xref: /openbmc/qemu/block/mirror.c (revision 745a4f5e)
1 /*
2  * Image mirroring
3  *
4  * Copyright Red Hat, Inc. 2012
5  *
6  * Authors:
7  *  Paolo Bonzini  <pbonzini@redhat.com>
8  *
9  * This work is licensed under the terms of the GNU LGPL, version 2 or later.
10  * See the COPYING.LIB file in the top-level directory.
11  *
12  */
13 
14 #include "qemu/osdep.h"
15 #include "qemu/cutils.h"
16 #include "qemu/coroutine.h"
17 #include "qemu/range.h"
18 #include "trace.h"
19 #include "block/blockjob_int.h"
20 #include "block/block_int.h"
21 #include "sysemu/block-backend.h"
22 #include "qapi/error.h"
23 #include "qapi/qmp/qerror.h"
24 #include "qemu/ratelimit.h"
25 #include "qemu/bitmap.h"
26 
27 #define MAX_IN_FLIGHT 16
28 #define MAX_IO_BYTES (1 << 20) /* 1 Mb */
29 #define DEFAULT_MIRROR_BUF_SIZE (MAX_IN_FLIGHT * MAX_IO_BYTES)
30 
31 /* The mirroring buffer is a list of granularity-sized chunks.
32  * Free chunks are organized in a list.
33  */
34 typedef struct MirrorBuffer {
35     QSIMPLEQ_ENTRY(MirrorBuffer) next;
36 } MirrorBuffer;
37 
38 typedef struct MirrorOp MirrorOp;
39 
40 typedef struct MirrorBlockJob {
41     BlockJob common;
42     BlockBackend *target;
43     BlockDriverState *mirror_top_bs;
44     BlockDriverState *base;
45 
46     /* The name of the graph node to replace */
47     char *replaces;
48     /* The BDS to replace */
49     BlockDriverState *to_replace;
50     /* Used to block operations on the drive-mirror-replace target */
51     Error *replace_blocker;
52     bool is_none_mode;
53     BlockMirrorBackingMode backing_mode;
54     MirrorCopyMode copy_mode;
55     BlockdevOnError on_source_error, on_target_error;
56     bool synced;
57     /* Set when the target is synced (dirty bitmap is clean, nothing
58      * in flight) and the job is running in active mode */
59     bool actively_synced;
60     bool should_complete;
61     int64_t granularity;
62     size_t buf_size;
63     int64_t bdev_length;
64     unsigned long *cow_bitmap;
65     BdrvDirtyBitmap *dirty_bitmap;
66     BdrvDirtyBitmapIter *dbi;
67     uint8_t *buf;
68     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
69     int buf_free_count;
70 
71     uint64_t last_pause_ns;
72     unsigned long *in_flight_bitmap;
73     int in_flight;
74     int64_t bytes_in_flight;
75     QTAILQ_HEAD(MirrorOpList, MirrorOp) ops_in_flight;
76     int ret;
77     bool unmap;
78     int target_cluster_size;
79     int max_iov;
80     bool initial_zeroing_ongoing;
81     int in_active_write_counter;
82     bool prepared;
83 } MirrorBlockJob;
84 
85 typedef struct MirrorBDSOpaque {
86     MirrorBlockJob *job;
87 } MirrorBDSOpaque;
88 
89 struct MirrorOp {
90     MirrorBlockJob *s;
91     QEMUIOVector qiov;
92     int64_t offset;
93     uint64_t bytes;
94 
95     /* The pointee is set by mirror_co_read(), mirror_co_zero(), and
96      * mirror_co_discard() before yielding for the first time */
97     int64_t *bytes_handled;
98 
99     bool is_pseudo_op;
100     bool is_active_write;
101     CoQueue waiting_requests;
102 
103     QTAILQ_ENTRY(MirrorOp) next;
104 };
105 
106 typedef enum MirrorMethod {
107     MIRROR_METHOD_COPY,
108     MIRROR_METHOD_ZERO,
109     MIRROR_METHOD_DISCARD,
110 } MirrorMethod;
111 
112 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
113                                             int error)
114 {
115     s->synced = false;
116     s->actively_synced = false;
117     if (read) {
118         return block_job_error_action(&s->common, s->on_source_error,
119                                       true, error);
120     } else {
121         return block_job_error_action(&s->common, s->on_target_error,
122                                       false, error);
123     }
124 }
125 
126 static void coroutine_fn mirror_wait_on_conflicts(MirrorOp *self,
127                                                   MirrorBlockJob *s,
128                                                   uint64_t offset,
129                                                   uint64_t bytes)
130 {
131     uint64_t self_start_chunk = offset / s->granularity;
132     uint64_t self_end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
133     uint64_t self_nb_chunks = self_end_chunk - self_start_chunk;
134 
135     while (find_next_bit(s->in_flight_bitmap, self_end_chunk,
136                          self_start_chunk) < self_end_chunk &&
137            s->ret >= 0)
138     {
139         MirrorOp *op;
140 
141         QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
142             uint64_t op_start_chunk = op->offset / s->granularity;
143             uint64_t op_nb_chunks = DIV_ROUND_UP(op->offset + op->bytes,
144                                                  s->granularity) -
145                                     op_start_chunk;
146 
147             if (op == self) {
148                 continue;
149             }
150 
151             if (ranges_overlap(self_start_chunk, self_nb_chunks,
152                                op_start_chunk, op_nb_chunks))
153             {
154                 qemu_co_queue_wait(&op->waiting_requests, NULL);
155                 break;
156             }
157         }
158     }
159 }
160 
161 static void coroutine_fn mirror_iteration_done(MirrorOp *op, int ret)
162 {
163     MirrorBlockJob *s = op->s;
164     struct iovec *iov;
165     int64_t chunk_num;
166     int i, nb_chunks;
167 
168     trace_mirror_iteration_done(s, op->offset, op->bytes, ret);
169 
170     s->in_flight--;
171     s->bytes_in_flight -= op->bytes;
172     iov = op->qiov.iov;
173     for (i = 0; i < op->qiov.niov; i++) {
174         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
175         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
176         s->buf_free_count++;
177     }
178 
179     chunk_num = op->offset / s->granularity;
180     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
181 
182     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
183     QTAILQ_REMOVE(&s->ops_in_flight, op, next);
184     if (ret >= 0) {
185         if (s->cow_bitmap) {
186             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
187         }
188         if (!s->initial_zeroing_ongoing) {
189             job_progress_update(&s->common.job, op->bytes);
190         }
191     }
192     qemu_iovec_destroy(&op->qiov);
193 
194     qemu_co_queue_restart_all(&op->waiting_requests);
195     g_free(op);
196 }
197 
198 static void coroutine_fn mirror_write_complete(MirrorOp *op, int ret)
199 {
200     MirrorBlockJob *s = op->s;
201 
202     if (ret < 0) {
203         BlockErrorAction action;
204 
205         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
206         action = mirror_error_action(s, false, -ret);
207         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
208             s->ret = ret;
209         }
210     }
211 
212     mirror_iteration_done(op, ret);
213 }
214 
215 static void coroutine_fn mirror_read_complete(MirrorOp *op, int ret)
216 {
217     MirrorBlockJob *s = op->s;
218 
219     if (ret < 0) {
220         BlockErrorAction action;
221 
222         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
223         action = mirror_error_action(s, true, -ret);
224         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
225             s->ret = ret;
226         }
227 
228         mirror_iteration_done(op, ret);
229         return;
230     }
231 
232     ret = blk_co_pwritev(s->target, op->offset, op->qiov.size, &op->qiov, 0);
233     mirror_write_complete(op, ret);
234 }
235 
236 /* Clip bytes relative to offset to not exceed end-of-file */
237 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s,
238                                         int64_t offset,
