xref: /openbmc/qemu/block/mirror.c (revision e09484ef)
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 
24 #define SLICE_TIME    100000000ULL /* ns */
25 #define MAX_IN_FLIGHT 16
26 #define MAX_IO_SECTORS ((1 << 20) >> BDRV_SECTOR_BITS) /* 1 Mb */
27 #define DEFAULT_MIRROR_BUF_SIZE \
28     (MAX_IN_FLIGHT * MAX_IO_SECTORS * BDRV_SECTOR_SIZE)
29 
30 /* The mirroring buffer is a list of granularity-sized chunks.
31  * Free chunks are organized in a list.
32  */
33 typedef struct MirrorBuffer {
34     QSIMPLEQ_ENTRY(MirrorBuffer) next;
35 } MirrorBuffer;
36 
37 typedef struct MirrorBlockJob {
38     BlockJob common;
39     RateLimit limit;
40     BlockBackend *target;
41     BlockDriverState *base;
42     /* The name of the graph node to replace */
43     char *replaces;
44     /* The BDS to replace */
45     BlockDriverState *to_replace;
46     /* Used to block operations on the drive-mirror-replace target */
47     Error *replace_blocker;
48     bool is_none_mode;
49     BlockMirrorBackingMode backing_mode;
50     BlockdevOnError on_source_error, on_target_error;
51     bool synced;
52     bool should_complete;
53     int64_t granularity;
54     size_t buf_size;
55     int64_t bdev_length;
56     unsigned long *cow_bitmap;
57     BdrvDirtyBitmap *dirty_bitmap;
58     HBitmapIter hbi;
59     uint8_t *buf;
60     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
61     int buf_free_count;
62 
63     uint64_t last_pause_ns;
64     unsigned long *in_flight_bitmap;
65     int in_flight;
66     int64_t sectors_in_flight;
67     int ret;
68     bool unmap;
69     bool waiting_for_io;
70     int target_cluster_sectors;
71     int max_iov;
72 } MirrorBlockJob;
73 
74 typedef struct MirrorOp {
75     MirrorBlockJob *s;
76     QEMUIOVector qiov;
77     int64_t sector_num;
78     int nb_sectors;
79 } MirrorOp;
80 
81 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
82                                             int error)
83 {
84     s->synced = false;
85     if (read) {
86         return block_job_error_action(&s->common, s->on_source_error,
87                                       true, error);
88     } else {
89         return block_job_error_action(&s->common, s->on_target_error,
90                                       false, error);
91     }
92 }
93 
94 static void mirror_iteration_done(MirrorOp *op, int ret)
95 {
96     MirrorBlockJob *s = op->s;
97     struct iovec *iov;
98     int64_t chunk_num;
99     int i, nb_chunks, sectors_per_chunk;
100 
101     trace_mirror_iteration_done(s, op->sector_num, op->nb_sectors, ret);
102 
103     s->in_flight--;
104     s->sectors_in_flight -= op->nb_sectors;
105     iov = op->qiov.iov;
106     for (i = 0; i < op->qiov.niov; i++) {
107         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
108         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
109         s->buf_free_count++;
110     }
111 
112     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
113     chunk_num = op->sector_num / sectors_per_chunk;
114     nb_chunks = DIV_ROUND_UP(op->nb_sectors, sectors_per_chunk);
115     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
116     if (ret >= 0) {
117         if (s->cow_bitmap) {
118             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
119         }
120         s->common.offset += (uint64_t)op->nb_sectors * BDRV_SECTOR_SIZE;
121     }
122 
123     qemu_iovec_destroy(&op->qiov);
124     g_free(op);
125 
126     if (s->waiting_for_io) {
127         qemu_coroutine_enter(s->common.co);
128     }
129 }
130 
131 static void mirror_write_complete(void *opaque, int ret)
132 {
133     MirrorOp *op = opaque;
134     MirrorBlockJob *s = op->s;
135     if (ret < 0) {
136         BlockErrorAction action;
137 
138         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->sector_num, op->nb_sectors);
139         action = mirror_error_action(s, false, -ret);
140         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
141             s->ret = ret;
142         }
143     }
144     mirror_iteration_done(op, ret);
145 }
146 
147 static void mirror_read_complete(void *opaque, int ret)
148 {
149     MirrorOp *op = opaque;
150     MirrorBlockJob *s = op->s;
151     if (ret < 0) {
152         BlockErrorAction action;
153 
154         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->sector_num, op->nb_sectors);
155         action = mirror_error_action(s, true, -ret);
156         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
157             s->ret = ret;
