xref: /openbmc/qemu/block/mirror.c (revision 4b9fa0b4)
1 /*
2  * Image mirroring
3  *
4  * Copyright Red Hat, Inc. 2012
5  *
6  * Authors:
7  *  Paolo Bonzini  <pbonzini@redhat.com>
8  *
9  * This work is licensed under the terms of the GNU LGPL, version 2 or later.
10  * See the COPYING.LIB file in the top-level directory.
11  *
12  */
13 
14 #include "qemu/osdep.h"
15 #include "qemu/cutils.h"
16 #include "qemu/coroutine.h"
17 #include "qemu/range.h"
18 #include "trace.h"
19 #include "block/blockjob_int.h"
20 #include "block/block_int.h"
21 #include "sysemu/block-backend.h"
22 #include "qapi/error.h"
23 #include "qapi/qmp/qerror.h"
24 #include "qemu/ratelimit.h"
25 #include "qemu/bitmap.h"
26 
27 #define MAX_IN_FLIGHT 16
28 #define MAX_IO_BYTES (1 << 20) /* 1 Mb */
29 #define DEFAULT_MIRROR_BUF_SIZE (MAX_IN_FLIGHT * MAX_IO_BYTES)
30 
31 /* The mirroring buffer is a list of granularity-sized chunks.
32  * Free chunks are organized in a list.
33  */
34 typedef struct MirrorBuffer {
35     QSIMPLEQ_ENTRY(MirrorBuffer) next;
36 } MirrorBuffer;
37 
38 typedef struct MirrorOp MirrorOp;
39 
40 typedef struct MirrorBlockJob {
41     BlockJob common;
42     BlockBackend *target;
43     BlockDriverState *mirror_top_bs;
44     BlockDriverState *base;
45 
46     /* The name of the graph node to replace */
47     char *replaces;
48     /* The BDS to replace */
49     BlockDriverState *to_replace;
50     /* Used to block operations on the drive-mirror-replace target */
51     Error *replace_blocker;
52     bool is_none_mode;
53     BlockMirrorBackingMode backing_mode;
54     /* Whether the target image requires explicit zero-initialization */
55     bool zero_target;
56     MirrorCopyMode copy_mode;
57     BlockdevOnError on_source_error, on_target_error;
58     bool synced;
59     /* Set when the target is synced (dirty bitmap is clean, nothing
60      * in flight) and the job is running in active mode */
61     bool actively_synced;
62     bool should_complete;
63     int64_t granularity;
64     size_t buf_size;
65     int64_t bdev_length;
66     unsigned long *cow_bitmap;
67     BdrvDirtyBitmap *dirty_bitmap;
68     BdrvDirtyBitmapIter *dbi;
69     uint8_t *buf;
70     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
71     int buf_free_count;
72 
73     uint64_t last_pause_ns;
74     unsigned long *in_flight_bitmap;
75     int in_flight;
76     int64_t bytes_in_flight;
77     QTAILQ_HEAD(, MirrorOp) ops_in_flight;
78     int ret;
79     bool unmap;
80     int target_cluster_size;
81     int max_iov;
82     bool initial_zeroing_ongoing;
83     int in_active_write_counter;
84     bool prepared;
85     bool in_drain;
86 } MirrorBlockJob;
87 
88 typedef struct MirrorBDSOpaque {
89     MirrorBlockJob *job;
90     bool stop;
91 } MirrorBDSOpaque;
92 
93 struct MirrorOp {
94     MirrorBlockJob *s;
95     QEMUIOVector qiov;
96     int64_t offset;
97     uint64_t bytes;
98 
99     /* The pointee is set by mirror_co_read(), mirror_co_zero(), and
100      * mirror_co_discard() before yielding for the first time */
101     int64_t *bytes_handled;
102 
103     bool is_pseudo_op;
104     bool is_active_write;
105     CoQueue waiting_requests;
106 
107     QTAILQ_ENTRY(MirrorOp) next;
108 };
109 
110 typedef enum MirrorMethod {
111     MIRROR_METHOD_COPY,
112     MIRROR_METHOD_ZERO,
113     MIRROR_METHOD_DISCARD,
114 } MirrorMethod;
115 
116 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
117                                             int error)
118 {
119     s->synced = false;
120     s->actively_synced = false;
121     if (read) {
122         return block_job_error_action(&s->common, s->on_source_error,
123                                       true, error);
124     } else {
125         return block_job_error_action(&s->common, s->on_target_error,
126                                       false, error);
127     }
128 }
129 
130 static void coroutine_fn mirror_wait_on_conflicts(MirrorOp *self,
131                                                   MirrorBlockJob *s,
132                                                   uint64_t offset,
133                                                   uint64_t bytes)
134 {
135     uint64_t self_start_chunk = offset / s->granularity;
136     uint64_t self_end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
137     uint64_t self_nb_chunks = self_end_chunk - self_start_chunk;
138 
139     while (find_next_bit(s->in_flight_bitmap, self_end_chunk,
140                          self_start_chunk) < self_end_chunk &&
141            s->ret >= 0)
142     {
143         MirrorOp *op;
144 
145         QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
146             uint64_t op_start_chunk = op->offset / s->granularity;
147             uint64_t op_nb_chunks = DIV_ROUND_UP(op->offset + op->bytes,
148                                                  s->granularity) -
149                                     op_start_chunk;
150 
151             if (op == self) {
152                 continue;
153             }
154 
155             if (ranges_overlap(self_start_chunk, self_nb_chunks,
156                                op_start_chunk, op_nb_chunks))
157             {
158                 qemu_co_queue_wait(&op->waiting_requests, NULL);
159                 break;
160             }
161         }
162     }
163 }
164 
165 static void coroutine_fn mirror_iteration_done(MirrorOp *op, int ret)
166 {
167     MirrorBlockJob *s = op->s;
168     struct iovec *iov;
169     int64_t chunk_num;
170     int i, nb_chunks;
171 
172     trace_mirror_iteration_done(s, op->offset, op->bytes, ret);
173 
174     s->in_flight--;
175     s->bytes_in_flight -= op->bytes;
176     iov = op->qiov.iov;
177     for (i = 0; i < op->qiov.niov; i++) {
178         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
179         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
180         s->buf_free_count++;
181     }
182 
183     chunk_num = op->offset / s->granularity;
184     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
185 
186     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
187     QTAILQ_REMOVE(&s->ops_in_flight, op, next);
188     if (ret >= 0) {
189         if (s->cow_bitmap) {
190             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
191         }
192         if (!s->initial_zeroing_ongoing) {
193             job_progress_update(&s->common.job, op->bytes);
194         }
195     }
196     qemu_iovec_destroy(&op->qiov);
197 
198     qemu_co_queue_restart_all(&op->waiting_requests);
199     g_free(op);
200 }
201 
202 static void coroutine_fn mirror_write_complete(MirrorOp *op, int ret)
203 {
204     MirrorBlockJob *s = op->s;
205 
206     if (ret < 0) {
207         BlockErrorAction action;
208 
209         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
210         action = mirror_error_action(s, false, -ret);
211         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
212             s->ret = ret;
213         }
214     }
215 
216     mirror_iteration_done(op, ret);
217 }
218 
219 static void coroutine_fn mirror_read_complete(MirrorOp *op, int ret)
220 {
221     MirrorBlockJob *s = op->s;
222 
223     if (ret < 0) {
224         BlockErrorAction action;
225 
226         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
227         action = mirror_error_action(s, true, -ret);
228         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
229             s->ret = ret;
230         }
231 
232         mirror_iteration_done(op, ret);
233         return;
234     }
235 
236     ret = blk_co_pwritev(s->target, op->offset, op->qiov.size, &op->qiov, 0);
237     mirror_write_complete(op, ret);
238 }
239 
240 /* Clip bytes relative to offset to not exceed end-of-file */
241 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s,
242                                         int64_t offset,