239                                         int64_t bytes)
240 {
241     return MIN(bytes, s->bdev_length - offset);
242 }
243 
244 /* Round offset and/or bytes to target cluster if COW is needed, and
245  * return the offset of the adjusted tail against original. */
246 static int mirror_cow_align(MirrorBlockJob *s, int64_t *offset,
247                             uint64_t *bytes)
248 {
249     bool need_cow;
250     int ret = 0;
251     int64_t align_offset = *offset;
252     int64_t align_bytes = *bytes;
253     int max_bytes = s->granularity * s->max_iov;
254 
255     need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap);
256     need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity,
257                           s->cow_bitmap);
258     if (need_cow) {
259         bdrv_round_to_clusters(blk_bs(s->target), *offset, *bytes,
260                                &align_offset, &align_bytes);
261     }
262 
263     if (align_bytes > max_bytes) {
264         align_bytes = max_bytes;
265         if (need_cow) {
266             align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size);
267         }
268     }
269     /* Clipping may result in align_bytes unaligned to chunk boundary, but
270      * that doesn't matter because it's already the end of source image. */
271     align_bytes = mirror_clip_bytes(s, align_offset, align_bytes);
272 
273     ret = align_offset + align_bytes - (*offset + *bytes);
274     *offset = align_offset;
275     *bytes = align_bytes;
276     assert(ret >= 0);
277     return ret;
278 }
279 
280 static inline void mirror_wait_for_any_operation(MirrorBlockJob *s, bool active)
281 {
282     MirrorOp *op;
283 
284     QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
285         /* Do not wait on pseudo ops, because it may in turn wait on
286          * some other operation to start, which may in fact be the
287          * caller of this function.  Since there is only one pseudo op
288          * at any given time, we will always find some real operation
289          * to wait on. */
290         if (!op->is_pseudo_op && op->is_active_write == active) {
291             qemu_co_queue_wait(&op->waiting_requests, NULL);
292             return;
293         }
294     }
295     abort();
296 }
297 
298 static inline void mirror_wait_for_free_in_flight_slot(MirrorBlockJob *s)
299 {
300     /* Only non-active operations use up in-flight slots */
301     mirror_wait_for_any_operation(s, false);
302 }
303 
304 /* Perform a mirror copy operation.
305  *
306  * *op->bytes_handled is set to the number of bytes copied after and
307  * including offset, excluding any bytes copied prior to offset due
308  * to alignment.  This will be op->bytes if no alignment is necessary,
309  * or (new_end - op->offset) if the tail is rounded up or down due to
310  * alignment or buffer limit.
311  */
312 static void coroutine_fn mirror_co_read(void *opaque)
313 {
314     MirrorOp *op = opaque;
315     MirrorBlockJob *s = op->s;
316     int nb_chunks;
317     uint64_t ret;
318     uint64_t max_bytes;
319 
320     max_bytes = s->granularity * s->max_iov;
321 
322     /* We can only handle as much as buf_size at a time. */
323     op->bytes = MIN(s->buf_size, MIN(max_bytes, op->bytes));
324     assert(op->bytes);
325     assert(op->bytes < BDRV_REQUEST_MAX_BYTES);
326     *op->bytes_handled = op->bytes;
327 
328     if (s->cow_bitmap) {
329         *op->bytes_handled += mirror_cow_align(s, &op->offset, &op->bytes);
330     }
331     /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */
332     assert(*op->bytes_handled <= UINT_MAX);
333     assert(op->bytes <= s->buf_size);
334     /* The offset is granularity-aligned because:
335      * 1) Caller passes in aligned values;
336      * 2) mirror_cow_align is used only when target cluster is larger. */
337     assert(QEMU_IS_ALIGNED(op->offset, s->granularity));
338     /* The range is sector-aligned, since bdrv_getlength() rounds up. */
339     assert(QEMU_IS_ALIGNED(op->bytes, BDRV_SECTOR_SIZE));
340     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
341 
342     while (s->buf_free_count < nb_chunks) {
343         trace_mirror_yield_in_flight(s, op->offset, s->in_flight);
344         mirror_wait_for_free_in_flight_slot(s);
345     }
346 
347     /* Now make a QEMUIOVector taking enough granularity-sized chunks
348      * from s->buf_free.
349      */
350     qemu_iovec_init(&op->qiov, nb_chunks);
351     while (nb_chunks-- > 0) {
352         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
353         size_t remaining = op->bytes - op->qiov.size;
354 
355         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
356         s->buf_free_count--;
357         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
358     }
359 
360     /* Copy the dirty cluster.  */
361     s->in_flight++;
362     s->bytes_in_flight += op->bytes;
363     trace_mirror_one_iteration(s, op->offset, op->bytes);
364 
365     ret = bdrv_co_preadv(s->mirror_top_bs->backing, op->offset, op->bytes,
366                          &op->qiov, 0);
367     mirror_read_complete(op, ret);
368 }
369 
370 static void coroutine_fn mirror_co_zero(void *opaque)
371 {
372     MirrorOp *op = opaque;
373     int ret;
374 
375     op->s->in_flight++;
376     op->s->bytes_in_flight += op->bytes;
377     *op->bytes_handled = op->bytes;
378 
379     ret = blk_co_pwrite_zeroes(op->s->target, op->offset, op->bytes,
380                                op->s->unmap ? BDRV_REQ_MAY_UNMAP : 0);
381     mirror_write_complete(op, ret);
382 }
383 
384 static void coroutine_fn mirror_co_discard(void *opaque)
385 {
386     MirrorOp *op = opaque;
387     int ret;
388 
389     op->s->in_flight++;
390     op->s->bytes_in_flight += op->bytes;
391     *op->bytes_handled = op->bytes;
392 
393     ret = blk_co_pdiscard(op->s->target, op->offset, op->bytes);
394     mirror_write_complete(op, ret);
395 }
396 
397 static unsigned mirror_perform(MirrorBlockJob *s, int64_t offset,
398                                unsigned bytes, MirrorMethod mirror_method)
399 {
400     MirrorOp *op;
401     Coroutine *co;
402     int64_t bytes_handled = -1;
403 
404     op = g_new(MirrorOp, 1);
405     *op = (MirrorOp){
406         .s              = s,
407         .offset         = offset,
408         .bytes          = bytes,
409         .bytes_handled  = &bytes_handled,
410     };
411     qemu_co_queue_init(&op->waiting_requests);
412 
413     switch (mirror_method) {
414     case MIRROR_METHOD_COPY:
415         co = qemu_coroutine_create(mirror_co_read, op);
416         break;
417     case MIRROR_METHOD_ZERO:
418         co = qemu_coroutine_create(mirror_co_zero, op);
419         break;
420     case MIRROR_METHOD_DISCARD:
421         co = qemu_coroutine_create(mirror_co_discard, op);
422         break;
423     default:
424         abort();
425     }
426 
427     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
428     qemu_coroutine_enter(co);
429     /* At this point, ownership of op has been moved to the coroutine
430      * and the object may already be freed */
431 
432     /* Assert that this value has been set */
433     assert(bytes_handled >= 0);
434 
435     /* Same assertion as in mirror_co_read() (and for mirror_co_read()
436      * and mirror_co_discard(), bytes_handled == op->bytes, which
437      * is the @bytes parameter given to this function) */
438     assert(bytes_handled <= UINT_MAX);
439     return bytes_handled;
440 }
441 
442 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
443 {
444     BlockDriverState *source = s->mirror_top_bs->backing->bs;
445     MirrorOp *pseudo_op;
446     int64_t offset;
447     uint64_t delay_ns = 0, ret = 0;
