xref: /openbmc/qemu/block/mirror.c (revision 762c280d)
1 /*
2  * Image mirroring
3  *
4  * Copyright Red Hat, Inc. 2012
5  *
6  * Authors:
7  *  Paolo Bonzini  <pbonzini@redhat.com>
8  *
9  * This work is licensed under the terms of the GNU LGPL, version 2 or later.
10  * See the COPYING.LIB file in the top-level directory.
11  *
12  */
13 
14 #include "qemu/osdep.h"
15 #include "qemu/cutils.h"
16 #include "qemu/coroutine.h"
17 #include "qemu/range.h"
18 #include "trace.h"
19 #include "block/blockjob_int.h"
20 #include "block/block_int.h"
21 #include "sysemu/block-backend.h"
22 #include "qapi/error.h"
23 #include "qapi/qmp/qerror.h"
24 #include "qemu/ratelimit.h"
25 #include "qemu/bitmap.h"
26 
27 #define MAX_IN_FLIGHT 16
28 #define MAX_IO_BYTES (1 << 20) /* 1 Mb */
29 #define DEFAULT_MIRROR_BUF_SIZE (MAX_IN_FLIGHT * MAX_IO_BYTES)
30 
31 /* The mirroring buffer is a list of granularity-sized chunks.
32  * Free chunks are organized in a list.
33  */
34 typedef struct MirrorBuffer {
35     QSIMPLEQ_ENTRY(MirrorBuffer) next;
36 } MirrorBuffer;
37 
38 typedef struct MirrorOp MirrorOp;
39 
40 typedef struct MirrorBlockJob {
41     BlockJob common;
42     BlockBackend *target;
43     BlockDriverState *mirror_top_bs;
44     BlockDriverState *base;
45     BlockDriverState *base_overlay;
46 
47     /* The name of the graph node to replace */
48     char *replaces;
49     /* The BDS to replace */
50     BlockDriverState *to_replace;
51     /* Used to block operations on the drive-mirror-replace target */
52     Error *replace_blocker;
53     bool is_none_mode;
54     BlockMirrorBackingMode backing_mode;
55     /* Whether the target image requires explicit zero-initialization */
56     bool zero_target;
57     MirrorCopyMode copy_mode;
58     BlockdevOnError on_source_error, on_target_error;
59     /* 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     bool is_commit;
92 } MirrorBDSOpaque;
93 
94 struct MirrorOp {
95     MirrorBlockJob *s;
96     QEMUIOVector qiov;
97     int64_t offset;
98     uint64_t bytes;
99 
100     /* The pointee is set by mirror_co_read(), mirror_co_zero(), and
101      * mirror_co_discard() before yielding for the first time */
102     int64_t *bytes_handled;
103 
104     bool is_pseudo_op;
105     bool is_active_write;
106     bool is_in_flight;
107     CoQueue waiting_requests;
108     Coroutine *co;
109     MirrorOp *waiting_for_op;
110 
111     QTAILQ_ENTRY(MirrorOp) next;
112 };
113 
114 typedef enum MirrorMethod {
115     MIRROR_METHOD_COPY,
116     MIRROR_METHOD_ZERO,
117     MIRROR_METHOD_DISCARD,
118 } MirrorMethod;
119 
120 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
121                                             int error)
122 {
123     s->actively_synced = false;
124     if (read) {
125         return block_job_error_action(&s->common, s->on_source_error,
126                                       true, error);
127     } else {
128         return block_job_error_action(&s->common, s->on_target_error,
129                                       false, error);
130     }
131 }
132 
133 static void coroutine_fn mirror_wait_on_conflicts(MirrorOp *self,
134                                                   MirrorBlockJob *s,
135                                                   uint64_t offset,
136                                                   uint64_t bytes)
137 {
138     uint64_t self_start_chunk = offset / s->granularity;
139     uint64_t self_end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
140     uint64_t self_nb_chunks = self_end_chunk - self_start_chunk;
141 
142     while (find_next_bit(s->in_flight_bitmap, self_end_chunk,
143                          self_start_chunk) < self_end_chunk &&
144            s->ret >= 0)
145     {
146         MirrorOp *op;
147 
148         QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
149             uint64_t op_start_chunk = op->offset / s->granularity;
150             uint64_t op_nb_chunks = DIV_ROUND_UP(op->offset + op->bytes,
151                                                  s->granularity) -
152                                     op_start_chunk;
153 
154             if (op == self) {
155                 continue;
156             }
157 
158             if (ranges_overlap(self_start_chunk, self_nb_chunks,
159                                op_start_chunk, op_nb_chunks))
160             {
161                 if (self) {
162                     /*
163                      * If the operation is already (indirectly) waiting for us,
164                      * or will wait for us as soon as it wakes up, then just go
165                      * on (instead of producing a deadlock in the former case).
166                      */
167                     if (op->waiting_for_op) {
168                         continue;
169                     }
170 
171                     self->waiting_for_op = op;
172                 }
173 
174                 qemu_co_queue_wait(&op->waiting_requests, NULL);
175 
176                 if (self) {
177                     self->waiting_for_op = NULL;
178                 }
179 
180                 break;
181             }
182         }
183     }
184 }
185 
186 static void coroutine_fn mirror_iteration_done(MirrorOp *op, int ret)
187 {
188     MirrorBlockJob *s = op->s;
189     struct iovec *iov;
190     int64_t chunk_num;
191     int i, nb_chunks;
192 
193     trace_mirror_iteration_done(s, op->offset, op->bytes, ret);
194 
195     s->in_flight--;
196     s->bytes_in_flight -= op->bytes;
197     iov = op->qiov.iov;
198     for (i = 0; i < op->qiov.niov; i++) {
199         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
200         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
201         s->buf_free_count++;
202     }
203 
204     chunk_num = op->offset / s->granularity;
205     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
206 
207     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
208     QTAILQ_REMOVE(&s->ops_in_flight, op, next);
209     if (ret >= 0) {
210         if (s->cow_bitmap) {
211             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
212         }
213         if (!s->initial_zeroing_ongoing) {
214             job_progress_update(&s->common.job, op->bytes);
215         }
216     }
217     qemu_iovec_destroy(&op->qiov);
218 
219     qemu_co_queue_restart_all(&op->waiting_requests);
220     g_free(op);
221 }
222 
223 static void coroutine_fn mirror_write_complete(MirrorOp *op, int ret)
224 {
225     MirrorBlockJob *s = op->s;
226 
227     if (ret < 0) {
228         BlockErrorAction action;
229 
230         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
231         action = mirror_error_action(s, false, -ret);
232         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
233             s->ret = ret;
234         }
235     }
236 
237     mirror_iteration_done(op, ret);
238 }
239 
240 static void coroutine_fn mirror_read_complete(MirrorOp *op, int ret)
241 {
242     MirrorBlockJob *s = op->s;
243 
244     if (ret < 0) {
245         BlockErrorAction action;
246 
247         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
248         action = mirror_error_action(s, true, -ret);
249         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
250             s->ret = ret;
251         }
252 
253         mirror_iteration_done(op, ret);
254         return;
255     }
256 
257     ret = blk_co_pwritev(s->target, op->offset, op->qiov.size, &op->qiov, 0);
258     mirror_write_complete(op, ret);
259 }
260 
261 /* Clip bytes relative to offset to not exceed end-of-file */
262 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s,
263                                         int64_t offset,
264                                         int64_t bytes)
265 {
266     return MIN(bytes, s->bdev_length - offset);
267 }
268 
269 /* Round offset and/or bytes to target cluster if COW is needed, and
270  * return the offset of the adjusted tail against original. */
271 static int mirror_cow_align(MirrorBlockJob *s, int64_t *offset,
272                             uint64_t *bytes)
273 {
274     bool need_cow;
275     int ret = 0;
276     int64_t align_offset = *offset;
277     int64_t align_bytes = *bytes;
278     int max_bytes = s->granularity * s->max_iov;
279 
280     need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap);
281     need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity,
282                           s->cow_bitmap);
283     if (need_cow) {
284         bdrv_round_to_clusters(blk_bs(s->target), *offset, *bytes,
285                                &align_offset, &align_bytes);
286     }
287 
288     if (align_bytes > max_bytes) {
289         align_bytes = max_bytes;
290         if (need_cow) {
291             align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size);
292         }
293     }
294     /* Clipping may result in align_bytes unaligned to chunk boundary, but
295      * that doesn't matter because it's already the end of source image. */
296     align_bytes = mirror_clip_bytes(s, align_offset, align_bytes);
297 
298     ret = align_offset + align_bytes - (*offset + *bytes);
299     *offset = align_offset;
300     *bytes = align_bytes;
301     assert(ret >= 0);
302     return ret;
303 }
304 
305 static inline void coroutine_fn
306 mirror_wait_for_any_operation(MirrorBlockJob *s, bool active)
307 {
308     MirrorOp *op;
309 
310     QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
311         /* Do not wait on pseudo ops, because it may in turn wait on
312          * some other operation to start, which may in fact be the
313          * caller of this function.  Since there is only one pseudo op
314          * at any given time, we will always find some real operation
315          * to wait on. */
316         if (!op->is_pseudo_op && op->is_in_flight &&
317             op->is_active_write == active)
318         {
319             qemu_co_queue_wait(&op->waiting_requests, NULL);
320             return;
321         }
322     }
323     abort();
324 }
325 
326 static inline void coroutine_fn
327 mirror_wait_for_free_in_flight_slot(MirrorBlockJob *s)
328 {
329     /* Only non-active operations use up in-flight slots */
330     mirror_wait_for_any_operation(s, false);
331 }
332 
333 /* Perform a mirror copy operation.
