes, which
453      * is the @bytes parameter given to this function) */
454     assert(bytes_handled <= UINT_MAX);
455     return bytes_handled;
456 }
457 
458 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
459 {
460     BlockDriverState *source = s->mirror_top_bs->backing->bs;
461     MirrorOp *pseudo_op;
462     int64_t offset;
463     uint64_t delay_ns = 0, ret = 0;
464     /* At least the first dirty chunk is mirrored in one iteration. */
465     int nb_chunks = 1;
466     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
467     int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES);
468 
469     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
470     offset = bdrv_dirty_iter_next(s->dbi);
471     if (offset < 0) {
472         bdrv_set_dirty_iter(s->dbi, 0);
473         offset = bdrv_dirty_iter_next(s->dbi);
474         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
475         assert(offset >= 0);
476     }
477     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
478 
479     mirror_wait_on_conflicts(NULL, s, offset, 1);
480 
481     job_pause_point(&s->common.job);
482 
483     /* Find the number of consective dirty chunks following the first dirty
484      * one, and wait for in flight requests in them. */
485     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
486     while (nb_chunks * s->granularity < s->buf_size) {
487         int64_t next_dirty;
488         int64_t next_offset = offset + nb_chunks * s->granularity;
489         int64_t next_chunk = next_offset / s->granularity;
490         if (next_offset >= s->bdev_length ||
491             !bdrv_dirty_bitmap_get_locked(s->dirty_bitmap, next_offset)) {
492             break;
493         }
494         if (test_bit(next_chunk, s->in_flight_bitmap)) {
495             break;
496         }
497 
498         next_dirty = bdrv_dirty_iter_next(s->dbi);
499         if (next_dirty > next_offset || next_dirty < 0) {
500             /* The bitmap iterator's cache is stale, refresh it */
501             bdrv_set_dirty_iter(s->dbi, next_offset);
502             next_dirty = bdrv_dirty_iter_next(s->dbi);
503         }
504         assert(next_dirty == next_offset);
505         nb_chunks++;
506     }
507 
508     /* Clear dirty bits before querying the block status, because
509      * calling bdrv_block_status_above could yield - if some blocks are
510      * marked dirty in this window, we need to know.
511      */
512     bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset,
513                                    nb_chunks * s->granularity);
514     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
515 
516     /* Before claiming an area in the in-flight bitmap, we have to
517      * create a MirrorOp for it so that conflicting requests can wait
518      * for it.  mirror_perform() will create the real MirrorOps later,
519      * for now we just create a pseudo operation that will wake up all
520      * conflicting requests once all real operations have been
521      * launched. */
522     pseudo_op = g_new(MirrorOp, 1);
523     *pseudo_op = (MirrorOp){
524         .offset         = offset,
525         .bytes          = nb_chunks * s->granularity,
526         .is_pseudo_op   = true,
527     };
528     qemu_co_queue_init(&pseudo_op->waiting_requests);
529     QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next);
530 
531     bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks);
532     while (nb_chunks > 0 && offset < s->bdev_length) {
533         int ret;
534         int64_t io_bytes;
535         int64_t io_bytes_acct;
536         MirrorMethod mirror_method = MIRROR_METHOD_COPY;
537 
538         assert(!(offset % s->granularity));
539         ret = bdrv_block_status_above(source, NULL, offset,
540                                       nb_chunks * s->granularity,
541                                       &io_bytes, NULL, NULL);
542         if (ret < 0) {
543             io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes);
544         } else if (ret & BDRV_BLOCK_DATA) {
545             io_bytes = MIN(io_bytes, max_io_bytes);
546         }
547 
548         io_bytes -= io_bytes % s->granularity;
549         if (io_bytes < s->granularity) {
550             io_bytes = s->granularity;
551         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
552             int64_t target_offset;
553             int64_t target_bytes;
554             bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes,
555                                    &target_offset, &target_bytes);
556             if (target_offset == offset &&
557                 target_bytes == io_bytes) {
558                 mirror_method = ret & BDRV_BLOCK_ZERO ?
559                                     MIRROR_METHOD_ZERO :
560                                     MIRROR_METHOD_DISCARD;
561             }
562         }
563 
564         while (s->in_flight >= MAX_IN_FLIGHT) {
565             trace_mirror_yield_in_flight(s, offset, s->in_flight);
566             mirror_wait_for_free_in_flight_slot(s);
567         }
568 
569         if (s->ret < 0) {
570             ret = 0;
571             goto fail;
572         }
573 
574         io_bytes = mirror_clip_bytes(s, offset, io_bytes);
575         io_bytes = mirror_perform(s, offset, io_bytes, mirror_method);
576         if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) {
577             io_bytes_acct = 0;
578         } else {
579             io_bytes_acct = io_bytes;
580         }
581         assert(io_bytes);
582         offset += io_bytes;
583         nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity);
584         delay_ns = block_job_ratelimit_get_delay(&s->common, io_bytes_acct);
585     }
586 
587     ret = delay_ns;
588 fail:
589     QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next);
590     qemu_co_queue_restart_all(&pseudo_op->waiting_requests);
591     g_free(pseudo_op);
592 
593     return ret;
594 }
595 
596 static void mirror_free_init(MirrorBlockJob *s)
597 {
598     int granularity = s->granularity;
599     size_t buf_size = s->buf_size;
600     uint8_t *buf = s->buf;
601 
602     assert(s->buf_free_count == 0);
603     QSIMPLEQ_INIT(&s->buf_free);
604     while (buf_size != 0) {
605         MirrorBuffer *cur = (MirrorBuffer *)buf;
606         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
607         s->buf_free_count++;
608         buf_size -= granularity;
609         buf += granularity;
610     }
611 }
612 
613 /* This is also used for the .pause callback. There is no matching
614  * mirror_resume() because mirror_run() will begin iterating again
615  * when the job is resumed.
616  */
617 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s)
618 {
619     while (s->in_flight > 0) {
620         mirror_wait_for_free_in_flight_slot(s);
621     }
622 }
623 
624 /**
625  * mirror_exit_common: handle both abort() and prepare() cases.
626  * for .prepare, returns 0 on success and -errno on failure.
627  * for .abort cases, denoted by abort = true, MUST return 0.
