xref: /openbmc/qemu/block/mirror.c (revision f085ba29)
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 #include "qemu/memalign.h"
27 
28 #define MAX_IN_FLIGHT 16
29 #define MAX_IO_BYTES (1 << 20) /* 1 Mb */
30 #define DEFAULT_MIRROR_BUF_SIZE (MAX_IN_FLIGHT * MAX_IO_BYTES)
31 
32 /* The mirroring buffer is a list of granularity-sized chunks.
33  * Free chunks are organized in a list.
34  */
35 typedef struct MirrorBuffer {
36     QSIMPLEQ_ENTRY(MirrorBuffer) next;
37 } MirrorBuffer;
38 
39 typedef struct MirrorOp MirrorOp;
40 
41 typedef struct MirrorBlockJob {
42     BlockJob common;
43     BlockBackend *target;
44     BlockDriverState *mirror_top_bs;
45     BlockDriverState *base;
46     BlockDriverState *base_overlay;
47 
48     /* The name of the graph node to replace */
49     char *replaces;
50     /* The BDS to replace */
51     BlockDriverState *to_replace;
52     /* Used to block operations on the drive-mirror-replace target */
53     Error *replace_blocker;
54     bool is_none_mode;
55     BlockMirrorBackingMode backing_mode;
56     /* Whether the target image requires explicit zero-initialization */
57     bool zero_target;
58     MirrorCopyMode copy_mode;
59     BlockdevOnError on_source_error, on_target_error;
60     /* Set when the target is synced (dirty bitmap is clean, nothing
61      * in flight) and the job is running in active mode */
62     bool actively_synced;
63     bool should_complete;
64     int64_t granularity;
65     size_t buf_size;
66     int64_t bdev_length;
67     unsigned long *cow_bitmap;
68     BdrvDirtyBitmap *dirty_bitmap;
69     BdrvDirtyBitmapIter *dbi;
70     uint8_t *buf;
71     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
72     int buf_free_count;
73 
74     uint64_t last_pause_ns;
75     unsigned long *in_flight_bitmap;
76     unsigned in_flight;
77     int64_t bytes_in_flight;
78     QTAILQ_HEAD(, MirrorOp) ops_in_flight;
79     int ret;
80     bool unmap;
81     int target_cluster_size;
82     int max_iov;
83     bool initial_zeroing_ongoing;
84     int in_active_write_counter;
85     bool prepared;
86     bool in_drain;
87 } MirrorBlockJob;
88 
89 typedef struct MirrorBDSOpaque {
90     MirrorBlockJob *job;
91     bool stop;
92     bool is_commit;
93 } MirrorBDSOpaque;
94 
95 struct MirrorOp {
96     MirrorBlockJob *s;
97     QEMUIOVector qiov;
98     int64_t offset;
99     uint64_t bytes;
100 
101     /* The pointee is set by mirror_co_read(), mirror_co_zero(), and
102      * mirror_co_discard() before yielding for the first time */
103     int64_t *bytes_handled;
104 
105     bool is_pseudo_op;
106     bool is_active_write;
107     bool is_in_flight;
108     CoQueue waiting_requests;
109     Coroutine *co;
110     MirrorOp *waiting_for_op;
111 
112     QTAILQ_ENTRY(MirrorOp) next;
113 };
114 
115 typedef enum MirrorMethod {
116     MIRROR_METHOD_COPY,
117     MIRROR_METHOD_ZERO,
118     MIRROR_METHOD_DISCARD,
119 } MirrorMethod;
120 
121 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
122                                             int error)
123 {
124     s->actively_synced = false;
125     if (read) {
126         return block_job_error_action(&s->common, s->on_source_error,
127                                       true, error);
128     } else {
129         return block_job_error_action(&s->common, s->on_target_error,
130                                       false, error);
131     }
132 }
133 
134 static void coroutine_fn mirror_wait_on_conflicts(MirrorOp *self,
135                                                   MirrorBlockJob *s,
136                                                   uint64_t offset,
137                                                   uint64_t bytes)
138 {
139     uint64_t self_start_chunk = offset / s->granularity;
140     uint64_t self_end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
141     uint64_t self_nb_chunks = self_end_chunk - self_start_chunk;
142 
143     while (find_next_bit(s->in_flight_bitmap, self_end_chunk,
144                          self_start_chunk) < self_end_chunk &&
145            s->ret >= 0)
146     {
147         MirrorOp *op;
148 
149         QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
150             uint64_t op_start_chunk = op->offset / s->granularity;
151             uint64_t op_nb_chunks = DIV_ROUND_UP(op->offset + op->bytes,
152                                                  s->granularity) -
153                                     op_start_chunk;
154 
155             if (op == self) {
156                 continue;
157             }
158 
159             if (ranges_overlap(self_start_chunk, self_nb_chunks,
160                                op_start_chunk, op_nb_chunks))
161             {
162                 if (self) {
163                     /*
164                      * If the operation is already (indirectly) waiting for us,
165                      * or will wait for us as soon as it wakes up, then just go
166                      * on (instead of producing a deadlock in the former case).
167                      */
168                     if (op->waiting_for_op) {
169                         continue;
170                     }
171 
172                     self->waiting_for_op = op;
173                 }
174 
175                 qemu_co_queue_wait(&op->waiting_requests, NULL);
176 
177                 if (self) {
178                     self->waiting_for_op = NULL;
179                 }
180 
181                 break;
182             }
183         }
184     }
185 }
186 
187 static void coroutine_fn mirror_iteration_done(MirrorOp *op, int ret)
188 {
189     MirrorBlockJob *s = op->s;
190     struct iovec *iov;
191     int64_t chunk_num;
192     int i, nb_chunks;
193 
194     trace_mirror_iteration_done(s, op->offset, op->bytes, ret);
195 
196     s->in_flight--;
197     s->bytes_in_flight -= op->bytes;
198     iov = op->qiov.iov;
199     for (i = 0; i < op->qiov.niov; i++) {
200         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
201         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
202         s->buf_free_count++;
203     }
204 
205     chunk_num = op->offset / s->granularity;
206     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
207 
208     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
209     QTAILQ_REMOVE(&s->ops_in_flight, op, next);
210     if (ret >= 0) {
211         if (s->cow_bitmap) {
212             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
213         }
214         if (!s->initial_zeroing_ongoing) {
215             job_progress_update(&s->common.job, op->bytes);
216         }
217     }
218     qemu_iovec_destroy(&op->qiov);
219 
220     qemu_co_queue_restart_all(&op->waiting_requests);
221     g_free(op);
222 }
223 
224 static void coroutine_fn mirror_write_complete(MirrorOp *op, int ret)
225 {
226     MirrorBlockJob *s = op->s;
227 
228     if (ret < 0) {
229         BlockErrorAction action;
230 
231         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
232         action = mirror_error_action(s, false, -ret);
233         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
234             s->ret = ret;
235         }
236     }
237 
238     mirror_iteration_done(op, ret);
239 }
240 
241 static void coroutine_fn mirror_read_complete(MirrorOp *op, int ret)
242 {
243     MirrorBlockJob *s = op->s;
244 
245     if (ret < 0) {
246         BlockErrorAction action;
247 
248         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->offset, op->bytes);
249         action = mirror_error_action(s, true, -ret);
250         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
251             s->ret = ret;
252         }
253 
254         mirror_iteration_done(op, ret);
255         return;
256     }
257 
258     ret = blk_co_pwritev(s->target, op->offset, op->qiov.size, &op->qiov, 0);
259     mirror_write_complete(op, ret);
260 }
261 
262 /* Clip bytes relative to offset to not exceed end-of-file */
263 static inline int64_t mirror_clip_bytes(MirrorBlockJob *s,
264                                         int64_t offset,
265                                         int64_t bytes)
266 {
267     return MIN(bytes, s->bdev_length - offset);
268 }
269 
270 /* Round offset and/or bytes to target cluster if COW is needed, and
271  * return the offset of the adjusted tail against original. */
272 static int mirror_cow_align(MirrorBlockJob *s, int64_t *offset,
273                             uint64_t *bytes)
274 {
275     bool need_cow;
276     int ret = 0;
277     int64_t align_offset = *offset;
278     int64_t align_bytes = *bytes;
279     int max_bytes = s->granularity * s->max_iov;
280 
281     need_cow = !test_bit(*offset / s->granularity, s->cow_bitmap);
282     need_cow |= !test_bit((*offset + *bytes - 1) / s->granularity,
283                           s->cow_bitmap);
284     if (need_cow) {
285         bdrv_round_to_clusters(blk_bs(s->target), *offset, *bytes,
286                                &align_offset, &align_bytes);
287     }
288 
289     if (align_bytes > max_bytes) {
290         align_bytes = max_bytes;
291         if (need_cow) {
292             align_bytes = QEMU_ALIGN_DOWN(align_bytes, s->target_cluster_size);
293         }
294     }
295     /* Clipping may result in align_bytes unaligned to chunk boundary, but
296      * that doesn't matter because it's already the end of source image. */
297     align_bytes = mirror_clip_bytes(s, align_offset, align_bytes);
298 
299     ret = align_offset + align_bytes - (*offset + *bytes);
300     *offset = align_offset;
301     *bytes = align_bytes;
302     assert(ret >= 0);
303     return ret;
304 }
305 
306 static inline void coroutine_fn
307 mirror_wait_for_any_operation(MirrorBlockJob *s, bool active)
308 {
309     MirrorOp *op;
310 
311     QTAILQ_FOREACH(op, &s->ops_in_flight, next) {
312         /* Do not wait on pseudo ops, because it may in turn wait on
313          * some other operation to start, which may in fact be the
314          * caller of this function.  Since there is only one pseudo op
315          * at any given time, we will always find some real operation
316          * to wait on. */
317         if (!op->is_pseudo_op && op->is_in_flight &&
318             op->is_active_write == active)
319         {
320             qemu_co_queue_wait(&op->waiting_requests, NULL);
321             return;
322         }
323     }
324     abort();
325 }
326 
327 static inline void coroutine_fn
328 mirror_wait_for_free_in_flight_slot(MirrorBlockJob *s)
329 {
330     /* Only non-active operations use up in-flight slots */
331     mirror_wait_for_any_operation(s, false);
332 }
333 
334 /* Perform a mirror copy operation.