448     /* At least the first dirty chunk is mirrored in one iteration. */
449     int nb_chunks = 1;
450     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
451     int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES);
452 
453     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
454     offset = bdrv_dirty_iter_next(s->dbi);
455     if (offset < 0) {
456         bdrv_set_dirty_iter(s->dbi, 0);
457         offset = bdrv_dirty_iter_next(s->dbi);
458         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
459         assert(offset >= 0);
460     }
461     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
462 
463     mirror_wait_on_conflicts(NULL, s, offset, 1);
464 
465     job_pause_point(&s->common.job);
466 
467     /* Find the number of consective dirty chunks following the first dirty
468      * one, and wait for in flight requests in them. */
469     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
470     while (nb_chunks * s->granularity < s->buf_size) {
471         int64_t next_dirty;
472         int64_t next_offset = offset + nb_chunks * s->granularity;
473         int64_t next_chunk = next_offset / s->granularity;
474         if (next_offset >= s->bdev_length ||
475             !bdrv_get_dirty_locked(source, s->dirty_bitmap, next_offset)) {
476             break;
477         }
478         if (test_bit(next_chunk, s->in_flight_bitmap)) {
479             break;
480         }
481 
482         next_dirty = bdrv_dirty_iter_next(s->dbi);
483         if (next_dirty > next_offset || next_dirty < 0) {
484             /* The bitmap iterator's cache is stale, refresh it */
485             bdrv_set_dirty_iter(s->dbi, next_offset);
486             next_dirty = bdrv_dirty_iter_next(s->dbi);
487         }
488         assert(next_dirty == next_offset);
489         nb_chunks++;
490     }
491 
492     /* Clear dirty bits before querying the block status, because
493      * calling bdrv_block_status_above could yield - if some blocks are
494      * marked dirty in this window, we need to know.
495      */
496     bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset,
497                                    nb_chunks * s->granularity);
498     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
499 
500     /* Before claiming an area in the in-flight bitmap, we have to
501      * create a MirrorOp for it so that conflicting requests can wait
502      * for it.  mirror_perform() will create the real MirrorOps later,
503      * for now we just create a pseudo operation that will wake up all
504      * conflicting requests once all real operations have been
505      * launched. */
506     pseudo_op = g_new(MirrorOp, 1);
507     *pseudo_op = (MirrorOp){
508         .offset         = offset,
509         .bytes          = nb_chunks * s->granularity,
510         .is_pseudo_op   = true,
511     };
512     qemu_co_queue_init(&pseudo_op->waiting_requests);
513     QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next);
514 
515     bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks);
516     while (nb_chunks > 0 && offset < s->bdev_length) {
517         int ret;
518         int64_t io_bytes;
519         int64_t io_bytes_acct;
520         MirrorMethod mirror_method = MIRROR_METHOD_COPY;
521 
522         assert(!(offset % s->granularity));
523         ret = bdrv_block_status_above(source, NULL, offset,
524                                       nb_chunks * s->granularity,
525                                       &io_bytes, NULL, NULL);
526         if (ret < 0) {
527             io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes);
528         } else if (ret & BDRV_BLOCK_DATA) {
529             io_bytes = MIN(io_bytes, max_io_bytes);
530         }
531 
532         io_bytes -= io_bytes % s->granularity;
533         if (io_bytes < s->granularity) {
534             io_bytes = s->granularity;
535         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
536             int64_t target_offset;
537             int64_t target_bytes;
538             bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes,
539                                    &target_offset, &target_bytes);
540             if (target_offset == offset &&
541                 target_bytes == io_bytes) {
542                 mirror_method = ret & BDRV_BLOCK_ZERO ?
543                                     MIRROR_METHOD_ZERO :
544                                     MIRROR_METHOD_DISCARD;
545             }
546         }
547 
548         while (s->in_flight >= MAX_IN_FLIGHT) {
549             trace_mirror_yield_in_flight(s, offset, s->in_flight);
550             mirror_wait_for_free_in_flight_slot(s);
551         }
552 
553         if (s->ret < 0) {
554             ret = 0;
555             goto fail;
556         }
557 
558         io_bytes = mirror_clip_bytes(s, offset, io_bytes);
559         io_bytes = mirror_perform(s, offset, io_bytes, mirror_method);
560         if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) {
561             io_bytes_acct = 0;
562         } else {
563             io_bytes_acct = io_bytes;
564         }
565         assert(io_bytes);
566         offset += io_bytes;
567         nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity);
568         delay_ns = block_job_ratelimit_get_delay(&s->common, io_bytes_acct);
569     }
570 
571     ret = delay_ns;
572 fail:
573     QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next);
574     qemu_co_queue_restart_all(&pseudo_op->waiting_requests);
575     g_free(pseudo_op);
576 
577     return ret;
578 }
579 
580 static void mirror_free_init(MirrorBlockJob *s)
581 {
582     int granularity = s->granularity;
583     size_t buf_size = s->buf_size;
584     uint8_t *buf = s->buf;
585 
586     assert(s->buf_free_count == 0);
587     QSIMPLEQ_INIT(&s->buf_free);
588     while (buf_size != 0) {
589         MirrorBuffer *cur = (MirrorBuffer *)buf;
590         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
591         s->buf_free_count++;
592         buf_size -= granularity;
593         buf += granularity;
594     }
595 }
596 
597 /* This is also used for the .pause callback. There is no matching
598  * mirror_resume() because mirror_run() will begin iterating again
599  * when the job is resumed.
600  */
601 static void mirror_wait_for_all_io(MirrorBlockJob *s)
602 {
603     while (s->in_flight > 0) {
604         mirror_wait_for_free_in_flight_slot(s);
605     }
606 }
607 
608 /**
609  * mirror_exit_common: handle both abort() and prepare() cases.
610  * for .prepare, returns 0 on success and -errno on failure.
611  * for .abort cases, denoted by abort = true, MUST return 0.
612  */
613 static int mirror_exit_common(Job *job)
614 {
615     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
616     BlockJob *bjob = &s->common;
617     MirrorBDSOpaque *bs_opaque = s->mirror_top_bs->opaque;
618     AioContext *replace_aio_context = NULL;
619     BlockDriverState *src = s->mirror_top_bs->backing->bs;
620     BlockDriverState *target_bs = blk_bs(s->target);
621     BlockDriverState *mirror_top_bs = s->mirror_top_bs;
622     Error *local_err = NULL;
623     bool abort = job->ret < 0;
624     int ret = 0;
625 
626     if (s->prepared) {
627         return 0;
628     }
629     s->prepared = true;
630 
631     bdrv_release_dirty_bitmap(src, s->dirty_bitmap);
632 
633     /* Make sure that the source BDS doesn't go away during bdrv_replace_node,
634      * before we can call bdrv_drained_end */
635     bdrv_ref(src);
636     bdrv_ref(mirror_top_bs);
637     bdrv_ref(target_bs);
638 
639     /* Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
640      * inserting target_bs at s->to_replace, where we might not be able to get
641      * these permissions.