628  */
629 static int mirror_exit_common(Job *job)
630 {
631     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
632     BlockJob *bjob = &s->common;
633     MirrorBDSOpaque *bs_opaque;
634     AioContext *replace_aio_context = NULL;
635     BlockDriverState *src;
636     BlockDriverState *target_bs;
637     BlockDriverState *mirror_top_bs;
638     Error *local_err = NULL;
639     bool abort = job->ret < 0;
640     int ret = 0;
641 
642     if (s->prepared) {
643         return 0;
644     }
645     s->prepared = true;
646 
647     mirror_top_bs = s->mirror_top_bs;
648     bs_opaque = mirror_top_bs->opaque;
649     src = mirror_top_bs->backing->bs;
650     target_bs = blk_bs(s->target);
651 
652     if (bdrv_chain_contains(src, target_bs)) {
653         bdrv_unfreeze_backing_chain(mirror_top_bs, target_bs);
654     }
655 
656     bdrv_release_dirty_bitmap(s->dirty_bitmap);
657 
658     /* Make sure that the source BDS doesn't go away during bdrv_replace_node,
659      * before we can call bdrv_drained_end */
660     bdrv_ref(src);
661     bdrv_ref(mirror_top_bs);
662     bdrv_ref(target_bs);
663 
664     /*
665      * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
666      * inserting target_bs at s->to_replace, where we might not be able to get
667      * these permissions.
668      */
669     blk_unref(s->target);
670     s->target = NULL;
671 
672     /* We don't access the source any more. Dropping any WRITE/RESIZE is
673      * required before it could become a backing file of target_bs. Not having
674      * these permissions any more means that we can't allow any new requests on
675      * mirror_top_bs from now on, so keep it drained. */
676     bdrv_drained_begin(mirror_top_bs);
677     bs_opaque->stop = true;
678     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
679                              &error_abort);
680     if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
681         BlockDriverState *backing = s->is_none_mode ? src : s->base;
682         BlockDriverState *unfiltered_target = bdrv_skip_filters(target_bs);
683 
684         if (bdrv_cow_bs(unfiltered_target) != backing) {
685             bdrv_set_backing_hd(unfiltered_target, backing, &local_err);
686             if (local_err) {
687                 error_report_err(local_err);
688                 local_err = NULL;
689                 ret = -EPERM;
690             }
691         }
692     } else if (!abort && s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
693         assert(!bdrv_backing_chain_next(target_bs));
694         ret = bdrv_open_backing_file(bdrv_skip_filters(target_bs), NULL,
695                                      "backing", &local_err);
696         if (ret < 0) {
697             error_report_err(local_err);
698             local_err = NULL;
699         }
700     }
701 
702     if (s->to_replace) {
703         replace_aio_context = bdrv_get_aio_context(s->to_replace);
704         aio_context_acquire(replace_aio_context);
705     }
706 
707     if (s->should_complete && !abort) {
708         BlockDriverState *to_replace = s->to_replace ?: src;
709         bool ro = bdrv_is_read_only(to_replace);
710 
711         if (ro != bdrv_is_read_only(target_bs)) {
712             bdrv_reopen_set_read_only(target_bs, ro, NULL);
713         }
714 
715         /* The mirror job has no requests in flight any more, but we need to
716          * drain potential other users of the BDS before changing the graph. */
717         assert(s->in_drain);
718         bdrv_drained_begin(target_bs);
719         /*
720          * Cannot use check_to_replace_node() here, because that would
721          * check for an op blocker on @to_replace, and we have our own
722          * there.
723          */
724         if (bdrv_recurse_can_replace(src, to_replace)) {
725             bdrv_replace_node(to_replace, target_bs, &local_err);
726         } else {
727             error_setg(&local_err, "Can no longer replace '%s' by '%s', "
728                        "because it can no longer be guaranteed that doing so "
729                        "would not lead to an abrupt change of visible data",
730                        to_replace->node_name, target_bs->node_name);
731         }
732         bdrv_drained_end(target_bs);
733         if (local_err) {
734             error_report_err(local_err);
735             ret = -EPERM;
736         }
737     }
738     if (s->to_replace) {
739         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
740         error_free(s->replace_blocker);
741         bdrv_unref(s->to_replace);
742     }
743     if (replace_aio_context) {
744         aio_context_release(replace_aio_context);
745     }
746     g_free(s->replaces);
747     bdrv_unref(target_bs);
748 
749     /*
750      * Remove the mirror filter driver from the graph. Before this, get rid of
751      * the blockers on the intermediate nodes so that the resulting state is
752      * valid.
753      */
754     block_job_remove_all_bdrv(bjob);
755     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
756 
757     /* We just changed the BDS the job BB refers to (with either or both of the
758      * bdrv_replace_node() calls), so switch the BB back so the cleanup does
759      * the right thing. We don't need any permissions any more now. */
760     blk_remove_bs(bjob->blk);
761     blk_set_perm(bjob->blk, 0, BLK_PERM_ALL, &error_abort);
762     blk_insert_bs(bjob->blk, mirror_top_bs, &error_abort);
763 
764     bs_opaque->job = NULL;
765 
766     bdrv_drained_end(src);
767     bdrv_drained_end(mirror_top_bs);
768     s->in_drain = false;
769     bdrv_unref(mirror_top_bs);
770     bdrv_unref(src);
771 
772     return ret;
773 }
774 
775 static int mirror_prepare(Job *job)
776 {
777     return mirror_exit_common(job);
778 }
779 
780 static void mirror_abort(Job *job)
781 {
782     int ret = mirror_exit_common(job);
783     assert(ret == 0);
784 }
785 
786 static void coroutine_fn mirror_throttle(MirrorBlockJob *s)
787 {
788     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
789 
790     if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) {
791         s->last_pause_ns = now;
792         job_sleep_ns(&s->common.job, 0);
793     } else {
794         job_pause_point(&s->common.job);
795     }
796 }
797 
798 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
799 {
800     int64_t offset;
801     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
802     BlockDriverState *target_bs = blk_bs(s->target);
803     int ret;
804     int64_t count;
805 
806     if (s->zero_target) {
807         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
808             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
809             return 0;
810         }
811 
812         s->initial_zeroing_ongoing = true;
813         for (offset = 0; offset < s->bdev_length; ) {
814             int bytes = MIN(s->bdev_length - offset,
815                             QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
816 
817             mirror_throttle(s);
818 
819             if (job_is_cancelled(&s->common.job)) {
820                 s->initial_zeroing_ongoing = false;
821                 return 0;
822             }
823 
824             if (s->in_flight >= MAX_IN_FLIGHT) {
825                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
826                                    s->in_flight);
827                 mirror_wait_for_free_in_flight_slot(s);
828                 continue;
829             }
830 
831             mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO);
832             offset += bytes;
833         }
834 
835         mirror_wait_for_all_io(s);
836         s->initial_zeroing_ongoing = false;
837     }
838 
839     /* First part, loop on the sectors and initialize the dirty bitmap.  */
840     for (offset = 0; offset < s->bdev_length; ) {
841         /* Just to make sure we are not exceeding int limit. */
842         int bytes = MIN(s->bdev_length - offset,
843                         QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
844 
845         mirror_throttle(s);
846 
847         if (job_is_cancelled(&s->common.job)) {
848             return 0;
849         }
850 
851         ret = bdrv_is_allocated_above(bs, s->base_overlay, true, offset, bytes,
852                                       &count);
853         if (ret < 0) {
854             return ret;
855         }
856 
857         assert(count);
858         if (ret > 0) {
859             bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
860         }
861         offset += count;
862     }
863     return 0;
864 }
865 
866 /* Called when going out of the streaming phase to flush the bulk of the
867  * data to the medium, or just before completing.