335  *
336  * *op->bytes_handled is set to the number of bytes copied after and
337  * including offset, excluding any bytes copied prior to offset due
338  * to alignment.  This will be op->bytes if no alignment is necessary,
339  * or (new_end - op->offset) if the tail is rounded up or down due to
340  * alignment or buffer limit.
341  */
342 static void coroutine_fn mirror_co_read(void *opaque)
343 {
344     MirrorOp *op = opaque;
345     MirrorBlockJob *s = op->s;
346     int nb_chunks;
347     uint64_t ret;
348     uint64_t max_bytes;
349 
350     max_bytes = s->granularity * s->max_iov;
351 
352     /* We can only handle as much as buf_size at a time. */
353     op->bytes = MIN(s->buf_size, MIN(max_bytes, op->bytes));
354     assert(op->bytes);
355     assert(op->bytes < BDRV_REQUEST_MAX_BYTES);
356     *op->bytes_handled = op->bytes;
357 
358     if (s->cow_bitmap) {
359         *op->bytes_handled += mirror_cow_align(s, &op->offset, &op->bytes);
360     }
361     /* Cannot exceed BDRV_REQUEST_MAX_BYTES + INT_MAX */
362     assert(*op->bytes_handled <= UINT_MAX);
363     assert(op->bytes <= s->buf_size);
364     /* The offset is granularity-aligned because:
365      * 1) Caller passes in aligned values;
366      * 2) mirror_cow_align is used only when target cluster is larger. */
367     assert(QEMU_IS_ALIGNED(op->offset, s->granularity));
368     /* The range is sector-aligned, since bdrv_getlength() rounds up. */
369     assert(QEMU_IS_ALIGNED(op->bytes, BDRV_SECTOR_SIZE));
370     nb_chunks = DIV_ROUND_UP(op->bytes, s->granularity);
371 
372     while (s->buf_free_count < nb_chunks) {
373         trace_mirror_yield_in_flight(s, op->offset, s->in_flight);
374         mirror_wait_for_free_in_flight_slot(s);
375     }
376 
377     /* Now make a QEMUIOVector taking enough granularity-sized chunks
378      * from s->buf_free.
379      */
380     qemu_iovec_init(&op->qiov, nb_chunks);
381     while (nb_chunks-- > 0) {
382         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
383         size_t remaining = op->bytes - op->qiov.size;
384 
385         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
386         s->buf_free_count--;
387         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
388     }
389 
390     /* Copy the dirty cluster.  */
391     s->in_flight++;
392     s->bytes_in_flight += op->bytes;
393     op->is_in_flight = true;
394     trace_mirror_one_iteration(s, op->offset, op->bytes);
395 
396     ret = bdrv_co_preadv(s->mirror_top_bs->backing, op->offset, op->bytes,
397                          &op->qiov, 0);
398     mirror_read_complete(op, ret);
399 }
400 
401 static void coroutine_fn mirror_co_zero(void *opaque)
402 {
403     MirrorOp *op = opaque;
404     int ret;
405 
406     op->s->in_flight++;
407     op->s->bytes_in_flight += op->bytes;
408     *op->bytes_handled = op->bytes;
409     op->is_in_flight = true;
410 
411     ret = blk_co_pwrite_zeroes(op->s->target, op->offset, op->bytes,
412                                op->s->unmap ? BDRV_REQ_MAY_UNMAP : 0);
413     mirror_write_complete(op, ret);
414 }
415 
416 static void coroutine_fn mirror_co_discard(void *opaque)
417 {
418     MirrorOp *op = opaque;
419     int ret;
420 
421     op->s->in_flight++;
422     op->s->bytes_in_flight += op->bytes;
423     *op->bytes_handled = op->bytes;
424     op->is_in_flight = true;
425 
426     ret = blk_co_pdiscard(op->s->target, op->offset, op->bytes);
427     mirror_write_complete(op, ret);
428 }
429 
430 static unsigned mirror_perform(MirrorBlockJob *s, int64_t offset,
431                                unsigned bytes, MirrorMethod mirror_method)
432 {
433     MirrorOp *op;
434     Coroutine *co;
435     int64_t bytes_handled = -1;
436 
437     op = g_new(MirrorOp, 1);
438     *op = (MirrorOp){
439         .s              = s,
440         .offset         = offset,
441         .bytes          = bytes,
442         .bytes_handled  = &bytes_handled,
443     };
444     qemu_co_queue_init(&op->waiting_requests);
445 
446     switch (mirror_method) {
447     case MIRROR_METHOD_COPY:
448         co = qemu_coroutine_create(mirror_co_read, op);
449         break;
450     case MIRROR_METHOD_ZERO:
451         co = qemu_coroutine_create(mirror_co_zero, op);
452         break;
453     case MIRROR_METHOD_DISCARD:
454         co = qemu_coroutine_create(mirror_co_discard, op);
455         break;
456     default:
457         abort();
458     }
459     op->co = co;
460 
461     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
462     qemu_coroutine_enter(co);
463     /* At this point, ownership of op has been moved to the coroutine
464      * and the object may already be freed */
465 
466     /* Assert that this value has been set */
467     assert(bytes_handled >= 0);
468 
469     /* Same assertion as in mirror_co_read() (and for mirror_co_read()
470      * and mirror_co_discard(), bytes_handled == op->bytes, which
471      * is the @bytes parameter given to this function) */
472     assert(bytes_handled <= UINT_MAX);
473     return bytes_handled;
474 }
475 
476 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
477 {
478     BlockDriverState *source = s->mirror_top_bs->backing->bs;
479     MirrorOp *pseudo_op;
480     int64_t offset;
481     uint64_t delay_ns = 0, ret = 0;
482     /* At least the first dirty chunk is mirrored in one iteration. */
483     int nb_chunks = 1;
484     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
485     int max_io_bytes = MAX(s->buf_size / MAX_IN_FLIGHT, MAX_IO_BYTES);
486 
487     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
488     offset = bdrv_dirty_iter_next(s->dbi);
489     if (offset < 0) {
490         bdrv_set_dirty_iter(s->dbi, 0);
491         offset = bdrv_dirty_iter_next(s->dbi);
492         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
493         assert(offset >= 0);
494     }
495     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
496 
497     mirror_wait_on_conflicts(NULL, s, offset, 1);
498 
499     job_pause_point(&s->common.job);
500 
501     /* Find the number of consective dirty chunks following the first dirty
502      * one, and wait for in flight requests in them. */
503     bdrv_dirty_bitmap_lock(s->dirty_bitmap);
504     while (nb_chunks * s->granularity < s->buf_size) {
505         int64_t next_dirty;
506         int64_t next_offset = offset + nb_chunks * s->granularity;
507         int64_t next_chunk = next_offset / s->granularity;
508         if (next_offset >= s->bdev_length ||
509             !bdrv_dirty_bitmap_get_locked(s->dirty_bitmap, next_offset)) {
510             break;
511         }
512         if (test_bit(next_chunk, s->in_flight_bitmap)) {
513             break;
514         }
515 
516         next_dirty = bdrv_dirty_iter_next(s->dbi);
517         if (next_dirty > next_offset || next_dirty < 0) {
518             /* The bitmap iterator's cache is stale, refresh it */
519             bdrv_set_dirty_iter(s->dbi, next_offset);
520             next_dirty = bdrv_dirty_iter_next(s->dbi);
521         }
522         assert(next_dirty == next_offset);
523         nb_chunks++;
524     }
525 
526     /* Clear dirty bits before querying the block status, because
527      * calling bdrv_block_status_above could yield - if some blocks are
528      * marked dirty in this window, we need to know.
