xref: /openbmc/qemu/block/mirror.c (revision 8cf108c5)
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(, 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     bool in_drain;
84 } MirrorBlockJob;
85 
86 typedef struct MirrorBDSOpaque {
87     MirrorBlockJob *job;
88 } MirrorBDSOpaque;
89 
90 struct MirrorOp {
91     MirrorBlockJob *s;
92     QEMUIOVector qiov;
93     int64_t offset;
94     uint64_t bytes;
95 
96     /* The pointee is set by mirror_co_read(), mirror_co_zero(), and
97      * mirror_co_discard() before yielding for the first time */
98     int64_t *bytes_handled;
99 
100     bool is_pseudo_op;
101     bool is_active_write;
102     CoQueue waiting_requests;
103 
104     QTAILQ_ENTRY(MirrorOp) next;
105 };
106 
107 typedef enum MirrorMethod {
108     MIRROR_METHOD_COPY,
109     MIRROR_METHOD_ZERO,
110     MIRROR_METHOD_DISCARD,
111 } MirrorMethod;
112 
113 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
114                                             int error)
115 {
116     s->synced = false;
117     s->actively_synced = false;
118     if (read) {
119         return block_job_error_action(&s->common, s->on_source_error,
120                                       true, error);
121     } else {
122         return block_job_error_action(&s->common, s->on_target_error,
123                                       false, error);
124     }
125 }
126 
127 static void coroutine_fn mirror_wait_on_conflicts(MirrorOp *self,
128                                                   MirrorBlockJob *s,
129                                                   uint64_t offset,
130                                                   uint64_t bytes)
131 {
132     uint64_t self_start_chunk = offset / s->granularity;
133     uint64_t self_end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
134     uint64_t self_nb_chunks = self_end_chunk - self_start_chunk;
135 
136     while (find_next_bit(s->in_flight_bitmap, self_end_chunk,
137                          self_start_chunk) < self_end_chunk &&
138            s->ret >= 0)
139     {
140         MirrorOp *op;
141 
142         QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
143             uint64_t op_start_chunk = op->offset / s->granularity;
144             uint64_t op_nb_chunks = DIV_ROUND_UP(op->offset + op->bytes,
145                                                  s->granularity) -
146                                     op_start_chunk;
147 
148             if (op == self) {
149                 continue;
150             }
151 
152             if (ranges_overlap(self_start_chunk, self_nb_chunks,
153                                op_start_chunk, op_nb_chunks))
154             {
155                 qemu_co_queue_wait(&op->waiting_requests, NULL);
156                 break;
157             }
158         }
159     }
160 }
161 
162 static void coroutine_fn mirror_iteration_done(MirrorOp *op, int ret)
163 {
164     MirrorBlockJob *s = op->s;
165     struct iovec *iov;
166     int64_t chunk_num;
167     int i, nb_chunks;
168 
169     trace_mirror_iteration_done(s, op->offset, op->bytes, ret);
170 
171     s->in_flight--;
172     s->bytes_in_flight -= op->bytes;
173     iov = op->qiov.iov;
174     for (i = 0; i < op->qiov.niov; i++) {
175         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
176         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
177         s->buf_free_count++;
178     }
179 
180     chunk_num = op->offset / s->granularity;
181     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
182 
183     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
184     QTAILQ_REMOVE(&s->ops_in_flight, op, next);
185     if (ret >= 0) {
186         if (s->cow_bitmap) {
187             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
188         }
189         if (!s->initial_zeroing_ongoing) {
190             job_progress_update(&s->common.job, op->bytes);
191         }
192     }
193     qemu_iovec_destroy(&op->qiov);
194 
195     qemu_co_queue_restart_all(&op->waiting_requests);
196     g_free(op);
197 }
198 
199 static void coroutine_fn mirror_write_complete(MirrorOp *op, int ret)
200 {
201     MirrorBlockJob *s = op->s;
202 
203     if (ret < 0) {
204         BlockErrorAction action;
205 
206         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
207         action = mirror_error_action(s, false, -ret);
208         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
209             s->ret = ret;
210         }
211     }
212 
213     mirror_iteration_done(op, ret);
214 }
215 
216 static void coroutine_fn mirror_read_complete(MirrorOp *op, int ret)
217 {
218     MirrorBlockJob *s = op->s;
219 
220     if (ret < 0) {
221         BlockErrorAction action;
222 
223         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
224         action = mirror_error_action(s, true, -ret);
225         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
226             s->ret = ret;
227         }
228 
229         mirror_iteration_done(op, ret);
230         return;
231     }
232 
233     ret = blk_co_pwritev(s->target, op->offset, op->qiov.size, &op->qiov, 0);
234     mirror_write_complete(op, ret);
235 }
236 
237 /* Clip bytes relative to offset to not exceed end-of-file */
238 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s,
239                                         int64_t offset,
240                                         int64_t bytes)
241 {
242     return MIN(bytes, s->bdev_length - offset);
243 }
244 
245 /* Round offset and/or bytes to target cluster if COW is needed, and
246  * return the offset of the adjusted tail against original. */
247 static int mirror_cow_align(MirrorBlockJob *s, int64_t *offset,
248                             uint64_t *bytes)
249 {
250     bool need_cow;
251     int ret = 0;
252     int64_t align_offset = *offset;
253     int64_t align_bytes = *bytes;
254     int max_bytes = s->granularity * s->max_iov;
255 
256     need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap);
257     need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity,
258                           s->cow_bitmap);
259     if (need_cow) {
260         bdrv_round_to_clusters(blk_bs(s->target), *offset, *bytes,
261                                &align_offset, &align_bytes);
262     }
263 
264     if (align_bytes > max_bytes) {
265         align_bytes = max_bytes;
266         if (need_cow) {
267             align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size);
268         }
269     }
270     /* Clipping may result in align_bytes unaligned to chunk boundary, but
271      * that doesn't matter because it's already the end of source image. */
272     align_bytes = mirror_clip_bytes(s, align_offset, align_bytes);
273 
274     ret = align_offset + align_bytes - (*offset + *bytes);
275     *offset = align_offset;
276     *bytes = align_bytes;
277     assert(ret >= 0);
278     return ret;
279 }
280 
281 static inline void coroutine_fn
282 mirror_wait_for_any_operation(MirrorBlockJob *s, bool active)
283 {
284     MirrorOp *op;
285 
286     QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
287         /* Do not wait on pseudo ops, because it may in turn wait on
288          * some other operation to start, which may in fact be the
289          * caller of this function.  Since there is only one pseudo op
290          * at any given time, we will always find some real operation
291          * to wait on. */
292         if (!op->is_pseudo_op && op->is_active_write == active) {
293             qemu_co_queue_wait(&op->waiting_requests, NULL);
294             return;
295         }
296     }
297     abort();
298 }
299 
300 static inline void coroutine_fn
301 mirror_wait_for_free_in_flight_slot(MirrorBlockJob *s)
302 {
303     /* Only non-active operations use up in-flight slots */
304     mirror_wait_for_any_operation(s, false);
305 }
306 
307 /* Perform a mirror copy operation.