868  */
869 static int mirror_flush(MirrorBlockJob *s)
870 {
871     int ret = blk_flush(s->target);
872     if (ret < 0) {
873         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
874             s->ret = ret;
875         }
876     }
877     return ret;
878 }
879 
880 static int coroutine_fn mirror_run(Job *job, Error **errp)
881 {
882     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
883     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
884     BlockDriverState *target_bs = blk_bs(s->target);
885     bool need_drain = true;
886     int64_t length;
887     int64_t target_length;
888     BlockDriverInfo bdi;
889     char backing_filename[2]; /* we only need 2 characters because we are only
890                                  checking for a NULL string */
891     int ret = 0;
892 
893     if (job_is_cancelled(&s->common.job)) {
894         goto immediate_exit;
895     }
896 
897     s->bdev_length = bdrv_getlength(bs);
898     if (s->bdev_length < 0) {
899         ret = s->bdev_length;
900         goto immediate_exit;
901     }
902 
903     target_length = blk_getlength(s->target);
904     if (target_length < 0) {
905         ret = target_length;
906         goto immediate_exit;
907     }
908 
909     /* Active commit must resize the base image if its size differs from the
910      * active layer. */
911     if (s->base == blk_bs(s->target)) {
912         if (s->bdev_length > target_length) {
913             ret = blk_truncate(s->target, s->bdev_length, false,
914                                PREALLOC_MODE_OFF, 0, NULL);
915             if (ret < 0) {
916                 goto immediate_exit;
917             }
918         }
919     } else if (s->bdev_length != target_length) {
920         error_setg(errp, "Source and target image have different sizes");
921         ret = -EINVAL;
922         goto immediate_exit;
923     }
924 
925     if (s->bdev_length == 0) {
926         /* Transition to the READY state and wait for complete. */
927         job_transition_to_ready(&s->common.job);
928         s->synced = true;
929         s->actively_synced = true;
930         while (!job_is_cancelled(&s->common.job) && !s->should_complete) {
931             job_yield(&s->common.job);
932         }
933         s->common.job.cancelled = false;
934         goto immediate_exit;
935     }
936 
937     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
938     s->in_flight_bitmap = bitmap_new(length);
939 
940     /* If we have no backing file yet in the destination, we cannot let
941      * the destination do COW.  Instead, we copy sectors around the
942      * dirty data if needed.  We need a bitmap to do that.
943      */
944     bdrv_get_backing_filename(target_bs, backing_filename,
945                               sizeof(backing_filename));
946     if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) {
947         s->target_cluster_size = bdi.cluster_size;
948     } else {
949         s->target_cluster_size = BDRV_SECTOR_SIZE;
950     }
951     if (backing_filename[0] && !bdrv_backing_chain_next(target_bs) &&
952         s->granularity < s->target_cluster_size) {
953         s->buf_size = MAX(s->buf_size, s->target_cluster_size);
954         s->cow_bitmap = bitmap_new(length);
955     }
956     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
957 
958     s->buf = qemu_try_blockalign(bs, s->buf_size);
959     if (s->buf == NULL) {
960         ret = -ENOMEM;
961         goto immediate_exit;
962     }
963 
964     mirror_free_init(s);
965 
966     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
967     if (!s->is_none_mode) {
968         ret = mirror_dirty_init(s);
969         if (ret < 0 || job_is_cancelled(&s->common.job)) {
970             goto immediate_exit;
971         }
972     }
973 
974     assert(!s->dbi);
975     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
976     for (;;) {
977         uint64_t delay_ns = 0;
978         int64_t cnt, delta;
979         bool should_complete;
980 
981         /* Do not start passive operations while there are active
982          * writes in progress */
983         while (s->in_active_write_counter) {
984             mirror_wait_for_any_operation(s, true);
985         }
986 
987         if (s->ret < 0) {
988             ret = s->ret;
989             goto immediate_exit;
990         }
991 
992         job_pause_point(&s->common.job);
993 
994         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
995         /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
996          * the number of bytes currently being processed; together those are
997          * the current remaining operation length */
998         job_progress_set_remaining(&s->common.job, s->bytes_in_flight + cnt);
999 
1000         /* Note that even when no rate limit is applied we need to yield
1001          * periodically with no pending I/O so that bdrv_drain_all() returns.