642      *
643      * Note that blk_unref() alone doesn't necessarily drop permissions because
644      * we might be running nested inside mirror_drain(), which takes an extra
645      * reference, so use an explicit blk_set_perm() first. */
646     blk_set_perm(s->target, 0, BLK_PERM_ALL, &error_abort);
647     blk_unref(s->target);
648     s->target = NULL;
649 
650     /* We don't access the source any more. Dropping any WRITE/RESIZE is
651      * required before it could become a backing file of target_bs. */
652     bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
653                             &error_abort);
654     if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
655         BlockDriverState *backing = s->is_none_mode ? src : s->base;
656         if (backing_bs(target_bs) != backing) {
657             bdrv_set_backing_hd(target_bs, backing, &local_err);
658             if (local_err) {
659                 error_report_err(local_err);
660                 ret = -EPERM;
661             }
662         }
663     }
664 
665     if (s->to_replace) {
666         replace_aio_context = bdrv_get_aio_context(s->to_replace);
667         aio_context_acquire(replace_aio_context);
668     }
669 
670     if (s->should_complete && !abort) {
671         BlockDriverState *to_replace = s->to_replace ?: src;
672 
673         if (bdrv_get_flags(target_bs) != bdrv_get_flags(to_replace)) {
674             bdrv_reopen(target_bs, bdrv_get_flags(to_replace), NULL);
675         }
676 
677         /* The mirror job has no requests in flight any more, but we need to
678          * drain potential other users of the BDS before changing the graph. */
679         bdrv_drained_begin(target_bs);
680         bdrv_replace_node(to_replace, target_bs, &local_err);
681         bdrv_drained_end(target_bs);
682         if (local_err) {
683             error_report_err(local_err);
684             ret = -EPERM;
685         }
686     }
687     if (s->to_replace) {
688         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
689         error_free(s->replace_blocker);
690         bdrv_unref(s->to_replace);
691     }
692     if (replace_aio_context) {
693         aio_context_release(replace_aio_context);
694     }
695     g_free(s->replaces);
696     bdrv_unref(target_bs);
697 
698     /* Remove the mirror filter driver from the graph. Before this, get rid of
699      * the blockers on the intermediate nodes so that the resulting state is
700      * valid. Also give up permissions on mirror_top_bs->backing, which might
701      * block the removal. */
702     block_job_remove_all_bdrv(bjob);
703     bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
704                             &error_abort);
705     bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
706 
707     /* We just changed the BDS the job BB refers to (with either or both of the
708      * bdrv_replace_node() calls), so switch the BB back so the cleanup does
709      * the right thing. We don't need any permissions any more now. */
710     blk_remove_bs(bjob->blk);
711     blk_set_perm(bjob->blk, 0, BLK_PERM_ALL, &error_abort);
712     blk_insert_bs(bjob->blk, mirror_top_bs, &error_abort);
713 
714     bs_opaque->job = NULL;
715 
716     bdrv_drained_end(src);
717     bdrv_unref(mirror_top_bs);
718     bdrv_unref(src);
719 
720     return ret;
721 }
722 
723 static int mirror_prepare(Job *job)
724 {
725     return mirror_exit_common(job);
726 }
727 
728 static void mirror_abort(Job *job)
729 {
730     int ret = mirror_exit_common(job);
731     assert(ret == 0);
732 }
733 
734 static void mirror_throttle(MirrorBlockJob *s)
735 {
736     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
737 
738     if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) {
739         s->last_pause_ns = now;
740         job_sleep_ns(&s->common.job, 0);
741     } else {
742         job_pause_point(&s->common.job);
743     }
744 }
745 
746 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
747 {
748     int64_t offset;
749     BlockDriverState *base = s->base;
750     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
751     BlockDriverState *target_bs = blk_bs(s->target);
752     int ret;
753     int64_t count;
754 
755     if (base == NULL && !bdrv_has_zero_init(target_bs)) {
756         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
757             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
758             return 0;
759         }
760 
761         s->initial_zeroing_ongoing = true;
762         for (offset = 0; offset < s->bdev_length; ) {
763             int bytes = MIN(s->bdev_length - offset,
764                             QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
765 
766             mirror_throttle(s);
767 
768             if (job_is_cancelled(&s->common.job)) {
769                 s->initial_zeroing_ongoing = false;
770                 return 0;
771             }
772 
773             if (s->in_flight >= MAX_IN_FLIGHT) {
774                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
775                                    s->in_flight);
776                 mirror_wait_for_free_in_flight_slot(s);
777                 continue;
778             }
779 
780             mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO);
781             offset += bytes;
782         }
783 
784         mirror_wait_for_all_io(s);
785         s->initial_zeroing_ongoing = false;
786     }
787 
788     /* First part, loop on the sectors and initialize the dirty bitmap.  */
789     for (offset = 0; offset < s->bdev_length; ) {
790         /* Just to make sure we are not exceeding int limit. */
791         int bytes = MIN(s->bdev_length - offset,
792                         QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
793 
794         mirror_throttle(s);
795 
796         if (job_is_cancelled(&s->common.job)) {
797             return 0;
798         }
799 
800         ret = bdrv_is_allocated_above(bs, base, offset, bytes, &count);
801         if (ret < 0) {
802             return ret;
803         }
804 
805         assert(count);
806         if (ret == 1) {
807             bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
808         }
809         offset += count;
810     }
811     return 0;
812 }
813 
814 /* Called when going out of the streaming phase to flush the bulk of the
815  * data to the medium, or just before completing.
816  */
817 static int mirror_flush(MirrorBlockJob *s)
818 {
819     int ret = blk_flush(s->target);
820     if (ret < 0) {
821         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
822             s->ret = ret;
823         }
824     }
825     return ret;
826 }
827 
828 static int coroutine_fn mirror_run(Job *job, Error **errp)
829 {
830     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
831     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
832     BlockDriverState *target_bs = blk_bs(s->target);
833     bool need_drain = true;
834     int64_t length;
835     BlockDriverInfo bdi;
836     char backing_filename[2]; /* we only need 2 characters because we are only
837                                  checking for a NULL string */
838     int ret = 0;
839 
840     if (job_is_cancelled(&s->common.job)) {
841         goto immediate_exit;
842     }
843 
844     s->bdev_length = bdrv_getlength(bs);
845     if (s->bdev_length < 0) {
846         ret = s->bdev_length;
847         goto immediate_exit;
848     }
849 
850     /* Active commit must resize the base image if its size differs from the
851      * active layer. */
852     if (s->base == blk_bs(s->target)) {
853         int64_t base_length;
854 
855         base_length = blk_getlength(s->target);
856         if (base_length < 0) {
857             ret = base_length;
858             goto immediate_exit;
859         }
860 
861         if (s->bdev_length > base_length) {
862             ret = blk_truncate(s->target, s->bdev_length, PREALLOC_MODE_OFF,
863                                NULL);
864             if (ret < 0) {
865                 goto immediate_exit;
866             }
867         }
868     }
869 
870     if (s->bdev_length == 0) {
871         /* Transition to the READY state and wait for complete. */
872         job_transition_to_ready(&s->common.job);
873         s->synced = true;
874         s->actively_synced = true;
875         while (!job_is_cancelled(&s->common.job) && !s->should_complete) {
876             job_yield(&s->common.job);
877         }
878         s->common.job.cancelled = false;
879         goto immediate_exit;
880     }
881 
882     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
883     s->in_flight_bitmap = bitmap_new(length);
884 
885     /* If we have no backing file yet in the destination, we cannot let
886      * the destination do COW.  Instead, we copy sectors around the
887      * dirty data if needed.  We need a bitmap to do that.
