xref: /openbmc/qemu/block/mirror.c (revision 4b4629d9)
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 "trace.h"
16 #include "block/blockjob.h"
17 #include "block/block_int.h"
18 #include "sysemu/block-backend.h"
19 #include "qapi/error.h"
20 #include "qapi/qmp/qerror.h"
21 #include "qemu/ratelimit.h"
22 #include "qemu/bitmap.h"
23 #include "qemu/error-report.h"
24 
25 #define SLICE_TIME    100000000ULL /* ns */
26 #define MAX_IN_FLIGHT 16
27 #define DEFAULT_MIRROR_BUF_SIZE   (10 << 20)
28 
29 /* The mirroring buffer is a list of granularity-sized chunks.
30  * Free chunks are organized in a list.
31  */
32 typedef struct MirrorBuffer {
33     QSIMPLEQ_ENTRY(MirrorBuffer) next;
34 } MirrorBuffer;
35 
36 typedef struct MirrorBlockJob {
37     BlockJob common;
38     RateLimit limit;
39     BlockDriverState *target;
40     BlockDriverState *base;
41     /* The name of the graph node to replace */
42     char *replaces;
43     /* The BDS to replace */
44     BlockDriverState *to_replace;
45     /* Used to block operations on the drive-mirror-replace target */
46     Error *replace_blocker;
47     bool is_none_mode;
48     BlockdevOnError on_source_error, on_target_error;
49     bool synced;
50     bool should_complete;
51     int64_t granularity;
52     size_t buf_size;
53     int64_t bdev_length;
54     unsigned long *cow_bitmap;
55     BdrvDirtyBitmap *dirty_bitmap;
56     HBitmapIter hbi;
57     uint8_t *buf;
58     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
59     int buf_free_count;
60 
61     unsigned long *in_flight_bitmap;
62     int in_flight;
63     int sectors_in_flight;
64     int ret;
65     bool unmap;
66     bool waiting_for_io;
67     int target_cluster_sectors;
68     int max_iov;
69 } MirrorBlockJob;
70 
71 typedef struct MirrorOp {
72     MirrorBlockJob *s;
73     QEMUIOVector qiov;
74     int64_t sector_num;
75     int nb_sectors;
76 } MirrorOp;
77 
78 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
79                                             int error)
80 {
81     s->synced = false;
82     if (read) {
83         return block_job_error_action(&s->common, s->common.bs,
84                                       s->on_source_error, true, error);
85     } else {
86         return block_job_error_action(&s->common, s->target,
87                                       s->on_target_error, false, error);
88     }
89 }
90 
91 static void mirror_iteration_done(MirrorOp *op, int ret)
92 {
93     MirrorBlockJob *s = op->s;
94     struct iovec *iov;
95     int64_t chunk_num;
96     int i, nb_chunks, sectors_per_chunk;
97 
98     trace_mirror_iteration_done(s, op->sector_num, op->nb_sectors, ret);
99 
100     s->in_flight--;
101     s->sectors_in_flight -= op->nb_sectors;
102     iov = op->qiov.iov;
103     for (i = 0; i < op->qiov.niov; i++) {
104         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
105         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
106         s->buf_free_count++;
107     }
108 
109     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
110     chunk_num = op->sector_num / sectors_per_chunk;
111     nb_chunks = DIV_ROUND_UP(op->nb_sectors, sectors_per_chunk);
112     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
113     if (ret >= 0) {
114         if (s->cow_bitmap) {
115             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
116         }
117         s->common.offset += (uint64_t)op->nb_sectors * BDRV_SECTOR_SIZE;
118     }
119 
120     qemu_iovec_destroy(&op->qiov);
121     g_free(op);
122 
123     if (s->waiting_for_io) {
124         qemu_coroutine_enter(s->common.co, NULL);
125     }
126 }
127 
128 static void mirror_write_complete(void *opaque, int ret)
129 {
130     MirrorOp *op = opaque;
131     MirrorBlockJob *s = op->s;
132     if (ret < 0) {
133         BlockErrorAction action;
134 
135         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->sector_num, op->nb_sectors);
136         action = mirror_error_action(s, false, -ret);
137         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
138             s->ret = ret;
139         }
140     }
141     mirror_iteration_done(op, ret);
142 }
143 
144 static void mirror_read_complete(void *opaque, int ret)
145 {
146     MirrorOp *op = opaque;
147     MirrorBlockJob *s = op->s;
148     if (ret < 0) {
149         BlockErrorAction action;
150 