1002          * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
1003          * an error, or when the source is clean, whichever comes first. */
1004         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
1005         if (delta < BLOCK_JOB_SLICE_TIME &&
1006             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
1007             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
1008                 (cnt == 0 && s->in_flight > 0)) {
1009                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
1010                 mirror_wait_for_free_in_flight_slot(s);
1011                 continue;
1012             } else if (cnt != 0) {
1013                 delay_ns = mirror_iteration(s);
1014             }
1015         }
1016 
1017         should_complete = false;
1018         if (s->in_flight == 0 && cnt == 0) {
1019             trace_mirror_before_flush(s);
1020             if (!s->synced) {
1021                 if (mirror_flush(s) < 0) {
1022                     /* Go check s->ret.  */
1023                     continue;
1024                 }
1025                 /* We're out of the streaming phase.  From now on, if the job
1026                  * is cancelled we will actually complete all pending I/O and
1027                  * report completion.  This way, block-job-cancel will leave
1028                  * the target in a consistent state.
1029                  */
1030                 job_transition_to_ready(&s->common.job);
1031                 s->synced = true;
1032                 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) {
1033                     s->actively_synced = true;
1034                 }
1035             }
1036 
1037             should_complete = s->should_complete ||
1038                 job_is_cancelled(&s->common.job);
1039             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1040         }
1041 
1042         if (cnt == 0 && should_complete) {
1043             /* The dirty bitmap is not updated while operations are pending.
1044              * If we're about to exit, wait for pending operations before
1045              * calling bdrv_get_dirty_count(bs), or we may exit while the
1046              * source has dirty data to copy!
1047              *
1048              * Note that I/O can be submitted by the guest while
1049              * mirror_populate runs, so pause it now.  Before deciding
1050              * whether to switch to target check one last time if I/O has
1051              * come in the meanwhile, and if not flush the data to disk.
1052              */
1053             trace_mirror_before_drain(s, cnt);
1054 
1055             s->in_drain = true;
1056             bdrv_drained_begin(bs);
1057             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1058             if (cnt > 0 || mirror_flush(s) < 0) {
1059                 bdrv_drained_end(bs);
1060                 s->in_drain = false;
1061                 continue;
1062             }
1063 
1064             /* The two disks are in sync.  Exit and report successful
1065              * completion.
1066              */
1067             assert(QLIST_EMPTY(&bs->tracked_requests));
1068             s->common.job.cancelled = false;
1069             need_drain = false;
1070             break;
1071         }
1072 
1073         ret = 0;
1074 
1075         if (s->synced && !should_complete) {
1076             delay_ns = (s->in_flight == 0 &&
1077                         cnt == 0 ? BLOCK_JOB_SLICE_TIME : 0);
1078         }
1079         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
1080         job_sleep_ns(&s->common.job, delay_ns);
1081         if (job_is_cancelled(&s->common.job) &&
1082             (!s->synced || s->common.job.force_cancel))
1083         {
1084             break;
1085         }
1086         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1087     }
1088 
1089 immediate_exit:
1090     if (s->in_flight > 0) {
1091         /* We get here only if something went wrong.  Either the job failed,
1092          * or it was cancelled prematurely so that we do not guarantee that
1093          * the target is a copy of the source.
1094          */
1095         assert(ret < 0 || ((s->common.job.force_cancel || !s->synced) &&
1096                job_is_cancelled(&s->common.job)));
1097         assert(need_drain);
1098         mirror_wait_for_all_io(s);
1099     }
1100 
1101     assert(s->in_flight == 0);
1102     qemu_vfree(s->buf);
1103     g_free(s->cow_bitmap);
1104     g_free(s->in_flight_bitmap);
1105     bdrv_dirty_iter_free(s->dbi);
1106 
1107     if (need_drain) {
1108         s->in_drain = true;
1109         bdrv_drained_begin(bs);
1110     }
1111 
1112     return ret;
1113 }
1114 
1115 static void mirror_complete(Job *job, Error **errp)
1116 {
1117     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1118 
1119     if (!s->synced) {
1120         error_setg(errp, "The active block job '%s' cannot be completed",
1121                    job->id);
1122         return;
1123     }
1124 
1125     /* block all operations on to_replace bs */
1126     if (s->replaces) {
1127         AioContext *replace_aio_context;
1128 
1129         s->to_replace = bdrv_find_node(s->replaces);
1130         if (!s->to_replace) {
1131             error_setg(errp, "Node name '%s' not found", s->replaces);
1132             return;
1133         }
1134 
1135         replace_aio_context = bdrv_get_aio_context(s->to_replace);
1136         aio_context_acquire(replace_aio_context);
1137 
1138         /* TODO Translate this into permission system. Current definition of
1139          * GRAPH_MOD would require to request it for the parents; they might
1140          * not even be BlockDriverStates, however, so a BdrvChild can't address
1141          * them. May need redefinition of GRAPH_MOD. */
1142         error_setg(&s->replace_blocker,
1143                    "block device is in use by block-job-complete");
1144         bdrv_op_block_all(s->to_replace, s->replace_blocker);
1145         bdrv_ref(s->to_replace);
1146 
1147         aio_context_release(replace_aio_context);
1148     }
1149 
1150     s->should_complete = true;
1151 
1152     /* If the job is paused, it will be re-entered when it is resumed */
1153     if (!job->paused) {
1154         job_enter(job);
1155     }
1156 }
1157 
1158 static void coroutine_fn mirror_pause(Job *job)
1159 {
1160     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1161 
1162     mirror_wait_for_all_io(s);
1163 }
1164 
1165 static bool mirror_drained_poll(BlockJob *job)
1166 {
1167     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1168 
1169     /* If the job isn't paused nor cancelled, we can't be sure that it won't
1170      * issue more requests. We make an exception if we've reached this point
1171      * from one of our own drain sections, to avoid a deadlock waiting for
1172      * ourselves.