243                                         int64_t bytes)
244 {
245     return MIN(bytes, s->bdev_length - offset);
246 }
247 
248 /* Round offset and/or bytes to target cluster if COW is needed, and
249  * return the offset of the adjusted tail against original. */
250 static int mirror_cow_align(MirrorBlockJob *s, int64_t *offset,
251                             uint64_t *bytes)
252 {
253     bool need_cow;
254     int ret = 0;
255     int64_t align_offset = *offset;
256     int64_t align_bytes = *bytes;
257     int max_bytes = s->granularity * s->max_iov;
258 
259     need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap);
260     need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity,
261                           s->cow_bitmap);
262     if (need_cow) {
263         bdrv_round_to_clusters(blk_bs(s->target), *offset, *bytes,
264                                &align_offset, &align_bytes);
265     }
266 
267     if (align_bytes > max_bytes) {
268         align_bytes = max_bytes;
269         if (need_cow) {
270             align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size);
271         }
272     }
273     /* Clipping may result in align_bytes unaligned to chunk boundary, but
274      * that doesn't matter because it's already the end of source image. */
275     align_bytes = mirror_clip_bytes(s, align_offset, align_bytes);
276 
277     ret = align_offset + align_bytes - (*offset + *bytes);
278     *offset = align_offset;
279     *bytes = align_bytes;
280     assert(ret >= 0);
281     return ret;
282 }
283 
284 static inline void coroutine_fn
285 mirror_wait_for_any_operation(MirrorBlockJob *s, bool active)
286 {
287     MirrorOp *op;
288 
289     QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
290         /* Do not wait on pseudo ops, because it may in turn wait on
291          * some other operation to start, which may in fact be the
292          * caller of this function.  Since there is only one pseudo op
293          * at any given time, we will always find some real operation
294          * to wait on. */
295         if (!op->is_pseudo_op && op->is_active_write == active) {
296             qemu_co_queue_wait(&op->waiting_requests, NULL);
297             return;
298         }
299     }
300     abort();
301 }
302 
303 static inline void coroutine_fn
304 mirror_wait_for_free_in_flight_slot(MirrorBlockJob *s)
305 {
306     /* Only non-active operations use up in-flight slots */
307     mirror_wait_for_any_operation(s, false);
308 }
309 
310 /* Perform a mirror copy operation.
311  *
312  * *op->bytes_handled is set to the number of bytes copied after and
313  * including offset, excluding any bytes copied prior to offset due
314  * to alignment.  This will be op->bytes if no alignment is necessary,
315  * or (new_end - op->offset) if the tail is rounded up or down due to
316  * alignment or buffer limit.
317  */
318 static void coroutine_fn mirror_co_read(void *opaque)
319 {
320     MirrorOp *op = opaque;
321     MirrorBlockJob *s = op->s;
322     int nb_chunks;
323     uint64_t ret;
324     uint64_t max_bytes;
325 
326     max_bytes = s->granularity * s->max_iov;
327 
328     /* We can only handle as much as buf_size at a time. */
329     op->bytes = MIN(s->buf_size, MIN(max_bytes, op->bytes));
330     assert(op->bytes);
331     assert(op->bytes < BDRV_REQUEST_MAX_BYTES);
332     *op->bytes_handled = op->bytes;
333 
334     if (s->cow_bitmap) {
335         *op->bytes_handled += mirror_cow_align(s, &op->offset, &op->bytes);
336     }
337     /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */
338     assert(*op->bytes_handled <= UINT_MAX);
339     assert(op->bytes <= s->buf_size);
340     /* The offset is granularity-aligned because:
341      * 1) Caller passes in aligned values;
342      * 2) mirror_cow_align is used only when target cluster is larger. */
343     assert(QEMU_IS_ALIGNED(op->offset, s->granularity));
344     /* The range is sector-aligned, since bdrv_getlength() rounds up. */
345     assert(QEMU_IS_ALIGNED(op->bytes, BDRV_SECTOR_SIZE));
346     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
347 
348     while (s->buf_free_count < nb_chunks) {
349         trace_mirror_yield_in_flight(s, op->offset, s->in_flight);
350         mirror_wait_for_free_in_flight_slot(s);
351     }
352 
353     /* Now make a QEMUIOVector taking enough granularity-sized chunks
354      * from s->buf_free.
355      */
356     qemu_iovec_init(&op->qiov, nb_chunks);
357     while (nb_chunks-- > 0) {
358         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
359         size_t remaining = op->bytes - op->qiov.size;
360 
361         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
362         s->buf_free_count--;
363         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
364     }
365 
366     /* Copy the dirty cluster.  */
367     s->in_flight++;
368     s->bytes_in_flight += op->bytes;
369     trace_mirror_one_iteration(s, op->offset, op->bytes);
370 
371     ret = bdrv_co_preadv(s->mirror_top_bs->backing, op->offset, op->bytes,
372                          &op->qiov, 0);
373     mirror_read_complete(op, ret);
374 }
375 
376 static void coroutine_fn mirror_co_zero(void *opaque)
377 {
378     MirrorOp *op = opaque;
379     int ret;
380 
381     op->s->in_flight++;
382     op->s->bytes_in_flight += op->bytes;
383     *op->bytes_handled = op->bytes;
384 
385     ret = blk_co_pwrite_zeroes(op->s->target, op->offset, op->bytes,
386                                op->s->unmap ? BDRV_REQ_MAY_UNMAP : 0);
387     mirror_write_complete(op, ret);
388 }
389 
390 static void coroutine_fn mirror_co_discard(void *opaque)
391 {
392     MirrorOp *op = opaque;
393     int ret;
394 
395     op->s->in_flight++;
396     op->s->bytes_in_flight += op->bytes;
397     *op->bytes_handled = op->bytes;
398 
399     ret = blk_co_pdiscard(op->s->target, op->offset, op->bytes);
400     mirror_write_complete(op, ret);
401 }
402 
403 static unsigned mirror_perform(MirrorBlockJob *s, int64_t offset,
404                                unsigned bytes, MirrorMethod mirror_method)
405 {
406     MirrorOp *op;
407     Coroutine *co;
408     int64_t bytes_handled = -1;
409 
410     op = g_new(MirrorOp, 1);
411     *op = (MirrorOp){
412         .s              = s,
413         .offset         = offset,
414         .bytes          = bytes,
415         .bytes_handled  = &bytes_handled,
416     };
417     qemu_co_queue_init(&op->waiting_requests);
418 
419     switch (mirror_method) {
420     case MIRROR_METHOD_COPY:
421         co = qemu_coroutine_create(mirror_co_read, op);
422         break;
423     case MIRROR_METHOD_ZERO:
424         co = qemu_coroutine_create(mirror_co_zero, op);
425         break;
426     case MIRROR_METHOD_DISCARD:
427         co = qemu_coroutine_create(mirror_co_discard, op);
428         break;
429     default:
430         abort();
431     }
432 
433     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
434     qemu_coroutine_enter(co);
435     /* At this point, ownership of op has been moved to the coroutine
436      * and the object may already be freed */
437 
438     /* Assert that this value has been set */
439     assert(bytes_handled >= 0);
440 
441     /* Same assertion as in mirror_co_read() (and for mirror_co_read()
442      * and mirror_co_discard(), bytes_handled == op->bytes, which
443      * is the @bytes parameter given to this function) */
444     assert(bytes_handled <= UINT_MAX);
445     return bytes_handled;
446 }
447 
448 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
449 {
450     BlockDriverState *source = s->mirror_top_bs->backing->bs;
451     MirrorOp *pseudo_op;
452     int64_t offset;
453     uint64_t delay_ns = 0, ret = 0;
454     /* At least the first dirty chunk is mirrored in one iteration. */