308  *
309  * *op->bytes_handled is set to the number of bytes copied after and
310  * including offset, excluding any bytes copied prior to offset due
311  * to alignment.  This will be op->bytes if no alignment is necessary,
312  * or (new_end - op->offset) if the tail is rounded up or down due to
313  * alignment or buffer limit.
314  */
315 static void coroutine_fn mirror_co_read(void *opaque)
316 {
317     MirrorOp *op = opaque;
318     MirrorBlockJob *s = op->s;
319     int nb_chunks;
320     uint64_t ret;
321     uint64_t max_bytes;
322 
323     max_bytes = s->granularity * s->max_iov;
324 
325     /* We can only handle as much as buf_size at a time. */
326     op->bytes = MIN(s->buf_size, MIN(max_bytes, op->bytes));
327     assert(op->bytes);
328     assert(op->bytes < BDRV_REQUEST_MAX_BYTES);
329     *op->bytes_handled = op->bytes;
330 
331     if (s->cow_bitmap) {
332         *op->bytes_handled += mirror_cow_align(s, &op->offset, &op->bytes);
333     }
334     /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */
335     assert(*op->bytes_handled <= UINT_MAX);
336     assert(op->bytes <= s->buf_size);
337     /* The offset is granularity-aligned because:
338      * 1) Caller passes in aligned values;
339      * 2) mirror_cow_align is used only when target cluster is larger. */
340     assert(QEMU_IS_ALIGNED(op->offset, s->granularity));
341     /* The range is sector-aligned, since bdrv_getlength() rounds up. */
342     assert(QEMU_IS_ALIGNED(op->bytes, BDRV_SECTOR_SIZE));
343     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
344 
345     while (s->buf_free_count < nb_chunks) {
346         trace_mirror_yield_in_flight(s, op->offset, s->in_flight);
347         mirror_wait_for_free_in_flight_slot(s);
348     }
349 
350     /* Now make a QEMUIOVector taking enough granularity-sized chunks
351      * from s->buf_free.
352      */
353     qemu_iovec_init(&op->qiov, nb_chunks);
354     while (nb_chunks-- > 0) {
355         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
356         size_t remaining = op->bytes - op->qiov.size;
357 
358         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
359         s->buf_free_count--;
360         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
361     }
362 
363     /* Copy the dirty cluster.  */
364     s->in_flight++;
365     s->bytes_in_flight += op->bytes;
366     trace_mirror_one_iteration(s, op->offset, op->bytes);
367 
368     ret = bdrv_co_preadv(s->mirror_top_bs->backing, op->offset, op->bytes,
369                          &op->qiov, 0);
370     mirror_read_complete(op, ret);
371 }
372 
373 static void coroutine_fn mirror_co_zero(void *opaque)
374 {
375     MirrorOp *op = opaque;
376     int ret;
377 
378     op->s->in_flight++;
379     op->s->bytes_in_flight += op->bytes;
380     *op->bytes_handled = op->bytes;
381 
382     ret = blk_co_pwrite_zeroes(op->s->target, op->offset, op->bytes,
383                                op->s->unmap ? BDRV_REQ_MAY_UNMAP : 0);
384     mirror_write_complete(op, ret);
385 }
386 
387 static void coroutine_fn mirror_co_discard(void *opaque)
388 {
389     MirrorOp *op = opaque;
390     int ret;
391 
392     op->s->in_flight++;
393     op->s->bytes_in_flight += op->bytes;
394     *op->bytes_handled = op->bytes;
395 
396     ret = blk_co_pdiscard(op->s->target, op->offset, op->bytes);
397     mirror_write_complete(op, ret);
398 }
399 
400 static unsigned mirror_perform(MirrorBlockJob *s, int64_t offset,
401                                unsigned bytes, MirrorMethod mirror_method)
402 {
403     MirrorOp *op;
404     Coroutine *co;
405     int64_t bytes_handled = -1;
406 
407     op = g_new(MirrorOp, 1);
408     *op = (MirrorOp){
409         .s              = s,
410         .offset         = offset,
411         .bytes          = bytes,
412         .bytes_handled  = &bytes_handled,
413     };
414     qemu_co_queue_init(&op->waiting_requests);
415 
416     switch (mirror_method) {
417     case MIRROR_METHOD_COPY:
418         co = qemu_coroutine_create(mirror_co_read, op);
419         break;
420     case MIRROR_METHOD_ZERO:
421         co = qemu_coroutine_create(mirror_co_zero, op);
422         break;
423     case MIRROR_METHOD_DISCARD:
424         co = qemu_coroutine_create(mirror_co_discard, op);
425         break;
426     default:
427         abort();
428     }
429 
430     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
431     qemu_coroutine_enter(co);
432     /* At this point, ownership of op has been moved to the coroutine
433      * and the object may already be freed */
434 
435     /* Assert that this value has been set */
436     assert(bytes_handled >= 0);
437 
438     /* Same assertion as in mirror_co_read() (and for mirror_co_read()
439      * and mirror_co_discard(), bytes_handled == op->bytes, which
440      * is the @bytes parameter given to this function) */
441     assert(bytes_handled <= UINT_MAX);
442     return bytes_handled;
443 }
444 
445 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
446 {
447     BlockDriverState *source = s->mirror_top_bs->backing->bs;
448     MirrorOp *pseudo_op;
449     int64_t offset;
450     uint64_t delay_ns = 0, ret = 0;
451     /* At least the first dirty chunk is mirrored in one iteration. */
452     int nb_chunks = 1;
453     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
454     int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES);
455 
456     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
457     offset = bdrv_dirty_iter_next(s->dbi);
458     if (offset < 0) {
459         bdrv_set_dirty_iter(s->dbi, 0);
460         offset = bdrv_dirty_iter_next(s->dbi);
461         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
462         assert(offset >= 0);
463     }
464     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
465 
466     mirror_wait_on_conflicts(NULL, s, offset, 1);
467 
468     job_pause_point(&s->common.job);
469 
470     /* Find the number of consective dirty chunks following the first dirty
471      * one, and wait for in flight requests in them. */
472     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
473     while (nb_chunks * s->granularity < s->buf_size) {
474         int64_t next_dirty;
475         int64_t next_offset = offset + nb_chunks * s->granularity;
476         int64_t next_chunk = next_offset / s->granularity;
477         if (next_offset >= s->bdev_length ||
478             !bdrv_get_dirty_locked(source, s->dirty_bitmap, next_offset)) {
479             break;
480         }
481         if (test_bit(next_chunk, s->in_flight_bitmap)) {
482             break;
483         }
484 
485         next_dirty = bdrv_dirty_iter_next(s->dbi);
486         if (next_dirty > next_offset || next_dirty < 0) {
487             /* The bitmap iterator's cache is stale, refresh it */
488             bdrv_set_dirty_iter(s->dbi, next_offset);
489             next_dirty = bdrv_dirty_iter_next(s->dbi);
490         }
491         assert(next_dirty == next_offset);
492         nb_chunks++;
493     }
494 
495     /* Clear dirty bits before querying the block status, because
496      * calling bdrv_block_status_above could yield - if some blocks are
497      * marked dirty in this window, we need to know.