151         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->sector_num, op->nb_sectors);
152         action = mirror_error_action(s, true, -ret);
153         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
154             s->ret = ret;
155         }
156 
157         mirror_iteration_done(op, ret);
158         return;
159     }
160     bdrv_aio_writev(s->target, op->sector_num, &op->qiov, op->nb_sectors,
161                     mirror_write_complete, op);
162 }
163 
164 static inline void mirror_clip_sectors(MirrorBlockJob *s,
165                                        int64_t sector_num,
166                                        int *nb_sectors)
167 {
168     *nb_sectors = MIN(*nb_sectors,
169                       s->bdev_length / BDRV_SECTOR_SIZE - sector_num);
170 }
171 
172 /* Round sector_num and/or nb_sectors to target cluster if COW is needed, and
173  * return the offset of the adjusted tail sector against original. */
174 static int mirror_cow_align(MirrorBlockJob *s,
175                             int64_t *sector_num,
176                             int *nb_sectors)
177 {
178     bool need_cow;
179     int ret = 0;
180     int chunk_sectors = s->granularity >> BDRV_SECTOR_BITS;
181     int64_t align_sector_num = *sector_num;
182     int align_nb_sectors = *nb_sectors;
183     int max_sectors = chunk_sectors * s->max_iov;
184 
185     need_cow = !test_bit(*sector_num / chunk_sectors, s->cow_bitmap);
186     need_cow |= !test_bit((*sector_num + *nb_sectors - 1) / chunk_sectors,
187                           s->cow_bitmap);
188     if (need_cow) {
189         bdrv_round_to_clusters(s->target, *sector_num, *nb_sectors,
190                                &align_sector_num, &align_nb_sectors);
191     }
192 
193     if (align_nb_sectors > max_sectors) {
194         align_nb_sectors = max_sectors;
195         if (need_cow) {
196             align_nb_sectors = QEMU_ALIGN_DOWN(align_nb_sectors,
197                                                s->target_cluster_sectors);
198         }
199     }
200     /* Clipping may result in align_nb_sectors unaligned to chunk boundary, but
201      * that doesn't matter because it's already the end of source image. */
202     mirror_clip_sectors(s, align_sector_num, &align_nb_sectors);
203 
204     ret = align_sector_num + align_nb_sectors - (*sector_num + *nb_sectors);
205     *sector_num = align_sector_num;
206     *nb_sectors = align_nb_sectors;
207     assert(ret >= 0);
208     return ret;
209 }
210 
211 static inline void mirror_wait_for_io(MirrorBlockJob *s)
212 {
213     assert(!s->waiting_for_io);
214     s->waiting_for_io = true;
215     qemu_coroutine_yield();
216     s->waiting_for_io = false;
217 }
218 
219 /* Submit async read while handling COW.
220  * Returns: nb_sectors if no alignment is necessary, or
221  *          (new_end - sector_num) if tail is rounded up or down due to
222  *          alignment or buffer limit.
223  */
224 static int mirror_do_read(MirrorBlockJob *s, int64_t sector_num,
225                           int nb_sectors)
226 {
227     BlockDriverState *source = s->common.bs;
228     int sectors_per_chunk, nb_chunks;
229     int ret = nb_sectors;
230     MirrorOp *op;
231 
232     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
233 
234     /* We can only handle as much as buf_size at a time. */
235     nb_sectors = MIN(s->buf_size >> BDRV_SECTOR_BITS, nb_sectors);
236     assert(nb_sectors);
237 
238     if (s->cow_bitmap) {
239         ret += mirror_cow_align(s, &sector_num, &nb_sectors);
240     }
241     assert(nb_sectors << BDRV_SECTOR_BITS <= s->buf_size);
242     /* The sector range must meet granularity because:
243      * 1) Caller passes in aligned values;
244      * 2) mirror_cow_align is used only when target cluster is larger. */
245     assert(!(sector_num % sectors_per_chunk));
246     nb_chunks = DIV_ROUND_UP(nb_sectors, sectors_per_chunk);
247 
248     while (s->buf_free_count < nb_chunks) {
249         trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
250         mirror_wait_for_io(s);
251     }
252 
253     /* Allocate a MirrorOp that is used as an AIO callback.  */
254     op = g_new(MirrorOp, 1);
255     op->s = s;
256     op->sector_num = sector_num;
257     op->nb_sectors = nb_sectors;
258 
259     /* Now make a QEMUIOVector taking enough granularity-sized chunks
260      * from s->buf_free.