334  *
335  * *op->bytes_handled is set to the number of bytes copied after and
336  * including offset, excluding any bytes copied prior to offset due
337  * to alignment.  This will be op->bytes if no alignment is necessary,
338  * or (new_end - op->offset) if the tail is rounded up or down due to
339  * alignment or buffer limit.
340  */
341 static void coroutine_fn mirror_co_read(void *opaque)
342 {
343     MirrorOp *op = opaque;
344     MirrorBlockJob *s = op->s;
345     int nb_chunks;
346     uint64_t ret;
347     uint64_t max_bytes;
348 
349     max_bytes = s->granularity * s->max_iov;
350 
351     /* We can only handle as much as buf_size at a time. */
352     op->bytes = MIN(s->buf_size, MIN(max_bytes, op->bytes));
353     assert(op->bytes);
354     assert(op->bytes < BDRV_REQUEST_MAX_BYTES);
355     *op->bytes_handled = op->bytes;
356 
357     if (s->cow_bitmap) {
358         *op->bytes_handled += mirror_cow_align(s, &op->offset, &op->bytes);
359     }
360     /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */
361     assert(*op->bytes_handled <= UINT_MAX);
362     assert(op->bytes <= s->buf_size);
363     /* The offset is granularity-aligned because:
364      * 1) Caller passes in aligned values;
365      * 2) mirror_cow_align is used only when target cluster is larger. */
366     assert(QEMU_IS_ALIGNED(op->offset, s->granularity));
367     /* The range is sector-aligned, since bdrv_getlength() rounds up. */
368     assert(QEMU_IS_ALIGNED(op->bytes, BDRV_SECTOR_SIZE));
369     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
370 
371     while (s->buf_free_count < nb_chunks) {
372         trace_mirror_yield_in_flight(s, op->offset, s->in_flight);
373         mirror_wait_for_free_in_flight_slot(s);
374     }
375 
376     /* Now make a QEMUIOVector taking enough granularity-sized chunks
377      * from s->buf_free.
378      */
379     qemu_iovec_init(&op->qiov, nb_chunks);
380     while (nb_chunks-- > 0) {
381         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
382         size_t remaining = op->bytes - op->qiov.size;
383 
384         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
385         s->buf_free_count--;
386         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
387     }
388 
389     /* Copy the dirty cluster.  */
390     s->in_flight++;
391     s->bytes_in_flight += op->bytes;
392     op->is_in_flight = true;
393     trace_mirror_one_iteration(s, op->offset, op->bytes);
394 
395     ret = bdrv_co_preadv(s->mirror_top_bs->backing, op->offset, op->bytes,
396                          &op->qiov, 0);
397     mirror_read_complete(op, ret);
398 }
399 
400 static void coroutine_fn mirror_co_zero(void *opaque)
401 {
402     MirrorOp *op = opaque;
403     int ret;
404 
405     op->s->in_flight++;
406     op->s->bytes_in_flight += op->bytes;
407     *op->bytes_handled = op->bytes;
408     op->is_in_flight = true;
409 
410     ret = blk_co_pwrite_zeroes(op->s->target, op->offset, op->bytes,
411                                op->s->unmap ? BDRV_REQ_MAY_UNMAP : 0);
412     mirror_write_complete(op, ret);
413 }
414 
415 static void coroutine_fn mirror_co_discard(void *opaque)
416 {
417     MirrorOp *op = opaque;
418     int ret;
419 
420     op->s->in_flight++;
421     op->s->bytes_in_flight += op->bytes;
422     *op->bytes_handled = op->bytes;
423     op->is_in_flight = true;
424 
425     ret = blk_co_pdiscard(op->s->target, op->offset, op->bytes);
426     mirror_write_complete(op, ret);
427 }
428 
429 static unsigned mirror_perform(MirrorBlockJob *s, int64_t offset,
430                                unsigned bytes, MirrorMethod mirror_method)
431 {
432     MirrorOp *op;
433     Coroutine *co;
434     int64_t bytes_handled = -1;
435 
436     op = g_new(MirrorOp, 1);
437     *op = (MirrorOp){
438         .s              = s,
439         .offset         = offset,
440         .bytes          = bytes,
441         .bytes_handled  = &bytes_handled,
442     };
443     qemu_co_queue_init(&op->waiting_requests);
444 
445     switch (mirror_method) {
446     case MIRROR_METHOD_COPY:
447         co = qemu_coroutine_create(mirror_co_read, op);
448         break;
449     case MIRROR_METHOD_ZERO:
450         co = qemu_coroutine_create(mirror_co_zero, op);
451         break;
452     case MIRROR_METHOD_DISCARD:
453         co = qemu_coroutine_create(mirror_co_discard, op);
454         break;
455     default:
456         abort();
457     }
458     op->co = co;
459 
460     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
461     qemu_coroutine_enter(co);
462     /* At this point, ownership of op has been moved to the coroutine
463      * and the object may already be freed */
464 
465     /* Assert that this value has been set */
466     assert(bytes_handled >= 0);
467 
468     /* Same assertion as in mirror_co_read() (and for mirror_co_read()
469      * and mirror_co_discard(), bytes_handled == op->bytes, which
470      * is the @bytes parameter given to this function) */
471     assert(bytes_handled <= UINT_MAX);
472     return bytes_handled;
473 }
474 
475 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
476 {
477     BlockDriverState *source = s->mirror_top_bs->backing->bs;
478     MirrorOp *pseudo_op;
479     int64_t offset;
480     uint64_t delay_ns = 0, ret = 0;
481     /* At least the first dirty chunk is mirrored in one iteration. */
482     int nb_chunks = 1;
483     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
484     int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES);
485 
486     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
487     offset = bdrv_dirty_iter_next(s->dbi);
488     if (offset < 0) {
489         bdrv_set_dirty_iter(s->dbi, 0);
490         offset = bdrv_dirty_iter_next(s->dbi);
491         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
492         assert(offset >= 0);
493     }
494     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
495 
496     mirror_wait_on_conflicts(NULL, s, offset, 1);
497 
498     job_pause_point(&s->common.job);
499 
500     /* Find the number of consective dirty chunks following the first dirty
501      * one, and wait for in flight requests in them. */
502     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
503     while (nb_chunks * s->granularity < s->buf_size) {
504         int64_t next_dirty;
505         int64_t next_offset = offset + nb_chunks * s->granularity;
506         int64_t next_chunk = next_offset / s->granularity;
507         if (next_offset >= s->bdev_length ||
508             !bdrv_dirty_bitmap_get_locked(s->dirty_bitmap, next_offset)) {
509             break;
510         }
511         if (test_bit(next_chunk, s->in_flight_bitmap)) {
512             break;
513         }
514 
515         next_dirty = bdrv_dirty_iter_next(s->dbi);
516         if (next_dirty > next_offset || next_dirty < 0) {
517             /* The bitmap iterator's cache is stale, refresh it */
518             bdrv_set_dirty_iter(s->dbi, next_offset);
519             next_dirty = bdrv_dirty_iter_next(s->dbi);
520         }
521         assert(next_dirty == next_offset);
522         nb_chunks++;
523     }
524 
525     /* Clear dirty bits before querying the block status, because
526      * calling bdrv_block_status_above could yield - if some blocks are
527      * marked dirty in this window, we need to know.