498      */
499     bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset,
500                                    nb_chunks * s->granularity);
501     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
502 
503     /* Before claiming an area in the in-flight bitmap, we have to
504      * create a MirrorOp for it so that conflicting requests can wait
505      * for it.  mirror_perform() will create the real MirrorOps later,
506      * for now we just create a pseudo operation that will wake up all
507      * conflicting requests once all real operations have been
508      * launched. */
509     pseudo_op = g_new(MirrorOp, 1);
510     *pseudo_op = (MirrorOp){
511         .offset         = offset,
512         .bytes          = nb_chunks * s->granularity,
513         .is_pseudo_op   = true,
514     };
515     qemu_co_queue_init(&pseudo_op->waiting_requests);
516     QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next);
517 
518     bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks);
519     while (nb_chunks > 0 && offset < s->bdev_length) {
520         int ret;
521         int64_t io_bytes;
522         int64_t io_bytes_acct;
523         MirrorMethod mirror_method = MIRROR_METHOD_COPY;
524 
525         assert(!(offset % s->granularity));
526         ret = bdrv_block_status_above(source, NULL, offset,
527                                       nb_chunks * s->granularity,
528                                       &io_bytes, NULL, NULL);
529         if (ret < 0) {
530             io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes);
531         } else if (ret & BDRV_BLOCK_DATA) {
532             io_bytes = MIN(io_bytes, max_io_bytes);
533         }
534 
535         io_bytes -= io_bytes % s->granularity;
536         if (io_bytes < s->granularity) {
537             io_bytes = s->granularity;
538         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
539             int64_t target_offset;
540             int64_t target_bytes;
541             bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes,
542                                    &target_offset, &target_bytes);
543             if (target_offset == offset &&
544                 target_bytes == io_bytes) {
545                 mirror_method = ret & BDRV_BLOCK_ZERO ?
546                                     MIRROR_METHOD_ZERO :
547                                     MIRROR_METHOD_DISCARD;
548             }
549         }
550 
551         while (s->in_flight >= MAX_IN_FLIGHT) {
552             trace_mirror_yield_in_flight(s, offset, s->in_flight);
553             mirror_wait_for_free_in_flight_slot(s);
554         }
555 
556         if (s->ret < 0) {
557             ret = 0;
558             goto fail;
559         }
560 
561         io_bytes = mirror_clip_bytes(s, offset, io_bytes);
562         io_bytes = mirror_perform(s, offset, io_bytes, mirror_method);
563         if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) {
564             io_bytes_acct = 0;
565         } else {
566             io_bytes_acct = io_bytes;
567         }
568         assert(io_bytes);
569         offset += io_bytes;
570         nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity);
571         delay_ns = block_job_ratelimit_get_delay(&s->common, io_bytes_acct);
572     }
573 
574     ret = delay_ns;
575 fail:
576     QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next);
577     qemu_co_queue_restart_all(&pseudo_op->waiting_requests);
578     g_free(pseudo_op);
579 
580     return ret;
581 }
582 
583 static void mirror_free_init(MirrorBlockJob *s)
584 {
585     int granularity = s->granularity;
586     size_t buf_size = s->buf_size;
587     uint8_t *buf = s->buf;
588 
589     assert(s->buf_free_count == 0);
590     QSIMPLEQ_INIT(&s->buf_free);
591     while (buf_size != 0) {
592         MirrorBuffer *cur = (MirrorBuffer *)buf;
593         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
594         s->buf_free_count++;
595         buf_size -= granularity;
596         buf += granularity;
597     }
598 }
599 
600 /* This is also used for the .pause callback. There is no matching
601  * mirror_resume() because mirror_run() will begin iterating again
602  * when the job is resumed.
603  */
604 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s)
605 {
606     while (s->in_flight > 0) {
607         mirror_wait_for_free_in_flight_slot(s);
608     }
609 }
610 
611 /**
612  * mirror_exit_common: handle both abort() and prepare() cases.
613  * for .prepare, returns 0 on success and -errno on failure.
614  * for .abort cases, denoted by abort = true, MUST return 0.
615  */
616 static int mirror_exit_common(Job *job)
617 {
618     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
619     BlockJob *bjob = &s->common;
620     MirrorBDSOpaque *bs_opaque = s->mirror_top_bs->opaque;
621     AioContext *replace_aio_context = NULL;
622     BlockDriverState *src = s->mirror_top_bs->backing->bs;
623     BlockDriverState *target_bs = blk_bs(s->target);
624     BlockDriverState *mirror_top_bs = s->mirror_top_bs;
625     Error *local_err = NULL;
626     bool abort = job->ret < 0;
627     int ret = 0;
628 
629     if (s->prepared) {
630         return 0;
631     }
632     s->prepared = true;
633 
634     if (bdrv_chain_contains(src, target_bs)) {
635         bdrv_unfreeze_backing_chain(mirror_top_bs, target_bs);
636     }
637 
638     bdrv_release_dirty_bitmap(src, s->dirty_bitmap);
639 
640     /* Make sure that the source BDS doesn't go away during bdrv_replace_node,
641      * before we can call bdrv_drained_end */
642     bdrv_ref(src);
643     bdrv_ref(mirror_top_bs);
644     bdrv_ref(target_bs);
645 
646     /* Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
647      * inserting target_bs at s->to_replace, where we might not be able to get
648      * these permissions.