261      */
262     qemu_iovec_init(&op->qiov, nb_chunks);
263     while (nb_chunks-- > 0) {
264         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
265         size_t remaining = nb_sectors * BDRV_SECTOR_SIZE - op->qiov.size;
266 
267         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
268         s->buf_free_count--;
269         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
270     }
271 
272     /* Copy the dirty cluster.  */
273     s->in_flight++;
274     s->sectors_in_flight += nb_sectors;
275     trace_mirror_one_iteration(s, sector_num, nb_sectors);
276 
277     bdrv_aio_readv(source, sector_num, &op->qiov, nb_sectors,
278                    mirror_read_complete, op);
279     return ret;
280 }
281 
282 static void mirror_do_zero_or_discard(MirrorBlockJob *s,
283                                       int64_t sector_num,
284                                       int nb_sectors,
285                                       bool is_discard)
286 {
287     MirrorOp *op;
288 
289     /* Allocate a MirrorOp that is used as an AIO callback. The qiov is zeroed
290      * so the freeing in mirror_iteration_done is nop. */
291     op = g_new0(MirrorOp, 1);
292     op->s = s;
293     op->sector_num = sector_num;
294     op->nb_sectors = nb_sectors;
295 
296     s->in_flight++;
297     s->sectors_in_flight += nb_sectors;
298     if (is_discard) {
299         bdrv_aio_discard(s->target, sector_num, op->nb_sectors,
300                          mirror_write_complete, op);
301     } else {
302         bdrv_aio_write_zeroes(s->target, sector_num, op->nb_sectors,
303                               s->unmap ? BDRV_REQ_MAY_UNMAP : 0,
304                               mirror_write_complete, op);
305     }
306 }
307 
308 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
309 {
310     BlockDriverState *source = s->common.bs;
311     int64_t sector_num, first_chunk;
312     uint64_t delay_ns = 0;
313     /* At least the first dirty chunk is mirrored in one iteration. */
314     int nb_chunks = 1;
315     int64_t end = s->bdev_length / BDRV_SECTOR_SIZE;
316     int sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
317 
318     sector_num = hbitmap_iter_next(&s->hbi);
319     if (sector_num < 0) {
320         bdrv_dirty_iter_init(s->dirty_bitmap, &s->hbi);
321         sector_num = hbitmap_iter_next(&s->hbi);
322         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
323         assert(sector_num >= 0);
324     }
325 
326     first_chunk = sector_num / sectors_per_chunk;
327     while (test_bit(first_chunk, s->in_flight_bitmap)) {
328         trace_mirror_yield_in_flight(s, first_chunk, s->in_flight);
329         mirror_wait_for_io(s);
330     }
331 
332     /* Find the number of consective dirty chunks following the first dirty
333      * one, and wait for in flight requests in them. */
334     while (nb_chunks * sectors_per_chunk < (s->buf_size >> BDRV_SECTOR_BITS)) {
335         int64_t hbitmap_next;
336         int64_t next_sector = sector_num + nb_chunks * sectors_per_chunk;
337         int64_t next_chunk = next_sector / sectors_per_chunk;
338         if (next_sector >= end ||
339             !bdrv_get_dirty(source, s->dirty_bitmap, next_sector)) {
340             break;
341         }
342         if (test_bit(next_chunk, s->in_flight_bitmap)) {
343             break;
344         }
345 
346         hbitmap_next = hbitmap_iter_next(&s->hbi);
347         if (hbitmap_next > next_sector || hbitmap_next < 0) {
348             /* The bitmap iterator's cache is stale, refresh it */
349             bdrv_set_dirty_iter(&s->hbi, next_sector);
350             hbitmap_next = hbitmap_iter_next(&s->hbi);
351         }
352         assert(hbitmap_next == next_sector);
353         nb_chunks++;
354     }
355 
356     /* Clear dirty bits before querying the block status, because
357      * calling bdrv_get_block_status_above could yield - if some blocks are
358      * marked dirty in this window, we need to know.