528      */
529     bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset,
530                                    nb_chunks * s->granularity);
531     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
532 
533     /* Before claiming an area in the in-flight bitmap, we have to
534      * create a MirrorOp for it so that conflicting requests can wait
535      * for it.  mirror_perform() will create the real MirrorOps later,
536      * for now we just create a pseudo operation that will wake up all
537      * conflicting requests once all real operations have been
538      * launched. */
539     pseudo_op = g_new(MirrorOp, 1);
540     *pseudo_op = (MirrorOp){
541         .offset         = offset,
542         .bytes          = nb_chunks * s->granularity,
543         .is_pseudo_op   = true,
544     };
545     qemu_co_queue_init(&pseudo_op->waiting_requests);
546     QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next);
547 
548     bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks);
549     while (nb_chunks > 0 && offset < s->bdev_length) {
550         int ret;
551         int64_t io_bytes;
552         int64_t io_bytes_acct;
553         MirrorMethod mirror_method = MIRROR_METHOD_COPY;
554 
555         assert(!(offset % s->granularity));
556         ret = bdrv_block_status_above(source, NULL, offset,
557                                       nb_chunks * s->granularity,
558                                       &io_bytes, NULL, NULL);
559         if (ret < 0) {
560             io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes);
561         } else if (ret & BDRV_BLOCK_DATA) {
562             io_bytes = MIN(io_bytes, max_io_bytes);
563         }
564 
565         io_bytes -= io_bytes % s->granularity;
566         if (io_bytes < s->granularity) {
567             io_bytes = s->granularity;
568         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
569             int64_t target_offset;
570             int64_t target_bytes;
571             bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes,
572                                    &target_offset, &target_bytes);
573             if (target_offset == offset &&
574                 target_bytes == io_bytes) {
575                 mirror_method = ret & BDRV_BLOCK_ZERO ?
576                                     MIRROR_METHOD_ZERO :
577                                     MIRROR_METHOD_DISCARD;
578             }
579         }
580 
581         while (s->in_flight >= MAX_IN_FLIGHT) {
582             trace_mirror_yield_in_flight(s, offset, s->in_flight);
583             mirror_wait_for_free_in_flight_slot(s);
584         }
585 
586         if (s->ret < 0) {
587             ret = 0;
588             goto fail;
589         }
590 
591         io_bytes = mirror_clip_bytes(s, offset, io_bytes);
592         io_bytes = mirror_perform(s, offset, io_bytes, mirror_method);
593         if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) {
594             io_bytes_acct = 0;
595         } else {
596             io_bytes_acct = io_bytes;
597         }
598         assert(io_bytes);
599         offset += io_bytes;
600         nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity);
601         delay_ns = block_job_ratelimit_get_delay(&s->common, io_bytes_acct);
602     }
603 
604     ret = delay_ns;
605 fail:
606     QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next);
607     qemu_co_queue_restart_all(&pseudo_op->waiting_requests);
608     g_free(pseudo_op);
609 
610     return ret;
611 }
612 
613 static void mirror_free_init(MirrorBlockJob *s)
614 {
615     int granularity = s->granularity;
616     size_t buf_size = s->buf_size;
617     uint8_t *buf = s->buf;
618 
619     assert(s->buf_free_count == 0);
620     QSIMPLEQ_INIT(&s->buf_free);
621     while (buf_size != 0) {
622         MirrorBuffer *cur = (MirrorBuffer *)buf;
623         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
624         s->buf_free_count++;
625         buf_size -= granularity;
626         buf += granularity;
627     }
628 }
629 
630 /* This is also used for the .pause callback. There is no matching
631  * mirror_resume() because mirror_run() will begin iterating again
632  * when the job is resumed.
633  */
634 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s)
635 {
636     while (s->in_flight > 0) {
637         mirror_wait_for_free_in_flight_slot(s);
638     }
639 }
640 
641 /**
642  * mirror_exit_common: handle both abort() and prepare() cases.
643  * for .prepare, returns 0 on success and -errno on failure.
644  * for .abort cases, denoted by abort = true, MUST return 0.
645  */
646 static int mirror_exit_common(Job *job)
647 {
648     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
649     BlockJob *bjob = &s->common;
650     MirrorBDSOpaque *bs_opaque;
651     AioContext *replace_aio_context = NULL;
652     BlockDriverState *src;
653     BlockDriverState *target_bs;
654     BlockDriverState *mirror_top_bs;
655     Error *local_err = NULL;
656     bool abort = job->ret < 0;
657     int ret = 0;
658 
659     if (s->prepared) {
660         return 0;
661     }
662     s->prepared = true;
663 
664     mirror_top_bs = s->mirror_top_bs;
665     bs_opaque = mirror_top_bs->opaque;
666     src = mirror_top_bs->backing->bs;
667     target_bs = blk_bs(s->target);
668 
669     if (bdrv_chain_contains(src, target_bs)) {
670         bdrv_unfreeze_backing_chain(mirror_top_bs, target_bs);
671     }
672 
673     bdrv_release_dirty_bitmap(s->dirty_bitmap);
674 
675     /* Make sure that the source BDS doesn't go away during bdrv_replace_node,
676      * before we can call bdrv_drained_end */
677     bdrv_ref(src);
678     bdrv_ref(mirror_top_bs);
679     bdrv_ref(target_bs);
680 
681     /*
682      * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
683      * inserting target_bs at s->to_replace, where we might not be able to get
684      * these permissions.
685      */
686     blk_unref(s->target);
687     s->target = NULL;
688 
689     /* We don't access the source any more. Dropping any WRITE/RESIZE is
690      * required before it could become a backing file of target_bs. Not having
691      * these permissions any more means that we can't allow any new requests on
692      * mirror_top_bs from now on, so keep it drained. */
693     bdrv_drained_begin(mirror_top_bs);
694     bs_opaque->stop = true;
695     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
696                              &error_abort);
697     if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
698         BlockDriverState *backing = s->is_none_mode ? src : s->base;
699         BlockDriverState *unfiltered_target = bdrv_skip_filters(target_bs);
700 
701         if (bdrv_cow_bs(unfiltered_target) != backing) {
702             bdrv_set_backing_hd(unfiltered_target, backing, &local_err);
703             if (local_err) {
704                 error_report_err(local_err);
705                 local_err = NULL;
706                 ret = -EPERM;
707             }
708         }
709     } else if (!abort && s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
710         assert(!bdrv_backing_chain_next(target_bs));
711         ret = bdrv_open_backing_file(bdrv_skip_filters(target_bs), NULL,
712                                      "backing", &local_err);
713         if (ret < 0) {
714             error_report_err(local_err);
715             local_err = NULL;
716         }
717     }
718 
719     if (s->to_replace) {
720         replace_aio_context = bdrv_get_aio_context(s->to_replace);
721         aio_context_acquire(replace_aio_context);
722     }
723 
724     if (s->should_complete && !abort) {
725         BlockDriverState *to_replace = s->to_replace ?: src;
726         bool ro = bdrv_is_read_only(to_replace);
727 
728         if (ro != bdrv_is_read_only(target_bs)) {
729             bdrv_reopen_set_read_only(target_bs, ro, NULL);
730         }
731 
732         /* The mirror job has no requests in flight any more, but we need to
733          * drain potential other users of the BDS before changing the graph. */
734         assert(s->in_drain);
735         bdrv_drained_begin(target_bs);
736         /*
737          * Cannot use check_to_replace_node() here, because that would
738          * check for an op blocker on @to_replace, and we have our own
739          * there.