529      */
530     bdrv_reset_dirty_bitmap_locked(s->dirty_bitmap, offset,
531                                    nb_chunks * s->granularity);
532     bdrv_dirty_bitmap_unlock(s->dirty_bitmap);
533 
534     /* Before claiming an area in the in-flight bitmap, we have to
535      * create a MirrorOp for it so that conflicting requests can wait
536      * for it.  mirror_perform() will create the real MirrorOps later,
537      * for now we just create a pseudo operation that will wake up all
538      * conflicting requests once all real operations have been
539      * launched. */
540     pseudo_op = g_new(MirrorOp, 1);
541     *pseudo_op = (MirrorOp){
542         .offset         = offset,
543         .bytes          = nb_chunks * s->granularity,
544         .is_pseudo_op   = true,
545     };
546     qemu_co_queue_init(&pseudo_op->waiting_requests);
547     QTAILQ_INSERT_TAIL(&s->ops_in_flight, pseudo_op, next);
548 
549     bitmap_set(s->in_flight_bitmap, offset / s->granularity, nb_chunks);
550     while (nb_chunks > 0 && offset < s->bdev_length) {
551         int ret;
552         int64_t io_bytes;
553         int64_t io_bytes_acct;
554         MirrorMethod mirror_method = MIRROR_METHOD_COPY;
555 
556         assert(!(offset % s->granularity));
557         ret = bdrv_block_status_above(source, NULL, offset,
558                                       nb_chunks * s->granularity,
559                                       &io_bytes, NULL, NULL);
560         if (ret < 0) {
561             io_bytes = MIN(nb_chunks * s->granularity, max_io_bytes);
562         } else if (ret & BDRV_BLOCK_DATA) {
563             io_bytes = MIN(io_bytes, max_io_bytes);
564         }
565 
566         io_bytes -= io_bytes % s->granularity;
567         if (io_bytes < s->granularity) {
568             io_bytes = s->granularity;
569         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
570             int64_t target_offset;
571             int64_t target_bytes;
572             bdrv_round_to_clusters(blk_bs(s->target), offset, io_bytes,
573                                    &target_offset, &target_bytes);
574             if (target_offset == offset &&
575                 target_bytes == io_bytes) {
576                 mirror_method = ret & BDRV_BLOCK_ZERO ?
577                                     MIRROR_METHOD_ZERO :
578                                     MIRROR_METHOD_DISCARD;
579             }
580         }
581 
582         while (s->in_flight >= MAX_IN_FLIGHT) {
583             trace_mirror_yield_in_flight(s, offset, s->in_flight);
584             mirror_wait_for_free_in_flight_slot(s);
585         }
586 
587         if (s->ret < 0) {
588             ret = 0;
589             goto fail;
590         }
591 
592         io_bytes = mirror_clip_bytes(s, offset, io_bytes);
593         io_bytes = mirror_perform(s, offset, io_bytes, mirror_method);
594         if (mirror_method != MIRROR_METHOD_COPY && write_zeroes_ok) {
595             io_bytes_acct = 0;
596         } else {
597             io_bytes_acct = io_bytes;
598         }
599         assert(io_bytes);
600         offset += io_bytes;
601         nb_chunks -= DIV_ROUND_UP(io_bytes, s->granularity);
602         delay_ns = block_job_ratelimit_get_delay(&s->common, io_bytes_acct);
603     }
604 
605     ret = delay_ns;
606 fail:
607     QTAILQ_REMOVE(&s->ops_in_flight, pseudo_op, next);
608     qemu_co_queue_restart_all(&pseudo_op->waiting_requests);
609     g_free(pseudo_op);
610 
611     return ret;
612 }
613 
614 static void mirror_free_init(MirrorBlockJob *s)
615 {
616     int granularity = s->granularity;
617     size_t buf_size = s->buf_size;
618     uint8_t *buf = s->buf;
619 
620     assert(s->buf_free_count == 0);
621     QSIMPLEQ_INIT(&s->buf_free);
622     while (buf_size != 0) {
623         MirrorBuffer *cur = (MirrorBuffer *)buf;
624         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
625         s->buf_free_count++;
626         buf_size -= granularity;
627         buf += granularity;
628     }
629 }
630 
631 /* This is also used for the .pause callback. There is no matching
632  * mirror_resume() because mirror_run() will begin iterating again
633  * when the job is resumed.
634  */
635 static void coroutine_fn mirror_wait_for_all_io(MirrorBlockJob *s)
636 {
637     while (s->in_flight > 0) {
638         mirror_wait_for_free_in_flight_slot(s);
639     }
640 }
641 
642 /**
643  * mirror_exit_common: handle both abort() and prepare() cases.
644  * for .prepare, returns 0 on success and -errno on failure.
645  * for .abort cases, denoted by abort = true, MUST return 0.
646  */
647 static int mirror_exit_common(Job *job)
648 {
649     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
650     BlockJob *bjob = &s->common;
651     MirrorBDSOpaque *bs_opaque;
652     AioContext *replace_aio_context = NULL;
653     BlockDriverState *src;
654     BlockDriverState *target_bs;
655     BlockDriverState *mirror_top_bs;
656     Error *local_err = NULL;
657     bool abort = job->ret < 0;
658     int ret = 0;
659 
660     if (s->prepared) {
661         return 0;
662     }
663     s->prepared = true;
664 
665     mirror_top_bs = s->mirror_top_bs;
666     bs_opaque = mirror_top_bs->opaque;
667     src = mirror_top_bs->backing->bs;
668     target_bs = blk_bs(s->target);
669 
670     if (bdrv_chain_contains(src, target_bs)) {
671         bdrv_unfreeze_backing_chain(mirror_top_bs, target_bs);
672     }
673 
674     bdrv_release_dirty_bitmap(s->dirty_bitmap);
675 
676     /* Make sure that the source BDS doesn't go away during bdrv_replace_node,
677      * before we can call bdrv_drained_end */
678     bdrv_ref(src);
679     bdrv_ref(mirror_top_bs);
680     bdrv_ref(target_bs);
681 
682     /*
683      * Remove target parent that still uses BLK_PERM_WRITE/RESIZE before
684      * inserting target_bs at s->to_replace, where we might not be able to get
685      * these permissions.
686      */
687     blk_unref(s->target);
688     s->target = NULL;
689 
690     /* We don't access the source any more. Dropping any WRITE/RESIZE is
691      * required before it could become a backing file of target_bs. Not having
692      * these permissions any more means that we can't allow any new requests on
693      * mirror_top_bs from now on, so keep it drained. */
694     bdrv_drained_begin(mirror_top_bs);
695     bs_opaque->stop = true;
696     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
697                              &error_abort);
698     if (!abort && s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
699         BlockDriverState *backing = s->is_none_mode ? src : s->base;
700         BlockDriverState *unfiltered_target = bdrv_skip_filters(target_bs);
701 
702         if (bdrv_cow_bs(unfiltered_target) != backing) {
703             bdrv_set_backing_hd(unfiltered_target, backing, &local_err);
704             if (local_err) {
705                 error_report_err(local_err);
706                 local_err = NULL;
707                 ret = -EPERM;
708             }
709         }
710     } else if (!abort && s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
711         assert(!bdrv_backing_chain_next(target_bs));
712         ret = bdrv_open_backing_file(bdrv_skip_filters(target_bs), NULL,
713                                      "backing", &local_err);
714         if (ret < 0) {
715             error_report_err(local_err);
716             local_err = NULL;
717         }
718     }
719 
720     if (s->to_replace) {
721         replace_aio_context = bdrv_get_aio_context(s->to_replace);
722         aio_context_acquire(replace_aio_context);
723     }
724 
725     if (s->should_complete && !abort) {
726         BlockDriverState *to_replace = s->to_replace ?: src;
727         bool ro = bdrv_is_read_only(to_replace);
728 
729         if (ro != bdrv_is_read_only(target_bs)) {
730             bdrv_reopen_set_read_only(target_bs, ro, NULL);
731         }
732 
733         /* The mirror job has no requests in flight any more, but we need to
734          * drain potential other users of the BDS before changing the graph. */
735         assert(s->in_drain);
736         bdrv_drained_begin(target_bs);
737         /*
738          * Cannot use check_to_replace_node() here, because that would
739          * check for an op blocker on @to_replace, and we have our own
740          * there.