1173      */
1174     if (!s->common.job.paused && !s->common.job.cancelled && !s->in_drain) {
1175         return true;
1176     }
1177 
1178     return !!s->in_flight;
1179 }
1180 
1181 static void mirror_cancel(Job *job, bool force)
1182 {
1183     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1184     BlockDriverState *target = blk_bs(s->target);
1185 
1186     if (force || !job_is_ready(job)) {
1187         bdrv_cancel_in_flight(target);
1188     }
1189 }
1190 
1191 static const BlockJobDriver mirror_job_driver = {
1192     .job_driver = {
1193         .instance_size          = sizeof(MirrorBlockJob),
1194         .job_type               = JOB_TYPE_MIRROR,
1195         .free                   = block_job_free,
1196         .user_resume            = block_job_user_resume,
1197         .run                    = mirror_run,
1198         .prepare                = mirror_prepare,
1199         .abort                  = mirror_abort,
1200         .pause                  = mirror_pause,
1201         .complete               = mirror_complete,
1202         .cancel                 = mirror_cancel,
1203     },
1204     .drained_poll           = mirror_drained_poll,
1205 };
1206 
1207 static const BlockJobDriver commit_active_job_driver = {
1208     .job_driver = {
1209         .instance_size          = sizeof(MirrorBlockJob),
1210         .job_type               = JOB_TYPE_COMMIT,
1211         .free                   = block_job_free,
1212         .user_resume            = block_job_user_resume,
1213         .run                    = mirror_run,
1214         .prepare                = mirror_prepare,
1215         .abort                  = mirror_abort,
1216         .pause                  = mirror_pause,
1217         .complete               = mirror_complete,
1218     },
1219     .drained_poll           = mirror_drained_poll,
1220 };
1221 
1222 static void coroutine_fn
1223 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method,
1224                      uint64_t offset, uint64_t bytes,
1225                      QEMUIOVector *qiov, int flags)
1226 {
1227     int ret;
1228     size_t qiov_offset = 0;
1229     int64_t bitmap_offset, bitmap_end;
1230 
1231     if (!QEMU_IS_ALIGNED(offset, job->granularity) &&
1232         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset))
1233     {
1234             /*
1235              * Dirty unaligned padding: ignore it.
1236              *
1237              * Reasoning:
1238              * 1. If we copy it, we can't reset corresponding bit in
1239              *    dirty_bitmap as there may be some "dirty" bytes still not
1240              *    copied.
1241              * 2. It's already dirty, so skipping it we don't diverge mirror
1242              *    progress.
1243              *
1244              * Note, that because of this, guest write may have no contribution
1245              * into mirror converge, but that's not bad, as we have background
1246              * process of mirroring. If under some bad circumstances (high guest
1247              * IO load) background process starve, we will not converge anyway,
1248              * even if each write will contribute, as guest is not guaranteed to
1249              * rewrite the whole disk.
1250              */
1251             qiov_offset = QEMU_ALIGN_UP(offset, job->granularity) - offset;
1252             if (bytes <= qiov_offset) {
1253                 /* nothing to do after shrink */
1254                 return;
1255             }
1256             offset += qiov_offset;
1257             bytes -= qiov_offset;
1258     }
1259 
1260     if (!QEMU_IS_ALIGNED(offset + bytes, job->granularity) &&
1261         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset + bytes - 1))
1262     {
1263         uint64_t tail = (offset + bytes) % job->granularity;
1264 
1265         if (bytes <= tail) {
1266             /* nothing to do after shrink */
1267             return;
1268         }
1269         bytes -= tail;
1270     }
1271 
1272     /*
1273      * Tails are either clean or shrunk, so for bitmap resetting
1274      * we safely align the range down.
1275      */
1276     bitmap_offset = QEMU_ALIGN_UP(offset, job->granularity);
1277     bitmap_end = QEMU_ALIGN_DOWN(offset + bytes, job->granularity);
1278     if (bitmap_offset < bitmap_end) {
1279         bdrv_reset_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1280                                 bitmap_end - bitmap_offset);
1281     }
1282 
1283     job_progress_increase_remaining(&job->common.job, bytes);
1284 
1285     switch (method) {
1286     case MIRROR_METHOD_COPY:
1287         ret = blk_co_pwritev_part(job->target, offset, bytes,
1288                                   qiov, qiov_offset, flags);
1289         break;
1290 
1291     case MIRROR_METHOD_ZERO:
1292         assert(!qiov);
1293         ret = blk_co_pwrite_zeroes(job->target, offset, bytes, flags);
1294         break;
1295 
1296     case MIRROR_METHOD_DISCARD:
1297         assert(!qiov);
1298         ret = blk_co_pdiscard(job->target, offset, bytes);
1299         break;
1300 
1301     default:
1302         abort();
1303     }
1304 
1305     if (ret >= 0) {
1306         job_progress_update(&job->common.job, bytes);
1307     } else {
1308         BlockErrorAction action;
1309 
1310         /*
1311          * We failed, so we should mark dirty the whole area, aligned up.
1312          * Note that we don't care about shrunk tails if any: they were dirty
1313          * at function start, and they must be still dirty, as we've locked
1314          * the region for in-flight op.
1315          */
1316         bitmap_offset = QEMU_ALIGN_DOWN(offset, job->granularity);
1317         bitmap_end = QEMU_ALIGN_UP(offset + bytes, job->granularity);
1318         bdrv_set_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1319                               bitmap_end - bitmap_offset);
1320         job->actively_synced = false;
1321 
1322         action = mirror_error_action(job, false, -ret);
1323         if (action == BLOCK_ERROR_ACTION_REPORT) {
1324             if (!job->ret) {
1325                 job->ret = ret;
1326             }
1327         }
1328     }
1329 }
1330 
1331 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s,
1332                                                    uint64_t offset,
1333                                                    uint64_t bytes)
1334 {
1335     MirrorOp *op;
1336     uint64_t start_chunk = offset / s->granularity;
1337     uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
1338 
1339     op = g_new(MirrorOp, 1);
1340     *op = (MirrorOp){
1341         .s                  = s,
1342         .offset             = offset,
1343         .bytes              = bytes,
1344         .is_active_write    = true,
1345         .is_in_flight       = true,
1346     };
1347     qemu_co_queue_init(&op->waiting_requests);
1348     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
1349 
1350     s->in_active_write_counter++;
1351 
1352     mirror_wait_on_conflicts(op, s, offset, bytes);
1353 
1354     bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1355 
1356     return op;
1357 }
1358 
1359 static void coroutine_fn active_write_settle(MirrorOp *op)
1360 {
1361     uint64_t start_chunk = op->offset / op->s->granularity;
1362     uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes,
1363                                       op->s->granularity);
1364 
1365     if (!--op->s->in_active_write_counter && op->s->actively_synced) {
1366         BdrvChild *source = op->s->mirror_top_bs->backing;
1367 
1368         if (QLIST_FIRST(&source->bs->parents) == source &&
1369             QLIST_NEXT(source, next_parent) == NULL)
1370         {
1371             /* Assert that we are back in sync once all active write
1372              * operations are settled.