455     int nb_chunks = 1;
456     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
457     int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES);
458 
459     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
460     offset = bdrv_dirty_iter_next(s->dbi);
461     if (offset < 0) {
462         bdrv_set_dirty_iter(s->dbi, 0);
463         offset = bdrv_dirty_iter_next(s->dbi);
464         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
465         assert(offset >= 0);
466     }
467     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
468 
469     mirror_wait_on_conflicts(NULL, s, offset, 1);
470 
471     job_pause_point(&s->common.job);
472 
473     /* Find the number of consective dirty chunks following the first dirty
474      * one, and wait for in flight requests in them. */
475     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
476     while (nb_chunks * s->granularity < s->buf_size) {
477         int64_t next_dirty;
478         int64_t next_offset = offset + nb_chunks * s->granularity;
479         int64_t next_chunk = next_offset / s->granularity;
480         if (next_offset >= s->bdev_length ||
481             !bdrv_dirty_bitmap_get_locked(s->dirty_bitmap, next_offset)) {
482             break;
483         }
484         if (test_bit(next_chunk, s->in_flight_bitmap)) {
485             break;
486         }
487 
488         next_dirty = bdrv_dirty_iter_next(s->dbi);
489         if (next_dirty > next_offset || next_dirty < 0) {
490             /* The bitmap iterator's cache is stale, refresh it */
491             bdrv_set_dirty_iter(s->dbi, next_offset);
492             next_dirty = bdrv_dirty_iter_next(s->dbi);
493         }
494         assert(next_dirty == next_offset);
495         nb_chunks++;
496     }
497 
498     /* Clear dirty bits before querying the block status, because
499      * calling bdrv_block_status_above could yield - if some blocks are
500      * marked dirty in this window, we need to know.
501      */
502     bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset,
503                                    nb_chunks * s->granularity);
504     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
505 
506     /* Before claiming an area in the in-flight bitmap, we have to
507      * create a MirrorOp for it so that conflicting requests can wait
508      * for it.  mirror_perform() will create the real MirrorOps later,
509      * for now we just create a pseudo operation that will wake up all
510      * conflicting requests once all real operations have been
511      * launched. */
512     pseudo_op = g_new(MirrorOp, 1);
513     *pseudo_op = (MirrorOp){
514         .offset         = offset,
515         .bytes          = nb_chunks * s->granularity,
516         .is_pseudo_op   = true,
517     };
518     qemu_co_queue_init(&pseudo_op->waiting_requests);
519     QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next);
520 
521     bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks);
522     while (nb_chunks > 0 && offset < s->bdev_length) {
523         int ret;
524         int64_t io_bytes;
525         int64_t io_bytes_acct;
526         MirrorMethod mirror_method = MIRROR_METHOD_COPY;
527 
528         assert(!(offset % s->granularity));
529         ret = bdrv_block_status_above(source, NULL, offset,
530                                       nb_chunks * s->granularity,
531                                       &io_bytes, NULL, NULL);
532         if (ret < 0) {
533             io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes);
534         } else if (ret & BDRV_BLOCK_DATA) {
535             io_bytes = MIN(io_bytes, max_io_bytes);
536         }
537 
538         io_bytes -= io_bytes % s->granularity;
539         if (io_bytes < s->granularity) {
540             io_bytes = s->granularity;
541         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
542             int64_t target_offset;
543             int64_t target_bytes;
544             bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes,
545                                    &target_offset, &target_bytes);
546             if (target_offset == offset &&
547                 target_bytes == io_bytes) {
548                 mirror_method = ret & BDRV_BLOCK_ZERO ?
549                                     MIRROR_METHOD_ZERO :
550                                     MIRROR_METHOD_DISCARD;
551             }
552         }
553 
554         while (s->in_flight >= MAX_IN_FLIGHT) {
555             trace_mirror_yield_in_flight(s, offset, s->in_flight);
556             mirror_wait_for_free_in_flight_slot(s);
557         }
558 
559         if (s->ret < 0) {
560             ret = 0;
561             goto fail;
562         }
563 
564         io_bytes = mirror_clip_bytes(s, offset, io_bytes);
565         io_bytes = mirror_perform(s, offset, io_bytes, mirror_method);
566         if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) {
567             io_bytes_acct = 0;
568         } else {
569             io_bytes_acct = io_bytes;
570         }
571         assert(io_bytes);
572         offset += io_bytes;
573         nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity);
574         delay_ns = block_job_ratelimit_get_delay(&s->common, io_bytes_acct);
575     }
576 
577     ret = delay_ns;
578 fail:
579     QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next);
580     qemu_co_queue_restart_all(&pseudo_op->waiting_requests);
581     g_free(pseudo_op);
582 
583     return ret;
584 }
585 
586 static void mirror_free_init(MirrorBlockJob *s)
587 {
588     int granularity = s->granularity;
589     size_t buf_size = s->buf_size;
590     uint8_t *buf = s->buf;
591 
592     assert(s->buf_free_count == 0);
593     QSIMPLEQ_INIT(&s->buf_free);
594     while (buf_size != 0) {
595         MirrorBuffer *cur = (MirrorBuffer *)buf;
596         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
597         s->buf_free_count++;
598         buf_size -= granularity;
599         buf += granularity;
600     }
601 }
602 
603 /* This is also used for the .pause callback. There is no matching
604  * mirror_resume() because mirror_run() will begin iterating again
605  * when the job is resumed.
606  */
607 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s)
608 {
609     while (s->in_flight > 0) {
610         mirror_wait_for_free_in_flight_slot(s);
611     }
612 }
613 
614 /**
615  * mirror_exit_common: handle both abort() and prepare() cases.
616  * for .prepare, returns 0 on success and -errno on failure.
617  * for .abort cases, denoted by abort = true, MUST return 0.
618  */
619 static int mirror_exit_common(Job *job)
620 {
621     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
622     BlockJob *bjob = &s->common;
623     MirrorBDSOpaque *bs_opaque = s->mirror_top_bs->opaque;
624     AioContext *replace_aio_context = NULL;
625     BlockDriverState *src = s->mirror_top_bs->backing->bs;
626     BlockDriverState *target_bs = blk_bs(s->target);
627     BlockDriverState *mirror_top_bs = s->mirror_top_bs;
628     Error *local_err = NULL;
629     bool abort = job->ret < 0;
630     int ret = 0;
631 
632     if (s->prepared) {
633         return 0;
634     }
635     s->prepared = true;
636 
637     if (bdrv_chain_contains(src, target_bs)) {
638         bdrv_unfreeze_backing_chain(mirror_top_bs, target_bs);
639     }
640 
641     bdrv_release_dirty_bitmap(s->dirty_bitmap);
642 
643     /* Make sure that the source BDS doesn't go away during bdrv_replace_node,
644      * before we can call bdrv_drained_end */
645     bdrv_ref(src);
646     bdrv_ref(mirror_top_bs);
647     bdrv_ref(target_bs);
648 
649     /*
650      * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
651      * inserting target_bs at s->to_replace, where we might not be able to get
652      * these permissions.