741          */
742         if (bdrv_recurse_can_replace(src, to_replace)) {
743             bdrv_replace_node(to_replace, target_bs, &local_err);
744         } else {
745             error_setg(&local_err, "Can no longer replace '%s' by '%s', "
746                        "because it can no longer be guaranteed that doing so "
747                        "would not lead to an abrupt change of visible data",
748                        to_replace->node_name, target_bs->node_name);
749         }
750         bdrv_drained_end(target_bs);
751         if (local_err) {
752             error_report_err(local_err);
753             ret = -EPERM;
754         }
755     }
756     if (s->to_replace) {
757         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
758         error_free(s->replace_blocker);
759         bdrv_unref(s->to_replace);
760     }
761     if (replace_aio_context) {
762         aio_context_release(replace_aio_context);
763     }
764     g_free(s->replaces);
765     bdrv_unref(target_bs);
766 
767     /*
768      * Remove the mirror filter driver from the graph. Before this, get rid of
769      * the blockers on the intermediate nodes so that the resulting state is
770      * valid.
771      */
772     block_job_remove_all_bdrv(bjob);
773     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
774 
775     bs_opaque->job = NULL;
776 
777     bdrv_drained_end(src);
778     bdrv_drained_end(mirror_top_bs);
779     s->in_drain = false;
780     bdrv_unref(mirror_top_bs);
781     bdrv_unref(src);
782 
783     return ret;
784 }
785 
786 static int mirror_prepare(Job *job)
787 {
788     return mirror_exit_common(job);
789 }
790 
791 static void mirror_abort(Job *job)
792 {
793     int ret = mirror_exit_common(job);
794     assert(ret == 0);
795 }
796 
797 static void coroutine_fn mirror_throttle(MirrorBlockJob *s)
798 {
799     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
800 
801     if (now - s->last_pause_ns > BLOCK_JOB_SLICE_TIME) {
802         s->last_pause_ns = now;
803         job_sleep_ns(&s->common.job, 0);
804     } else {
805         job_pause_point(&s->common.job);
806     }
807 }
808 
809 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
810 {
811     int64_t offset;
812     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
813     BlockDriverState *target_bs = blk_bs(s->target);
814     int ret;
815     int64_t count;
816 
817     if (s->zero_target) {
818         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
819             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, s->bdev_length);
820             return 0;
821         }
822 
823         s->initial_zeroing_ongoing = true;
824         for (offset = 0; offset < s->bdev_length; ) {
825             int bytes = MIN(s->bdev_length - offset,
826                             QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
827 
828             mirror_throttle(s);
829 
830             if (job_is_cancelled(&s->common.job)) {
831                 s->initial_zeroing_ongoing = false;
832                 return 0;
833             }
834 
835             if (s->in_flight >= MAX_IN_FLIGHT) {
836                 trace_mirror_yield(s, UINT64_MAX, s->buf_free_count,
837                                    s->in_flight);
838                 mirror_wait_for_free_in_flight_slot(s);
839                 continue;
840             }
841 
842             mirror_perform(s, offset, bytes, MIRROR_METHOD_ZERO);
843             offset += bytes;
844         }
845 
846         mirror_wait_for_all_io(s);
847         s->initial_zeroing_ongoing = false;
848     }
849 
850     /* First part, loop on the sectors and initialize the dirty bitmap.  */
851     for (offset = 0; offset < s->bdev_length; ) {
852         /* Just to make sure we are not exceeding int limit. */
853         int bytes = MIN(s->bdev_length - offset,
854                         QEMU_ALIGN_DOWN(INT_MAX, s->granularity));
855 
856         mirror_throttle(s);
857 
858         if (job_is_cancelled(&s->common.job)) {
859             return 0;
860         }
861 
862         ret = bdrv_is_allocated_above(bs, s->base_overlay, true, offset, bytes,
863                                       &count);
864         if (ret < 0) {
865             return ret;
866         }
867 
868         assert(count);
869         if (ret > 0) {
870             bdrv_set_dirty_bitmap(s->dirty_bitmap, offset, count);
871         }
872         offset += count;
873     }
874     return 0;
875 }
876 
877 /* Called when going out of the streaming phase to flush the bulk of the
878  * data to the medium, or just before completing.
879  */
880 static int mirror_flush(MirrorBlockJob *s)
881 {
882     int ret = blk_flush(s->target);
883     if (ret < 0) {
884         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
885             s->ret = ret;
886         }
887     }
888     return ret;
889 }
890 
891 static int coroutine_fn mirror_run(Job *job, Error **errp)
892 {
893     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
894     BlockDriverState *bs = s->mirror_top_bs->backing->bs;
895     BlockDriverState *target_bs = blk_bs(s->target);
896     bool need_drain = true;
897     int64_t length;
898     int64_t target_length;
899     BlockDriverInfo bdi;
900     char backing_filename[2]; /* we only need 2 characters because we are only
901                                  checking for a NULL string */
902     int ret = 0;
903 
904     if (job_is_cancelled(&s->common.job)) {
905         goto immediate_exit;
906     }
907 
908     s->bdev_length = bdrv_getlength(bs);
909     if (s->bdev_length < 0) {
910         ret = s->bdev_length;
911         goto immediate_exit;
912     }
913 
914     target_length = blk_getlength(s->target);
915     if (target_length < 0) {
916         ret = target_length;
917         goto immediate_exit;
918     }
919 
920     /* Active commit must resize the base image if its size differs from the
921      * active layer. */
922     if (s->base == blk_bs(s->target)) {
923         if (s->bdev_length > target_length) {
924             ret = blk_truncate(s->target, s->bdev_length, false,
925                                PREALLOC_MODE_OFF, 0, NULL);
926             if (ret < 0) {
927                 goto immediate_exit;
928             }
929         }
930     } else if (s->bdev_length != target_length) {
931         error_setg(errp, "Source and target image have different sizes");
932         ret = -EINVAL;
933         goto immediate_exit;
934     }
935 
936     if (s->bdev_length == 0) {
937         /* Transition to the READY state and wait for complete. */
938         job_transition_to_ready(&s->common.job);
939         s->actively_synced = true;
940         while (!job_cancel_requested(&s->common.job) && !s->should_complete) {
941             job_yield(&s->common.job);
942         }
943         goto immediate_exit;
944     }
945 
946     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
947     s->in_flight_bitmap = bitmap_new(length);
948 
949     /* If we have no backing file yet in the destination, we cannot let
950      * the destination do COW.  Instead, we copy sectors around the
951      * dirty data if needed.  We need a bitmap to do that.
952      */
953     bdrv_get_backing_filename(target_bs, backing_filename,
954                               sizeof(backing_filename));
955     if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) {
956         s->target_cluster_size = bdi.cluster_size;
957     } else {
958         s->target_cluster_size = BDRV_SECTOR_SIZE;
959     }
960     if (backing_filename[0] && !bdrv_backing_chain_next(target_bs) &&
961         s->granularity < s->target_cluster_size) {
962         s->buf_size = MAX(s->buf_size, s->target_cluster_size);
963         s->cow_bitmap = bitmap_new(length);
964     }
965     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
966 
967     s->buf = qemu_try_blockalign(bs, s->buf_size);
968     if (s->buf == NULL) {
969         ret = -ENOMEM;
970         goto immediate_exit;
971     }
972 
973     mirror_free_init(s);
974 
975     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
976     if (!s->is_none_mode) {
977         ret = mirror_dirty_init(s);
978         if (ret < 0 || job_is_cancelled(&s->common.job)) {
979             goto immediate_exit;
980         }
981     }
982 
983     assert(!s->dbi);
984     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap);
985     for (;;) {
986         uint64_t delay_ns = 0;
987         int64_t cnt, delta;
988         bool should_complete;
989 
990         /* Do not start passive operations while there are active
991          * writes in progress */
992         while (s->in_active_write_counter) {
993             mirror_wait_for_any_operation(s, true);
994         }
995 
996         if (s->ret < 0) {
997             ret = s->ret;
998             goto immediate_exit;
999         }
1000 
1001         job_pause_point(&s->common.job);
1002 
1003         if (job_is_cancelled(&s->common.job)) {
1004             ret = 0;
1005             goto immediate_exit;
1006         }
1007 
1008         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1009         /* cnt is the number of dirty bytes remaining and s->bytes_in_flight is
1010          * the number of bytes currently being processed; together those are
1011          * the current remaining operation length */
1012         job_progress_set_remaining(&s->common.job, s->bytes_in_flight + cnt);
1013 
1014         /* Note that even when no rate limit is applied we need to yield
1015          * periodically with no pending I/O so that bdrv_drain_all() returns.