888      */
889     bdrv_get_backing_filename(target_bs, backing_filename,
890                               sizeof(backing_filename));
891     if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) {
892         s->target_cluster_size = bdi.cluster_size;
893     } else {
894         s->target_cluster_size = BDRV_SECTOR_SIZE;
895     }
896     if (backing_filename[0] && !target_bs->backing &&
897         s->granularity < s->target_cluster_size) {
898         s->buf_size = MAX(s->buf_size, s->target_cluster_size);
899         s->cow_bitmap = bitmap_new(length);
900     }
901     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
902 
903     s->buf = qemu_try_blockalign(bs, s->buf_size);
904     if (s->buf == NULL) {
905         ret = -ENOMEM;
906         goto immediate_exit;
907     }
908 
909     mirror_free_init(s);
910 
911     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
912     if (!s->is_none_mode) {
913         ret = mirror_dirty_init(s);
914         if (ret < 0 || job_is_cancelled(&s->common.job)) {
915             goto immediate_exit;
916         }
917     }
918 
919     assert(!s->dbi);
920     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
921     for (;;) {
922         uint64_t delay_ns = 0;
923         int64_t cnt, delta;
924         bool should_complete;
925 
926         /* Do not start passive operations while there are active
927          * writes in progress */
928         while (s->in_active_write_counter) {
929             mirror_wait_for_any_operation(s, true);
930         }
931 
932         if (s->ret < 0) {
933             ret = s->ret;
934             goto immediate_exit;
935         }
936 
937         job_pause_point(&s->common.job);
938 
939         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
940         /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
941          * the number of bytes currently being processed; together those are
942          * the current remaining operation length */
943         job_progress_set_remaining(&s->common.job, s->bytes_in_flight + cnt);
944 
945         /* Note that even when no rate limit is applied we need to yield
946          * periodically with no pending I/O so that bdrv_drain_all() returns.
947          * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
948          * an error, or when the source is clean, whichever comes first. */
949         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
950         if (delta < BLOCK_JOB_SLICE_TIME &&
951             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
952             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
953                 (cnt == 0 && s->in_flight > 0)) {
954                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
955                 mirror_wait_for_free_in_flight_slot(s);
956                 continue;
957             } else if (cnt != 0) {
958                 delay_ns = mirror_iteration(s);
959             }
960         }
961 
962         should_complete = false;
963         if (s->in_flight == 0 && cnt == 0) {
964             trace_mirror_before_flush(s);
965             if (!s->synced) {
966                 if (mirror_flush(s) < 0) {
967                     /* Go check s->ret.  */
968                     continue;
969                 }
970                 /* We're out of the streaming phase.  From now on, if the job
971                  * is cancelled we will actually complete all pending I/O and
972                  * report completion.  This way, block-job-cancel will leave
973                  * the target in a consistent state.
974                  */
975                 job_transition_to_ready(&s->common.job);
976                 s->synced = true;
977                 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) {
978                     s->actively_synced = true;
979                 }
980             }
981 
982             should_complete = s->should_complete ||
983                 job_is_cancelled(&s->common.job);
984             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
985         }
986 
987         if (cnt == 0 && should_complete) {
988             /* The dirty bitmap is not updated while operations are pending.
989              * If we're about to exit, wait for pending operations before
990              * calling bdrv_get_dirty_count(bs), or we may exit while the
991              * source has dirty data to copy!
992              *
993              * Note that I/O can be submitted by the guest while
994              * mirror_populate runs, so pause it now.  Before deciding
995              * whether to switch to target check one last time if I/O has
996              * come in the meanwhile, and if not flush the data to disk.
997              */
998             trace_mirror_before_drain(s, cnt);
999 
1000             bdrv_drained_begin(bs);
1001             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1002             if (cnt > 0 || mirror_flush(s) < 0) {
1003                 bdrv_drained_end(bs);
1004                 continue;
1005             }
1006 
1007             /* The two disks are in sync.  Exit and report successful
1008              * completion.
1009              */
1010             assert(QLIST_EMPTY(&bs->tracked_requests));
1011             s->common.job.cancelled = false;
1012             need_drain = false;
1013             break;
1014         }
1015 
1016         ret = 0;
1017 
1018         if (s->synced && !should_complete) {
1019             delay_ns = (s->in_flight == 0 &&
1020                         cnt == 0 ? BLOCK_JOB_SLICE_TIME : 0);
1021         }
1022         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
1023         job_sleep_ns(&s->common.job, delay_ns);
1024         if (job_is_cancelled(&s->common.job) &&
1025             (!s->synced || s->common.job.force_cancel))
1026         {
1027             break;
1028         }
1029         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1030     }
1031 
1032 immediate_exit:
1033     if (s->in_flight > 0) {
1034         /* We get here only if something went wrong.  Either the job failed,
1035          * or it was cancelled prematurely so that we do not guarantee that
1036          * the target is a copy of the source.
1037          */
1038         assert(ret < 0 || ((s->common.job.force_cancel || !s->synced) &&
1039                job_is_cancelled(&s->common.job)));
1040         assert(need_drain);
1041         mirror_wait_for_all_io(s);
1042     }
1043 
1044     assert(s->in_flight == 0);
1045     qemu_vfree(s->buf);
1046     g_free(s->cow_bitmap);
1047     g_free(s->in_flight_bitmap);
1048     bdrv_dirty_iter_free(s->dbi);
1049 
1050     if (need_drain) {
1051         bdrv_drained_begin(bs);
1052     }
1053 
1054     return ret;
1055 }
1056 
1057 static void mirror_complete(Job *job, Error **errp)
1058 {
1059     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1060     BlockDriverState *target;
1061 
1062     target = blk_bs(s->target);
1063 
1064     if (!s->synced) {
1065         error_setg(errp, "The active block job '%s' cannot be completed",
1066                    job->id);
1067         return;
1068     }
1069 
1070     if (s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
1071         int ret;
1072 
1073         assert(!target->backing);
1074         ret = bdrv_open_backing_file(target, NULL, "backing", errp);
1075         if (ret < 0) {
1076             return;
1077         }
1078     }
1079 
1080     /* block all operations on to_replace bs */
1081     if (s->replaces) {
1082         AioContext *replace_aio_context;
1083 
1084         s->to_replace = bdrv_find_node(s->replaces);
1085         if (!s->to_replace) {
1086             error_setg(errp, "Node name '%s' not found", s->replaces);
1087             return;
1088         }
1089 
1090         replace_aio_context = bdrv_get_aio_context(s->to_replace);
1091         aio_context_acquire(replace_aio_context);
1092 
1093         /* TODO Translate this into permission system. Current definition of
1094          * GRAPH_MOD would require to request it for the parents; they might
1095          * not even be BlockDriverStates, however, so a BdrvChild can't address
1096          * them. May need redefinition of GRAPH_MOD. */
1097         error_setg(&s->replace_blocker,
1098                    "block device is in use by block-job-complete");
1099         bdrv_op_block_all(s->to_replace, s->replace_blocker);
1100         bdrv_ref(s->to_replace);
1101 
1102         aio_context_release(replace_aio_context);
1103     }
1104 
1105     s->should_complete = true;
1106     job_enter(job);
1107 }
1108 
1109 static void mirror_pause(Job *job)
1110 {
1111     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1112 
1113     mirror_wait_for_all_io(s);
1114 }
1115 
1116 static bool mirror_drained_poll(BlockJob *job)
1117 {
1118     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1119     return !!s->in_flight;
1120 }
1121 
1122 static void mirror_attached_aio_context(BlockJob *job, AioContext *new_context)
1123 {
1124     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1125 
1126     blk_set_aio_context(s->target, new_context);
1127 }
1128 
1129 static void mirror_drain(BlockJob *job)
1130 {
1131     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1132 
1133     /* Need to keep a reference in case blk_drain triggers execution
1134      * of mirror_complete...