653      */
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. Not having
659      * these permissions any more means that we can't allow any new requests on
660      * mirror_top_bs from now on, so keep it drained. */
661     bdrv_drained_begin(mirror_top_bs);
662     bs_opaque->stop = true;
663     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
664                              &error_abort);
665     if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
666         BlockDriverState *backing = s->is_none_mode ? src : s->base;
667         if (backing_bs(target_bs) != backing) {
668             bdrv_set_backing_hd(target_bs, backing, &local_err);
669             if (local_err) {
670                 error_report_err(local_err);
671                 ret = -EPERM;
672             }
673         }
674     }
675 
676     if (s->to_replace) {
677         replace_aio_context = bdrv_get_aio_context(s->to_replace);
678         aio_context_acquire(replace_aio_context);
679     }
680 
681     if (s->should_complete && !abort) {
682         BlockDriverState *to_replace = s->to_replace ?: src;
683         bool ro = bdrv_is_read_only(to_replace);
684 
685         if (ro != bdrv_is_read_only(target_bs)) {
686             bdrv_reopen_set_read_only(target_bs, ro, NULL);
687         }
688 
689         /* The mirror job has no requests in flight any more, but we need to
690          * drain potential other users of the BDS before changing the graph. */
691         assert(s->in_drain);
692         bdrv_drained_begin(target_bs);
693         bdrv_replace_node(to_replace, target_bs, &local_err);
694         bdrv_drained_end(target_bs);
695         if (local_err) {
696             error_report_err(local_err);
697             ret = -EPERM;
698         }
699     }
700     if (s->to_replace) {
701         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
702         error_free(s->replace_blocker);
703         bdrv_unref(s->to_replace);
704     }
705     if (replace_aio_context) {
706         aio_context_release(replace_aio_context);
707     }
708     g_free(s->replaces);
709     bdrv_unref(target_bs);
710 
711     /*
712      * Remove the mirror filter driver from the graph. Before this, get rid of
713      * the blockers on the intermediate nodes so that the resulting state is
714      * valid.
715      */
716     block_job_remove_all_bdrv(bjob);
717     bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
718 
719     /* We just changed the BDS the job BB refers to (with either or both of the
720      * bdrv_replace_node() calls), so switch the BB back so the cleanup does
721      * the right thing. We don't need any permissions any more now. */
722     blk_remove_bs(bjob->blk);
723     blk_set_perm(bjob->blk, 0, BLK_PERM_ALL, &error_abort);
724     blk_insert_bs(bjob->blk, mirror_top_bs, &error_abort);
725 
726     bs_opaque->job = NULL;
727 
728     bdrv_drained_end(src);
729     bdrv_drained_end(mirror_top_bs);
730     s->in_drain = false;
731     bdrv_unref(mirror_top_bs);
732     bdrv_unref(src);
733 
734     return ret;
735 }
736 
737 static int mirror_prepare(Job *job)
738 {
739     return mirror_exit_common(job);
740 }
741 
742 static void mirror_abort(Job *job)
743 {
744     int ret = mirror_exit_common(job);
745     assert(ret == 0);
746 }
747 
748 static void coroutine_fn mirror_throttle(MirrorBlockJob *s)
749 {
750     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
751 
752     if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) {
753         s->last_pause_ns = now;
754         job_sleep_ns(&s->common.job, 0);
755     } else {
756         job_pause_point(&s->common.job);
757     }
758 }
759 
760 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
761 {
762     int64_t offset;
763     BlockDriverState *base = s->base;
764     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
765     BlockDriverState *target_bs = blk_bs(s->target);
766     int ret;
767     int64_t count;
768 
769     if (s->zero_target) {
770         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
771             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
772             return 0;
773         }
774 
775         s->initial_zeroing_ongoing = true;
776         for (offset = 0; offset < s->bdev_length; ) {
777             int bytes = MIN(s->bdev_length - offset,
778                             QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
779 
780             mirror_throttle(s);
781 
782             if (job_is_cancelled(&s->common.job)) {
783                 s->initial_zeroing_ongoing = false;
784                 return 0;
785             }
786 
787             if (s->in_flight >= MAX_IN_FLIGHT) {
788                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
789                                    s->in_flight);
790                 mirror_wait_for_free_in_flight_slot(s);
791                 continue;
792             }
793 
794             mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO);
795             offset += bytes;
796         }
797 
798         mirror_wait_for_all_io(s);
799         s->initial_zeroing_ongoing = false;
800     }
801 
802     /* First part, loop on the sectors and initialize the dirty bitmap.  */
803     for (offset = 0; offset < s->bdev_length; ) {
804         /* Just to make sure we are not exceeding int limit. */
805         int bytes = MIN(s->bdev_length - offset,
806                         QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
807 
808         mirror_throttle(s);
809 
810         if (job_is_cancelled(&s->common.job)) {
811             return 0;
812         }
813 
814         ret = bdrv_is_allocated_above(bs, base, false, offset, bytes, &count);
815         if (ret < 0) {
816             return ret;
817         }
818 
819         assert(count);
820         if (ret == 1) {
821             bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
822         }
823         offset += count;
824     }
825     return 0;
826 }
827 
828 /* Called when going out of the streaming phase to flush the bulk of the
829  * data to the medium, or just before completing.
830  */
831 static int mirror_flush(MirrorBlockJob *s)
832 {
833     int ret = blk_flush(s->target);
834     if (ret < 0) {
835         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
836             s->ret = ret;
837         }
838     }
839     return ret;
840 }
841 
842 static int coroutine_fn mirror_run(Job *job, Error **errp)
843 {
844     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
845     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
846     BlockDriverState *target_bs = blk_bs(s->target);
847     bool need_drain = true;
848     int64_t length;
849     BlockDriverInfo bdi;
850     char backing_filename[2]; /* we only need 2 characters because we are only
851                                  checking for a NULL string */
852     int ret = 0;
853 
854     if (job_is_cancelled(&s->common.job)) {
855         goto immediate_exit;
856     }
857 
858     s->bdev_length = bdrv_getlength(bs);
859     if (s->bdev_length < 0) {
860         ret = s->bdev_length;
861         goto immediate_exit;
862     }
863 
864     /* Active commit must resize the base image if its size differs from the
865      * active layer. */
866     if (s->base == blk_bs(s->target)) {
867         int64_t base_length;
868 
869         base_length = blk_getlength(s->target);
870         if (base_length < 0) {
871             ret = base_length;
872             goto immediate_exit;
873         }
874 
875         if (s->bdev_length > base_length) {
876             ret = blk_truncate(s->target, s->bdev_length, PREALLOC_MODE_OFF,
877                                NULL);
878             if (ret < 0) {
879                 goto immediate_exit;
880             }
881         }
882     }
883 
884     if (s->bdev_length == 0) {
885         /* Transition to the READY state and wait for complete. */
886         job_transition_to_ready(&s->common.job);
887         s->synced = true;
888         s->actively_synced = true;
889         while (!job_is_cancelled(&s->common.job) && !s->should_complete) {
890             job_yield(&s->common.job);
891         }
892         s->common.job.cancelled = false;
893         goto immediate_exit;
894     }
895 
896     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
897     s->in_flight_bitmap = bitmap_new(length);
898 
899     /* If we have no backing file yet in the destination, we cannot let
900      * the destination do COW.  Instead, we copy sectors around the
901      * dirty data if needed.  We need a bitmap to do that.