649      *
650      * Note that blk_unref() alone doesn't necessarily drop permissions because
651      * we might be running nested inside mirror_drain(), which takes an extra
652      * reference, so use an explicit blk_set_perm() first. */
653     blk_set_perm(s->target, 0, BLK_PERM_ALL, &error_abort);
654     blk_unref(s->target);
655     s->target = NULL;
656 
657     /* We don't access the source any more. Dropping any WRITE/RESIZE is
658      * required before it could become a backing file of target_bs. */
659     bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
660                             &error_abort);
661     if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
662         BlockDriverState *backing = s->is_none_mode ? src : s->base;
663         if (backing_bs(target_bs) != backing) {
664             bdrv_set_backing_hd(target_bs, backing, &local_err);
665             if (local_err) {
666                 error_report_err(local_err);
667                 ret = -EPERM;
668             }
669         }
670     }
671 
672     if (s->to_replace) {
673         replace_aio_context = bdrv_get_aio_context(s->to_replace);
674         aio_context_acquire(replace_aio_context);
675     }
676 
677     if (s->should_complete && !abort) {
678         BlockDriverState *to_replace = s->to_replace ?: src;
679         bool ro = bdrv_is_read_only(to_replace);
680 
681         if (ro != bdrv_is_read_only(target_bs)) {
682             bdrv_reopen_set_read_only(target_bs, ro, NULL);
683         }
684 
685         /* The mirror job has no requests in flight any more, but we need to
686          * drain potential other users of the BDS before changing the graph. */
687         assert(s->in_drain);
688         bdrv_drained_begin(target_bs);
689         bdrv_replace_node(to_replace, target_bs, &local_err);
690         bdrv_drained_end(target_bs);
691         if (local_err) {
692             error_report_err(local_err);
693             ret = -EPERM;
694         }
695     }
696     if (s->to_replace) {
697         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
698         error_free(s->replace_blocker);
699         bdrv_unref(s->to_replace);
700     }
701     if (replace_aio_context) {
702         aio_context_release(replace_aio_context);
703     }
704     g_free(s->replaces);
705     bdrv_unref(target_bs);
706 
707     /* Remove the mirror filter driver from the graph. Before this, get rid of
708      * the blockers on the intermediate nodes so that the resulting state is
709      * valid. Also give up permissions on mirror_top_bs->backing, which might
710      * block the removal. */
711     block_job_remove_all_bdrv(bjob);
712     bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
713                             &error_abort);
714     bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
715 
716     /* We just changed the BDS the job BB refers to (with either or both of the
717      * bdrv_replace_node() calls), so switch the BB back so the cleanup does
718      * the right thing. We don't need any permissions any more now. */
719     blk_remove_bs(bjob->blk);
720     blk_set_perm(bjob->blk, 0, BLK_PERM_ALL, &error_abort);
721     blk_insert_bs(bjob->blk, mirror_top_bs, &error_abort);
722 
723     bs_opaque->job = NULL;
724 
725     bdrv_drained_end(src);
726     s->in_drain = false;
727     bdrv_unref(mirror_top_bs);
728     bdrv_unref(src);
729 
730     return ret;
731 }
732 
733 static int mirror_prepare(Job *job)
734 {
735     return mirror_exit_common(job);
736 }
737 
738 static void mirror_abort(Job *job)
739 {
740     int ret = mirror_exit_common(job);
741     assert(ret == 0);
742 }
743 
744 static void coroutine_fn mirror_throttle(MirrorBlockJob *s)
745 {
746     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
747 
748     if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) {
749         s->last_pause_ns = now;
750         job_sleep_ns(&s->common.job, 0);
751     } else {
752         job_pause_point(&s->common.job);
753     }
754 }
755 
756 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
757 {
758     int64_t offset;
759     BlockDriverState *base = s->base;
760     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
761     BlockDriverState *target_bs = blk_bs(s->target);
762     int ret;
763     int64_t count;
764 
765     if (base == NULL && !bdrv_has_zero_init(target_bs)) {
766         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
767             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
768             return 0;
769         }
770 
771         s->initial_zeroing_ongoing = true;
772         for (offset = 0; offset < s->bdev_length; ) {
773             int bytes = MIN(s->bdev_length - offset,
774                             QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
775 
776             mirror_throttle(s);
777 
778             if (job_is_cancelled(&s->common.job)) {
779                 s->initial_zeroing_ongoing = false;
780                 return 0;
781             }
782 
783             if (s->in_flight >= MAX_IN_FLIGHT) {
784                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
785                                    s->in_flight);
786                 mirror_wait_for_free_in_flight_slot(s);
787                 continue;
788             }
789 
790             mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO);
791             offset += bytes;
792         }
793 
794         mirror_wait_for_all_io(s);
795         s->initial_zeroing_ongoing = false;
796     }
797 
798     /* First part, loop on the sectors and initialize the dirty bitmap.  */
799     for (offset = 0; offset < s->bdev_length; ) {
800         /* Just to make sure we are not exceeding int limit. */
801         int bytes = MIN(s->bdev_length - offset,
802                         QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
803 
804         mirror_throttle(s);
805 
806         if (job_is_cancelled(&s->common.job)) {
807             return 0;
808         }
809 
810         ret = bdrv_is_allocated_above(bs, base, offset, bytes, &count);
811         if (ret < 0) {
812             return ret;
813         }
814 
815         assert(count);
816         if (ret == 1) {
817             bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
818         }
819         offset += count;
820     }
821     return 0;
822 }
823 
824 /* Called when going out of the streaming phase to flush the bulk of the
825  * data to the medium, or just before completing.
826  */
827 static int mirror_flush(MirrorBlockJob *s)
828 {
829     int ret = blk_flush(s->target);
830     if (ret < 0) {
831         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
832             s->ret = ret;
833         }
834     }
835     return ret;
836 }
837 
838 static int coroutine_fn mirror_run(Job *job, Error **errp)
839 {
840     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
841     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
842     BlockDriverState *target_bs = blk_bs(s->target);
843     bool need_drain = true;
844     int64_t length;
845     BlockDriverInfo bdi;
846     char backing_filename[2]; /* we only need 2 characters because we are only
847                                  checking for a NULL string */
848     int ret = 0;
849 
850     if (job_is_cancelled(&s->common.job)) {
851         goto immediate_exit;
852     }
853 
854     s->bdev_length = bdrv_getlength(bs);
855     if (s->bdev_length < 0) {
856         ret = s->bdev_length;
857         goto immediate_exit;
858     }
859 
860     /* Active commit must resize the base image if its size differs from the
861      * active layer. */
862     if (s->base == blk_bs(s->target)) {
863         int64_t base_length;
864 
865         base_length = blk_getlength(s->target);
866         if (base_length < 0) {
867             ret = base_length;
868             goto immediate_exit;
869         }
870 
871         if (s->bdev_length > base_length) {
872             ret = blk_truncate(s->target, s->bdev_length, PREALLOC_MODE_OFF,
873                                NULL);
874             if (ret < 0) {
875                 goto immediate_exit;
876             }
877         }
878     }
879 
880     if (s->bdev_length == 0) {
881         /* Transition to the READY state and wait for complete. */
882         job_transition_to_ready(&s->common.job);
883         s->synced = true;
884         s->actively_synced = true;
885         while (!job_is_cancelled(&s->common.job) && !s->should_complete) {
886             job_yield(&s->common.job);
887         }
888         s->common.job.cancelled = false;
889         goto immediate_exit;
890     }
891 
892     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
893     s->in_flight_bitmap = bitmap_new(length);
894 
895     /* If we have no backing file yet in the destination, we cannot let
896      * the destination do COW.  Instead, we copy sectors around the
897      * dirty data if needed.  We need a bitmap to do that.
898      */
899     bdrv_get_backing_filename(target_bs, backing_filename,
900                               sizeof(backing_filename));
901     if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) {
902         s->target_cluster_size = bdi.cluster_size;
903     } else {
904         s->target_cluster_size = BDRV_SECTOR_SIZE;
905     }
906     if (backing_filename[0] && !target_bs->backing &&
907         s->granularity < s->target_cluster_size) {
908         s->buf_size = MAX(s->buf_size, s->target_cluster_size);
909         s->cow_bitmap = bitmap_new(length);
910     }
911     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
912 
913     s->buf = qemu_try_blockalign(bs, s->buf_size);
914     if (s->buf == NULL) {
915         ret = -ENOMEM;
916         goto immediate_exit;
917     }
918 
919     mirror_free_init(s);
920 
921     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
922     if (!s->is_none_mode) {
923         ret = mirror_dirty_init(s);
924         if (ret < 0 || job_is_cancelled(&s->common.job)) {
925             goto immediate_exit;
926         }
927     }
928 
929     assert(!s->dbi);
930     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
931     for (;;) {
932         uint64_t delay_ns = 0;
933         int64_t cnt, delta;
934         bool should_complete;
935 
936         /* Do not start passive operations while there are active
937          * writes in progress */
938         while (s->in_active_write_counter) {
939             mirror_wait_for_any_operation(s, true);
940         }
941 
942         if (s->ret < 0) {
943             ret = s->ret;
944             goto immediate_exit;
945         }
946 
947         job_pause_point(&s->common.job);
948 
949         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
950         /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
951          * the number of bytes currently being processed; together those are
952          * the current remaining operation length */
953         job_progress_set_remaining(&s->common.job, s->bytes_in_flight + cnt);
954 
955         /* Note that even when no rate limit is applied we need to yield
956          * periodically with no pending I/O so that bdrv_drain_all() returns.