1135      */
1136     if (s->target) {
1137         BlockBackend *target = s->target;
1138         blk_ref(target);
1139         blk_drain(target);
1140         blk_unref(target);
1141     }
1142 }
1143 
1144 static const BlockJobDriver mirror_job_driver = {
1145     .job_driver = {
1146         .instance_size          = sizeof(MirrorBlockJob),
1147         .job_type               = JOB_TYPE_MIRROR,
1148         .free                   = block_job_free,
1149         .user_resume            = block_job_user_resume,
1150         .drain                  = block_job_drain,
1151         .run                    = mirror_run,
1152         .prepare                = mirror_prepare,
1153         .abort                  = mirror_abort,
1154         .pause                  = mirror_pause,
1155         .complete               = mirror_complete,
1156     },
1157     .drained_poll           = mirror_drained_poll,
1158     .attached_aio_context   = mirror_attached_aio_context,
1159     .drain                  = mirror_drain,
1160 };
1161 
1162 static const BlockJobDriver commit_active_job_driver = {
1163     .job_driver = {
1164         .instance_size          = sizeof(MirrorBlockJob),
1165         .job_type               = JOB_TYPE_COMMIT,
1166         .free                   = block_job_free,
1167         .user_resume            = block_job_user_resume,
1168         .drain                  = block_job_drain,
1169         .run                    = mirror_run,
1170         .prepare                = mirror_prepare,
1171         .abort                  = mirror_abort,
1172         .pause                  = mirror_pause,
1173         .complete               = mirror_complete,
1174     },
1175     .drained_poll           = mirror_drained_poll,
1176     .attached_aio_context   = mirror_attached_aio_context,
1177     .drain                  = mirror_drain,
1178 };
1179 
1180 static void do_sync_target_write(MirrorBlockJob *job, MirrorMethod method,
1181                                  uint64_t offset, uint64_t bytes,
1182                                  QEMUIOVector *qiov, int flags)
1183 {
1184     BdrvDirtyBitmapIter *iter;
1185     QEMUIOVector target_qiov;
1186     uint64_t dirty_offset;
1187     int dirty_bytes;
1188 
1189     if (qiov) {
1190         qemu_iovec_init(&target_qiov, qiov->niov);
1191     }
1192 
1193     iter = bdrv_dirty_iter_new(job->dirty_bitmap);
1194     bdrv_set_dirty_iter(iter, offset);
1195 
1196     while (true) {
1197         bool valid_area;
1198         int ret;
1199 
1200         bdrv_dirty_bitmap_lock(job->dirty_bitmap);
1201         valid_area = bdrv_dirty_iter_next_area(iter, offset + bytes,
1202                                                &dirty_offset, &dirty_bytes);
1203         if (!valid_area) {
1204             bdrv_dirty_bitmap_unlock(job->dirty_bitmap);
1205             break;
1206         }
1207 
1208         bdrv_reset_dirty_bitmap_locked(job->dirty_bitmap,
1209                                        dirty_offset, dirty_bytes);
1210         bdrv_dirty_bitmap_unlock(job->dirty_bitmap);
1211 
1212         job_progress_increase_remaining(&job->common.job, dirty_bytes);
1213 
1214         assert(dirty_offset - offset <= SIZE_MAX);
1215         if (qiov) {
1216             qemu_iovec_reset(&target_qiov);
1217             qemu_iovec_concat(&target_qiov, qiov,
1218                               dirty_offset - offset, dirty_bytes);
1219         }
1220 
1221         switch (method) {
1222         case MIRROR_METHOD_COPY:
1223             ret = blk_co_pwritev(job->target, dirty_offset, dirty_bytes,
1224                                  qiov ? &target_qiov : NULL, flags);
1225             break;
1226 
1227         case MIRROR_METHOD_ZERO:
1228             assert(!qiov);
1229             ret = blk_co_pwrite_zeroes(job->target, dirty_offset, dirty_bytes,
1230                                        flags);
1231             break;
1232 
1233         case MIRROR_METHOD_DISCARD:
1234             assert(!qiov);
1235             ret = blk_co_pdiscard(job->target, dirty_offset, dirty_bytes);
1236             break;
1237 
1238         default:
1239             abort();
1240         }
1241 
1242         if (ret >= 0) {
1243             job_progress_update(&job->common.job, dirty_bytes);
1244         } else {
1245             BlockErrorAction action;
1246 
1247             bdrv_set_dirty_bitmap(job->dirty_bitmap, dirty_offset, dirty_bytes);
1248             job->actively_synced = false;
1249 
1250             action = mirror_error_action(job, false, -ret);
1251             if (action == BLOCK_ERROR_ACTION_REPORT) {
1252                 if (!job->ret) {
1253                     job->ret = ret;
1254                 }
1255                 break;
1256             }
1257         }
1258     }
1259 
1260     bdrv_dirty_iter_free(iter);
1261     if (qiov) {
1262         qemu_iovec_destroy(&target_qiov);
1263     }
1264 }
1265 
1266 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s,
1267                                                    uint64_t offset,
1268                                                    uint64_t bytes)
1269 {
1270     MirrorOp *op;
1271     uint64_t start_chunk = offset / s->granularity;
1272     uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
1273 
1274     op = g_new(MirrorOp, 1);
1275     *op = (MirrorOp){
1276         .s                  = s,
1277         .offset             = offset,
1278         .bytes              = bytes,
1279         .is_active_write    = true,
1280     };
1281     qemu_co_queue_init(&op->waiting_requests);
1282     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
1283 
1284     s->in_active_write_counter++;
1285 
1286     mirror_wait_on_conflicts(op, s, offset, bytes);
1287 
1288     bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1289 
1290     return op;
1291 }
1292 
1293 static void coroutine_fn active_write_settle(MirrorOp *op)
1294 {
1295     uint64_t start_chunk = op->offset / op->s->granularity;
1296     uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes,
1297                                       op->s->granularity);
1298 
1299     if (!--op->s->in_active_write_counter && op->s->actively_synced) {
1300         BdrvChild *source = op->s->mirror_top_bs->backing;
1301 
1302         if (QLIST_FIRST(&source->bs->parents) == source &&
1303             QLIST_NEXT(source, next_parent) == NULL)
1304         {
1305             /* Assert that we are back in sync once all active write
1306              * operations are settled.