902      */
903     bdrv_get_backing_filename(target_bs, backing_filename,
904                               sizeof(backing_filename));
905     if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) {
906         s->target_cluster_size = bdi.cluster_size;
907     } else {
908         s->target_cluster_size = BDRV_SECTOR_SIZE;
909     }
910     if (backing_filename[0] && !target_bs->backing &&
911         s->granularity < s->target_cluster_size) {
912         s->buf_size = MAX(s->buf_size, s->target_cluster_size);
913         s->cow_bitmap = bitmap_new(length);
914     }
915     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
916 
917     s->buf = qemu_try_blockalign(bs, s->buf_size);
918     if (s->buf == NULL) {
919         ret = -ENOMEM;
920         goto immediate_exit;
921     }
922 
923     mirror_free_init(s);
924 
925     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
926     if (!s->is_none_mode) {
927         ret = mirror_dirty_init(s);
928         if (ret < 0 || job_is_cancelled(&s->common.job)) {
929             goto immediate_exit;
930         }
931     }
932 
933     assert(!s->dbi);
934     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
935     for (;;) {
936         uint64_t delay_ns = 0;
937         int64_t cnt, delta;
938         bool should_complete;
939 
940         /* Do not start passive operations while there are active
941          * writes in progress */
942         while (s->in_active_write_counter) {
943             mirror_wait_for_any_operation(s, true);
944         }
945 
946         if (s->ret < 0) {
947             ret = s->ret;
948             goto immediate_exit;
949         }
950 
951         job_pause_point(&s->common.job);
952 
953         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
954         /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
955          * the number of bytes currently being processed; together those are
956          * the current remaining operation length */
957         job_progress_set_remaining(&s->common.job, s->bytes_in_flight + cnt);
958 
959         /* Note that even when no rate limit is applied we need to yield
960          * periodically with no pending I/O so that bdrv_drain_all() returns.
961          * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
962          * an error, or when the source is clean, whichever comes first. */
963         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
964         if (delta < BLOCK_JOB_SLICE_TIME &&
965             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
966             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
967                 (cnt == 0 && s->in_flight > 0)) {
968                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
969                 mirror_wait_for_free_in_flight_slot(s);
970                 continue;
971             } else if (cnt != 0) {
972                 delay_ns = mirror_iteration(s);
973             }
974         }
975 
976         should_complete = false;
977         if (s->in_flight == 0 && cnt == 0) {
978             trace_mirror_before_flush(s);
979             if (!s->synced) {
980                 if (mirror_flush(s) < 0) {
981                     /* Go check s->ret.  */
982                     continue;
983                 }
984                 /* We're out of the streaming phase.  From now on, if the job
985                  * is cancelled we will actually complete all pending I/O and
986                  * report completion.  This way, block-job-cancel will leave
987                  * the target in a consistent state.
988                  */
989                 job_transition_to_ready(&s->common.job);
990                 s->synced = true;
991                 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) {
992                     s->actively_synced = true;
993                 }
994             }
995 
996             should_complete = s->should_complete ||
997                 job_is_cancelled(&s->common.job);
998             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
999         }
1000 
1001         if (cnt == 0 && should_complete) {
1002             /* The dirty bitmap is not updated while operations are pending.
1003              * If we're about to exit, wait for pending operations before
1004              * calling bdrv_get_dirty_count(bs), or we may exit while the
1005              * source has dirty data to copy!
1006              *
1007              * Note that I/O can be submitted by the guest while
1008              * mirror_populate runs, so pause it now.  Before deciding
1009              * whether to switch to target check one last time if I/O has
1010              * come in the meanwhile, and if not flush the data to disk.
1011              */
1012             trace_mirror_before_drain(s, cnt);
1013 
1014             s->in_drain = true;
1015             bdrv_drained_begin(bs);
1016             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1017             if (cnt > 0 || mirror_flush(s) < 0) {
1018                 bdrv_drained_end(bs);
1019                 s->in_drain = false;
1020                 continue;
1021             }
1022 
1023             /* The two disks are in sync.  Exit and report successful
1024              * completion.
1025              */
1026             assert(QLIST_EMPTY(&bs->tracked_requests));
1027             s->common.job.cancelled = false;
1028             need_drain = false;
1029             break;
1030         }
1031 
1032         ret = 0;
1033 
1034         if (s->synced && !should_complete) {
1035             delay_ns = (s->in_flight == 0 &&
1036                         cnt == 0 ? BLOCK_JOB_SLICE_TIME : 0);
1037         }
1038         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
1039         job_sleep_ns(&s->common.job, delay_ns);
1040         if (job_is_cancelled(&s->common.job) &&
1041             (!s->synced || s->common.job.force_cancel))
1042         {
1043             break;
1044         }
1045         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1046     }
1047 
1048 immediate_exit:
1049     if (s->in_flight > 0) {
1050         /* We get here only if something went wrong.  Either the job failed,
1051          * or it was cancelled prematurely so that we do not guarantee that
1052          * the target is a copy of the source.
1053          */
1054         assert(ret < 0 || ((s->common.job.force_cancel || !s->synced) &&
1055                job_is_cancelled(&s->common.job)));
1056         assert(need_drain);
1057         mirror_wait_for_all_io(s);
1058     }
1059 
1060     assert(s->in_flight == 0);
1061     qemu_vfree(s->buf);
1062     g_free(s->cow_bitmap);
1063     g_free(s->in_flight_bitmap);
1064     bdrv_dirty_iter_free(s->dbi);
1065 
1066     if (need_drain) {
1067         s->in_drain = true;
1068         bdrv_drained_begin(bs);
1069     }
1070 
1071     return ret;
1072 }
1073 
1074 static void mirror_complete(Job *job, Error **errp)
1075 {
1076     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1077     BlockDriverState *target;
1078 
1079     target = blk_bs(s->target);
1080 
1081     if (!s->synced) {
1082         error_setg(errp, "The active block job '%s' cannot be completed",
1083                    job->id);
1084         return;
1085     }
1086 
1087     if (s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
1088         int ret;
1089 
1090         assert(!target->backing);
1091         ret = bdrv_open_backing_file(target, NULL, "backing", errp);
1092         if (ret < 0) {
1093             return;
1094         }
1095     }
1096 
1097     /* block all operations on to_replace bs */
1098     if (s->replaces) {
1099         AioContext *replace_aio_context;
1100 
1101         s->to_replace = bdrv_find_node(s->replaces);
1102         if (!s->to_replace) {
1103             error_setg(errp, "Node name '%s' not found", s->replaces);
1104             return;
1105         }
1106 
1107         replace_aio_context = bdrv_get_aio_context(s->to_replace);
1108         aio_context_acquire(replace_aio_context);
1109 
1110         /* TODO Translate this into permission system. Current definition of
1111          * GRAPH_MOD would require to request it for the parents; they might
1112          * not even be BlockDriverStates, however, so a BdrvChild can't address
1113          * them. May need redefinition of GRAPH_MOD. */
1114         error_setg(&s->replace_blocker,
1115                    "block device is in use by block-job-complete");
1116         bdrv_op_block_all(s->to_replace, s->replace_blocker);
1117         bdrv_ref(s->to_replace);
1118 
1119         aio_context_release(replace_aio_context);
1120     }
1121 
1122     s->should_complete = true;
1123     job_enter(job);
1124 }
1125 
1126 static void coroutine_fn mirror_pause(Job *job)
1127 {
1128     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1129 
1130     mirror_wait_for_all_io(s);
1131 }
1132 
1133 static bool mirror_drained_poll(BlockJob *job)
1134 {
1135     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1136 
1137     /* If the job isn't paused nor cancelled, we can't be sure that it won't
1138      * issue more requests. We make an exception if we've reached this point
1139      * from one of our own drain sections, to avoid a deadlock waiting for
1140      * ourselves.