957          * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
958          * an error, or when the source is clean, whichever comes first. */
959         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
960         if (delta < BLOCK_JOB_SLICE_TIME &&
961             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
962             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
963                 (cnt == 0 && s->in_flight > 0)) {
964                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
965                 mirror_wait_for_free_in_flight_slot(s);
966                 continue;
967             } else if (cnt != 0) {
968                 delay_ns = mirror_iteration(s);
969             }
970         }
971 
972         should_complete = false;
973         if (s->in_flight == 0 && cnt == 0) {
974             trace_mirror_before_flush(s);
975             if (!s->synced) {
976                 if (mirror_flush(s) < 0) {
977                     /* Go check s->ret.  */
978                     continue;
979                 }
980                 /* We're out of the streaming phase.  From now on, if the job
981                  * is cancelled we will actually complete all pending I/O and
982                  * report completion.  This way, block-job-cancel will leave
983                  * the target in a consistent state.
984                  */
985                 job_transition_to_ready(&s->common.job);
986                 s->synced = true;
987                 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) {
988                     s->actively_synced = true;
989                 }
990             }
991 
992             should_complete = s->should_complete ||
993                 job_is_cancelled(&s->common.job);
994             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
995         }
996 
997         if (cnt == 0 && should_complete) {
998             /* The dirty bitmap is not updated while operations are pending.
999              * If we're about to exit, wait for pending operations before
1000              * calling bdrv_get_dirty_count(bs), or we may exit while the
1001              * source has dirty data to copy!
1002              *
1003              * Note that I/O can be submitted by the guest while
1004              * mirror_populate runs, so pause it now.  Before deciding
1005              * whether to switch to target check one last time if I/O has
1006              * come in the meanwhile, and if not flush the data to disk.
1007              */
1008             trace_mirror_before_drain(s, cnt);
1009 
1010             s->in_drain = true;
1011             bdrv_drained_begin(bs);
1012             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1013             if (cnt > 0 || mirror_flush(s) < 0) {
1014                 bdrv_drained_end(bs);
1015                 s->in_drain = false;
1016                 continue;
1017             }
1018 
1019             /* The two disks are in sync.  Exit and report successful
1020              * completion.
1021              */
1022             assert(QLIST_EMPTY(&bs->tracked_requests));
1023             s->common.job.cancelled = false;
1024             need_drain = false;
1025             break;
1026         }
1027 
1028         ret = 0;
1029 
1030         if (s->synced && !should_complete) {
1031             delay_ns = (s->in_flight == 0 &&
1032                         cnt == 0 ? BLOCK_JOB_SLICE_TIME : 0);
1033         }
1034         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
1035         job_sleep_ns(&s->common.job, delay_ns);
1036         if (job_is_cancelled(&s->common.job) &&
1037             (!s->synced || s->common.job.force_cancel))
1038         {
1039             break;
1040         }
1041         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1042     }
1043 
1044 immediate_exit:
1045     if (s->in_flight > 0) {
1046         /* We get here only if something went wrong.  Either the job failed,
1047          * or it was cancelled prematurely so that we do not guarantee that
1048          * the target is a copy of the source.
1049          */
1050         assert(ret < 0 || ((s->common.job.force_cancel || !s->synced) &&
1051                job_is_cancelled(&s->common.job)));
1052         assert(need_drain);
1053         mirror_wait_for_all_io(s);
1054     }
1055 
1056     assert(s->in_flight == 0);
1057     qemu_vfree(s->buf);
1058     g_free(s->cow_bitmap);
1059     g_free(s->in_flight_bitmap);
1060     bdrv_dirty_iter_free(s->dbi);
1061 
1062     if (need_drain) {
1063         s->in_drain = true;
1064         bdrv_drained_begin(bs);
1065     }
1066 
1067     return ret;
1068 }
1069 
1070 static void mirror_complete(Job *job, Error **errp)
1071 {
1072     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1073     BlockDriverState *target;
1074 
1075     target = blk_bs(s->target);
1076 
1077     if (!s->synced) {
1078         error_setg(errp, "The active block job '%s' cannot be completed",
1079                    job->id);
1080         return;
1081     }
1082 
1083     if (s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
1084         int ret;
1085 
1086         assert(!target->backing);
1087         ret = bdrv_open_backing_file(target, NULL, "backing", errp);
1088         if (ret < 0) {
1089             return;
1090         }
1091     }
1092 
1093     /* block all operations on to_replace bs */
1094     if (s->replaces) {
1095         AioContext *replace_aio_context;
1096 
1097         s->to_replace = bdrv_find_node(s->replaces);
1098         if (!s->to_replace) {
1099             error_setg(errp, "Node name '%s' not found", s->replaces);
1100             return;
1101         }
1102 
1103         replace_aio_context = bdrv_get_aio_context(s->to_replace);
1104         aio_context_acquire(replace_aio_context);
1105 
1106         /* TODO Translate this into permission system. Current definition of
1107          * GRAPH_MOD would require to request it for the parents; they might
1108          * not even be BlockDriverStates, however, so a BdrvChild can't address
1109          * them. May need redefinition of GRAPH_MOD. */
1110         error_setg(&s->replace_blocker,
1111                    "block device is in use by block-job-complete");
1112         bdrv_op_block_all(s->to_replace, s->replace_blocker);
1113         bdrv_ref(s->to_replace);
1114 
1115         aio_context_release(replace_aio_context);
1116     }
1117 
1118     s->should_complete = true;
1119     job_enter(job);
1120 }
1121 
1122 static void coroutine_fn mirror_pause(Job *job)
1123 {
1124     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1125 
1126     mirror_wait_for_all_io(s);
1127 }
1128 
1129 static bool mirror_drained_poll(BlockJob *job)
1130 {
1131     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1132 
1133     /* If the job isn't paused nor cancelled, we can't be sure that it won't
1134      * issue more requests. We make an exception if we've reached this point
1135      * from one of our own drain sections, to avoid a deadlock waiting for
1136      * ourselves.
1137      */
1138     if (!s->common.job.paused && !s->common.job.cancelled && !s->in_drain) {
1139         return true;
1140     }
1141 
1142     return !!s->in_flight;
1143 }
1144 
1145 static void mirror_attached_aio_context(BlockJob *job, AioContext *new_context)
1146 {
1147     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1148 
1149     blk_set_aio_context(s->target, new_context);
1150 }
1151 
1152 static void mirror_drain(BlockJob *job)
1153 {
1154     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1155 
1156     /* Need to keep a reference in case blk_drain triggers execution
1157      * of mirror_complete...