740          */
741         if (bdrv_recurse_can_replace(src, to_replace)) {
742             bdrv_replace_node(to_replace, target_bs, &local_err);
743         } else {
744             error_setg(&local_err, "Can no longer replace '%s' by '%s', "
745                        "because it can no longer be guaranteed that doing so "
746                        "would not lead to an abrupt change of visible data",
747                        to_replace->node_name, target_bs->node_name);
748         }
749         bdrv_drained_end(target_bs);
750         if (local_err) {
751             error_report_err(local_err);
752             ret = -EPERM;
753         }
754     }
755     if (s->to_replace) {
756         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
757         error_free(s->replace_blocker);
758         bdrv_unref(s->to_replace);
759     }
760     if (replace_aio_context) {
761         aio_context_release(replace_aio_context);
762     }
763     g_free(s->replaces);
764     bdrv_unref(target_bs);
765 
766     /*
767      * Remove the mirror filter driver from the graph. Before this, get rid of
768      * the blockers on the intermediate nodes so that the resulting state is
769      * valid.
770      */
771     block_job_remove_all_bdrv(bjob);
772     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
773 
774     bs_opaque->job = NULL;
775 
776     bdrv_drained_end(src);
777     bdrv_drained_end(mirror_top_bs);
778     s->in_drain = false;
779     bdrv_unref(mirror_top_bs);
780     bdrv_unref(src);
781 
782     return ret;
783 }
784 
785 static int mirror_prepare(Job *job)
786 {
787     return mirror_exit_common(job);
788 }
789 
790 static void mirror_abort(Job *job)
791 {
792     int ret = mirror_exit_common(job);
793     assert(ret == 0);
794 }
795 
796 static void coroutine_fn mirror_throttle(MirrorBlockJob *s)
797 {
798     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
799 
800     if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) {
801         s->last_pause_ns = now;
802         job_sleep_ns(&s->common.job, 0);
803     } else {
804         job_pause_point(&s->common.job);
805     }
806 }
807 
808 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
809 {
810     int64_t offset;
811     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
812     BlockDriverState *target_bs = blk_bs(s->target);
813     int ret;
814     int64_t count;
815 
816     if (s->zero_target) {
817         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
818             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
819             return 0;
820         }
821 
822         s->initial_zeroing_ongoing = true;
823         for (offset = 0; offset < s->bdev_length; ) {
824             int bytes = MIN(s->bdev_length - offset,
825                             QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
826 
827             mirror_throttle(s);
828 
829             if (job_is_cancelled(&s->common.job)) {
830                 s->initial_zeroing_ongoing = false;
831                 return 0;
832             }
833 
834             if (s->in_flight >= MAX_IN_FLIGHT) {
835                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
836                                    s->in_flight);
837                 mirror_wait_for_free_in_flight_slot(s);
838                 continue;
839             }
840 
841             mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO);
842             offset += bytes;
843         }
844 
845         mirror_wait_for_all_io(s);
846         s->initial_zeroing_ongoing = false;
847     }
848 
849     /* First part, loop on the sectors and initialize the dirty bitmap.  */
850     for (offset = 0; offset < s->bdev_length; ) {
851         /* Just to make sure we are not exceeding int limit. */
852         int bytes = MIN(s->bdev_length - offset,
853                         QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
854 
855         mirror_throttle(s);
856 
857         if (job_is_cancelled(&s->common.job)) {
858             return 0;
859         }
860 
861         ret = bdrv_is_allocated_above(bs, s->base_overlay, true, offset, bytes,
862                                       &count);
863         if (ret < 0) {
864             return ret;
865         }
866 
867         assert(count);
868         if (ret > 0) {
869             bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
870         }
871         offset += count;
872     }
873     return 0;
874 }
875 
876 /* Called when going out of the streaming phase to flush the bulk of the
877  * data to the medium, or just before completing.
878  */
879 static int mirror_flush(MirrorBlockJob *s)
880 {
881     int ret = blk_flush(s->target);
882     if (ret < 0) {
883         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
884             s->ret = ret;
885         }
886     }
887     return ret;
888 }
889 
890 static int coroutine_fn mirror_run(Job *job, Error **errp)
891 {
892     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
893     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
894     BlockDriverState *target_bs = blk_bs(s->target);
895     bool need_drain = true;
896     int64_t length;
897     int64_t target_length;
898     BlockDriverInfo bdi;
899     char backing_filename[2]; /* we only need 2 characters because we are only
900                                  checking for a NULL string */
901     int ret = 0;
902 
903     if (job_is_cancelled(&s->common.job)) {
904         goto immediate_exit;
905     }
906 
907     s->bdev_length = bdrv_getlength(bs);
908     if (s->bdev_length < 0) {
909         ret = s->bdev_length;
910         goto immediate_exit;
911     }
912 
913     target_length = blk_getlength(s->target);
914     if (target_length < 0) {
915         ret = target_length;
916         goto immediate_exit;
917     }
918 
919     /* Active commit must resize the base image if its size differs from the
920      * active layer. */
921     if (s->base == blk_bs(s->target)) {
922         if (s->bdev_length > target_length) {
923             ret = blk_truncate(s->target, s->bdev_length, false,
924                                PREALLOC_MODE_OFF, 0, NULL);
925             if (ret < 0) {
926                 goto immediate_exit;
927             }
928         }
929     } else if (s->bdev_length != target_length) {
930         error_setg(errp, "Source and target image have different sizes");
931         ret = -EINVAL;
932         goto immediate_exit;
933     }
934 
935     if (s->bdev_length == 0) {
936         /* Transition to the READY state and wait for complete. */
937         job_transition_to_ready(&s->common.job);
938         s->actively_synced = true;
939         while (!job_cancel_requested(&s->common.job) && !s->should_complete) {
940             job_yield(&s->common.job);
941         }
942         goto immediate_exit;
943     }
944 
945     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
946     s->in_flight_bitmap = bitmap_new(length);
947 
948     /* If we have no backing file yet in the destination, we cannot let
949      * the destination do COW.  Instead, we copy sectors around the
950      * dirty data if needed.  We need a bitmap to do that.
951      */
952     bdrv_get_backing_filename(target_bs, backing_filename,
953                               sizeof(backing_filename));
954     if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) {
955         s->target_cluster_size = bdi.cluster_size;
956     } else {
957         s->target_cluster_size = BDRV_SECTOR_SIZE;
958     }
959     if (backing_filename[0] && !bdrv_backing_chain_next(target_bs) &&
960         s->granularity < s->target_cluster_size) {
961         s->buf_size = MAX(s->buf_size, s->target_cluster_size);
962         s->cow_bitmap = bitmap_new(length);
963     }
964     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
965 
966     s->buf = qemu_try_blockalign(bs, s->buf_size);
967     if (s->buf == NULL) {
968         ret = -ENOMEM;
969         goto immediate_exit;
970     }
971 
972     mirror_free_init(s);
973 
974     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
975     if (!s->is_none_mode) {
976         ret = mirror_dirty_init(s);
977         if (ret < 0 || job_is_cancelled(&s->common.job)) {
978             goto immediate_exit;
979         }
980     }
981 
982     assert(!s->dbi);
983     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
984     for (;;) {
985         uint64_t delay_ns = 0;
986         int64_t cnt, delta;
987         bool should_complete;
988 
989         /* Do not start passive operations while there are active
990          * writes in progress */
991         while (s->in_active_write_counter) {
992             mirror_wait_for_any_operation(s, true);
993         }
994 
995         if (s->ret < 0) {
996             ret = s->ret;
997             goto immediate_exit;
998         }
999 
1000         job_pause_point(&s->common.job);
1001 
1002         if (job_is_cancelled(&s->common.job)) {
1003             ret = 0;
1004             goto immediate_exit;
1005         }
1006 
1007         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1008         /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
1009          * the number of bytes currently being processed; together those are
1010          * the current remaining operation length */
1011         job_progress_set_remaining(&s->common.job, s->bytes_in_flight + cnt);
1012 
1013         /* Note that even when no rate limit is applied we need to yield
1014          * periodically with no pending I/O so that bdrv_drain_all() returns.