158         }
159 
160         mirror_iteration_done(op, ret);
161         return;
162     }
163     blk_aio_pwritev(s->target, op->sector_num * BDRV_SECTOR_SIZE, &op->qiov,
164                     0, mirror_write_complete, op);
165 }
166 
167 static inline void mirror_clip_sectors(MirrorBlockJob *s,
168                                        int64_t sector_num,
169                                        int *nb_sectors)
170 {
171     *nb_sectors = MIN(*nb_sectors,
172                       s->bdev_length / BDRV_SECTOR_SIZE - sector_num);
173 }
174 
175 /* Round sector_num and/or nb_sectors to target cluster if COW is needed, and
176  * return the offset of the adjusted tail sector against original. */
177 static int mirror_cow_align(MirrorBlockJob *s,
178                             int64_t *sector_num,
179                             int *nb_sectors)
180 {
181     bool need_cow;
182     int ret = 0;
183     int chunk_sectors = s->granularity >> BDRV_SECTOR_BITS;
184     int64_t align_sector_num = *sector_num;
185     int align_nb_sectors = *nb_sectors;
186     int max_sectors = chunk_sectors * s->max_iov;
187 
188     need_cow = !test_bit(*sector_num / chunk_sectors, s->cow_bitmap);
189     need_cow |= !test_bit((*sector_num + *nb_sectors - 1) / chunk_sectors,
190                           s->cow_bitmap);
191     if (need_cow) {
192         bdrv_round_sectors_to_clusters(blk_bs(s->target), *sector_num,
193                                        *nb_sectors, &align_sector_num,
194                                        &align_nb_sectors);
195     }
196 
197     if (align_nb_sectors > max_sectors) {
198         align_nb_sectors = max_sectors;
199         if (need_cow) {
200             align_nb_sectors = QEMU_ALIGN_DOWN(align_nb_sectors,
201                                                s->target_cluster_sectors);
202         }
203     }
204     /* Clipping may result in align_nb_sectors unaligned to chunk boundary, but
205      * that doesn't matter because it's already the end of source image. */
206     mirror_clip_sectors(s, align_sector_num, &align_nb_sectors);
207 
208     ret = align_sector_num + align_nb_sectors - (*sector_num + *nb_sectors);
209     *sector_num = align_sector_num;
210     *nb_sectors = align_nb_sectors;
211     assert(ret >= 0);
212     return ret;
213 }
214 
215 static inline void mirror_wait_for_io(MirrorBlockJob *s)
216 {
217     assert(!s->waiting_for_io);
218     s->waiting_for_io = true;
219     qemu_coroutine_yield();
220     s->waiting_for_io = false;
221 }
222 
223 /* Submit async read while handling COW.
224  * Returns: The number of sectors copied after and including sector_num,
225  *          excluding any sectors copied prior to sector_num due to alignment.
226  *          This will be nb_sectors if no alignment is necessary, or
227  *          (new_end - sector_num) if tail is rounded up or down due to
228  *          alignment or buffer limit.
229  */
230 static int mirror_do_read(MirrorBlockJob *s, int64_t sector_num,
231                           int nb_sectors)
232 {
233     BlockBackend *source = s->common.blk;
234     int sectors_per_chunk, nb_chunks;
235     int ret;
236     MirrorOp *op;
237     int max_sectors;
238 
239     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
240     max_sectors = sectors_per_chunk * s->max_iov;
241 
242     /* We can only handle as much as buf_size at a time. */
243     nb_sectors = MIN(s->buf_size >> BDRV_SECTOR_BITS, nb_sectors);
244     nb_sectors = MIN(max_sectors, nb_sectors);
245     assert(nb_sectors);
246     ret = nb_sectors;
247 
248     if (s->cow_bitmap) {
249         ret += mirror_cow_align(s, &sector_num, &nb_sectors);
250     }
251     assert(nb_sectors << BDRV_SECTOR_BITS <= s->buf_size);
252     /* The sector range must meet granularity because:
253      * 1) Caller passes in aligned values;
254      * 2) mirror_cow_align is used only when target cluster is larger. */
255     assert(!(sector_num % sectors_per_chunk));
256     nb_chunks = DIV_ROUND_UP(nb_sectors, sectors_per_chunk);
257 
258     while (s->buf_free_count < nb_chunks) {
259         trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
260         mirror_wait_for_io(s);
261     }
262 
263     /* Allocate a MirrorOp that is used as an AIO callback.  */
264     op = g_new(MirrorOp, 1);
265     op->s = s;
266     op->sector_num = sector_num;
267     op->nb_sectors = nb_sectors;
268 
269     /* Now make a QEMUIOVector taking enough granularity-sized chunks
270      * from s->buf_free.