1158      */
1159     if (s->target) {
1160         BlockBackend *target = s->target;
1161         blk_ref(target);
1162         blk_drain(target);
1163         blk_unref(target);
1164     }
1165 }
1166 
1167 static const BlockJobDriver mirror_job_driver = {
1168     .job_driver = {
1169         .instance_size          = sizeof(MirrorBlockJob),
1170         .job_type               = JOB_TYPE_MIRROR,
1171         .free                   = block_job_free,
1172         .user_resume            = block_job_user_resume,
1173         .drain                  = block_job_drain,
1174         .run                    = mirror_run,
1175         .prepare                = mirror_prepare,
1176         .abort                  = mirror_abort,
1177         .pause                  = mirror_pause,
1178         .complete               = mirror_complete,
1179     },
1180     .drained_poll           = mirror_drained_poll,
1181     .attached_aio_context   = mirror_attached_aio_context,
1182     .drain                  = mirror_drain,
1183 };
1184 
1185 static const BlockJobDriver commit_active_job_driver = {
1186     .job_driver = {
1187         .instance_size          = sizeof(MirrorBlockJob),
1188         .job_type               = JOB_TYPE_COMMIT,
1189         .free                   = block_job_free,
1190         .user_resume            = block_job_user_resume,
1191         .drain                  = block_job_drain,
1192         .run                    = mirror_run,
1193         .prepare                = mirror_prepare,
1194         .abort                  = mirror_abort,
1195         .pause                  = mirror_pause,
1196         .complete               = mirror_complete,
1197     },
1198     .drained_poll           = mirror_drained_poll,
1199     .attached_aio_context   = mirror_attached_aio_context,
1200     .drain                  = mirror_drain,
1201 };
1202 
1203 static void coroutine_fn
1204 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method,
1205                      uint64_t offset, uint64_t bytes,
1206                      QEMUIOVector *qiov, int flags)
1207 {
1208     QEMUIOVector target_qiov;
1209     uint64_t dirty_offset = offset;
1210     uint64_t dirty_bytes;
1211 
1212     if (qiov) {
1213         qemu_iovec_init(&target_qiov, qiov->niov);
1214     }
1215 
1216     while (true) {
1217         bool valid_area;
1218         int ret;
1219 
1220         bdrv_dirty_bitmap_lock(job->dirty_bitmap);
1221         dirty_bytes = MIN(offset + bytes - dirty_offset, INT_MAX);
1222         valid_area = bdrv_dirty_bitmap_next_dirty_area(job->dirty_bitmap,
1223                                                        &dirty_offset,
1224                                                        &dirty_bytes);
1225         if (!valid_area) {
1226             bdrv_dirty_bitmap_unlock(job->dirty_bitmap);
1227             break;
1228         }
1229 
1230         bdrv_reset_dirty_bitmap_locked(job->dirty_bitmap,
1231                                        dirty_offset, dirty_bytes);
1232         bdrv_dirty_bitmap_unlock(job->dirty_bitmap);
1233 
1234         job_progress_increase_remaining(&job->common.job, dirty_bytes);
1235 
1236         assert(dirty_offset - offset <= SIZE_MAX);
1237         if (qiov) {
1238             qemu_iovec_reset(&target_qiov);
1239             qemu_iovec_concat(&target_qiov, qiov,
1240                               dirty_offset - offset, dirty_bytes);
1241         }
1242 
1243         switch (method) {
1244         case MIRROR_METHOD_COPY:
1245             ret = blk_co_pwritev(job->target, dirty_offset, dirty_bytes,
1246                                  qiov ? &target_qiov : NULL, flags);
1247             break;
1248 
1249         case MIRROR_METHOD_ZERO:
1250             assert(!qiov);
1251             ret = blk_co_pwrite_zeroes(job->target, dirty_offset, dirty_bytes,
1252                                        flags);
1253             break;
1254 
1255         case MIRROR_METHOD_DISCARD:
1256             assert(!qiov);
1257             ret = blk_co_pdiscard(job->target, dirty_offset, dirty_bytes);
1258             break;
1259 
1260         default:
1261             abort();
1262         }
1263 
1264         if (ret >= 0) {
1265             job_progress_update(&job->common.job, dirty_bytes);
1266         } else {
1267             BlockErrorAction action;
1268 
1269             bdrv_set_dirty_bitmap(job->dirty_bitmap, dirty_offset, dirty_bytes);
1270             job->actively_synced = false;
1271 
1272             action = mirror_error_action(job, false, -ret);
1273             if (action == BLOCK_ERROR_ACTION_REPORT) {
1274                 if (!job->ret) {
1275                     job->ret = ret;
1276                 }
1277                 break;
1278             }
1279         }
1280 
1281         dirty_offset += dirty_bytes;
1282     }
1283 
1284     if (qiov) {
1285         qemu_iovec_destroy(&target_qiov);
1286     }
1287 }
1288 
1289 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s,
1290                                                    uint64_t offset,
1291                                                    uint64_t bytes)
1292 {
1293     MirrorOp *op;
1294     uint64_t start_chunk = offset / s->granularity;
1295     uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
1296 
1297     op = g_new(MirrorOp, 1);
1298     *op = (MirrorOp){
1299         .s                  = s,
1300         .offset             = offset,
1301         .bytes              = bytes,
1302         .is_active_write    = true,
1303     };
1304     qemu_co_queue_init(&op->waiting_requests);
1305     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
1306 
1307     s->in_active_write_counter++;
1308 
1309     mirror_wait_on_conflicts(op, s, offset, bytes);
1310 
1311     bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1312 
1313     return op;
1314 }
1315 
1316 static void coroutine_fn active_write_settle(MirrorOp *op)
1317 {
1318     uint64_t start_chunk = op->offset / op->s->granularity;
1319     uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes,
1320                                       op->s->granularity);
1321 
1322     if (!--op->s->in_active_write_counter && op->s->actively_synced) {
1323         BdrvChild *source = op->s->mirror_top_bs->backing;
1324 
1325         if (QLIST_FIRST(&source->bs->parents) == source &&
1326             QLIST_NEXT(source, next_parent) == NULL)
1327         {
1328             /* Assert that we are back in sync once all active write
1329              * operations are settled.