1015          * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
1016          * an error, or when the source is clean, whichever comes first. */
1017         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
1018         if (delta < BLOCK_JOB_SLICE_TIME &&
1019             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
1020             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
1021                 (cnt == 0 && s->in_flight > 0)) {
1022                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
1023                 mirror_wait_for_free_in_flight_slot(s);
1024                 continue;
1025             } else if (cnt != 0) {
1026                 delay_ns = mirror_iteration(s);
1027             }
1028         }
1029 
1030         should_complete = false;
1031         if (s->in_flight == 0 && cnt == 0) {
1032             trace_mirror_before_flush(s);
1033             if (!job_is_ready(&s->common.job)) {
1034                 if (mirror_flush(s) < 0) {
1035                     /* Go check s->ret.  */
1036                     continue;
1037                 }
1038                 /* We're out of the streaming phase.  From now on, if the job
1039                  * is cancelled we will actually complete all pending I/O and
1040                  * report completion.  This way, block-job-cancel will leave
1041                  * the target in a consistent state.
1042                  */
1043                 job_transition_to_ready(&s->common.job);
1044                 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) {
1045                     s->actively_synced = true;
1046                 }
1047             }
1048 
1049             should_complete = s->should_complete ||
1050                 job_cancel_requested(&s->common.job);
1051             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1052         }
1053 
1054         if (cnt == 0 && should_complete) {
1055             /* The dirty bitmap is not updated while operations are pending.
1056              * If we're about to exit, wait for pending operations before
1057              * calling bdrv_get_dirty_count(bs), or we may exit while the
1058              * source has dirty data to copy!
1059              *
1060              * Note that I/O can be submitted by the guest while
1061              * mirror_populate runs, so pause it now.  Before deciding
1062              * whether to switch to target check one last time if I/O has
1063              * come in the meanwhile, and if not flush the data to disk.
1064              */
1065             trace_mirror_before_drain(s, cnt);
1066 
1067             s->in_drain = true;
1068             bdrv_drained_begin(bs);
1069             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1070             if (cnt > 0 || mirror_flush(s) < 0) {
1071                 bdrv_drained_end(bs);
1072                 s->in_drain = false;
1073                 continue;
1074             }
1075 
1076             /* The two disks are in sync.  Exit and report successful
1077              * completion.
1078              */
1079             assert(QLIST_EMPTY(&bs->tracked_requests));
1080             need_drain = false;
1081             break;
1082         }
1083 
1084         if (job_is_ready(&s->common.job) && !should_complete) {
1085             delay_ns = (s->in_flight == 0 &&
1086                         cnt == 0 ? BLOCK_JOB_SLICE_TIME : 0);
1087         }
1088         trace_mirror_before_sleep(s, cnt, job_is_ready(&s->common.job),
1089                                   delay_ns);
1090         job_sleep_ns(&s->common.job, delay_ns);
1091         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1092     }
1093 
1094 immediate_exit:
1095     if (s->in_flight > 0) {
1096         /* We get here only if something went wrong.  Either the job failed,
1097          * or it was cancelled prematurely so that we do not guarantee that
1098          * the target is a copy of the source.
1099          */
1100         assert(ret < 0 || job_is_cancelled(&s->common.job));
1101         assert(need_drain);
1102         mirror_wait_for_all_io(s);
1103     }
1104 
1105     assert(s->in_flight == 0);
1106     qemu_vfree(s->buf);
1107     g_free(s->cow_bitmap);
1108     g_free(s->in_flight_bitmap);
1109     bdrv_dirty_iter_free(s->dbi);
1110 
1111     if (need_drain) {
1112         s->in_drain = true;
1113         bdrv_drained_begin(bs);
1114     }
1115 
1116     return ret;
1117 }
1118 
1119 static void mirror_complete(Job *job, Error **errp)
1120 {
1121     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1122 
1123     if (!job_is_ready(job)) {
1124         error_setg(errp, "The active block job '%s' cannot be completed",
1125                    job->id);
1126         return;
1127     }
1128 
1129     /* block all operations on to_replace bs */
1130     if (s->replaces) {
1131         AioContext *replace_aio_context;
1132 
1133         s->to_replace = bdrv_find_node(s->replaces);
1134         if (!s->to_replace) {
1135             error_setg(errp, "Node name '%s' not found", s->replaces);
1136             return;
1137         }
1138 
1139         replace_aio_context = bdrv_get_aio_context(s->to_replace);
1140         aio_context_acquire(replace_aio_context);
1141 
1142         /* TODO Translate this into child freeze system. */
1143         error_setg(&s->replace_blocker,
1144                    "block device is in use by block-job-complete");
1145         bdrv_op_block_all(s->to_replace, s->replace_blocker);
1146         bdrv_ref(s->to_replace);
1147 
1148         aio_context_release(replace_aio_context);
1149     }
1150 
1151     s->should_complete = true;
1152 
1153     /* If the job is paused, it will be re-entered when it is resumed */
1154     if (!job->paused) {
1155         job_enter(job);
1156     }
1157 }
1158 
1159 static void coroutine_fn mirror_pause(Job *job)
1160 {
1161     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1162 
1163     mirror_wait_for_all_io(s);
1164 }
1165 
1166 static bool mirror_drained_poll(BlockJob *job)
1167 {
1168     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1169 
1170     /* If the job isn't paused nor cancelled, we can't be sure that it won't
1171      * issue more requests. We make an exception if we've reached this point
1172      * from one of our own drain sections, to avoid a deadlock waiting for
1173      * ourselves.
1174      */
1175     if (!s->common.job.paused && !job_is_cancelled(&job->job) && !s->in_drain) {
1176         return true;
1177     }
1178 
1179     return !!s->in_flight;
1180 }
1181 
1182 static bool mirror_cancel(Job *job, bool force)
1183 {
1184     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1185     BlockDriverState *target = blk_bs(s->target);
1186 
1187     /*
1188      * Before the job is READY, we treat any cancellation like a
1189      * force-cancellation.
1190      */
1191     force = force || !job_is_ready(job);
1192 
1193     if (force) {
1194         bdrv_cancel_in_flight(target);
1195     }
1196     return force;
1197 }
1198 
1199 static bool commit_active_cancel(Job *job, bool force)
1200 {
1201     /* Same as above in mirror_cancel() */
1202     return force || !job_is_ready(job);
1203 }
1204 
1205 static const BlockJobDriver mirror_job_driver = {
1206     .job_driver = {
1207         .instance_size          = sizeof(MirrorBlockJob),
1208         .job_type               = JOB_TYPE_MIRROR,
1209         .free                   = block_job_free,
1210         .user_resume            = block_job_user_resume,
1211         .run                    = mirror_run,
1212         .prepare                = mirror_prepare,
1213         .abort                  = mirror_abort,
1214         .pause                  = mirror_pause,
1215         .complete               = mirror_complete,
1216         .cancel                 = mirror_cancel,
1217     },
1218     .drained_poll           = mirror_drained_poll,
1219 };
1220 
1221 static const BlockJobDriver commit_active_job_driver = {
1222     .job_driver = {
1223         .instance_size          = sizeof(MirrorBlockJob),
1224         .job_type               = JOB_TYPE_COMMIT,
1225         .free                   = block_job_free,
1226         .user_resume            = block_job_user_resume,
1227         .run                    = mirror_run,
1228         .prepare                = mirror_prepare,
1229         .abort                  = mirror_abort,
1230         .pause                  = mirror_pause,
1231         .complete               = mirror_complete,
1232         .cancel                 = commit_active_cancel,
1233     },
1234     .drained_poll           = mirror_drained_poll,
1235 };
1236 
1237 static void coroutine_fn
1238 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method,
1239                      uint64_t offset, uint64_t bytes,
1240                      QEMUIOVector *qiov, int flags)
1241 {
1242     int ret;
1243     size_t qiov_offset = 0;
1244     int64_t bitmap_offset, bitmap_end;
1245 
1246     if (!QEMU_IS_ALIGNED(offset, job->granularity) &&
1247         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset))
1248     {
1249             /*
1250              * Dirty unaligned padding: ignore it.