1016          * We do so every BLKOCK_JOB_SLICE_TIME nanoseconds, or when there is
1017          * an error, or when the source is clean, whichever comes first. */
1018         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
1019         if (delta < BLOCK_JOB_SLICE_TIME &&
1020             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
1021             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
1022                 (cnt == 0 && s->in_flight > 0)) {
1023                 trace_mirror_yield(s, cnt, s->buf_free_count, s->in_flight);
1024                 mirror_wait_for_free_in_flight_slot(s);
1025                 continue;
1026             } else if (cnt != 0) {
1027                 delay_ns = mirror_iteration(s);
1028             }
1029         }
1030 
1031         should_complete = false;
1032         if (s->in_flight == 0 && cnt == 0) {
1033             trace_mirror_before_flush(s);
1034             if (!job_is_ready(&s->common.job)) {
1035                 if (mirror_flush(s) < 0) {
1036                     /* Go check s->ret.  */
1037                     continue;
1038                 }
1039                 /* We're out of the streaming phase.  From now on, if the job
1040                  * is cancelled we will actually complete all pending I/O and
1041                  * report completion.  This way, block-job-cancel will leave
1042                  * the target in a consistent state.
1043                  */
1044                 job_transition_to_ready(&s->common.job);
1045                 if (s->copy_mode != MIRROR_COPY_MODE_BACKGROUND) {
1046                     s->actively_synced = true;
1047                 }
1048             }
1049 
1050             should_complete = s->should_complete ||
1051                 job_cancel_requested(&s->common.job);
1052             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1053         }
1054 
1055         if (cnt == 0 && should_complete) {
1056             /* The dirty bitmap is not updated while operations are pending.
1057              * If we're about to exit, wait for pending operations before
1058              * calling bdrv_get_dirty_count(bs), or we may exit while the
1059              * source has dirty data to copy!
1060              *
1061              * Note that I/O can be submitted by the guest while
1062              * mirror_populate runs, so pause it now.  Before deciding
1063              * whether to switch to target check one last time if I/O has
1064              * come in the meanwhile, and if not flush the data to disk.
1065              */
1066             trace_mirror_before_drain(s, cnt);
1067 
1068             s->in_drain = true;
1069             bdrv_drained_begin(bs);
1070             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
1071             if (cnt > 0 || mirror_flush(s) < 0) {
1072                 bdrv_drained_end(bs);
1073                 s->in_drain = false;
1074                 continue;
1075             }
1076 
1077             /* The two disks are in sync.  Exit and report successful
1078              * completion.
1079              */
1080             assert(QLIST_EMPTY(&bs->tracked_requests));
1081             need_drain = false;
1082             break;
1083         }
1084 
1085         if (job_is_ready(&s->common.job) && !should_complete) {
1086             delay_ns = (s->in_flight == 0 &&
1087                         cnt == 0 ? BLOCK_JOB_SLICE_TIME : 0);
1088         }
1089         trace_mirror_before_sleep(s, cnt, job_is_ready(&s->common.job),
1090                                   delay_ns);
1091         job_sleep_ns(&s->common.job, delay_ns);
1092         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1093     }
1094 
1095 immediate_exit:
1096     if (s->in_flight > 0) {
1097         /* We get here only if something went wrong.  Either the job failed,
1098          * or it was cancelled prematurely so that we do not guarantee that
1099          * the target is a copy of the source.
1100          */
1101         assert(ret < 0 || job_is_cancelled(&s->common.job));
1102         assert(need_drain);
1103         mirror_wait_for_all_io(s);
1104     }
1105 
1106     assert(s->in_flight == 0);
1107     qemu_vfree(s->buf);
1108     g_free(s->cow_bitmap);
1109     g_free(s->in_flight_bitmap);
1110     bdrv_dirty_iter_free(s->dbi);
1111 
1112     if (need_drain) {
1113         s->in_drain = true;
1114         bdrv_drained_begin(bs);
1115     }
1116 
1117     return ret;
1118 }
1119 
1120 static void mirror_complete(Job *job, Error **errp)
1121 {
1122     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1123 
1124     if (!job_is_ready(job)) {
1125         error_setg(errp, "The active block job '%s' cannot be completed",
1126                    job->id);
1127         return;
1128     }
1129 
1130     /* block all operations on to_replace bs */
1131     if (s->replaces) {
1132         AioContext *replace_aio_context;
1133 
1134         s->to_replace = bdrv_find_node(s->replaces);
1135         if (!s->to_replace) {
1136             error_setg(errp, "Node name '%s' not found", s->replaces);
1137             return;
1138         }
1139 
1140         replace_aio_context = bdrv_get_aio_context(s->to_replace);
1141         aio_context_acquire(replace_aio_context);
1142 
1143         /* TODO Translate this into child freeze system. */
1144         error_setg(&s->replace_blocker,
1145                    "block device is in use by block-job-complete");
1146         bdrv_op_block_all(s->to_replace, s->replace_blocker);
1147         bdrv_ref(s->to_replace);
1148 
1149         aio_context_release(replace_aio_context);
1150     }
1151 
1152     s->should_complete = true;
1153 
1154     /* If the job is paused, it will be re-entered when it is resumed */
1155     if (!job->paused) {
1156         job_enter(job);
1157     }
1158 }
1159 
1160 static void coroutine_fn mirror_pause(Job *job)
1161 {
1162     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1163 
1164     mirror_wait_for_all_io(s);
1165 }
1166 
1167 static bool mirror_drained_poll(BlockJob *job)
1168 {
1169     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
1170 
1171     /* If the job isn't paused nor cancelled, we can't be sure that it won't
1172      * issue more requests. We make an exception if we've reached this point
1173      * from one of our own drain sections, to avoid a deadlock waiting for
1174      * ourselves.
1175      */
1176     if (!s->common.job.paused && !job_is_cancelled(&job->job) && !s->in_drain) {
1177         return true;
1178     }
1179 
1180     return !!s->in_flight;
1181 }
1182 
1183 static bool mirror_cancel(Job *job, bool force)
1184 {
1185     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common.job);
1186     BlockDriverState *target = blk_bs(s->target);
1187 
1188     /*
1189      * Before the job is READY, we treat any cancellation like a
1190      * force-cancellation.
1191      */
1192     force = force || !job_is_ready(job);
1193 
1194     if (force) {
1195         bdrv_cancel_in_flight(target);
1196     }
1197     return force;
1198 }
1199 
1200 static bool commit_active_cancel(Job *job, bool force)
1201 {
1202     /* Same as above in mirror_cancel() */
1203     return force || !job_is_ready(job);
1204 }
1205 
1206 static const BlockJobDriver mirror_job_driver = {
1207     .job_driver = {
1208         .instance_size          = sizeof(MirrorBlockJob),
1209         .job_type               = JOB_TYPE_MIRROR,
1210         .free                   = block_job_free,
1211         .user_resume            = block_job_user_resume,
1212         .run                    = mirror_run,
1213         .prepare                = mirror_prepare,
1214         .abort                  = mirror_abort,
1215         .pause                  = mirror_pause,
1216         .complete               = mirror_complete,
1217         .cancel                 = mirror_cancel,
1218     },
1219     .drained_poll           = mirror_drained_poll,
1220 };
1221 
1222 static const BlockJobDriver commit_active_job_driver = {
1223     .job_driver = {
1224         .instance_size          = sizeof(MirrorBlockJob),
1225         .job_type               = JOB_TYPE_COMMIT,
1226         .free                   = block_job_free,
1227         .user_resume            = block_job_user_resume,
1228         .run                    = mirror_run,
1229         .prepare                = mirror_prepare,
1230         .abort                  = mirror_abort,
1231         .pause                  = mirror_pause,
1232         .complete               = mirror_complete,
1233         .cancel                 = commit_active_cancel,
1234     },
1235     .drained_poll           = mirror_drained_poll,
1236 };
1237 
1238 static void coroutine_fn
1239 do_sync_target_write(MirrorBlockJob *job, MirrorMethod method,
1240                      uint64_t offset, uint64_t bytes,
1241                      QEMUIOVector *qiov, int flags)
1242 {
1243     int ret;
1244     size_t qiov_offset = 0;
1245     int64_t bitmap_offset, bitmap_end;
1246 
1247     if (!QEMU_IS_ALIGNED(offset, job->granularity) &&
1248         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset))
1249     {
1250             /*
1251              * Dirty unaligned padding: ignore it.
1252              *
1253              * Reasoning:
1254              * 1. If we copy it, we can't reset corresponding bit in
1255              *    dirty_bitmap as there may be some "dirty" bytes still not
1256              *    copied.
1257              * 2. It's already dirty, so skipping it we don't diverge mirror
1258              *    progress.