1141      */
1142     if (!s->common.job.paused && !s->common.job.cancelled && !s->in_drain) {
1143         return true;
1144     }
1145 
1146     return !!s->in_flight;
1147 }
1148 
1149 static const BlockJobDriver mirror_job_driver = {
1150     .job_driver = {
1151         .instance_size          = sizeof(MirrorBlockJob),
1152         .job_type               = JOB_TYPE_MIRROR,
1153         .free                   = block_job_free,
1154         .user_resume            = block_job_user_resume,
1155         .run                    = mirror_run,
1156         .prepare                = mirror_prepare,
1157         .abort                  = mirror_abort,
1158         .pause                  = mirror_pause,
1159         .complete               = mirror_complete,
1160     },
1161     .drained_poll           = mirror_drained_poll,
1162 };
1163 
1164 static const BlockJobDriver commit_active_job_driver = {
1165     .job_driver = {
1166         .instance_size          = sizeof(MirrorBlockJob),
1167         .job_type               = JOB_TYPE_COMMIT,
1168         .free                   = block_job_free,
1169         .user_resume            = block_job_user_resume,
1170         .run                    = mirror_run,
1171         .prepare                = mirror_prepare,
1172         .abort                  = mirror_abort,
1173         .pause                  = mirror_pause,
1174         .complete               = mirror_complete,
1175     },
1176     .drained_poll           = mirror_drained_poll,
1177 };
1178 
1179 static void coroutine_fn
1180 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method,
1181                      uint64_t offset, uint64_t bytes,
1182                      QEMUIOVector *qiov, int flags)
1183 {
1184     QEMUIOVector target_qiov;
1185     uint64_t dirty_offset = offset;
1186     uint64_t dirty_bytes;
1187 
1188     if (qiov) {
1189         qemu_iovec_init(&target_qiov, qiov->niov);
1190     }
1191 
1192     while (true) {
1193         bool valid_area;
1194         int ret;
1195 
1196         bdrv_dirty_bitmap_lock(job->dirty_bitmap);
1197         dirty_bytes = MIN(offset + bytes - dirty_offset, INT_MAX);
1198         valid_area = bdrv_dirty_bitmap_next_dirty_area(job->dirty_bitmap,
1199                                                        &dirty_offset,
1200                                                        &dirty_bytes);
1201         if (!valid_area) {
1202             bdrv_dirty_bitmap_unlock(job->dirty_bitmap);
1203             break;
1204         }
1205 
1206         bdrv_reset_dirty_bitmap_locked(job->dirty_bitmap,
1207                                        dirty_offset, dirty_bytes);
1208         bdrv_dirty_bitmap_unlock(job->dirty_bitmap);
1209 
1210         job_progress_increase_remaining(&job->common.job, dirty_bytes);
1211 
1212         assert(dirty_offset - offset <= SIZE_MAX);
1213         if (qiov) {
1214             qemu_iovec_reset(&target_qiov);
1215             qemu_iovec_concat(&target_qiov, qiov,
1216                               dirty_offset - offset, dirty_bytes);
1217         }
1218 
1219         switch (method) {
1220         case MIRROR_METHOD_COPY:
1221             ret = blk_co_pwritev(job->target, dirty_offset, dirty_bytes,
1222                                  qiov ? &target_qiov : NULL, flags);
1223             break;
1224 
1225         case MIRROR_METHOD_ZERO:
1226             assert(!qiov);
1227             ret = blk_co_pwrite_zeroes(job->target, dirty_offset, dirty_bytes,
1228                                        flags);
1229             break;
1230 
1231         case MIRROR_METHOD_DISCARD:
1232             assert(!qiov);
1233             ret = blk_co_pdiscard(job->target, dirty_offset, dirty_bytes);
1234             break;
1235 
1236         default:
1237             abort();
1238         }
1239 
1240         if (ret >= 0) {
1241             job_progress_update(&job->common.job, dirty_bytes);
1242         } else {
1243             BlockErrorAction action;
1244 
1245             bdrv_set_dirty_bitmap(job->dirty_bitmap, dirty_offset, dirty_bytes);
1246             job->actively_synced = false;
1247 
1248             action = mirror_error_action(job, false, -ret);
1249             if (action == BLOCK_ERROR_ACTION_REPORT) {
1250                 if (!job->ret) {
1251                     job->ret = ret;
1252                 }
1253                 break;
1254             }
1255         }
1256 
1257         dirty_offset += dirty_bytes;
1258     }
1259 
1260     if (qiov) {
1261         qemu_iovec_destroy(&target_qiov);
1262     }
1263 }
1264 
1265 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s,
1266                                                    uint64_t offset,
1267                                                    uint64_t bytes)
1268 {
1269     MirrorOp *op;
1270     uint64_t start_chunk = offset / s->granularity;
1271     uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
1272 
1273     op = g_new(MirrorOp, 1);
1274     *op = (MirrorOp){
1275         .s                  = s,
1276         .offset             = offset,
1277         .bytes              = bytes,
1278         .is_active_write    = true,
1279     };
1280     qemu_co_queue_init(&op->waiting_requests);
1281     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
1282 
1283     s->in_active_write_counter++;
1284 
1285     mirror_wait_on_conflicts(op, s, offset, bytes);
1286 
1287     bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1288 
1289     return op;
1290 }
1291 
1292 static void coroutine_fn active_write_settle(MirrorOp *op)
1293 {
1294     uint64_t start_chunk = op->offset / op->s->granularity;
1295     uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes,
1296                                       op->s->granularity);
1297 
1298     if (!--op->s->in_active_write_counter && op->s->actively_synced) {
1299         BdrvChild *source = op->s->mirror_top_bs->backing;
1300 
1301         if (QLIST_FIRST(&source->bs->parents) == source &&
1302             QLIST_NEXT(source, next_parent) == NULL)
1303         {
1304             /* Assert that we are back in sync once all active write
1305              * operations are settled.