1330              * Note that we can only assert this if the mirror node
1331              * is the source node's only parent. */
1332             assert(!bdrv_get_dirty_count(op->s->dirty_bitmap));
1333         }
1334     }
1335     bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1336     QTAILQ_REMOVE(&op->s->ops_in_flight, op, next);
1337     qemu_co_queue_restart_all(&op->waiting_requests);
1338     g_free(op);
1339 }
1340 
1341 static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs,
1342     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1343 {
1344     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1345 }
1346 
1347 static int coroutine_fn bdrv_mirror_top_do_write(BlockDriverState *bs,
1348     MirrorMethod method, uint64_t offset, uint64_t bytes, QEMUIOVector *qiov,
1349     int flags)
1350 {
1351     MirrorOp *op = NULL;
1352     MirrorBDSOpaque *s = bs->opaque;
1353     int ret = 0;
1354     bool copy_to_target;
1355 
1356     copy_to_target = s->job->ret >= 0 &&
1357                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1358 
1359     if (copy_to_target) {
1360         op = active_write_prepare(s->job, offset, bytes);
1361     }
1362 
1363     switch (method) {
1364     case MIRROR_METHOD_COPY:
1365         ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1366         break;
1367 
1368     case MIRROR_METHOD_ZERO:
1369         ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1370         break;
1371 
1372     case MIRROR_METHOD_DISCARD:
1373         ret = bdrv_co_pdiscard(bs->backing, offset, bytes);
1374         break;
1375 
1376     default:
1377         abort();
1378     }
1379 
1380     if (ret < 0) {
1381         goto out;
1382     }
1383 
1384     if (copy_to_target) {
1385         do_sync_target_write(s->job, method, offset, bytes, qiov, flags);
1386     }
1387 
1388 out:
1389     if (copy_to_target) {
1390         active_write_settle(op);
1391     }
1392     return ret;
1393 }
1394 
1395 static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs,
1396     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1397 {
1398     MirrorBDSOpaque *s = bs->opaque;
1399     QEMUIOVector bounce_qiov;
1400     void *bounce_buf;
1401     int ret = 0;
1402     bool copy_to_target;
1403 
1404     copy_to_target = s->job->ret >= 0 &&
1405                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1406 
1407     if (copy_to_target) {
1408         /* The guest might concurrently modify the data to write; but
1409          * the data on source and destination must match, so we have
1410          * to use a bounce buffer if we are going to write to the
1411          * target now. */
1412         bounce_buf = qemu_blockalign(bs, bytes);
1413         iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes);
1414 
1415         qemu_iovec_init(&bounce_qiov, 1);
1416         qemu_iovec_add(&bounce_qiov, bounce_buf, bytes);
1417         qiov = &bounce_qiov;
1418     }
1419 
1420     ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov,
1421                                    flags);
1422 
1423     if (copy_to_target) {
1424         qemu_iovec_destroy(&bounce_qiov);
1425         qemu_vfree(bounce_buf);
1426     }
1427 
1428     return ret;
1429 }
1430 
1431 static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs)
1432 {
1433     if (bs->backing == NULL) {
1434         /* we can be here after failed bdrv_append in mirror_start_job */
1435         return 0;
1436     }
1437     return bdrv_co_flush(bs->backing->bs);
1438 }
1439 
1440 static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs,
1441     int64_t offset, int bytes, BdrvRequestFlags flags)
1442 {
1443     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL,
1444                                     flags);
1445 }
1446 
1447 static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs,
1448     int64_t offset, int bytes)
1449 {
1450     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes,
1451                                     NULL, 0);
1452 }
1453 
1454 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs)
1455 {
1456     if (bs->backing == NULL) {
1457         /* we can be here after failed bdrv_attach_child in
1458          * bdrv_set_backing_hd */
1459         return;
1460     }
1461     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1462             bs->backing->bs->filename);
1463 }
1464 
1465 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1466                                        const BdrvChildRole *role,
1467                                        BlockReopenQueue *reopen_queue,
1468                                        uint64_t perm, uint64_t shared,
1469                                        uint64_t *nperm, uint64_t *nshared)
1470 {
1471     /* Must be able to forward guest writes to the real image */
1472     *nperm = 0;
1473     if (perm & BLK_PERM_WRITE) {
1474         *nperm |= BLK_PERM_WRITE;
1475     }
1476 
1477     *nshared = BLK_PERM_ALL;
1478 }
1479 
1480 /* Dummy node that provides consistent read to its users without requiring it
1481  * from its backing file and that allows writes on the backing file chain. */
1482 static BlockDriver bdrv_mirror_top = {
1483     .format_name                = "mirror_top",
1484     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1485     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1486     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1487     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1488     .bdrv_co_flush              = bdrv_mirror_top_flush,
1489     .bdrv_co_block_status       = bdrv_co_block_status_from_backing,
1490     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1491     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1492 };
1493 
1494 static void mirror_start_job(const char *job_id, BlockDriverState *bs,
1495                              int creation_flags, BlockDriverState *target,
1496                              const char *replaces, int64_t speed,
1497                              uint32_t granularity, int64_t buf_size,
1498                              BlockMirrorBackingMode backing_mode,
1499                              BlockdevOnError on_source_error,
1500                              BlockdevOnError on_target_error,
1501                              bool unmap,
1502                              BlockCompletionFunc *cb,
1503                              void *opaque,
1504                              const BlockJobDriver *driver,
1505                              bool is_none_mode, BlockDriverState *base,
1506                              bool auto_complete, const char *filter_node_name,
1507                              bool is_mirror, MirrorCopyMode copy_mode,
1508                              Error **errp)
1509 {
1510     MirrorBlockJob *s;
1511     MirrorBDSOpaque *bs_opaque;
1512     BlockDriverState *mirror_top_bs;
1513     bool target_graph_mod;
1514     bool target_is_backing;
1515     Error *local_err = NULL;
1516     int ret;
1517 
1518     if (granularity == 0) {
1519         granularity = bdrv_get_default_bitmap_granularity(target);
1520     }
1521 
1522     assert(is_power_of_2(granularity));
1523 
1524     if (buf_size < 0) {
1525         error_setg(errp, "Invalid parameter 'buf-size'");
1526         return;
1527     }
1528 
1529     if (buf_size == 0) {
1530         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1531     }
1532 
1533     if (bs == target) {
1534         error_setg(errp, "Can't mirror node into itself");
1535         return;
1536     }
1537 
1538     /* In the case of active commit, add dummy driver to provide consistent
1539      * reads on the top, while disabling it in the intermediate nodes, and make
1540      * the backing chain writable. */
1541     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1542                                          BDRV_O_RDWR, errp);
1543     if (mirror_top_bs == NULL) {
1544         return;
1545     }
1546     if (!filter_node_name) {
1547         mirror_top_bs->implicit = true;
1548     }
1549     mirror_top_bs->total_sectors = bs->total_sectors;
1550     mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
1551     mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
1552                                           BDRV_REQ_NO_FALLBACK;
1553     bs_opaque = g_new0(MirrorBDSOpaque, 1);
1554     mirror_top_bs->opaque = bs_opaque;
1555     bdrv_set_aio_context(mirror_top_bs, bdrv_get_aio_context(bs));
1556 
1557     /* bdrv_append takes ownership of the mirror_top_bs reference, need to keep
1558      * it alive until block_job_create() succeeds even if bs has no parent. */
1559     bdrv_ref(mirror_top_bs);
1560     bdrv_drained_begin(bs);
1561     bdrv_append(mirror_top_bs, bs, &local_err);
1562     bdrv_drained_end(bs);
1563 
1564     if (local_err) {
1565         bdrv_unref(mirror_top_bs);
1566         error_propagate(errp, local_err);
1567         return;
1568     }
1569 
1570     /* Make sure that the source is not resized while the job is running */
1571     s = block_job_create(job_id, driver, NULL, mirror_top_bs,
1572                          BLK_PERM_CONSISTENT_READ,
1573                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1574                          BLK_PERM_WRITE | BLK_PERM_GRAPH_MOD, speed,
1575                          creation_flags, cb, opaque, errp);
1576     if (!s) {
1577         goto fail;
1578     }
1579     bs_opaque->job = s;
1580 
1581     /* The block job now has a reference to this node */
1582     bdrv_unref(mirror_top_bs);
1583 
1584     s->mirror_top_bs = mirror_top_bs;
1585 
1586     /* No resize for the target either; while the mirror is still running, a
1587      * consistent read isn't necessarily possible. We could possibly allow
1588      * writes and graph modifications, though it would likely defeat the
1589      * purpose of a mirror, so leave them blocked for now.