1307              * Note that we can only assert this if the mirror node
1308              * is the source node's only parent. */
1309             assert(!bdrv_get_dirty_count(op->s->dirty_bitmap));
1310         }
1311     }
1312     bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1313     QTAILQ_REMOVE(&op->s->ops_in_flight, op, next);
1314     qemu_co_queue_restart_all(&op->waiting_requests);
1315     g_free(op);
1316 }
1317 
1318 static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs,
1319     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1320 {
1321     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1322 }
1323 
1324 static int coroutine_fn bdrv_mirror_top_do_write(BlockDriverState *bs,
1325     MirrorMethod method, uint64_t offset, uint64_t bytes, QEMUIOVector *qiov,
1326     int flags)
1327 {
1328     MirrorOp *op = NULL;
1329     MirrorBDSOpaque *s = bs->opaque;
1330     int ret = 0;
1331     bool copy_to_target;
1332 
1333     copy_to_target = s->job->ret >= 0 &&
1334                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1335 
1336     if (copy_to_target) {
1337         op = active_write_prepare(s->job, offset, bytes);
1338     }
1339 
1340     switch (method) {
1341     case MIRROR_METHOD_COPY:
1342         ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1343         break;
1344 
1345     case MIRROR_METHOD_ZERO:
1346         ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1347         break;
1348 
1349     case MIRROR_METHOD_DISCARD:
1350         ret = bdrv_co_pdiscard(bs->backing, offset, bytes);
1351         break;
1352 
1353     default:
1354         abort();
1355     }
1356 
1357     if (ret < 0) {
1358         goto out;
1359     }
1360 
1361     if (copy_to_target) {
1362         do_sync_target_write(s->job, method, offset, bytes, qiov, flags);
1363     }
1364 
1365 out:
1366     if (copy_to_target) {
1367         active_write_settle(op);
1368     }
1369     return ret;
1370 }
1371 
1372 static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs,
1373     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1374 {
1375     MirrorBDSOpaque *s = bs->opaque;
1376     QEMUIOVector bounce_qiov;
1377     void *bounce_buf;
1378     int ret = 0;
1379     bool copy_to_target;
1380 
1381     copy_to_target = s->job->ret >= 0 &&
1382                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1383 
1384     if (copy_to_target) {
1385         /* The guest might concurrently modify the data to write; but
1386          * the data on source and destination must match, so we have
1387          * to use a bounce buffer if we are going to write to the
1388          * target now. */
1389         bounce_buf = qemu_blockalign(bs, bytes);
1390         iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes);
1391 
1392         qemu_iovec_init(&bounce_qiov, 1);
1393         qemu_iovec_add(&bounce_qiov, bounce_buf, bytes);
1394         qiov = &bounce_qiov;
1395     }
1396 
1397     ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov,
1398                                    flags);
1399 
1400     if (copy_to_target) {
1401         qemu_iovec_destroy(&bounce_qiov);
1402         qemu_vfree(bounce_buf);
1403     }
1404 
1405     return ret;
1406 }
1407 
1408 static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs)
1409 {
1410     if (bs->backing == NULL) {
1411         /* we can be here after failed bdrv_append in mirror_start_job */
1412         return 0;
1413     }
1414     return bdrv_co_flush(bs->backing->bs);
1415 }
1416 
1417 static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs,
1418     int64_t offset, int bytes, BdrvRequestFlags flags)
1419 {
1420     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL,
1421                                     flags);
1422 }
1423 
1424 static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs,
1425     int64_t offset, int bytes)
1426 {
1427     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes,
1428                                     NULL, 0);
1429 }
1430 
1431 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs, QDict *opts)
1432 {
1433     if (bs->backing == NULL) {
1434         /* we can be here after failed bdrv_attach_child in
1435          * bdrv_set_backing_hd */
1436         return;
1437     }
1438     bdrv_refresh_filename(bs->backing->bs);
1439     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1440             bs->backing->bs->filename);
1441 }
1442 
1443 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1444                                        const BdrvChildRole *role,
1445                                        BlockReopenQueue *reopen_queue,
1446                                        uint64_t perm, uint64_t shared,
1447                                        uint64_t *nperm, uint64_t *nshared)
1448 {
1449     /* Must be able to forward guest writes to the real image */
1450     *nperm = 0;
1451     if (perm & BLK_PERM_WRITE) {
1452         *nperm |= BLK_PERM_WRITE;
1453     }
1454 
1455     *nshared = BLK_PERM_ALL;
1456 }
1457 
1458 /* Dummy node that provides consistent read to its users without requiring it
1459  * from its backing file and that allows writes on the backing file chain. */
1460 static BlockDriver bdrv_mirror_top = {
1461     .format_name                = "mirror_top",
1462     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1463     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1464     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1465     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1466     .bdrv_co_flush              = bdrv_mirror_top_flush,
1467     .bdrv_co_block_status       = bdrv_co_block_status_from_backing,
1468     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1469     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1470 };
1471 
1472 static void mirror_start_job(const char *job_id, BlockDriverState *bs,
1473                              int creation_flags, BlockDriverState *target,
1474                              const char *replaces, int64_t speed,
1475                              uint32_t granularity, int64_t buf_size,
1476                              BlockMirrorBackingMode backing_mode,
1477                              BlockdevOnError on_source_error,
1478                              BlockdevOnError on_target_error,
1479                              bool unmap,
1480                              BlockCompletionFunc *cb,
1481                              void *opaque,
1482                              const BlockJobDriver *driver,
1483                              bool is_none_mode, BlockDriverState *base,
1484                              bool auto_complete, const char *filter_node_name,
1485                              bool is_mirror, MirrorCopyMode copy_mode,
1486                              Error **errp)
1487 {
1488     MirrorBlockJob *s;
1489     MirrorBDSOpaque *bs_opaque;
1490     BlockDriverState *mirror_top_bs;
1491     bool target_graph_mod;
1492     bool target_is_backing;
1493     Error *local_err = NULL;
1494     int ret;
1495 
1496     if (granularity == 0) {
1497         granularity = bdrv_get_default_bitmap_granularity(target);
1498     }
1499 
1500     assert(is_power_of_2(granularity));
1501 
1502     if (buf_size < 0) {
1503         error_setg(errp, "Invalid parameter 'buf-size'");
1504         return;
1505     }
1506 
1507     if (buf_size == 0) {
1508         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1509     }
1510 
1511     if (bs == target) {
1512         error_setg(errp, "Can't mirror node into itself");
1513         return;
1514     }
1515 
1516     /* In the case of active commit, add dummy driver to provide consistent
1517      * reads on the top, while disabling it in the intermediate nodes, and make
1518      * the backing chain writable. */
1519     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1520                                          BDRV_O_RDWR, errp);
1521     if (mirror_top_bs == NULL) {
1522         return;
1523     }
1524     if (!filter_node_name) {
1525         mirror_top_bs->implicit = true;
1526     }
1527     mirror_top_bs->total_sectors = bs->total_sectors;
1528     mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
1529     mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED;
1530     bs_opaque = g_new0(MirrorBDSOpaque, 1);
1531     mirror_top_bs->opaque = bs_opaque;
1532     bdrv_set_aio_context(mirror_top_bs, bdrv_get_aio_context(bs));
1533 
1534     /* bdrv_append takes ownership of the mirror_top_bs reference, need to keep
1535      * it alive until block_job_create() succeeds even if bs has no parent. */
1536     bdrv_ref(mirror_top_bs);
1537     bdrv_drained_begin(bs);
1538     bdrv_append(mirror_top_bs, bs, &local_err);
1539     bdrv_drained_end(bs);
1540 
1541     if (local_err) {
1542         bdrv_unref(mirror_top_bs);
1543         error_propagate(errp, local_err);
1544         return;
1545     }
1546 
1547     /* Make sure that the source is not resized while the job is running */
1548     s = block_job_create(job_id, driver, NULL, mirror_top_bs,
1549                          BLK_PERM_CONSISTENT_READ,
1550                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1551                          BLK_PERM_WRITE | BLK_PERM_GRAPH_MOD, speed,
1552                          creation_flags, cb, opaque, errp);
1553     if (!s) {
1554         goto fail;
1555     }
1556     bs_opaque->job = s;
1557 
1558     /* The block job now has a reference to this node */
1559     bdrv_unref(mirror_top_bs);
1560 
1561     s->mirror_top_bs = mirror_top_bs;
1562 
1563     /* No resize for the target either; while the mirror is still running, a
1564      * consistent read isn't necessarily possible. We could possibly allow
1565      * writes and graph modifications, though it would likely defeat the
1566      * purpose of a mirror, so leave them blocked for now.