1373              * Note that we can only assert this if the mirror node
1374              * is the source node's only parent. */
1375             assert(!bdrv_get_dirty_count(op->s->dirty_bitmap));
1376         }
1377     }
1378     bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1379     QTAILQ_REMOVE(&op->s->ops_in_flight, op, next);
1380     qemu_co_queue_restart_all(&op->waiting_requests);
1381     g_free(op);
1382 }
1383 
1384 static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs,
1385     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1386 {
1387     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1388 }
1389 
1390 static int coroutine_fn bdrv_mirror_top_do_write(BlockDriverState *bs,
1391     MirrorMethod method, uint64_t offset, uint64_t bytes, QEMUIOVector *qiov,
1392     int flags)
1393 {
1394     MirrorOp *op = NULL;
1395     MirrorBDSOpaque *s = bs->opaque;
1396     int ret = 0;
1397     bool copy_to_target;
1398 
1399     copy_to_target = s->job->ret >= 0 &&
1400                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1401 
1402     if (copy_to_target) {
1403         op = active_write_prepare(s->job, offset, bytes);
1404     }
1405 
1406     switch (method) {
1407     case MIRROR_METHOD_COPY:
1408         ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1409         break;
1410 
1411     case MIRROR_METHOD_ZERO:
1412         ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1413         break;
1414 
1415     case MIRROR_METHOD_DISCARD:
1416         ret = bdrv_co_pdiscard(bs->backing, offset, bytes);
1417         break;
1418 
1419     default:
1420         abort();
1421     }
1422 
1423     if (ret < 0) {
1424         goto out;
1425     }
1426 
1427     if (copy_to_target) {
1428         do_sync_target_write(s->job, method, offset, bytes, qiov, flags);
1429     }
1430 
1431 out:
1432     if (copy_to_target) {
1433         active_write_settle(op);
1434     }
1435     return ret;
1436 }
1437 
1438 static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs,
1439     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1440 {
1441     MirrorBDSOpaque *s = bs->opaque;
1442     QEMUIOVector bounce_qiov;
1443     void *bounce_buf;
1444     int ret = 0;
1445     bool copy_to_target;
1446 
1447     copy_to_target = s->job->ret >= 0 &&
1448                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1449 
1450     if (copy_to_target) {
1451         /* The guest might concurrently modify the data to write; but
1452          * the data on source and destination must match, so we have
1453          * to use a bounce buffer if we are going to write to the
1454          * target now. */
1455         bounce_buf = qemu_blockalign(bs, bytes);
1456         iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes);
1457 
1458         qemu_iovec_init(&bounce_qiov, 1);
1459         qemu_iovec_add(&bounce_qiov, bounce_buf, bytes);
1460         qiov = &bounce_qiov;
1461     }
1462 
1463     ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov,
1464                                    flags);
1465 
1466     if (copy_to_target) {
1467         qemu_iovec_destroy(&bounce_qiov);
1468         qemu_vfree(bounce_buf);
1469     }
1470 
1471     return ret;
1472 }
1473 
1474 static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs)
1475 {
1476     if (bs->backing == NULL) {
1477         /* we can be here after failed bdrv_append in mirror_start_job */
1478         return 0;
1479     }
1480     return bdrv_co_flush(bs->backing->bs);
1481 }
1482 
1483 static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs,
1484     int64_t offset, int bytes, BdrvRequestFlags flags)
1485 {
1486     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL,
1487                                     flags);
1488 }
1489 
1490 static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs,
1491     int64_t offset, int bytes)
1492 {
1493     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes,
1494                                     NULL, 0);
1495 }
1496 
1497 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs)
1498 {
1499     if (bs->backing == NULL) {
1500         /* we can be here after failed bdrv_attach_child in
1501          * bdrv_set_backing_hd */
1502         return;
1503     }
1504     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1505             bs->backing->bs->filename);
1506 }
1507 
1508 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1509                                        BdrvChildRole role,
1510                                        BlockReopenQueue *reopen_queue,
1511                                        uint64_t perm, uint64_t shared,
1512                                        uint64_t *nperm, uint64_t *nshared)
1513 {
1514     MirrorBDSOpaque *s = bs->opaque;
1515 
1516     if (s->stop) {
1517         /*
1518          * If the job is to be stopped, we do not need to forward
1519          * anything to the real image.
1520          */
1521         *nperm = 0;
1522         *nshared = BLK_PERM_ALL;
1523         return;
1524     }
1525 
1526     bdrv_default_perms(bs, c, role, reopen_queue,
1527                        perm, shared, nperm, nshared);
1528 
1529     if (s->is_commit) {
1530         /*
1531          * For commit jobs, we cannot take CONSISTENT_READ, because
1532          * that permission is unshared for everything above the base
1533          * node (except for filters on the base node).
1534          * We also have to force-share the WRITE permission, or
1535          * otherwise we would block ourselves at the base node (if
1536          * writes are blocked for a node, they are also blocked for
1537          * its backing file).
1538          * (We could also share RESIZE, because it may be needed for
1539          * the target if its size is less than the top node's; but
1540          * bdrv_default_perms_for_cow() automatically shares RESIZE
1541          * for backing nodes if WRITE is shared, so there is no need
1542          * to do it here.)