1590      *
1591      * In the case of active commit, things look a bit different, though,
1592      * because the target is an already populated backing file in active use.
1593      * We can allow anything except resize there.*/
1594     target_is_backing = bdrv_chain_contains(bs, target);
1595     target_graph_mod = (backing_mode != MIRROR_LEAVE_BACKING_CHAIN);
1596     s->target = blk_new(BLK_PERM_WRITE | BLK_PERM_RESIZE |
1597                         (target_graph_mod ? BLK_PERM_GRAPH_MOD : 0),
1598                         BLK_PERM_WRITE_UNCHANGED |
1599                         (target_is_backing ? BLK_PERM_CONSISTENT_READ |
1600                                              BLK_PERM_WRITE |
1601                                              BLK_PERM_GRAPH_MOD : 0));
1602     ret = blk_insert_bs(s->target, target, errp);
1603     if (ret < 0) {
1604         goto fail;
1605     }
1606     if (is_mirror) {
1607         /* XXX: Mirror target could be a NBD server of target QEMU in the case
1608          * of non-shared block migration. To allow migration completion, we
1609          * have to allow "inactivate" of the target BB.  When that happens, we
1610          * know the job is drained, and the vcpus are stopped, so no write
1611          * operation will be performed. Block layer already has assertions to
1612          * ensure that. */
1613         blk_set_force_allow_inactivate(s->target);
1614     }
1615 
1616     s->replaces = g_strdup(replaces);
1617     s->on_source_error = on_source_error;
1618     s->on_target_error = on_target_error;
1619     s->is_none_mode = is_none_mode;
1620     s->backing_mode = backing_mode;
1621     s->copy_mode = copy_mode;
1622     s->base = base;
1623     s->granularity = granularity;
1624     s->buf_size = ROUND_UP(buf_size, granularity);
1625     s->unmap = unmap;
1626     if (auto_complete) {
1627         s->should_complete = true;
1628     }
1629 
1630     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1631     if (!s->dirty_bitmap) {
1632         goto fail;
1633     }
1634 
1635     ret = block_job_add_bdrv(&s->common, "source", bs, 0,
1636                              BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE |
1637                              BLK_PERM_CONSISTENT_READ,
1638                              errp);
1639     if (ret < 0) {
1640         goto fail;
1641     }
1642 
1643     /* Required permissions are already taken with blk_new() */
1644     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1645                        &error_abort);
1646 
1647     /* In commit_active_start() all intermediate nodes disappear, so
1648      * any jobs in them must be blocked */
1649     if (target_is_backing) {
1650         BlockDriverState *iter;
1651         for (iter = backing_bs(bs); iter != target; iter = backing_bs(iter)) {
1652             /* XXX BLK_PERM_WRITE needs to be allowed so we don't block
1653              * ourselves at s->base (if writes are blocked for a node, they are
1654              * also blocked for its backing file). The other options would be a
1655              * second filter driver above s->base (== target). */
1656             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1657                                      BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
1658                                      errp);
1659             if (ret < 0) {
1660                 goto fail;
1661             }
1662         }
1663 
1664         if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) {
1665             goto fail;
1666         }
1667     }
1668 
1669     QTAILQ_INIT(&s->ops_in_flight);
1670 
1671     trace_mirror_start(bs, s, opaque);
1672     job_start(&s->common.job);
1673     return;
1674 
1675 fail:
1676     if (s) {
1677         /* Make sure this BDS does not go away until we have completed the graph
1678          * changes below */
1679         bdrv_ref(mirror_top_bs);
1680 
1681         g_free(s->replaces);
1682         blk_unref(s->target);
1683         bs_opaque->job = NULL;
1684         if (s->dirty_bitmap) {
1685             bdrv_release_dirty_bitmap(bs, s->dirty_bitmap);
1686         }
1687         job_early_fail(&s->common.job);
1688     }
1689 
1690     bdrv_child_try_set_perm(mirror_top_bs->backing, 0, BLK_PERM_ALL,
1691                             &error_abort);
1692     bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
1693 
1694     bdrv_unref(mirror_top_bs);
1695 }
1696 
1697 void mirror_start(const char *job_id, BlockDriverState *bs,
1698                   BlockDriverState *target, const char *replaces,
1699                   int creation_flags, int64_t speed,
1700                   uint32_t granularity, int64_t buf_size,
1701                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1702                   BlockdevOnError on_source_error,
1703                   BlockdevOnError on_target_error,
1704                   bool unmap, const char *filter_node_name,
1705                   MirrorCopyMode copy_mode, Error **errp)
1706 {
1707     bool is_none_mode;
1708     BlockDriverState *base;
1709 
1710     if (mode == MIRROR_SYNC_MODE_INCREMENTAL) {
1711         error_setg(errp, "Sync mode 'incremental' not supported");
1712         return;
1713     }
1714     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1715     base = mode == MIRROR_SYNC_MODE_TOP ? backing_bs(bs) : NULL;
1716     mirror_start_job(job_id, bs, creation_flags, target, replaces,
1717                      speed, granularity, buf_size, backing_mode,
1718                      on_source_error, on_target_error, unmap, NULL, NULL,
1719                      &mirror_job_driver, is_none_mode, base, false,
1720                      filter_node_name, true, copy_mode, errp);
1721 }
1722 
1723 void commit_active_start(const char *job_id, BlockDriverState *bs,
1724                          BlockDriverState *base, int creation_flags,
1725                          int64_t speed, BlockdevOnError on_error,
1726                          const char *filter_node_name,
1727                          BlockCompletionFunc *cb, void *opaque,
1728                          bool auto_complete, Error **errp)
1729 {
1730     bool base_read_only;
1731     Error *local_err = NULL;
1732 
1733     base_read_only = bdrv_is_read_only(base);
1734 
1735     if (base_read_only) {
1736         if (bdrv_reopen_set_read_only(base, false, errp) < 0) {
1737             return;
1738         }
1739     }
1740 
1741     mirror_start_job(job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1742                      MIRROR_LEAVE_BACKING_CHAIN,
1743                      on_error, on_error, true, cb, opaque,
1744                      &commit_active_job_driver, false, base, auto_complete,
1745                      filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND,
1746                      &local_err);
1747     if (local_err) {
1748         error_propagate(errp, local_err);
1749         goto error_restore_flags;
1750     }
1751 
1752     return;
1753 
1754 error_restore_flags:
1755     /* ignore error and errp for bdrv_reopen, because we want to propagate
1756      * the original error */
1757     if (base_read_only) {
1758         bdrv_reopen_set_read_only(base, true, NULL);
1759     }
1760     return;
1761 }
1762