1259              *
1260              * Note, that because of this, guest write may have no contribution
1261              * into mirror converge, but that's not bad, as we have background
1262              * process of mirroring. If under some bad circumstances (high guest
1263              * IO load) background process starve, we will not converge anyway,
1264              * even if each write will contribute, as guest is not guaranteed to
1265              * rewrite the whole disk.
1266              */
1267             qiov_offset = QEMU_ALIGN_UP(offset, job->granularity) - offset;
1268             if (bytes <= qiov_offset) {
1269                 /* nothing to do after shrink */
1270                 return;
1271             }
1272             offset += qiov_offset;
1273             bytes -= qiov_offset;
1274     }
1275 
1276     if (!QEMU_IS_ALIGNED(offset + bytes, job->granularity) &&
1277         bdrv_dirty_bitmap_get(job->dirty_bitmap, offset + bytes - 1))
1278     {
1279         uint64_t tail = (offset + bytes) % job->granularity;
1280 
1281         if (bytes <= tail) {
1282             /* nothing to do after shrink */
1283             return;
1284         }
1285         bytes -= tail;
1286     }
1287 
1288     /*
1289      * Tails are either clean or shrunk, so for bitmap resetting
1290      * we safely align the range down.
1291      */
1292     bitmap_offset = QEMU_ALIGN_UP(offset, job->granularity);
1293     bitmap_end = QEMU_ALIGN_DOWN(offset + bytes, job->granularity);
1294     if (bitmap_offset < bitmap_end) {
1295         bdrv_reset_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1296                                 bitmap_end - bitmap_offset);
1297     }
1298 
1299     job_progress_increase_remaining(&job->common.job, bytes);
1300 
1301     switch (method) {
1302     case MIRROR_METHOD_COPY:
1303         ret = blk_co_pwritev_part(job->target, offset, bytes,
1304                                   qiov, qiov_offset, flags);
1305         break;
1306 
1307     case MIRROR_METHOD_ZERO:
1308         assert(!qiov);
1309         ret = blk_co_pwrite_zeroes(job->target, offset, bytes, flags);
1310         break;
1311 
1312     case MIRROR_METHOD_DISCARD:
1313         assert(!qiov);
1314         ret = blk_co_pdiscard(job->target, offset, bytes);
1315         break;
1316 
1317     default:
1318         abort();
1319     }
1320 
1321     if (ret >= 0) {
1322         job_progress_update(&job->common.job, bytes);
1323     } else {
1324         BlockErrorAction action;
1325 
1326         /*
1327          * We failed, so we should mark dirty the whole area, aligned up.
1328          * Note that we don't care about shrunk tails if any: they were dirty
1329          * at function start, and they must be still dirty, as we've locked
1330          * the region for in-flight op.
1331          */
1332         bitmap_offset = QEMU_ALIGN_DOWN(offset, job->granularity);
1333         bitmap_end = QEMU_ALIGN_UP(offset + bytes, job->granularity);
1334         bdrv_set_dirty_bitmap(job->dirty_bitmap, bitmap_offset,
1335                               bitmap_end - bitmap_offset);
1336         job->actively_synced = false;
1337 
1338         action = mirror_error_action(job, false, -ret);
1339         if (action == BLOCK_ERROR_ACTION_REPORT) {
1340             if (!job->ret) {
1341                 job->ret = ret;
1342             }
1343         }
1344     }
1345 }
1346 
1347 static MirrorOp *coroutine_fn active_write_prepare(MirrorBlockJob *s,
1348                                                    uint64_t offset,
1349                                                    uint64_t bytes)
1350 {
1351     MirrorOp *op;
1352     uint64_t start_chunk = offset / s->granularity;
1353     uint64_t end_chunk = DIV_ROUND_UP(offset + bytes, s->granularity);
1354 
1355     op = g_new(MirrorOp, 1);
1356     *op = (MirrorOp){
1357         .s                  = s,
1358         .offset             = offset,
1359         .bytes              = bytes,
1360         .is_active_write    = true,
1361         .is_in_flight       = true,
1362         .co                 = qemu_coroutine_self(),
1363     };
1364     qemu_co_queue_init(&op->waiting_requests);
1365     QTAILQ_INSERT_TAIL(&s->ops_in_flight, op, next);
1366 
1367     s->in_active_write_counter++;
1368 
1369     mirror_wait_on_conflicts(op, s, offset, bytes);
1370 
1371     bitmap_set(s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1372 
1373     return op;
1374 }
1375 
1376 static void coroutine_fn active_write_settle(MirrorOp *op)
1377 {
1378     uint64_t start_chunk = op->offset / op->s->granularity;
1379     uint64_t end_chunk = DIV_ROUND_UP(op->offset + op->bytes,
1380                                       op->s->granularity);
1381 
1382     if (!--op->s->in_active_write_counter && op->s->actively_synced) {
1383         BdrvChild *source = op->s->mirror_top_bs->backing;
1384 
1385         if (QLIST_FIRST(&source->bs->parents) == source &&
1386             QLIST_NEXT(source, next_parent) == NULL)
1387         {
1388             /* Assert that we are back in sync once all active write
1389              * operations are settled.
1390              * Note that we can only assert this if the mirror node
1391              * is the source node's only parent. */
1392             assert(!bdrv_get_dirty_count(op->s->dirty_bitmap));
1393         }
1394     }
1395     bitmap_clear(op->s->in_flight_bitmap, start_chunk, end_chunk - start_chunk);
1396     QTAILQ_REMOVE(&op->s->ops_in_flight, op, next);
1397     qemu_co_queue_restart_all(&op->waiting_requests);
1398     g_free(op);
1399 }
1400 
1401 static int coroutine_fn bdrv_mirror_top_preadv(BlockDriverState *bs,
1402     int64_t offset, int64_t bytes, QEMUIOVector *qiov, BdrvRequestFlags flags)
1403 {
1404     return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
1405 }
1406 
1407 static int coroutine_fn bdrv_mirror_top_do_write(BlockDriverState *bs,
1408     MirrorMethod method, uint64_t offset, uint64_t bytes, QEMUIOVector *qiov,
1409     int flags)
1410 {
1411     MirrorOp *op = NULL;
1412     MirrorBDSOpaque *s = bs->opaque;
1413     int ret = 0;
1414     bool copy_to_target;
1415 
1416     copy_to_target = s->job->ret >= 0 &&
1417                      !job_is_cancelled(&s->job->common.job) &&
1418                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1419 
1420     if (copy_to_target) {
1421         op = active_write_prepare(s->job, offset, bytes);
1422     }
1423 
1424     switch (method) {
1425     case MIRROR_METHOD_COPY:
1426         ret = bdrv_co_pwritev(bs->backing, offset, bytes, qiov, flags);
1427         break;
1428 
1429     case MIRROR_METHOD_ZERO:
1430         ret = bdrv_co_pwrite_zeroes(bs->backing, offset, bytes, flags);
1431         break;
1432 
1433     case MIRROR_METHOD_DISCARD:
1434         ret = bdrv_co_pdiscard(bs->backing, offset, bytes);
1435         break;
1436 
1437     default:
1438         abort();
1439     }
1440 
1441     if (ret < 0) {
1442         goto out;
1443     }
1444 
1445     if (copy_to_target) {
1446         do_sync_target_write(s->job, method, offset, bytes, qiov, flags);
1447     }
1448 
1449 out:
1450     if (copy_to_target) {
1451         active_write_settle(op);
1452     }
1453     return ret;
1454 }
1455 
1456 static int coroutine_fn bdrv_mirror_top_pwritev(BlockDriverState *bs,
1457     int64_t offset, int64_t bytes, QEMUIOVector *qiov, BdrvRequestFlags flags)
1458 {
1459     MirrorBDSOpaque *s = bs->opaque;
1460     QEMUIOVector bounce_qiov;
1461     void *bounce_buf;
1462     int ret = 0;
1463     bool copy_to_target;
1464 
1465     copy_to_target = s->job->ret >= 0 &&
1466                      !job_is_cancelled(&s->job->common.job) &&
1467                      s->job->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING;
1468 
1469     if (copy_to_target) {
1470         /* The guest might concurrently modify the data to write; but
1471          * the data on source and destination must match, so we have
1472          * to use a bounce buffer if we are going to write to the
1473          * target now. */
1474         bounce_buf = qemu_blockalign(bs, bytes);
1475         iov_to_buf_full(qiov->iov, qiov->niov, 0, bounce_buf, bytes);
1476 
1477         qemu_iovec_init(&bounce_qiov, 1);
1478         qemu_iovec_add(&bounce_qiov, bounce_buf, bytes);
1479         qiov = &bounce_qiov;
1480     }
1481 
1482     ret = bdrv_mirror_top_do_write(bs, MIRROR_METHOD_COPY, offset, bytes, qiov,
1483                                    flags);
1484 
1485     if (copy_to_target) {
1486         qemu_iovec_destroy(&bounce_qiov);
1487         qemu_vfree(bounce_buf);
1488     }
1489 
1490     return ret;
1491 }
1492 
1493 static int coroutine_fn bdrv_mirror_top_flush(BlockDriverState *bs)
1494 {
1495     if (bs->backing == NULL) {
1496         /* we can be here after failed bdrv_append in mirror_start_job */
1497         return 0;
1498     }
1499     return bdrv_co_flush(bs->backing->bs);
1500 }
1501 
1502 static int coroutine_fn bdrv_mirror_top_pwrite_zeroes(BlockDriverState *bs,
1503     int64_t offset, int64_t bytes, BdrvRequestFlags flags)
1504 {
1505     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_ZERO, offset, bytes, NULL,
1506                                     flags);
1507 }
1508 
1509 static int coroutine_fn bdrv_mirror_top_pdiscard(BlockDriverState *bs,
1510     int64_t offset, int64_t bytes)
1511 {
1512     return bdrv_mirror_top_do_write(bs, MIRROR_METHOD_DISCARD, offset, bytes,
1513                                     NULL, 0);
1514 }
1515 
1516 static void bdrv_mirror_top_refresh_filename(BlockDriverState *bs)
1517 {
1518     if (bs->backing == NULL) {
1519         /* we can be here after failed bdrv_attach_child in
1520          * bdrv_set_backing_hd */
1521         return;
1522     }
1523     pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
1524             bs->backing->bs->filename);
1525 }
1526 
1527 static void bdrv_mirror_top_child_perm(BlockDriverState *bs, BdrvChild *c,
1528                                        BdrvChildRole role,
1529                                        BlockReopenQueue *reopen_queue,
1530                                        uint64_t perm, uint64_t shared,
1531                                        uint64_t *nperm, uint64_t *nshared)
1532 {
1533     MirrorBDSOpaque *s = bs->opaque;
1534 
1535     if (s->stop) {
1536         /*
1537          * If the job is to be stopped, we do not need to forward
1538          * anything to the real image.