1543          */
1544         *nperm &= ~BLK_PERM_CONSISTENT_READ;
1545         *nshared |= BLK_PERM_WRITE;
1546     }
1547 }
1548 
1549 /* Dummy node that provides consistent read to its users without requiring it
1550  * from its backing file and that allows writes on the backing file chain. */
1551 static BlockDriver bdrv_mirror_top = {
1552     .format_name                = "mirror_top",
1553     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1554     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1555     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1556     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1557     .bdrv_co_flush              = bdrv_mirror_top_flush,
1558     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1559     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1560 
1561     .is_filter                  = true,
1562 };
1563 
1564 static BlockJob *mirror_start_job(
1565                              const char *job_id, BlockDriverState *bs,
1566                              int creation_flags, BlockDriverState *target,
1567                              const char *replaces, int64_t speed,
1568                              uint32_t granularity, int64_t buf_size,
1569                              BlockMirrorBackingMode backing_mode,
1570                              bool zero_target,
1571                              BlockdevOnError on_source_error,
1572                              BlockdevOnError on_target_error,
1573                              bool unmap,
1574                              BlockCompletionFunc *cb,
1575                              void *opaque,
1576                              const BlockJobDriver *driver,
1577                              bool is_none_mode, BlockDriverState *base,
1578                              bool auto_complete, const char *filter_node_name,
1579                              bool is_mirror, MirrorCopyMode copy_mode,
1580                              Error **errp)
1581 {
1582     MirrorBlockJob *s;
1583     MirrorBDSOpaque *bs_opaque;
1584     BlockDriverState *mirror_top_bs;
1585     bool target_is_backing;
1586     uint64_t target_perms, target_shared_perms;
1587     int ret;
1588 
1589     if (granularity == 0) {
1590         granularity = bdrv_get_default_bitmap_granularity(target);
1591     }
1592 
1593     assert(is_power_of_2(granularity));
1594 
1595     if (buf_size < 0) {
1596         error_setg(errp, "Invalid parameter 'buf-size'");
1597         return NULL;
1598     }
1599 
1600     if (buf_size == 0) {
1601         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1602     }
1603 
1604     if (bdrv_skip_filters(bs) == bdrv_skip_filters(target)) {
1605         error_setg(errp, "Can't mirror node into itself");
1606         return NULL;
1607     }
1608 
1609     target_is_backing = bdrv_chain_contains(bs, target);
1610 
1611     /* In the case of active commit, add dummy driver to provide consistent
1612      * reads on the top, while disabling it in the intermediate nodes, and make
1613      * the backing chain writable. */
1614     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1615                                          BDRV_O_RDWR, errp);
1616     if (mirror_top_bs == NULL) {
1617         return NULL;
1618     }
1619     if (!filter_node_name) {
1620         mirror_top_bs->implicit = true;
1621     }
1622 
1623     /* So that we can always drop this node */
1624     mirror_top_bs->never_freeze = true;
1625 
1626     mirror_top_bs->total_sectors = bs->total_sectors;
1627     mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
1628     mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
1629                                           BDRV_REQ_NO_FALLBACK;
1630     bs_opaque = g_new0(MirrorBDSOpaque, 1);
1631     mirror_top_bs->opaque = bs_opaque;
1632 
1633     bs_opaque->is_commit = target_is_backing;
1634 
1635     bdrv_drained_begin(bs);
1636     ret = bdrv_append(mirror_top_bs, bs, errp);
1637     bdrv_drained_end(bs);
1638 
1639     if (ret < 0) {
1640         bdrv_unref(mirror_top_bs);
1641         return NULL;
1642     }
1643 
1644     /* Make sure that the source is not resized while the job is running */
1645     s = block_job_create(job_id, driver, NULL, mirror_top_bs,
1646                          BLK_PERM_CONSISTENT_READ,
1647                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1648                          BLK_PERM_WRITE | BLK_PERM_GRAPH_MOD, speed,
1649                          creation_flags, cb, opaque, errp);
1650     if (!s) {
1651         goto fail;
1652     }
1653     bs_opaque->job = s;
1654 
1655     /* The block job now has a reference to this node */
1656     bdrv_unref(mirror_top_bs);
1657 
1658     s->mirror_top_bs = mirror_top_bs;
1659 
1660     /* No resize for the target either; while the mirror is still running, a
1661      * consistent read isn't necessarily possible. We could possibly allow
1662      * writes and graph modifications, though it would likely defeat the
1663      * purpose of a mirror, so leave them blocked for now.
1664      *
1665      * In the case of active commit, things look a bit different, though,
1666      * because the target is an already populated backing file in active use.
1667      * We can allow anything except resize there.*/
1668 
1669     target_perms = BLK_PERM_WRITE;
1670     target_shared_perms = BLK_PERM_WRITE_UNCHANGED;
1671 
1672     if (target_is_backing) {
1673         int64_t bs_size, target_size;
1674         bs_size = bdrv_getlength(bs);
1675         if (bs_size < 0) {
1676             error_setg_errno(errp, -bs_size,
1677                              "Could not inquire top image size");
1678             goto fail;
1679         }
1680 
1681         target_size = bdrv_getlength(target);
1682         if (target_size < 0) {
1683             error_setg_errno(errp, -target_size,
1684                              "Could not inquire base image size");
1685             goto fail;
1686         }
1687 
1688         if (target_size < bs_size) {
1689             target_perms |= BLK_PERM_RESIZE;
1690         }
1691 
1692         target_shared_perms |= BLK_PERM_CONSISTENT_READ
1693                             |  BLK_PERM_WRITE
1694                             |  BLK_PERM_GRAPH_MOD;
1695     } else if (bdrv_chain_contains(bs, bdrv_skip_filters(target))) {
1696         /*
1697          * We may want to allow this in the future, but it would
1698          * require taking some extra care.