1251              *
1252              * Reasoning:
1253              * 1. If we copy it, we can't reset corresponding bit in
1254              *    dirty_bitmap as there may be some "dirty" bytes still not
1255              *    copied.
1256              * 2. It's already dirty, so skipping it we don't diverge mirror
1257              *    progress.
1258              *
1259              * Note, that because of this, guest write may have no contribution
1260              * into mirror converge, but that's not bad, as we have background
1261              * process of mirroring. If under some bad circumstances (high guest
1262              * IO load) background process starve, we will not converge anyway,
1263              * even if each write will contribute, as guest is not guaranteed to
1264              * rewrite the whole disk.
1265              */
1266             qiov_offset = QEMU_ALIGN_UP(offset, job->granularity) - offset;
1267             if (bytes <= qiov_offset) {
1268                 /* nothing to do after shrink */
1269                 return;
1270             }
1271             offset += qiov_offset;
1272             bytes -= qiov_offset;
1273     }
1274 
1275     if (!QEMU_IS_ALIGNED(offset + bytes, job->granularity) &&
1276         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset + bytes - 1))
1277     {
1278         uint64_t tail = (offset + bytes) % job->granularity;
1279 
1280         if (bytes <= tail) {
1281             /* nothing to do after shrink */
1282             return;
1283         }
1284         bytes -= tail;
1285     }
1286 
1287     /*
1288      * Tails are either clean or shrunk, so for bitmap resetting
1289      * we safely align the range down.
1290      */
1291     bitmap_offset = QEMU_ALIGN_UP(offset, job->granularity);
1292     bitmap_end = QEMU_ALIGN_DOWN(offset + bytes, job->granularity);
1293     if (bitmap_offset < bitmap_end) {
1294         bdrv_reset_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1295                                 bitmap_end - bitmap_offset);
1296     }
1297 
1298     job_progress_increase_remaining(&job->common.job, bytes);
1299 
1300     switch (method) {
1301     case MIRROR_METHOD_COPY:
1302         ret = blk_co_pwritev_part(job->target, offset, bytes,
1303                                   qiov, qiov_offset, flags);
1304         break;
1305 
1306     case MIRROR_METHOD_ZERO:
1307         assert(!qiov);
1308         ret = blk_co_pwrite_zeroes(job->target, offset, bytes, flags);
1309         break;
1310 
1311     case MIRROR_METHOD_DISCARD:
1312         assert(!qiov);
1313         ret = blk_co_pdiscard(job->target, offset, bytes);
1314         break;
1315 
1316     default:
1317         abort();
1318     }
1319 
1320     if (ret >= 0) {
1321         job_progress_update(&job->common.job, bytes);
1322     } else {
1323         BlockErrorAction action;
1324 
1325         /*
1326          * We failed, so we should mark dirty the whole area, aligned up.
1327          * Note that we don't care about shrunk tails if any: they were dirty
1328          * at function start, and they must be still dirty, as we've locked
1329          * the region for in-flight op.
1330          */
1331         bitmap_offset = QEMU_ALIGN_DOWN(offset, job->granularity);
1332         bitmap_end = QEMU_ALIGN_UP(offset + bytes, job->granularity);
1333         bdrv_set_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1334                               bitmap_end - bitmap_offset);
1335         job->actively_synced = false;
1336 
1337         action = mirror_error_action(job, false, -ret);
1338         if (action == BLOCK_ERROR_ACTION_REPORT) {
1339             if (!job->ret) {
1340                 job->ret = ret;
1341             }
1342         }
1343     }
1344 }
1345 
1346 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s,
1347                                                    uint64_t offset,
1348                                                    uint64_t bytes)
1349 {
1350     MirrorOp *op;
1351     uint64_t start_chunk = offset / s->granularity;
1352     uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
1353 
1354     op = g_new(MirrorOp, 1);
1355     *op = (MirrorOp){
1356         .s                  = s,
1357         .offset             = offset,
1358         .bytes              = bytes,
1359         .is_active_write    = true,
1360         .is_in_flight       = true,
1361         .co                 = qemu_coroutine_self(),
1362     };
1363     qemu_co_queue_init(&op->waiting_requests);
1364     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
1365 
1366     s->in_active_write_counter++;
1367 
1368     mirror_wait_on_conflicts(op, s, offset, bytes);
1369 
1370     bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1371 
1372     return op;
1373 }
1374 
1375 static void coroutine_fn active_write_settle(MirrorOp *op)
1376 {
1377     uint64_t start_chunk = op->offset / op->s->granularity;
1378     uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes,
1379                                       op->s->granularity);
1380 
1381     if (!--op->s->in_active_write_counter && op->s->actively_synced) {
1382         BdrvChild *source = op->s->mirror_top_bs->backing;
1383 
1384         if (QLIST_FIRST(&source->bs->parents) == source &&
1385             QLIST_NEXT(source, next_parent) == NULL)
1386         {
1387             /* Assert that we are back in sync once all active write
1388              * operations are settled.
1389              * Note that we can only assert this if the mirror node
1390              * is the source node's only parent. */
1391             assert(!bdrv_get_dirty_count(op->s->dirty_bitmap));
1392         }
1393     }
1394     bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1395     QTAILQ_REMOVE(&op->s->ops_in_flight, op, next);
1396     qemu_co_queue_restart_all(&op->waiting_requests);
1397     g_free(op);
1398 }
1399 
1400 static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs,
1401     int64_t offset, int64_t bytes, QEMUIOVector *qiov, BdrvRequestFlags flags)
1402 {
1403     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1404 }
1405 
1406 static int coroutine_fn bdrv_mirror_top_do_write(BlockDriverState *bs,
1407     MirrorMethod method, uint64_t offset, uint64_t bytes, QEMUIOVector *qiov,
1408     int flags)
1409 {
1410     MirrorOp *op = NULL;
1411     MirrorBDSOpaque *s = bs->opaque;
1412     int ret = 0;
1413     bool copy_to_target;
1414 
1415     copy_to_target = s->job->ret >= 0 &&
1416                      !job_is_cancelled(&s->job->common.job) &&
1417                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1418 
1419     if (copy_to_target) {
1420         op = active_write_prepare(s->job, offset, bytes);
1421     }
1422 
1423     switch (method) {
1424     case MIRROR_METHOD_COPY:
1425         ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1426         break;
1427 
1428     case MIRROR_METHOD_ZERO:
1429         ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1430         break;
1431 
1432     case MIRROR_METHOD_DISCARD:
1433         ret = bdrv_co_pdiscard(bs->backing, offset, bytes);
1434         break;
1435 
1436     default:
1437         abort();
1438     }
1439 
1440     if (ret < 0) {
1441         goto out;
1442     }
1443 
1444     if (copy_to_target) {
1445         do_sync_target_write(s->job, method, offset, bytes, qiov, flags);
1446     }
1447 
1448 out:
1449     if (copy_to_target) {
1450         active_write_settle(op);
1451     }
1452     return ret;
1453 }
1454 
1455 static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs,
1456     int64_t offset, int64_t bytes, QEMUIOVector *qiov, BdrvRequestFlags flags)
1457 {
1458     MirrorBDSOpaque *s = bs->opaque;
1459     QEMUIOVector bounce_qiov;
1460     void *bounce_buf;
1461     int ret = 0;
1462     bool copy_to_target;
1463 
1464     copy_to_target = s->job->ret >= 0 &&
1465                      !job_is_cancelled(&s->job->common.job) &&
1466                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1467 
1468     if (copy_to_target) {
1469         /* The guest might concurrently modify the data to write; but
1470          * the data on source and destination must match, so we have
1471          * to use a bounce buffer if we are going to write to the
1472          * target now. */
1473         bounce_buf = qemu_blockalign(bs, bytes);
1474         iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes);
1475 
1476         qemu_iovec_init(&bounce_qiov, 1);
1477         qemu_iovec_add(&bounce_qiov, bounce_buf, bytes);
1478         qiov = &bounce_qiov;
1479     }
1480 
1481     ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov,
1482                                    flags);
1483 
1484     if (copy_to_target) {
1485         qemu_iovec_destroy(&bounce_qiov);
1486         qemu_vfree(bounce_buf);
1487     }
1488 
1489     return ret;
1490 }
1491 
1492 static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs)
1493 {
1494     if (bs->backing == NULL) {
1495         /* we can be here after failed bdrv_append in mirror_start_job */
1496         return 0;
1497     }
1498     return bdrv_co_flush(bs->backing->bs);
1499 }
1500 
1501 static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs,
1502     int64_t offset, int64_t bytes, BdrvRequestFlags flags)
1503 {
1504     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL,
1505                                     flags);
1506 }
1507 
1508 static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs,
1509     int64_t offset, int64_t bytes)
1510 {
1511     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes,
1512                                     NULL, 0);
1513 }
1514 
1515 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs)
1516 {
1517     if (bs->backing == NULL) {
1518         /* we can be here after failed bdrv_attach_child in
1519          * bdrv_set_backing_hd */
1520         return;
1521     }
1522     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1523             bs->backing->bs->filename);
1524 }
1525 
1526 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1527                                        BdrvChildRole role,
1528                                        BlockReopenQueue *reopen_queue,
1529                                        uint64_t perm, uint64_t shared,
1530                                        uint64_t *nperm, uint64_t *nshared)
1531 {
1532     MirrorBDSOpaque *s = bs->opaque;
1533 
1534     if (s->stop) {
1535         /*
1536          * If the job is to be stopped, we do not need to forward
1537          * anything to the real image.