1539          */
1540         *nperm = 0;
1541         *nshared = BLK_PERM_ALL;
1542         return;
1543     }
1544 
1545     bdrv_default_perms(bs, c, role, reopen_queue,
1546                        perm, shared, nperm, nshared);
1547 
1548     if (s->is_commit) {
1549         /*
1550          * For commit jobs, we cannot take CONSISTENT_READ, because
1551          * that permission is unshared for everything above the base
1552          * node (except for filters on the base node).
1553          * We also have to force-share the WRITE permission, or
1554          * otherwise we would block ourselves at the base node (if
1555          * writes are blocked for a node, they are also blocked for
1556          * its backing file).
1557          * (We could also share RESIZE, because it may be needed for
1558          * the target if its size is less than the top node's; but
1559          * bdrv_default_perms_for_cow() automatically shares RESIZE
1560          * for backing nodes if WRITE is shared, so there is no need
1561          * to do it here.)
1562          */
1563         *nperm &= ~BLK_PERM_CONSISTENT_READ;
1564         *nshared |= BLK_PERM_WRITE;
1565     }
1566 }
1567 
1568 /* Dummy node that provides consistent read to its users without requiring it
1569  * from its backing file and that allows writes on the backing file chain. */
1570 static BlockDriver bdrv_mirror_top = {
1571     .format_name                = "mirror_top",
1572     .bdrv_co_preadv             = bdrv_mirror_top_preadv,
1573     .bdrv_co_pwritev            = bdrv_mirror_top_pwritev,
1574     .bdrv_co_pwrite_zeroes      = bdrv_mirror_top_pwrite_zeroes,
1575     .bdrv_co_pdiscard           = bdrv_mirror_top_pdiscard,
1576     .bdrv_co_flush              = bdrv_mirror_top_flush,
1577     .bdrv_refresh_filename      = bdrv_mirror_top_refresh_filename,
1578     .bdrv_child_perm            = bdrv_mirror_top_child_perm,
1579 
1580     .is_filter                  = true,
1581 };
1582 
1583 static BlockJob *mirror_start_job(
1584                              const char *job_id, BlockDriverState *bs,
1585                              int creation_flags, BlockDriverState *target,
1586                              const char *replaces, int64_t speed,
1587                              uint32_t granularity, int64_t buf_size,
1588                              BlockMirrorBackingMode backing_mode,
1589                              bool zero_target,
1590                              BlockdevOnError on_source_error,
1591                              BlockdevOnError on_target_error,
1592                              bool unmap,
1593                              BlockCompletionFunc *cb,
1594                              void *opaque,
1595                              const BlockJobDriver *driver,
1596                              bool is_none_mode, BlockDriverState *base,
1597                              bool auto_complete, const char *filter_node_name,
1598                              bool is_mirror, MirrorCopyMode copy_mode,
1599                              Error **errp)
1600 {
1601     MirrorBlockJob *s;
1602     MirrorBDSOpaque *bs_opaque;
1603     BlockDriverState *mirror_top_bs;
1604     bool target_is_backing;
1605     uint64_t target_perms, target_shared_perms;
1606     int ret;
1607 
1608     if (granularity == 0) {
1609         granularity = bdrv_get_default_bitmap_granularity(target);
1610     }
1611 
1612     assert(is_power_of_2(granularity));
1613 
1614     if (buf_size < 0) {
1615         error_setg(errp, "Invalid parameter 'buf-size'");
1616         return NULL;
1617     }
1618 
1619     if (buf_size == 0) {
1620         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1621     }
1622 
1623     if (bdrv_skip_filters(bs) == bdrv_skip_filters(target)) {
1624         error_setg(errp, "Can't mirror node into itself");
1625         return NULL;
1626     }
1627 
1628     target_is_backing = bdrv_chain_contains(bs, target);
1629 
1630     /* In the case of active commit, add dummy driver to provide consistent
1631      * reads on the top, while disabling it in the intermediate nodes, and make
1632      * the backing chain writable. */
1633     mirror_top_bs = bdrv_new_open_driver(&bdrv_mirror_top, filter_node_name,
1634                                          BDRV_O_RDWR, errp);
1635     if (mirror_top_bs == NULL) {
1636         return NULL;
1637     }
1638     if (!filter_node_name) {
1639         mirror_top_bs->implicit = true;
1640     }
1641 
1642     /* So that we can always drop this node */
1643     mirror_top_bs->never_freeze = true;
1644 
1645     mirror_top_bs->total_sectors = bs->total_sectors;
1646     mirror_top_bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED;
1647     mirror_top_bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
1648                                           BDRV_REQ_NO_FALLBACK;
1649     bs_opaque = g_new0(MirrorBDSOpaque, 1);
1650     mirror_top_bs->opaque = bs_opaque;
1651 
1652     bs_opaque->is_commit = target_is_backing;
1653 
1654     bdrv_drained_begin(bs);
1655     ret = bdrv_append(mirror_top_bs, bs, errp);
1656     bdrv_drained_end(bs);
1657 
1658     if (ret < 0) {
1659         bdrv_unref(mirror_top_bs);
1660         return NULL;
1661     }
1662 
1663     /* Make sure that the source is not resized while the job is running */
1664     s = block_job_create(job_id, driver, NULL, mirror_top_bs,
1665                          BLK_PERM_CONSISTENT_READ,
1666                          BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED |
1667                          BLK_PERM_WRITE, speed,
1668                          creation_flags, cb, opaque, errp);
1669     if (!s) {
1670         goto fail;
1671     }
1672     bs_opaque->job = s;
1673 
1674     /* The block job now has a reference to this node */
1675     bdrv_unref(mirror_top_bs);
1676 
1677     s->mirror_top_bs = mirror_top_bs;
1678 
1679     /* No resize for the target either; while the mirror is still running, a
1680      * consistent read isn't necessarily possible. We could possibly allow
1681      * writes and graph modifications, though it would likely defeat the
1682      * purpose of a mirror, so leave them blocked for now.
1683      *
1684      * In the case of active commit, things look a bit different, though,
1685      * because the target is an already populated backing file in active use.