359      */
360     bdrv_reset_dirty_bitmap(s->dirty_bitmap, sector_num,
361                             nb_chunks * sectors_per_chunk);
362     bitmap_set(s->in_flight_bitmap, sector_num / sectors_per_chunk, nb_chunks);
363     while (nb_chunks > 0 && sector_num < end) {
364         int ret;
365         int io_sectors;
366         BlockDriverState *file;
367         enum MirrorMethod {
368             MIRROR_METHOD_COPY,
369             MIRROR_METHOD_ZERO,
370             MIRROR_METHOD_DISCARD
371         } mirror_method = MIRROR_METHOD_COPY;
372 
373         assert(!(sector_num % sectors_per_chunk));
374         ret = bdrv_get_block_status_above(source, NULL, sector_num,
375                                           nb_chunks * sectors_per_chunk,
376                                           &io_sectors, &file);
377         if (ret < 0) {
378             io_sectors = nb_chunks * sectors_per_chunk;
379         }
380 
381         io_sectors -= io_sectors % sectors_per_chunk;
382         if (io_sectors < sectors_per_chunk) {
383             io_sectors = sectors_per_chunk;
384         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
385             int64_t target_sector_num;
386             int target_nb_sectors;
387             bdrv_round_to_clusters(s->target, sector_num, io_sectors,
388                                    &target_sector_num, &target_nb_sectors);
389             if (target_sector_num == sector_num &&
390                 target_nb_sectors == io_sectors) {
391                 mirror_method = ret & BDRV_BLOCK_ZERO ?
392                                     MIRROR_METHOD_ZERO :
393                                     MIRROR_METHOD_DISCARD;
394             }
395         }
396 
397         mirror_clip_sectors(s, sector_num, &io_sectors);
398         switch (mirror_method) {
399         case MIRROR_METHOD_COPY:
400             io_sectors = mirror_do_read(s, sector_num, io_sectors);
401             break;
402         case MIRROR_METHOD_ZERO:
403             mirror_do_zero_or_discard(s, sector_num, io_sectors, false);
404             break;
405         case MIRROR_METHOD_DISCARD:
406             mirror_do_zero_or_discard(s, sector_num, io_sectors, true);
407             break;
408         default:
409             abort();
410         }
411         assert(io_sectors);
412         sector_num += io_sectors;
413         nb_chunks -= DIV_ROUND_UP(io_sectors, sectors_per_chunk);
414         delay_ns += ratelimit_calculate_delay(&s->limit, io_sectors);
415     }
416     return delay_ns;
417 }
418 
419 static void mirror_free_init(MirrorBlockJob *s)
420 {
421     int granularity = s->granularity;
422     size_t buf_size = s->buf_size;
423     uint8_t *buf = s->buf;
424 
425     assert(s->buf_free_count == 0);
426     QSIMPLEQ_INIT(&s->buf_free);
427     while (buf_size != 0) {
428         MirrorBuffer *cur = (MirrorBuffer *)buf;
429         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
430         s->buf_free_count++;
431         buf_size -= granularity;
432         buf += granularity;
433     }
434 }
435 
436 static void mirror_drain(MirrorBlockJob *s)
437 {
438     while (s->in_flight > 0) {
439         mirror_wait_for_io(s);
440     }
441 }
442 
443 typedef struct {
444     int ret;
445 } MirrorExitData;
446 
447 static void mirror_exit(BlockJob *job, void *opaque)
448 {
449     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
450     MirrorExitData *data = opaque;
451     AioContext *replace_aio_context = NULL;
452     BlockDriverState *src = s->common.bs;
453 
454     /* Make sure that the source BDS doesn't go away before we called
455      * block_job_completed(). */
456     bdrv_ref(src);
457 
458     if (s->to_replace) {
459         replace_aio_context = bdrv_get_aio_context(s->to_replace);
460         aio_context_acquire(replace_aio_context);
461     }
462 
463     if (s->should_complete && data->ret == 0) {
464         BlockDriverState *to_replace = s->common.bs;
465         if (s->to_replace) {
466             to_replace = s->to_replace;
467         }
468 
469         /* This was checked in mirror_start_job(), but meanwhile one of the
470          * nodes could have been newly attached to a BlockBackend. */
471         if (to_replace->blk && s->target->blk) {
472             error_report("block job: Can't create node with two BlockBackends");
473             data->ret = -EINVAL;
474             goto out;
475         }
476 
477         if (bdrv_get_flags(s->target) != bdrv_get_flags(to_replace)) {
478             bdrv_reopen(s->target, bdrv_get_flags(to_replace), NULL);
479         }
480         bdrv_replace_in_backing_chain(to_replace, s->target);
481     }
482 
483 out:
484     if (s->to_replace) {
485         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
486         error_free(s->replace_blocker);
487         bdrv_unref(s->to_replace);
488     }
489     if (replace_aio_context) {
490         aio_context_release(replace_aio_context);
491     }
492     g_free(s->replaces);
493     bdrv_op_unblock_all(s->target, s->common.blocker);