1306              * Note that we can only assert this if the mirror node
1307              * is the source node's only parent. */
1308             assert(!bdrv_get_dirty_count(op->s->dirty_bitmap));
1309         }
1310     }
1311     bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1312     QTAILQ_REMOVE(&op->s->ops_in_flight, op, next);
1313     qemu_co_queue_restart_all(&op->waiting_requests);
1314     g_free(op);
1315 }
1316 
1317 static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs,
1318     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1319 {
1320     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1321 }
1322 
1323 static int coroutine_fn bdrv_mirror_top_do_write(BlockDriverState *bs,
1324     MirrorMethod method, uint64_t offset, uint64_t bytes, QEMUIOVector *qiov,
1325     int flags)
1326 {
1327     MirrorOp *op = NULL;
1328     MirrorBDSOpaque *s = bs->opaque;
1329     int ret = 0;
1330     bool copy_to_target;
1331 
1332     copy_to_target = s->job->ret >= 0 &&
1333                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1334 
1335     if (copy_to_target) {
1336         op = active_write_prepare(s->job, offset, bytes);
1337     }
1338 
1339     switch (method) {
1340     case MIRROR_METHOD_COPY:
1341         ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1342         break;
1343 
1344     case MIRROR_METHOD_ZERO:
1345         ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1346         break;
1347 
1348     case MIRROR_METHOD_DISCARD:
1349         ret = bdrv_co_pdiscard(bs->backing, offset, bytes);
1350         break;
1351 
1352     default:
1353         abort();
1354     }
1355 
1356     if (ret < 0) {
1357         goto out;
1358     }
1359 
1360     if (copy_to_target) {
1361         do_sync_target_write(s->job, method, offset, bytes, qiov, flags);
1362     }
1363 
1364 out:
1365     if (copy_to_target) {
1366         active_write_settle(op);
1367     }
1368     return ret;
1369 }
1370 
1371 static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs,
1372     uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
1373 {
1374     MirrorBDSOpaque *s = bs->opaque;
1375     QEMUIOVector bounce_qiov;
1376     void *bounce_buf;
1377     int ret = 0;
1378     bool copy_to_target;
1379 
1380     copy_to_target = s->job->ret >= 0 &&
1381                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1382 
1383     if (copy_to_target) {
1384         /* The guest might concurrently modify the data to write; but
1385          * the data on source and destination must match, so we have
1386          * to use a bounce buffer if we are going to write to the
1387          * target now. */
1388         bounce_buf = qemu_blockalign(bs, bytes);
1389         iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes);
1390 
1391         qemu_iovec_init(&bounce_qiov, 1);
1392         qemu_iovec_add(&bounce_qiov, bounce_buf, bytes);
1393         qiov = &bounce_qiov;
1394     }
1395 
1396     ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov,
1397                                    flags);
1398 
1399     if (copy_to_target) {
1400         qemu_iovec_destroy(&bounce_qiov);
1401         qemu_vfree(bounce_buf);
1402     }
1403 
1404     return ret;
1405 }
1406 
1407 static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs)
1408 {
1409     if (bs->backing == NULL) {
1410         /* we can be here after failed bdrv_append in mirror_start_job */
1411         return 0;
1412     }
1413     return bdrv_co_flush(bs->backing->bs);
1414 }
1415 
1416 static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs,
1417     int64_t offset, int bytes, BdrvRequestFlags flags)
1418 {
1419     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL,
1420                                     flags);
1421 }
1422 
1423 static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs,
1424     int64_t offset, int bytes)
1425 {
1426     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes,
1427                                     NULL, 0);
1428 }
1429 
1430 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs)
1431 {
1432     if (bs->backing == NULL) {
1433         /* we can be here after failed bdrv_attach_child in
1434          * bdrv_set_backing_hd */
1435         return;
1436     }
1437     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1438             bs->backing->bs->filename);
1439 }
1440 
1441 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1442                                        const BdrvChildRole *role,
1443                                        BlockReopenQueue *reopen_queue,
1444                                        uint64_t perm, uint64_t shared,
1445                                        uint64_t *nperm, uint64_t *nshared)
1446 {
1447     MirrorBDSOpaque *s = bs->opaque;
1448 
1449     if (s->stop) {
1450         /*
1451          * If the job is to be stopped, we do not need to forward
1452          * anything to the real image.
1453          */
1454         *nperm = 0;
1455         *nshared = BLK_PERM_ALL;
1456         return;
1457     }
1458 
1459     /* Must be able to forward guest writes to the real image */
1460     *nperm = 0;
1461     if (perm & BLK_PERM_WRITE) {
1462         *nperm |= BLK_PERM_WRITE;
1463     }
1464 
1465     *nshared = BLK_PERM_ALL;
1466 }
1467 
1468 static void bdrv_mirror_top_refresh_limits(BlockDriverState *bs, Error **errp)
1469 {
1470     MirrorBDSOpaque *s = bs->opaque;
1471 
1472     if (s && s->job && s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING) {
1473         bs->bl.request_alignment = s->job->granularity;
1474     }
1475 }
1476 
1477 /* Dummy node that provides consistent read to its users without requiring it
1478  * from its backing file and that allows writes on the backing file chain. */
1479 static BlockDriver bdrv_mirror_top = {
1480     .format_name                = "mirror_top",
1481     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1482     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1483     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1484     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1485     .bdrv_co_flush              = bdrv_mirror_top_flush,
1486     .bdrv_co_block_status       = bdrv_co_block_status_from_backing,
1487     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1488     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1489     .bdrv_refresh_limits        = bdrv_mirror_top_refresh_limits,
1490 };
1491 
1492 static BlockJob *mirror_start_job(
1493                              const char *job_id, BlockDriverState *bs,
1494                              int creation_flags, BlockDriverState *target,
1495                              const char *replaces, int64_t speed,
1496                              uint32_t granularity, int64_t buf_size,
1497                              BlockMirrorBackingMode backing_mode,
1498                              bool zero_target,
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 NULL;
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 NULL;
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 NULL;
1545     }
1546     if (!filter_node_name) {
1547         mirror_top_bs->implicit = true;
1548     }
1549 
1550     /* So that we can always drop this node */
1551     mirror_top_bs->never_freeze = true;
1552 
1553     mirror_top_bs->total_sectors = bs->total_sectors;
1554     mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
1555     mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
1556                                           BDRV_REQ_NO_FALLBACK;
1557     bs_opaque = g_new0(MirrorBDSOpaque, 1);
1558     mirror_top_bs->opaque = bs_opaque;
1559 
1560     /* bdrv_append takes ownership of the mirror_top_bs reference, need to keep
1561      * it alive until block_job_create() succeeds even if bs has no parent. */
1562     bdrv_ref(mirror_top_bs);
1563     bdrv_drained_begin(bs);
1564     bdrv_append(mirror_top_bs, bs, &local_err);
1565     bdrv_drained_end(bs);
1566 
1567     if (local_err) {
1568         bdrv_unref(mirror_top_bs);
1569         error_propagate(errp, local_err);
1570         return NULL;
1571     }
1572 
1573     /* Make sure that the source is not resized while the job is running */
1574     s = block_job_create(job_id, driver, NULL, mirror_top_bs,
1575                          BLK_PERM_CONSISTENT_READ,
1576                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1577                          BLK_PERM_WRITE | BLK_PERM_GRAPH_MOD, speed,
1578                          creation_flags, cb, opaque, errp);
1579     if (!s) {
1580         goto fail;
1581     }
1582     bs_opaque->job = s;
1583 
1584     /* The block job now has a reference to this node */
1585     bdrv_unref(mirror_top_bs);
1586 
1587     s->mirror_top_bs = mirror_top_bs;
1588 
1589     /* No resize for the target either; while the mirror is still running, a
1590      * consistent read isn't necessarily possible. We could possibly allow
1591      * writes and graph modifications, though it would likely defeat the
1592      * purpose of a mirror, so leave them blocked for now.
1593      *
1594      * In the case of active commit, things look a bit different, though,
1595      * because the target is an already populated backing file in active use.