271      */
272     qemu_iovec_init(&op->qiov, nb_chunks);
273     while (nb_chunks-- > 0) {
274         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
275         size_t remaining = nb_sectors * BDRV_SECTOR_SIZE - op->qiov.size;
276 
277         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
278         s->buf_free_count--;
279         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
280     }
281 
282     /* Copy the dirty cluster.  */
283     s->in_flight++;
284     s->sectors_in_flight += nb_sectors;
285     trace_mirror_one_iteration(s, sector_num, nb_sectors);
286 
287     blk_aio_preadv(source, sector_num * BDRV_SECTOR_SIZE, &op->qiov, 0,
288                    mirror_read_complete, op);
289     return ret;
290 }
291 
292 static void mirror_do_zero_or_discard(MirrorBlockJob *s,
293                                       int64_t sector_num,
294                                       int nb_sectors,
295                                       bool is_discard)
296 {
297     MirrorOp *op;
298 
299     /* Allocate a MirrorOp that is used as an AIO callback. The qiov is zeroed
300      * so the freeing in mirror_iteration_done is nop. */
301     op = g_new0(MirrorOp, 1);
302     op->s = s;
303     op->sector_num = sector_num;
304     op->nb_sectors = nb_sectors;
305 
306     s->in_flight++;
307     s->sectors_in_flight += nb_sectors;
308     if (is_discard) {
309         blk_aio_pdiscard(s->target, sector_num << BDRV_SECTOR_BITS,
310                          op->nb_sectors << BDRV_SECTOR_BITS,
311                          mirror_write_complete, op);
312     } else {
313         blk_aio_pwrite_zeroes(s->target, sector_num * BDRV_SECTOR_SIZE,
314                               op->nb_sectors * BDRV_SECTOR_SIZE,
315                               s->unmap ? BDRV_REQ_MAY_UNMAP : 0,
316                               mirror_write_complete, op);
317     }
318 }
319 
320 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
321 {
322     BlockDriverState *source = blk_bs(s->common.blk);
323     int64_t sector_num, first_chunk;
324     uint64_t delay_ns = 0;
325     /* At least the first dirty chunk is mirrored in one iteration. */
326     int nb_chunks = 1;
327     int64_t end = s->bdev_length / BDRV_SECTOR_SIZE;
328     int sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
329     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
330     int max_io_sectors = MAX((s->buf_size >> BDRV_SECTOR_BITS) / MAX_IN_FLIGHT,
331                              MAX_IO_SECTORS);
332 
333     sector_num = hbitmap_iter_next(&s->hbi);
334     if (sector_num < 0) {
335         bdrv_dirty_iter_init(s->dirty_bitmap, &s->hbi);
336         sector_num = hbitmap_iter_next(&s->hbi);
337         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
338         assert(sector_num >= 0);
339     }
340 
341     first_chunk = sector_num / sectors_per_chunk;
342     while (test_bit(first_chunk, s->in_flight_bitmap)) {
343         trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
344         mirror_wait_for_io(s);
345     }
346 
347     block_job_pause_point(&s->common);
348 
349     /* Find the number of consective dirty chunks following the first dirty
350      * one, and wait for in flight requests in them. */
351     while (nb_chunks * sectors_per_chunk < (s->buf_size >> BDRV_SECTOR_BITS)) {
352         int64_t hbitmap_next;
353         int64_t next_sector = sector_num + nb_chunks * sectors_per_chunk;
354         int64_t next_chunk = next_sector / sectors_per_chunk;
355         if (next_sector >= end ||
356             !bdrv_get_dirty(source, s->dirty_bitmap, next_sector)) {
357             break;
358         }
359         if (test_bit(next_chunk, s->in_flight_bitmap)) {
360             break;
361         }
362 
363         hbitmap_next = hbitmap_iter_next(&s->hbi);
364         if (hbitmap_next > next_sector || hbitmap_next < 0) {
365             /* The bitmap iterator's cache is stale, refresh it */
366             bdrv_set_dirty_iter(&s->hbi, next_sector);
367             hbitmap_next = hbitmap_iter_next(&s->hbi);
368         }
369         assert(hbitmap_next == next_sector);
370         nb_chunks++;
371     }
372 
373     /* Clear dirty bits before querying the block status, because
374      * calling bdrv_get_block_status_above could yield - if some blocks are
375      * marked dirty in this window, we need to know.