494     bdrv_unref(s->target);
495     block_job_completed(&s->common, data->ret);
496     g_free(data);
497     bdrv_drained_end(src);
498     if (qemu_get_aio_context() == bdrv_get_aio_context(src)) {
499         aio_enable_external(iohandler_get_aio_context());
500     }
501     bdrv_unref(src);
502 }
503 
504 static void coroutine_fn mirror_run(void *opaque)
505 {
506     MirrorBlockJob *s = opaque;
507     MirrorExitData *data;
508     BlockDriverState *bs = s->common.bs;
509     int64_t sector_num, end, length;
510     uint64_t last_pause_ns;
511     BlockDriverInfo bdi;
512     char backing_filename[2]; /* we only need 2 characters because we are only
513                                  checking for a NULL string */
514     int ret = 0;
515     int n;
516     int target_cluster_size = BDRV_SECTOR_SIZE;
517 
518     if (block_job_is_cancelled(&s->common)) {
519         goto immediate_exit;
520     }
521 
522     s->bdev_length = bdrv_getlength(bs);
523     if (s->bdev_length < 0) {
524         ret = s->bdev_length;
525         goto immediate_exit;
526     } else if (s->bdev_length == 0) {
527         /* Report BLOCK_JOB_READY and wait for complete. */
528         block_job_event_ready(&s->common);
529         s->synced = true;
530         while (!block_job_is_cancelled(&s->common) && !s->should_complete) {
531             block_job_yield(&s->common);
532         }
533         s->common.cancelled = false;
534         goto immediate_exit;
535     }
536 
537     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
538     s->in_flight_bitmap = bitmap_new(length);
539 
540     /* If we have no backing file yet in the destination, we cannot let
541      * the destination do COW.  Instead, we copy sectors around the
542      * dirty data if needed.  We need a bitmap to do that.
543      */
544     bdrv_get_backing_filename(s->target, backing_filename,
545                               sizeof(backing_filename));
546     if (!bdrv_get_info(s->target, &bdi) && bdi.cluster_size) {
547         target_cluster_size = bdi.cluster_size;
548     }
549     if (backing_filename[0] && !s->target->backing
550         && s->granularity < target_cluster_size) {
551         s->buf_size = MAX(s->buf_size, target_cluster_size);
552         s->cow_bitmap = bitmap_new(length);
553     }
554     s->target_cluster_sectors = target_cluster_size >> BDRV_SECTOR_BITS;
555     s->max_iov = MIN(s->common.bs->bl.max_iov, s->target->bl.max_iov);
556 
557     end = s->bdev_length / BDRV_SECTOR_SIZE;
558     s->buf = qemu_try_blockalign(bs, s->buf_size);
559     if (s->buf == NULL) {
560         ret = -ENOMEM;
561         goto immediate_exit;
562     }
563 
564     mirror_free_init(s);
565 
566     last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
567     if (!s->is_none_mode) {
568         /* First part, loop on the sectors and initialize the dirty bitmap.  */
569         BlockDriverState *base = s->base;
570         bool mark_all_dirty = s->base == NULL && !bdrv_has_zero_init(s->target);
571 
572         for (sector_num = 0; sector_num < end; ) {
573             /* Just to make sure we are not exceeding int limit. */
574             int nb_sectors = MIN(INT_MAX >> BDRV_SECTOR_BITS,
575                                  end - sector_num);
576             int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
577 
578             if (now - last_pause_ns > SLICE_TIME) {
579                 last_pause_ns = now;
580                 block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, 0);
581             }
582 
583             if (block_job_is_cancelled(&s->common)) {
584                 goto immediate_exit;
585             }
586 
587             ret = bdrv_is_allocated_above(bs, base, sector_num, nb_sectors, &n);
588 
589             if (ret < 0) {
590                 goto immediate_exit;
591             }
592 
593             assert(n > 0);
594             if (ret == 1 || mark_all_dirty) {
595                 bdrv_set_dirty_bitmap(s->dirty_bitmap, sector_num, n);
596             }
597             sector_num += n;
598         }
599     }
600 
601     bdrv_dirty_iter_init(s->dirty_bitmap, &s->hbi);
602     for (;;) {
603         uint64_t delay_ns = 0;
604         int64_t cnt;
605         bool should_complete;
606 
607         if (s->ret < 0) {
608             ret = s->ret;
609             goto immediate_exit;
610         }
611 
612         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
613         /* s->common.offset contains the number of bytes already processed so
614          * far, cnt is the number of dirty sectors remaining and
615          * s->sectors_in_flight is the number of sectors currently being
616          * processed; together those are the current total operation length */
617         s->common.len = s->common.offset +
618                         (cnt + s->sectors_in_flight) * BDRV_SECTOR_SIZE;
619 
620         /* Note that even when no rate limit is applied we need to yield
621          * periodically with no pending I/O so that bdrv_drain_all() returns.