1538          */
1539         *nperm = 0;
1540         *nshared = BLK_PERM_ALL;
1541         return;
1542     }
1543 
1544     bdrv_default_perms(bs, c, role, reopen_queue,
1545                        perm, shared, nperm, nshared);
1546 
1547     if (s->is_commit) {
1548         /*
1549          * For commit jobs, we cannot take CONSISTENT_READ, because
1550          * that permission is unshared for everything above the base
1551          * node (except for filters on the base node).
1552          * We also have to force-share the WRITE permission, or
1553          * otherwise we would block ourselves at the base node (if
1554          * writes are blocked for a node, they are also blocked for
1555          * its backing file).
1556          * (We could also share RESIZE, because it may be needed for
1557          * the target if its size is less than the top node's; but
1558          * bdrv_default_perms_for_cow() automatically shares RESIZE
1559          * for backing nodes if WRITE is shared, so there is no need
1560          * to do it here.)
1561          */
1562         *nperm &= ~BLK_PERM_CONSISTENT_READ;
1563         *nshared |= BLK_PERM_WRITE;
1564     }
1565 }
1566 
1567 /* Dummy node that provides consistent read to its users without requiring it
1568  * from its backing file and that allows writes on the backing file chain. */
1569 static BlockDriver bdrv_mirror_top = {
1570     .format_name                = "mirror_top",
1571     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1572     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1573     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1574     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1575     .bdrv_co_flush              = bdrv_mirror_top_flush,
1576     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1577     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1578 
1579     .is_filter                  = true,
1580 };
1581 
1582 static BlockJob *mirror_start_job(
1583                              const char *job_id, BlockDriverState *bs,
1584                              int creation_flags, BlockDriverState *target,
1585                              const char *replaces, int64_t speed,
1586                              uint32_t granularity, int64_t buf_size,
1587                              BlockMirrorBackingMode backing_mode,
1588                              bool zero_target,
1589                              BlockdevOnError on_source_error,
1590                              BlockdevOnError on_target_error,
1591                              bool unmap,
1592                              BlockCompletionFunc *cb,
1593                              void *opaque,
1594                              const BlockJobDriver *driver,
1595                              bool is_none_mode, BlockDriverState *base,
1596                              bool auto_complete, const char *filter_node_name,
1597                              bool is_mirror, MirrorCopyMode copy_mode,
1598                              Error **errp)
1599 {
1600     MirrorBlockJob *s;
1601     MirrorBDSOpaque *bs_opaque;
1602     BlockDriverState *mirror_top_bs;
1603     bool target_is_backing;
1604     uint64_t target_perms, target_shared_perms;
1605     int ret;
1606 
1607     if (granularity == 0) {
1608         granularity = bdrv_get_default_bitmap_granularity(target);
1609     }
1610 
1611     assert(is_power_of_2(granularity));
1612 
1613     if (buf_size < 0) {
1614         error_setg(errp, "Invalid parameter 'buf-size'");
1615         return NULL;
1616     }
1617 
1618     if (buf_size == 0) {
1619         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1620     }
1621 
1622     if (bdrv_skip_filters(bs) == bdrv_skip_filters(target)) {
1623         error_setg(errp, "Can't mirror node into itself");
1624         return NULL;
1625     }
1626 
1627     target_is_backing = bdrv_chain_contains(bs, target);
1628 
1629     /* In the case of active commit, add dummy driver to provide consistent
1630      * reads on the top, while disabling it in the intermediate nodes, and make
1631      * the backing chain writable. */
1632     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1633                                          BDRV_O_RDWR, errp);
1634     if (mirror_top_bs == NULL) {
1635         return NULL;
1636     }
1637     if (!filter_node_name) {
1638         mirror_top_bs->implicit = true;
1639     }
1640 
1641     /* So that we can always drop this node */
1642     mirror_top_bs->never_freeze = true;
1643 
1644     mirror_top_bs->total_sectors = bs->total_sectors;
1645     mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
1646     mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
1647                                           BDRV_REQ_NO_FALLBACK;
1648     bs_opaque = g_new0(MirrorBDSOpaque, 1);
1649     mirror_top_bs->opaque = bs_opaque;
1650 
1651     bs_opaque->is_commit = target_is_backing;
1652 
1653     bdrv_drained_begin(bs);
1654     ret = bdrv_append(mirror_top_bs, bs, errp);
1655     bdrv_drained_end(bs);
1656 
1657     if (ret < 0) {
1658         bdrv_unref(mirror_top_bs);
1659         return NULL;
1660     }
1661 
1662     /* Make sure that the source is not resized while the job is running */
1663     s = block_job_create(job_id, driver, NULL, mirror_top_bs,
1664                          BLK_PERM_CONSISTENT_READ,
1665                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1666                          BLK_PERM_WRITE, speed,
1667                          creation_flags, cb, opaque, errp);
1668     if (!s) {
1669         goto fail;
1670     }
1671     bs_opaque->job = s;
1672 
1673     /* The block job now has a reference to this node */
1674     bdrv_unref(mirror_top_bs);
1675 
1676     s->mirror_top_bs = mirror_top_bs;
1677 
1678     /* No resize for the target either; while the mirror is still running, a
1679      * consistent read isn't necessarily possible. We could possibly allow
1680      * writes and graph modifications, though it would likely defeat the
1681      * purpose of a mirror, so leave them blocked for now.
1682      *
1683      * In the case of active commit, things look a bit different, though,
1684      * because the target is an already populated backing file in active use.