1686      * We can allow anything except resize there.*/
1687 
1688     target_perms = BLK_PERM_WRITE;
1689     target_shared_perms = BLK_PERM_WRITE_UNCHANGED;
1690 
1691     if (target_is_backing) {
1692         int64_t bs_size, target_size;
1693         bs_size = bdrv_getlength(bs);
1694         if (bs_size < 0) {
1695             error_setg_errno(errp, -bs_size,
1696                              "Could not inquire top image size");
1697             goto fail;
1698         }
1699 
1700         target_size = bdrv_getlength(target);
1701         if (target_size < 0) {
1702             error_setg_errno(errp, -target_size,
1703                              "Could not inquire base image size");
1704             goto fail;
1705         }
1706 
1707         if (target_size < bs_size) {
1708             target_perms |= BLK_PERM_RESIZE;
1709         }
1710 
1711         target_shared_perms |= BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE;
1712     } else if (bdrv_chain_contains(bs, bdrv_skip_filters(target))) {
1713         /*
1714          * We may want to allow this in the future, but it would
1715          * require taking some extra care.
1716          */
1717         error_setg(errp, "Cannot mirror to a filter on top of a node in the "
1718                    "source's backing chain");
1719         goto fail;
1720     }
1721 
1722     s->target = blk_new(s->common.job.aio_context,
1723                         target_perms, target_shared_perms);
1724     ret = blk_insert_bs(s->target, target, errp);
1725     if (ret < 0) {
1726         goto fail;
1727     }
1728     if (is_mirror) {
1729         /* XXX: Mirror target could be a NBD server of target QEMU in the case
1730          * of non-shared block migration. To allow migration completion, we
1731          * have to allow "inactivate" of the target BB.  When that happens, we
1732          * know the job is drained, and the vcpus are stopped, so no write
1733          * operation will be performed. Block layer already has assertions to
1734          * ensure that. */
1735         blk_set_force_allow_inactivate(s->target);
1736     }
1737     blk_set_allow_aio_context_change(s->target, true);
1738     blk_set_disable_request_queuing(s->target, true);
1739 
1740     s->replaces = g_strdup(replaces);
1741     s->on_source_error = on_source_error;
1742     s->on_target_error = on_target_error;
1743     s->is_none_mode = is_none_mode;
1744     s->backing_mode = backing_mode;
1745     s->zero_target = zero_target;
1746     s->copy_mode = copy_mode;
1747     s->base = base;
1748     s->base_overlay = bdrv_find_overlay(bs, base);
1749     s->granularity = granularity;
1750     s->buf_size = ROUND_UP(buf_size, granularity);
1751     s->unmap = unmap;
1752     if (auto_complete) {
1753         s->should_complete = true;
1754     }
1755 
1756     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1757     if (!s->dirty_bitmap) {
1758         goto fail;
1759     }
1760     if (s->copy_mode == MIRROR_COPY_MODE_WRITE_BLOCKING) {
1761         bdrv_disable_dirty_bitmap(s->dirty_bitmap);
1762     }
1763 
1764     ret = block_job_add_bdrv(&s->common, "source", bs, 0,
1765                              BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE |
1766                              BLK_PERM_CONSISTENT_READ,
1767                              errp);
1768     if (ret < 0) {
1769         goto fail;
1770     }
1771 
1772     /* Required permissions are already taken with blk_new() */
1773     block_job_add_bdrv(&s->common, "target", target, 0, BLK_PERM_ALL,
1774                        &error_abort);
1775 
1776     /* In commit_active_start() all intermediate nodes disappear, so
1777      * any jobs in them must be blocked */
1778     if (target_is_backing) {
1779         BlockDriverState *iter, *filtered_target;
1780         uint64_t iter_shared_perms;
1781 
1782         /*
1783          * The topmost node with
1784          * bdrv_skip_filters(filtered_target) == bdrv_skip_filters(target)
1785          */
1786         filtered_target = bdrv_cow_bs(bdrv_find_overlay(bs, target));
1787 
1788         assert(bdrv_skip_filters(filtered_target) ==
1789                bdrv_skip_filters(target));
1790 
1791         /*
1792          * XXX BLK_PERM_WRITE needs to be allowed so we don't block
1793          * ourselves at s->base (if writes are blocked for a node, they are
1794          * also blocked for its backing file). The other options would be a
1795          * second filter driver above s->base (== target).
1796          */
1797         iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE;
1798 
1799         for (iter = bdrv_filter_or_cow_bs(bs); iter != target;
1800              iter = bdrv_filter_or_cow_bs(iter))
1801         {
1802             if (iter == filtered_target) {
1803                 /*
1804                  * From here on, all nodes are filters on the base.
1805                  * This allows us to share BLK_PERM_CONSISTENT_READ.
1806                  */
1807                 iter_shared_perms |= BLK_PERM_CONSISTENT_READ;
1808             }
1809 
1810             ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
1811                                      iter_shared_perms, errp);
1812             if (ret < 0) {
1813                 goto fail;
1814             }
1815         }
1816 
1817         if (bdrv_freeze_backing_chain(mirror_top_bs, target, errp) < 0) {
1818             goto fail;
1819         }
1820     }
1821 
1822     QTAILQ_INIT(&s->ops_in_flight);
1823 
1824     trace_mirror_start(bs, s, opaque);
1825     job_start(&s->common.job);
1826 
1827     return &s->common;
1828 
1829 fail:
1830     if (s) {
1831         /* Make sure this BDS does not go away until we have completed the graph
1832          * changes below */
1833         bdrv_ref(mirror_top_bs);
1834 
1835         g_free(s->replaces);
1836         blk_unref(s->target);
1837         bs_opaque->job = NULL;
1838         if (s->dirty_bitmap) {
1839             bdrv_release_dirty_bitmap(s->dirty_bitmap);
1840         }
1841         job_early_fail(&s->common.job);
1842     }
1843 
1844     bs_opaque->stop = true;
1845     bdrv_child_refresh_perms(mirror_top_bs, mirror_top_bs->backing,
1846                              &error_abort);
1847     bdrv_replace_node(mirror_top_bs, mirror_top_bs->backing->bs, &error_abort);
1848 
1849     bdrv_unref(mirror_top_bs);
1850 
1851     return NULL;
1852 }
1853 
1854 void mirror_start(const char *job_id, BlockDriverState *bs,
1855                   BlockDriverState *target, const char *replaces,
1856                   int creation_flags, int64_t speed,
1857                   uint32_t granularity, int64_t buf_size,
1858                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1859                   bool zero_target,
1860                   BlockdevOnError on_source_error,
1861                   BlockdevOnError on_target_error,
1862                   bool unmap, const char *filter_node_name,
1863                   MirrorCopyMode copy_mode, Error **errp)
1864 {
1865     bool is_none_mode;
1866     BlockDriverState *base;
1867 
1868     GLOBAL_STATE_CODE();
1869 
1870     if ((mode == MIRROR_SYNC_MODE_INCREMENTAL) ||
1871         (mode == MIRROR_SYNC_MODE_BITMAP)) {
1872         error_setg(errp, "Sync mode '%s' not supported",
1873                    MirrorSyncMode_str(mode));
1874         return;
1875     }
1876     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1877     base = mode == MIRROR_SYNC_MODE_TOP ? bdrv_backing_chain_next(bs) : NULL;
1878     mirror_start_job(job_id, bs, creation_flags, target, replaces,
1879                      speed, granularity, buf_size, backing_mode, zero_target,
1880                      on_source_error, on_target_error, unmap, NULL, NULL,
1881                      &mirror_job_driver, is_none_mode, base, false,
1882                      filter_node_name, true, copy_mode, errp);
1883 }
1884 
1885 BlockJob *commit_active_start(const char *job_id, BlockDriverState *bs,
1886                               BlockDriverState *base, int creation_flags,
1887                               int64_t speed, BlockdevOnError on_error,
1888                               const char *filter_node_name,
1889                               BlockCompletionFunc *cb, void *opaque,
1890                               bool auto_complete, Error **errp)
1891 {
1892     bool base_read_only;
1893     BlockJob *job;
1894 
1895     GLOBAL_STATE_CODE();
1896 
1897     base_read_only = bdrv_is_read_only(base);
1898 
1899     if (base_read_only) {
1900         if (bdrv_reopen_set_read_only(base, false, errp) < 0) {
1901             return NULL;
1902         }
1903     }
1904 
1905     job = mirror_start_job(
1906                      job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1907                      MIRROR_LEAVE_BACKING_CHAIN, false,
1908                      on_error, on_error, true, cb, opaque,
1909                      &commit_active_job_driver, false, base, auto_complete,
1910                      filter_node_name, false, MIRROR_COPY_MODE_BACKGROUND,
1911                      errp);
1912     if (!job) {
1913         goto error_restore_flags;
1914     }
1915 
1916     return job;
1917 
1918 error_restore_flags:
1919     /* ignore error and errp for bdrv_reopen, because we want to propagate
1920      * the original error */
1921     if (base_read_only) {
1922         bdrv_reopen_set_read_only(base, true, NULL);
1923     }
1924     return NULL;
1925 }
1926