622          * We do so every SLICE_TIME nanoseconds, or when there is an error,
623          * or when the source is clean, whichever comes first.
624          */
625         if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - last_pause_ns < SLICE_TIME &&
626             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
627             if (s->in_flight == MAX_IN_FLIGHT || s->buf_free_count == 0 ||
628                 (cnt == 0 && s->in_flight > 0)) {
629                 trace_mirror_yield(s, s->in_flight, s->buf_free_count, cnt);
630                 mirror_wait_for_io(s);
631                 continue;
632             } else if (cnt != 0) {
633                 delay_ns = mirror_iteration(s);
634             }
635         }
636 
637         should_complete = false;
638         if (s->in_flight == 0 && cnt == 0) {
639             trace_mirror_before_flush(s);
640             ret = bdrv_flush(s->target);
641             if (ret < 0) {
642                 if (mirror_error_action(s, false, -ret) ==
643                     BLOCK_ERROR_ACTION_REPORT) {
644                     goto immediate_exit;
645                 }
646             } else {
647                 /* We're out of the streaming phase.  From now on, if the job
648                  * is cancelled we will actually complete all pending I/O and
649                  * report completion.  This way, block-job-cancel will leave
650                  * the target in a consistent state.
651                  */
652                 if (!s->synced) {
653                     block_job_event_ready(&s->common);
654                     s->synced = true;
655                 }
656 
657                 should_complete = s->should_complete ||
658                     block_job_is_cancelled(&s->common);
659                 cnt = bdrv_get_dirty_count(s->dirty_bitmap);
660             }
661         }
662 
663         if (cnt == 0 && should_complete) {
664             /* The dirty bitmap is not updated while operations are pending.
665              * If we're about to exit, wait for pending operations before
666              * calling bdrv_get_dirty_count(bs), or we may exit while the
667              * source has dirty data to copy!
668              *
669              * Note that I/O can be submitted by the guest while
670              * mirror_populate runs.
671              */
672             trace_mirror_before_drain(s, cnt);
673             bdrv_co_drain(bs);
674             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
675         }
676 
677         ret = 0;
678         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
679         if (!s->synced) {
680             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
681             if (block_job_is_cancelled(&s->common)) {
682                 break;
683             }
684         } else if (!should_complete) {
685             delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0);
686             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
687         } else if (cnt == 0) {
688             /* The two disks are in sync.  Exit and report successful
689              * completion.
690              */
691             assert(QLIST_EMPTY(&bs->tracked_requests));
692             s->common.cancelled = false;
693             break;
694         }
695         last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
696     }
697 
698 immediate_exit:
699     if (s->in_flight > 0) {
700         /* We get here only if something went wrong.  Either the job failed,
701          * or it was cancelled prematurely so that we do not guarantee that
702          * the target is a copy of the source.