1596      * We can allow anything except resize there.*/
1597     target_is_backing = bdrv_chain_contains(bs, target);
1598     target_graph_mod = (backing_mode != MIRROR_LEAVE_BACKING_CHAIN);
1599     s->target = blk_new(s->common.job.aio_context,
1600                         BLK_PERM_WRITE | BLK_PERM_RESIZE |
1601                         (target_graph_mod ? BLK_PERM_GRAPH_MOD : 0),
1602                         BLK_PERM_WRITE_UNCHANGED |
1603                         (target_is_backing ? BLK_PERM_CONSISTENT_READ |
1604                                              BLK_PERM_WRITE |
1605                                              BLK_PERM_GRAPH_MOD : 0));
1606     ret = blk_insert_bs(s->target, target, errp);
1607     if (ret < 0) {
1608         goto fail;
1609     }
1610     if (is_mirror) {
1611         /* XXX: Mirror target could be a NBD server of target QEMU in the case
1612          * of non-shared block migration. To allow migration completion, we
1613          * have to allow "inactivate" of the target BB.  When that happens, we
1614          * know the job is drained, and the vcpus are stopped, so no write
1615          * operation will be performed. Block layer already has assertions to
1616          * ensure that. */
1617         blk_set_force_allow_inactivate(s->target);
1618     }
1619     blk_set_allow_aio_context_change(s->target, true);
1620     blk_set_disable_request_queuing(s->target, true);
1621 
1622     s->replaces = g_strdup(replaces);
1623     s->on_source_error = on_source_error;
1624     s->on_target_error = on_target_error;
1625     s->is_none_mode = is_none_mode;
1626     s->backing_mode = backing_mode;
1627     s->zero_target = zero_target;
1628     s->copy_mode = copy_mode;
1629     s->base = base;
1630     s->granularity = granularity;
1631     s->buf_size = ROUND_UP(buf_size, granularity);
1632     s->unmap = unmap;
1633     if (auto_complete) {
1634         s->should_complete = true;
1635     }
1636 
1637     /*
1638      * Must be called before we start tracking writes, but after
1639      *
1640      *     ((MirrorBlockJob *)
1641      *         ((MirrorBDSOpaque *)
1642      *             mirror_top_bs->opaque
1643      *         )->job
1644      *     )->copy_mode
1645      *
1646      * has the correct value.
1647      * (We start tracking writes as of the following
1648      * bdrv_create_dirty_bitmap() call.)
1649      */
1650     bdrv_refresh_limits(mirror_top_bs, &local_err);
1651     if (local_err) {
1652         error_propagate(errp, local_err);
1653         goto fail;
1654     }
1655 
1656     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1657     if (!s->dirty_bitmap) {
1658         goto fail;
1659     }
1660 
1661     ret = block_job_add_bdrv(&s->common, "source", bs, 0,
1662                              BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE |
1663                              BLK_PERM_CONSISTENT_READ,
1664                              errp);
1665     if (ret < 0) {
1666         goto fail;
1667     }
1668 
1669     /* Required permissions are already taken with blk_new() */
1670     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1671                        &error_abort);
1672 
1673     /* In commit_active_start() all intermediate nodes disappear, so
1674      * any jobs in them must be blocked */
1675     if (target_is_backing) {
1676         BlockDriverState *iter;
1677         for (iter = backing_bs(bs); iter != target; iter = backing_bs(iter)) {
1678             /* XXX BLK_PERM_WRITE needs to be allowed so we don't block
1679              * ourselves at s->base (if writes are blocked for a node, they are
1680              * also blocked for its backing file). The other options would be a
1681              * second filter driver above s->base (== target). */
1682             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1683                                      BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
1684                                      errp);
1685             if (ret < 0) {
1686                 goto fail;
1687             }
1688         }
1689 
1690         if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) {
1691             goto fail;
1692         }
1693     }
1694 
1695     QTAILQ_INIT(&s->ops_in_flight);
1696 
1697     trace_mirror_start(bs, s, opaque);
1698     job_start(&s->common.job);
1699 
1700     return &s->common;
1701 
1702 fail:
1703     if (s) {
1704         /* Make sure this BDS does not go away until we have completed the graph
1705          * changes below */
1706         bdrv_ref(mirror_top_bs);
1707 
1708         g_free(s->replaces);
1709         blk_unref(s->target);
1710         bs_opaque->job = NULL;
1711         if (s->dirty_bitmap) {
1712             bdrv_release_dirty_bitmap(s->dirty_bitmap);
1713         }
1714         job_early_fail(&s->common.job);
1715     }
1716 
1717     bs_opaque->stop = true;
1718     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
1719                              &error_abort);
1720     bdrv_replace_node(mirror_top_bs, backing_bs(mirror_top_bs), &error_abort);
1721 
1722     bdrv_unref(mirror_top_bs);
1723 
1724     return NULL;
1725 }
1726 
1727 void mirror_start(const char *job_id, BlockDriverState *bs,
1728                   BlockDriverState *target, const char *replaces,
1729                   int creation_flags, int64_t speed,
1730                   uint32_t granularity, int64_t buf_size,
1731                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1732                   bool zero_target,
1733                   BlockdevOnError on_source_error,
1734                   BlockdevOnError on_target_error,
1735                   bool unmap, const char *filter_node_name,
1736                   MirrorCopyMode copy_mode, Error **errp)
1737 {
1738     bool is_none_mode;
1739     BlockDriverState *base;
1740 
1741     if ((mode == MIRROR_SYNC_MODE_INCREMENTAL) ||
1742         (mode == MIRROR_SYNC_MODE_BITMAP)) {
1743         error_setg(errp, "Sync mode '%s' not supported",
1744                    MirrorSyncMode_str(mode));
1745         return;
1746     }
1747     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1748     base = mode == MIRROR_SYNC_MODE_TOP ? backing_bs(bs) : NULL;
1749     mirror_start_job(job_id, bs, creation_flags, target, replaces,
1750                      speed, granularity, buf_size, backing_mode, zero_target,
1751                      on_source_error, on_target_error, unmap, NULL, NULL,
1752                      &mirror_job_driver, is_none_mode, base, false,
1753                      filter_node_name, true, copy_mode, errp);
1754 }
1755 
1756 BlockJob *commit_active_start(const char *job_id, BlockDriverState *bs,
1757                               BlockDriverState *base, int creation_flags,
1758                               int64_t speed, BlockdevOnError on_error,
1759                               const char *filter_node_name,
1760                               BlockCompletionFunc *cb, void *opaque,
1761                               bool auto_complete, Error **errp)
1762 {
1763     bool base_read_only;
1764     Error *local_err = NULL;
1765     BlockJob *ret;
1766 
1767     base_read_only = bdrv_is_read_only(base);
1768 
1769     if (base_read_only) {
1770         if (bdrv_reopen_set_read_only(base, false, errp) < 0) {
1771             return NULL;
1772         }
1773     }
1774 
1775     ret = mirror_start_job(
1776                      job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1777                      MIRROR_LEAVE_BACKING_CHAIN, false,
1778                      on_error, on_error, true, cb, opaque,
1779                      &commit_active_job_driver, false, base, auto_complete,
1780                      filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND,
1781                      &local_err);
1782     if (local_err) {
1783         error_propagate(errp, local_err);
1784         goto error_restore_flags;
1785     }
1786 
1787     return ret;
1788 
1789 error_restore_flags:
1790     /* ignore error and errp for bdrv_reopen, because we want to propagate
1791      * the original error */
1792     if (base_read_only) {
1793         bdrv_reopen_set_read_only(base, true, NULL);
1794     }
1795     return NULL;
1796 }
1797