1567      *
1568      * In the case of active commit, things look a bit different, though,
1569      * because the target is an already populated backing file in active use.
1570      * We can allow anything except resize there.*/
1571     target_is_backing = bdrv_chain_contains(bs, target);
1572     target_graph_mod = (backing_mode != MIRROR_LEAVE_BACKING_CHAIN);
1573     s->target = blk_new(BLK_PERM_WRITE | BLK_PERM_RESIZE |
1574                         (target_graph_mod ? BLK_PERM_GRAPH_MOD : 0),
1575                         BLK_PERM_WRITE_UNCHANGED |
1576                         (target_is_backing ? BLK_PERM_CONSISTENT_READ |
1577                                              BLK_PERM_WRITE |
1578                                              BLK_PERM_GRAPH_MOD : 0));
1579     ret = blk_insert_bs(s->target, target, errp);
1580     if (ret < 0) {
1581         goto fail;
1582     }
1583     if (is_mirror) {
1584         /* XXX: Mirror target could be a NBD server of target QEMU in the case
1585          * of non-shared block migration. To allow migration completion, we
1586          * have to allow "inactivate" of the target BB.  When that happens, we
1587          * know the job is drained, and the vcpus are stopped, so no write
1588          * operation will be performed. Block layer already has assertions to
1589          * ensure that. */
1590         blk_set_force_allow_inactivate(s->target);
1591     }
1592 
1593     s->replaces = g_strdup(replaces);
1594     s->on_source_error = on_source_error;
1595     s->on_target_error = on_target_error;
1596     s->is_none_mode = is_none_mode;
1597     s->backing_mode = backing_mode;
1598     s->copy_mode = copy_mode;
1599     s->base = base;
1600     s->granularity = granularity;
1601     s->buf_size = ROUND_UP(buf_size, granularity);
1602     s->unmap = unmap;
1603     if (auto_complete) {
1604         s->should_complete = true;
1605     }
1606 
1607     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1608     if (!s->dirty_bitmap) {
1609         goto fail;
1610     }
1611 
1612     /* Required permissions are already taken with blk_new() */
1613     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1614                        &error_abort);
1615 
1616     /* In commit_active_start() all intermediate nodes disappear, so
1617      * any jobs in them must be blocked */
1618     if (target_is_backing) {
1619         BlockDriverState *iter;
1620         for (iter = backing_bs(bs); iter != target; iter = backing_bs(iter)) {
1621             /* XXX BLK_PERM_WRITE needs to be allowed so we don't block
1622              * ourselves at s->base (if writes are blocked for a node, they are
1623              * also blocked for its backing file). The other options would be a
1624              * second filter driver above s->base (== target). */
1625             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1626                                      BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
1627                                      errp);
1628             if (ret < 0) {
1629                 goto fail;
1630             }
1631         }
1632     }
1633 
1634     QTAILQ_INIT(&s->ops_in_flight);
1635 
1636     trace_mirror_start(bs, s, opaque);
1637     job_start(&s->common.job);
1638     return;
1639 
1640 fail:
1641     if (s) {
1642         /* Make sure this BDS does not go away until we have completed the graph
1643          * changes below */
1644         bdrv_ref(mirror_top_bs);
1645 
1646         g_free(s->replaces);
1647         blk_unref(s->target);
1648         bs_opaque->job = NULL;
1649         job_early_fail(&s->common.job);
1650     }
1651 
1652     bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
1653                             &error_abort);
1654     bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
1655 
1656     bdrv_unref(mirror_top_bs);
1657 }
1658 
1659 void mirror_start(const char *job_id, BlockDriverState *bs,
1660                   BlockDriverState *target, const char *replaces,
1661                   int creation_flags, int64_t speed,
1662                   uint32_t granularity, int64_t buf_size,
1663                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1664                   BlockdevOnError on_source_error,
1665                   BlockdevOnError on_target_error,
1666                   bool unmap, const char *filter_node_name,
1667                   MirrorCopyMode copy_mode, Error **errp)
1668 {
1669     bool is_none_mode;
1670     BlockDriverState *base;
1671 
1672     if (mode == MIRROR_SYNC_MODE_INCREMENTAL) {
1673         error_setg(errp, "Sync mode 'incremental' not supported");
1674         return;
1675     }
1676     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1677     base = mode == MIRROR_SYNC_MODE_TOP ? backing_bs(bs) : NULL;
1678     mirror_start_job(job_id, bs, creation_flags, target, replaces,
1679                      speed, granularity, buf_size, backing_mode,
1680                      on_source_error, on_target_error, unmap, NULL, NULL,
1681                      &mirror_job_driver, is_none_mode, base, false,
1682                      filter_node_name, true, copy_mode, errp);
1683 }
1684 
1685 void commit_active_start(const char *job_id, BlockDriverState *bs,
1686                          BlockDriverState *base, int creation_flags,
1687                          int64_t speed, BlockdevOnError on_error,
1688                          const char *filter_node_name,
1689                          BlockCompletionFunc *cb, void *opaque,
1690                          bool auto_complete, Error **errp)
1691 {
1692     int orig_base_flags;
1693     Error *local_err = NULL;
1694 
1695     orig_base_flags = bdrv_get_flags(base);
1696 
1697     if (bdrv_reopen(base, bs->open_flags, errp)) {
1698         return;
1699     }
1700 
1701     mirror_start_job(job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1702                      MIRROR_LEAVE_BACKING_CHAIN,
1703                      on_error, on_error, true, cb, opaque,
1704                      &commit_active_job_driver, false, base, auto_complete,
1705                      filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND,
1706                      &local_err);
1707     if (local_err) {
1708         error_propagate(errp, local_err);
1709         goto error_restore_flags;
1710     }
1711 
1712     return;
1713 
1714 error_restore_flags:
1715     /* ignore error and errp for bdrv_reopen, because we want to propagate
1716      * the original error */
1717     bdrv_reopen(base, orig_base_flags, NULL);
1718     return;
1719 }
1720