1699          */
1700         error_setg(errp, "Cannot mirror to a filter on top of a node in the "
1701                    "source's backing chain");
1702         goto fail;
1703     }
1704 
1705     if (backing_mode != MIRROR_LEAVE_BACKING_CHAIN) {
1706         target_perms |= BLK_PERM_GRAPH_MOD;
1707     }
1708 
1709     s->target = blk_new(s->common.job.aio_context,
1710                         target_perms, target_shared_perms);
1711     ret = blk_insert_bs(s->target, target, errp);
1712     if (ret < 0) {
1713         goto fail;
1714     }
1715     if (is_mirror) {
1716         /* XXX: Mirror target could be a NBD server of target QEMU in the case
1717          * of non-shared block migration. To allow migration completion, we
1718          * have to allow "inactivate" of the target BB.  When that happens, we
1719          * know the job is drained, and the vcpus are stopped, so no write
1720          * operation will be performed. Block layer already has assertions to
1721          * ensure that. */
1722         blk_set_force_allow_inactivate(s->target);
1723     }
1724     blk_set_allow_aio_context_change(s->target, true);
1725     blk_set_disable_request_queuing(s->target, true);
1726 
1727     s->replaces = g_strdup(replaces);
1728     s->on_source_error = on_source_error;
1729     s->on_target_error = on_target_error;
1730     s->is_none_mode = is_none_mode;
1731     s->backing_mode = backing_mode;
1732     s->zero_target = zero_target;
1733     s->copy_mode = copy_mode;
1734     s->base = base;
1735     s->base_overlay = bdrv_find_overlay(bs, base);
1736     s->granularity = granularity;
1737     s->buf_size = ROUND_UP(buf_size, granularity);
1738     s->unmap = unmap;
1739     if (auto_complete) {
1740         s->should_complete = true;
1741     }
1742 
1743     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1744     if (!s->dirty_bitmap) {
1745         goto fail;
1746     }
1747     if (s->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING) {
1748         bdrv_disable_dirty_bitmap(s->dirty_bitmap);
1749     }
1750 
1751     ret = block_job_add_bdrv(&s->common, "source", bs, 0,
1752                              BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE |
1753                              BLK_PERM_CONSISTENT_READ,
1754                              errp);
1755     if (ret < 0) {
1756         goto fail;
1757     }
1758 
1759     /* Required permissions are already taken with blk_new() */
1760     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1761                        &error_abort);
1762 
1763     /* In commit_active_start() all intermediate nodes disappear, so
1764      * any jobs in them must be blocked */
1765     if (target_is_backing) {
1766         BlockDriverState *iter, *filtered_target;
1767         uint64_t iter_shared_perms;
1768 
1769         /*
1770          * The topmost node with
1771          * bdrv_skip_filters(filtered_target) == bdrv_skip_filters(target)
1772          */
1773         filtered_target = bdrv_cow_bs(bdrv_find_overlay(bs, target));
1774 
1775         assert(bdrv_skip_filters(filtered_target) ==
1776                bdrv_skip_filters(target));
1777 
1778         /*
1779          * XXX BLK_PERM_WRITE needs to be allowed so we don't block
1780          * ourselves at s->base (if writes are blocked for a node, they are
1781          * also blocked for its backing file). The other options would be a
1782          * second filter driver above s->base (== target).
1783          */
1784         iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE;
1785 
1786         for (iter = bdrv_filter_or_cow_bs(bs); iter != target;
1787              iter = bdrv_filter_or_cow_bs(iter))
1788         {
1789             if (iter == filtered_target) {
1790                 /*
1791                  * From here on, all nodes are filters on the base.
1792                  * This allows us to share BLK_PERM_CONSISTENT_READ.
1793                  */
1794                 iter_shared_perms |= BLK_PERM_CONSISTENT_READ;
1795             }
1796 
1797             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1798                                      iter_shared_perms, errp);
1799             if (ret < 0) {
1800                 goto fail;
1801             }
1802         }
1803 
1804         if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) {
1805             goto fail;
1806         }
1807     }
1808 
1809     QTAILQ_INIT(&s->ops_in_flight);
1810 
1811     trace_mirror_start(bs, s, opaque);
1812     job_start(&s->common.job);
1813 
1814     return &s->common;
1815 
1816 fail:
1817     if (s) {
1818         /* Make sure this BDS does not go away until we have completed the graph
1819          * changes below */
1820         bdrv_ref(mirror_top_bs);
1821 
1822         g_free(s->replaces);
1823         blk_unref(s->target);
1824         bs_opaque->job = NULL;
1825         if (s->dirty_bitmap) {
1826             bdrv_release_dirty_bitmap(s->dirty_bitmap);
1827         }
1828         job_early_fail(&s->common.job);
1829     }
1830 
1831     bs_opaque->stop = true;
1832     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
1833                              &error_abort);
1834     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
1835 
1836     bdrv_unref(mirror_top_bs);
1837 
1838     return NULL;
1839 }
1840 
1841 void mirror_start(const char *job_id, BlockDriverState *bs,
1842                   BlockDriverState *target, const char *replaces,
1843                   int creation_flags, int64_t speed,
1844                   uint32_t granularity, int64_t buf_size,
1845                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1846                   bool zero_target,
1847                   BlockdevOnError on_source_error,
1848                   BlockdevOnError on_target_error,
1849                   bool unmap, const char *filter_node_name,
1850                   MirrorCopyMode copy_mode, Error **errp)
1851 {
1852     bool is_none_mode;
1853     BlockDriverState *base;
1854 
1855     if ((mode == MIRROR_SYNC_MODE_INCREMENTAL) ||
1856         (mode == MIRROR_SYNC_MODE_BITMAP)) {
1857         error_setg(errp, "Sync mode '%s' not supported",
1858                    MirrorSyncMode_str(mode));
1859         return;
1860     }
1861     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1862     base = mode == MIRROR_SYNC_MODE_TOP ? bdrv_backing_chain_next(bs) : NULL;
1863     mirror_start_job(job_id, bs, creation_flags, target, replaces,
1864                      speed, granularity, buf_size, backing_mode, zero_target,
1865                      on_source_error, on_target_error, unmap, NULL, NULL,
1866                      &mirror_job_driver, is_none_mode, base, false,
1867                      filter_node_name, true, copy_mode, errp);
1868 }
1869 
1870 BlockJob *commit_active_start(const char *job_id, BlockDriverState *bs,
1871                               BlockDriverState *base, int creation_flags,
1872                               int64_t speed, BlockdevOnError on_error,
1873                               const char *filter_node_name,
1874                               BlockCompletionFunc *cb, void *opaque,
1875                               bool auto_complete, Error **errp)
1876 {
1877     bool base_read_only;
1878     BlockJob *job;
1879 
1880     base_read_only = bdrv_is_read_only(base);
1881 
1882     if (base_read_only) {
1883         if (bdrv_reopen_set_read_only(base, false, errp) < 0) {
1884             return NULL;
1885         }
1886     }
1887 
1888     job = mirror_start_job(
1889                      job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1890                      MIRROR_LEAVE_BACKING_CHAIN, false,
1891                      on_error, on_error, true, cb, opaque,
1892                      &commit_active_job_driver, false, base, auto_complete,
1893                      filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND,
1894                      errp);
1895     if (!job) {
1896         goto error_restore_flags;
1897     }
1898 
1899     return job;
1900 
1901 error_restore_flags:
1902     /* ignore error and errp for bdrv_reopen, because we want to propagate
1903      * the original error */
1904     if (base_read_only) {
1905         bdrv_reopen_set_read_only(base, true, NULL);
1906     }
1907     return NULL;
1908 }
1909