703          */
704         assert(ret < 0 || (!s->synced && block_job_is_cancelled(&s->common)));
705         mirror_drain(s);
706     }
707 
708     assert(s->in_flight == 0);
709     qemu_vfree(s->buf);
710     g_free(s->cow_bitmap);
711     g_free(s->in_flight_bitmap);
712     bdrv_release_dirty_bitmap(bs, s->dirty_bitmap);
713     if (s->target->blk) {
714         blk_iostatus_disable(s->target->blk);
715     }
716 
717     data = g_malloc(sizeof(*data));
718     data->ret = ret;
719     /* Before we switch to target in mirror_exit, make sure data doesn't
720      * change. */
721     bdrv_drained_begin(s->common.bs);
722     if (qemu_get_aio_context() == bdrv_get_aio_context(bs)) {
723         /* FIXME: virtio host notifiers run on iohandler_ctx, therefore the
724          * above bdrv_drained_end isn't enough to quiesce it. This is ugly, we
725          * need a block layer API change to achieve this. */
726         aio_disable_external(iohandler_get_aio_context());
727     }
728     block_job_defer_to_main_loop(&s->common, mirror_exit, data);
729 }
730 
731 static void mirror_set_speed(BlockJob *job, int64_t speed, Error **errp)
732 {
733     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
734 
735     if (speed < 0) {
736         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
737         return;
738     }
739     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
740 }
741 
742 static void mirror_iostatus_reset(BlockJob *job)
743 {
744     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
745 
746     if (s->target->blk) {
747         blk_iostatus_reset(s->target->blk);
748     }
749 }
750 
751 static void mirror_complete(BlockJob *job, Error **errp)
752 {
753     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
754     Error *local_err = NULL;
755     int ret;
756 
757     ret = bdrv_open_backing_file(s->target, NULL, "backing", &local_err);
758     if (ret < 0) {
759         error_propagate(errp, local_err);
760         return;
761     }
762     if (!s->synced) {
763         error_setg(errp, QERR_BLOCK_JOB_NOT_READY, job->id);
764         return;
765     }
766 
767     /* check the target bs is not blocked and block all operations on it */
768     if (s->replaces) {
769         AioContext *replace_aio_context;
770 
771         s->to_replace = bdrv_find_node(s->replaces);
772         if (!s->to_replace) {
773             error_setg(errp, "Node name '%s' not found", s->replaces);
774             return;
775         }
776 
777         replace_aio_context = bdrv_get_aio_context(s->to_replace);
778         aio_context_acquire(replace_aio_context);
779 
780         error_setg(&s->replace_blocker,
781                    "block device is in use by block-job-complete");
782         bdrv_op_block_all(s->to_replace, s->replace_blocker);
783         bdrv_ref(s->to_replace);
784 
785         aio_context_release(replace_aio_context);
786     }
787 
788     s->should_complete = true;
789     block_job_enter(&s->common);
790 }
791 
792 static const BlockJobDriver mirror_job_driver = {
793     .instance_size = sizeof(MirrorBlockJob),
794     .job_type      = BLOCK_JOB_TYPE_MIRROR,
795     .set_speed     = mirror_set_speed,
796     .iostatus_reset= mirror_iostatus_reset,
797     .complete      = mirror_complete,
798 };
799 
800 static const BlockJobDriver commit_active_job_driver = {
801     .instance_size = sizeof(MirrorBlockJob),
802     .job_type      = BLOCK_JOB_TYPE_COMMIT,
803     .set_speed     = mirror_set_speed,
804     .iostatus_reset
805                    = mirror_iostatus_reset,
806     .complete      = mirror_complete,
807 };
808 
809 static void mirror_start_job(BlockDriverState *bs, BlockDriverState *target,
810                              const char *replaces,
811                              int64_t speed, uint32_t granularity,
812                              int64_t buf_size,
813                              BlockdevOnError on_source_error,
814                              BlockdevOnError on_target_error,
815                              bool unmap,
816                              BlockCompletionFunc *cb,
817                              void *opaque, Error **errp,
818                              const BlockJobDriver *driver,
819                              bool is_none_mode, BlockDriverState *base)
820 {
821     MirrorBlockJob *s;
822     BlockDriverState *replaced_bs;
823 
824     if (granularity == 0) {
825         granularity = bdrv_get_default_bitmap_granularity(target);
826     }
827 
828     assert ((granularity & (granularity - 1)) == 0);
829 
830     if ((on_source_error == BLOCKDEV_ON_ERROR_STOP ||
831          on_source_error == BLOCKDEV_ON_ERROR_ENOSPC) &&
832         (!bs->blk || !blk_iostatus_is_enabled(bs->blk))) {
833         error_setg(errp, QERR_INVALID_PARAMETER, "on-source-error");
834         return;
835     }
836 
837     if (buf_size < 0) {