1685      * We can allow anything except resize there.*/
1686 
1687     target_perms = BLK_PERM_WRITE;
1688     target_shared_perms = BLK_PERM_WRITE_UNCHANGED;
1689 
1690     if (target_is_backing) {
1691         int64_t bs_size, target_size;
1692         bs_size = bdrv_getlength(bs);
1693         if (bs_size < 0) {
1694             error_setg_errno(errp, -bs_size,
1695                              "Could not inquire top image size");
1696             goto fail;
1697         }
1698 
1699         target_size = bdrv_getlength(target);
1700         if (target_size < 0) {
1701             error_setg_errno(errp, -target_size,
1702                              "Could not inquire base image size");
1703             goto fail;
1704         }
1705 
1706         if (target_size < bs_size) {
1707             target_perms |= BLK_PERM_RESIZE;
1708         }
1709 
1710         target_shared_perms |= BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE;
1711     } else if (bdrv_chain_contains(bs, bdrv_skip_filters(target))) {
1712         /*
1713          * We may want to allow this in the future, but it would
1714          * require taking some extra care.
1715          */
1716         error_setg(errp, "Cannot mirror to a filter on top of a node in the "
1717                    "source's backing chain");
1718         goto fail;
1719     }
1720 
1721     s->target = blk_new(s->common.job.aio_context,
1722                         target_perms, target_shared_perms);
1723     ret = blk_insert_bs(s->target, target, errp);
1724     if (ret < 0) {
1725         goto fail;
1726     }
1727     if (is_mirror) {
1728         /* XXX: Mirror target could be a NBD server of target QEMU in the case
1729          * of non-shared block migration. To allow migration completion, we
1730          * have to allow "inactivate" of the target BB.  When that happens, we
1731          * know the job is drained, and the vcpus are stopped, so no write
1732          * operation will be performed. Block layer already has assertions to
1733          * ensure that. */
1734         blk_set_force_allow_inactivate(s->target);
1735     }
1736     blk_set_allow_aio_context_change(s->target, true);
1737     blk_set_disable_request_queuing(s->target, true);
1738 
1739     s->replaces = g_strdup(replaces);
1740     s->on_source_error = on_source_error;
1741     s->on_target_error = on_target_error;
1742     s->is_none_mode = is_none_mode;
1743     s->backing_mode = backing_mode;
1744     s->zero_target = zero_target;
1745     s->copy_mode = copy_mode;
1746     s->base = base;
1747     s->base_overlay = bdrv_find_overlay(bs, base);
1748     s->granularity = granularity;
1749     s->buf_size = ROUND_UP(buf_size, granularity);
1750     s->unmap = unmap;
1751     if (auto_complete) {
1752         s->should_complete = true;
1753     }
1754 
1755     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1756     if (!s->dirty_bitmap) {
1757         goto fail;
1758     }
1759     if (s->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING) {
1760         bdrv_disable_dirty_bitmap(s->dirty_bitmap);
1761     }
1762 
1763     ret = block_job_add_bdrv(&s->common, "source", bs, 0,
1764                              BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE |
1765                              BLK_PERM_CONSISTENT_READ,
1766                              errp);
1767     if (ret < 0) {
1768         goto fail;
1769     }
1770 
1771     /* Required permissions are already taken with blk_new() */
1772     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1773                        &error_abort);
1774 
1775     /* In commit_active_start() all intermediate nodes disappear, so
1776      * any jobs in them must be blocked */
1777     if (target_is_backing) {
1778         BlockDriverState *iter, *filtered_target;
1779         uint64_t iter_shared_perms;
1780 
1781         /*
1782          * The topmost node with
1783          * bdrv_skip_filters(filtered_target) == bdrv_skip_filters(target)
1784          */
1785         filtered_target = bdrv_cow_bs(bdrv_find_overlay(bs, target));
1786 
1787         assert(bdrv_skip_filters(filtered_target) ==
1788                bdrv_skip_filters(target));
1789 
1790         /*
1791          * XXX BLK_PERM_WRITE needs to be allowed so we don't block
1792          * ourselves at s->base (if writes are blocked for a node, they are
1793          * also blocked for its backing file). The other options would be a
1794          * second filter driver above s->base (== target).
1795          */
1796         iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE;
1797 
1798         for (iter = bdrv_filter_or_cow_bs(bs); iter != target;
1799              iter = bdrv_filter_or_cow_bs(iter))
1800         {
1801             if (iter == filtered_target) {
1802                 /*
1803                  * From here on, all nodes are filters on the base.
1804                  * This allows us to share BLK_PERM_CONSISTENT_READ.
1805                  */
1806                 iter_shared_perms |= BLK_PERM_CONSISTENT_READ;
1807             }
1808 
1809             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1810                                      iter_shared_perms, errp);
1811             if (ret < 0) {
1812                 goto fail;
1813             }
1814         }
1815 
1816         if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) {
1817             goto fail;
1818         }
1819     }
1820 
1821     QTAILQ_INIT(&s->ops_in_flight);
1822 
1823     trace_mirror_start(bs, s, opaque);
1824     job_start(&s->common.job);
1825 
1826     return &s->common;
1827 
1828 fail:
1829     if (s) {
1830         /* Make sure this BDS does not go away until we have completed the graph
1831          * changes below */
1832         bdrv_ref(mirror_top_bs);
1833 
1834         g_free(s->replaces);
1835         blk_unref(s->target);
1836         bs_opaque->job = NULL;
1837         if (s->dirty_bitmap) {
1838             bdrv_release_dirty_bitmap(s->dirty_bitmap);
1839         }
1840         job_early_fail(&s->common.job);
1841     }
1842 
1843     bs_opaque->stop = true;
1844     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
1845                              &error_abort);
1846     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
1847 
1848     bdrv_unref(mirror_top_bs);
1849 
1850     return NULL;
1851 }
1852 
1853 void mirror_start(const char *job_id, BlockDriverState *bs,
1854                   BlockDriverState *target, const char *replaces,
1855                   int creation_flags, int64_t speed,
1856                   uint32_t granularity, int64_t buf_size,
1857                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1858                   bool zero_target,
1859                   BlockdevOnError on_source_error,
1860                   BlockdevOnError on_target_error,
1861                   bool unmap, const char *filter_node_name,
1862                   MirrorCopyMode copy_mode, Error **errp)
1863 {
1864     bool is_none_mode;
1865     BlockDriverState *base;
1866 
1867     if ((mode == MIRROR_SYNC_MODE_INCREMENTAL) ||
1868         (mode == MIRROR_SYNC_MODE_BITMAP)) {
1869         error_setg(errp, "Sync mode '%s' not supported",
1870                    MirrorSyncMode_str(mode));
1871         return;
1872     }
1873     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1874     base = mode == MIRROR_SYNC_MODE_TOP ? bdrv_backing_chain_next(bs) : NULL;
1875     mirror_start_job(job_id, bs, creation_flags, target, replaces,
1876                      speed, granularity, buf_size, backing_mode, zero_target,
1877                      on_source_error, on_target_error, unmap, NULL, NULL,
1878                      &mirror_job_driver, is_none_mode, base, false,
1879                      filter_node_name, true, copy_mode, errp);
1880 }
1881 
1882 BlockJob *commit_active_start(const char *job_id, BlockDriverState *bs,
1883                               BlockDriverState *base, int creation_flags,
1884                               int64_t speed, BlockdevOnError on_error,
1885                               const char *filter_node_name,
1886                               BlockCompletionFunc *cb, void *opaque,
1887                               bool auto_complete, Error **errp)
1888 {
1889     bool base_read_only;
1890     BlockJob *job;
1891 
1892     base_read_only = bdrv_is_read_only(base);
1893 
1894     if (base_read_only) {
1895         if (bdrv_reopen_set_read_only(base, false, errp) < 0) {
1896             return NULL;
1897         }
1898     }
1899 
1900     job = mirror_start_job(
1901                      job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1902                      MIRROR_LEAVE_BACKING_CHAIN, false,
1903                      on_error, on_error, true, cb, opaque,
1904                      &commit_active_job_driver, false, base, auto_complete,
1905                      filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND,
1906                      errp);
1907     if (!job) {
1908         goto error_restore_flags;
1909     }
1910 
1911     return job;
1912 
1913 error_restore_flags:
1914     /* ignore error and errp for bdrv_reopen, because we want to propagate
1915      * the original error */
1916     if (base_read_only) {
1917         bdrv_reopen_set_read_only(base, true, NULL);
1918     }
1919     return NULL;
1920 }
1921