838         error_setg(errp, "Invalid parameter 'buf-size'");
839         return;
840     }
841 
842     if (buf_size == 0) {
843         buf_size = DEFAULT_MIRROR_BUF_SIZE;
844     }
845 
846     /* We can't support this case as long as the block layer can't handle
847      * multiple BlockBackends per BlockDriverState. */
848     if (replaces) {
849         replaced_bs = bdrv_lookup_bs(replaces, replaces, errp);
850         if (replaced_bs == NULL) {
851             return;
852         }
853     } else {
854         replaced_bs = bs;
855     }
856     if (replaced_bs->blk && target->blk) {
857         error_setg(errp, "Can't create node with two BlockBackends");
858         return;
859     }
860 
861     s = block_job_create(driver, bs, speed, cb, opaque, errp);
862     if (!s) {
863         return;
864     }
865 
866     s->replaces = g_strdup(replaces);
867     s->on_source_error = on_source_error;
868     s->on_target_error = on_target_error;
869     s->target = target;
870     s->is_none_mode = is_none_mode;
871     s->base = base;
872     s->granularity = granularity;
873     s->buf_size = ROUND_UP(buf_size, granularity);
874     s->unmap = unmap;
875 
876     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
877     if (!s->dirty_bitmap) {
878         g_free(s->replaces);
879         block_job_unref(&s->common);
880         return;
881     }
882 
883     bdrv_op_block_all(s->target, s->common.blocker);
884 
885     if (s->target->blk) {
886         blk_set_on_error(s->target->blk, on_target_error, on_target_error);
887         blk_iostatus_enable(s->target->blk);
888     }
889     s->common.co = qemu_coroutine_create(mirror_run);
890     trace_mirror_start(bs, s, s->common.co, opaque);
891     qemu_coroutine_enter(s->common.co, s);
892 }
893 
894 void mirror_start(BlockDriverState *bs, BlockDriverState *target,
895                   const char *replaces,
896                   int64_t speed, uint32_t granularity, int64_t buf_size,
897                   MirrorSyncMode mode, BlockdevOnError on_source_error,
898                   BlockdevOnError on_target_error,
899                   bool unmap,
900                   BlockCompletionFunc *cb,
901                   void *opaque, Error **errp)
902 {
903     bool is_none_mode;
904     BlockDriverState *base;
905 
906     if (mode == MIRROR_SYNC_MODE_INCREMENTAL) {
907         error_setg(errp, "Sync mode 'incremental' not supported");
908         return;
909     }
910     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
911     base = mode == MIRROR_SYNC_MODE_TOP ? backing_bs(bs) : NULL;
912     mirror_start_job(bs, target, replaces,
913                      speed, granularity, buf_size,
914                      on_source_error, on_target_error, unmap, cb, opaque, errp,
915                      &mirror_job_driver, is_none_mode, base);
916 }
917 
918 void commit_active_start(BlockDriverState *bs, BlockDriverState *base,
919                          int64_t speed,
920                          BlockdevOnError on_error,
921                          BlockCompletionFunc *cb,
922                          void *opaque, Error **errp)
923 {
924     int64_t length, base_length;
925     int orig_base_flags;
926     int ret;
927     Error *local_err = NULL;
928 
929     orig_base_flags = bdrv_get_flags(base);
930 
931     if (bdrv_reopen(base, bs->open_flags, errp)) {
932         return;
933     }
934 
935     length = bdrv_getlength(bs);
936     if (length < 0) {
937         error_setg_errno(errp, -length,
938                          "Unable to determine length of %s", bs->filename);
939         goto error_restore_flags;
940     }
941 
942     base_length = bdrv_getlength(base);
943     if (base_length < 0) {
944         error_setg_errno(errp, -base_length,
945                          "Unable to determine length of %s", base->filename);
946         goto error_restore_flags;
947     }
948 
949     if (length > base_length) {
950         ret = bdrv_truncate(base, length);
951         if (ret < 0) {
952             error_setg_errno(errp, -ret,
953                             "Top image %s is larger than base image %s, and "
954                              "resize of base image failed",
955                              bs->filename, base->filename);
956             goto error_restore_flags;
957         }
958     }
959 
960     bdrv_ref(base);
961     mirror_start_job(bs, base, NULL, speed, 0, 0,
962                      on_error, on_error, false, cb, opaque, &local_err,
963                      &commit_active_job_driver, false, base);
964     if (local_err) {
965         error_propagate(errp, local_err);
966         goto error_restore_flags;
967     }
968 
969     return;
970 
971 error_restore_flags:
972     /* ignore error and errp for bdrv_reopen, because we want to propagate
973      * the original error */
974     bdrv_reopen(base, orig_base_flags, NULL);
975     return;
976 }
977