376      */
377     bdrv_reset_dirty_bitmap(s->dirty_bitmap, sector_num,
378                             nb_chunks * sectors_per_chunk);
379     bitmap_set(s->in_flight_bitmap, sector_num / sectors_per_chunk, nb_chunks);
380     while (nb_chunks > 0 && sector_num < end) {
381         int ret;
382         int io_sectors, io_sectors_acct;
383         BlockDriverState *file;
384         enum MirrorMethod {
385             MIRROR_METHOD_COPY,
386             MIRROR_METHOD_ZERO,
387             MIRROR_METHOD_DISCARD
388         } mirror_method = MIRROR_METHOD_COPY;
389 
390         assert(!(sector_num % sectors_per_chunk));
391         ret = bdrv_get_block_status_above(source, NULL, sector_num,
392                                           nb_chunks * sectors_per_chunk,
393                                           &io_sectors, &file);
394         if (ret < 0) {
395             io_sectors = MIN(nb_chunks * sectors_per_chunk, max_io_sectors);
396         } else if (ret & BDRV_BLOCK_DATA) {
397             io_sectors = MIN(io_sectors, max_io_sectors);
398         }
399 
400         io_sectors -= io_sectors % sectors_per_chunk;
401         if (io_sectors < sectors_per_chunk) {
402             io_sectors = sectors_per_chunk;
403         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
404             int64_t target_sector_num;
405             int target_nb_sectors;
406             bdrv_round_sectors_to_clusters(blk_bs(s->target), sector_num,
407                                            io_sectors,  &target_sector_num,
408                                            &target_nb_sectors);
409             if (target_sector_num == sector_num &&
410                 target_nb_sectors == io_sectors) {
411                 mirror_method = ret & BDRV_BLOCK_ZERO ?
412                                     MIRROR_METHOD_ZERO :
413                                     MIRROR_METHOD_DISCARD;
414             }
415         }
416 
417         while (s->in_flight >= MAX_IN_FLIGHT) {
418             trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
419             mirror_wait_for_io(s);
420         }
421 
422         if (s->ret < 0) {
423             return 0;
424         }
425 
426         mirror_clip_sectors(s, sector_num, &io_sectors);
427         switch (mirror_method) {
428         case MIRROR_METHOD_COPY:
429             io_sectors = mirror_do_read(s, sector_num, io_sectors);
430             io_sectors_acct = io_sectors;
431             break;
432         case MIRROR_METHOD_ZERO:
433         case MIRROR_METHOD_DISCARD:
434             mirror_do_zero_or_discard(s, sector_num, io_sectors,
435                                       mirror_method == MIRROR_METHOD_DISCARD);
436             if (write_zeroes_ok) {
437                 io_sectors_acct = 0;
438             } else {
439                 io_sectors_acct = io_sectors;
440             }
441             break;
442         default:
443             abort();
444         }
445         assert(io_sectors);
446         sector_num += io_sectors;
447         nb_chunks -= DIV_ROUND_UP(io_sectors, sectors_per_chunk);
448         if (s->common.speed) {
449             delay_ns = ratelimit_calculate_delay(&s->limit, io_sectors_acct);
450         }
451     }
452     return delay_ns;
453 }
454 
455 static void mirror_free_init(MirrorBlockJob *s)
456 {
457     int granularity = s->granularity;
458     size_t buf_size = s->buf_size;
459     uint8_t *buf = s->buf;
460 
461     assert(s->buf_free_count == 0);
462     QSIMPLEQ_INIT(&s->buf_free);
463     while (buf_size != 0) {
464         MirrorBuffer *cur = (MirrorBuffer *)buf;
465         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
466         s->buf_free_count++;
467         buf_size -= granularity;
468         buf += granularity;
469     }
470 }
471 
472 static void mirror_drain(MirrorBlockJob *s)
473 {
474     while (s->in_flight > 0) {
475         mirror_wait_for_io(s);
476     }
477 }
478 
479 typedef struct {
480     int ret;
481 } MirrorExitData;
482 
483 static void mirror_exit(BlockJob *job, void *opaque)
484 {
485     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
486     MirrorExitData *data = opaque;
487     AioContext *replace_aio_context = NULL;
488     BlockDriverState *src = blk_bs(s->common.blk);
489     BlockDriverState *target_bs = blk_bs(s->target);
490 
491     /* Make sure that the source BDS doesn't go away before we called
492      * block_job_completed(). */
493     bdrv_ref(src);
494 
495     if (s->to_replace) {
496         replace_aio_context = bdrv_get_aio_context(s->to_replace);
497         aio_context_acquire(replace_aio_context);
498     }
499 
500     if (s->should_complete && data->ret == 0) {
501         BlockDriverState *to_replace = src;
502         if (s->to_replace) {
503             to_replace = s->to_replace;
504         }
505 
506         if (bdrv_get_flags(target_bs) != bdrv_get_flags(to_replace)) {
507             bdrv_reopen(target_bs, bdrv_get_flags(to_replace), NULL);
508         }
509 
510         /* The mirror job has no requests in flight any more, but we need to
511          * drain potential other users of the BDS before changing the graph. */
512         bdrv_drained_begin(target_bs);
513         bdrv_replace_in_backing_chain(to_replace, target_bs);
514         bdrv_drained_end(target_bs);
515 
516         /* We just changed the BDS the job BB refers to */
517         blk_remove_bs(job->blk);
518         blk_insert_bs(job->blk, src);
519     }
520     if (s->to_replace) {
521         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
522         error_free(s->replace_blocker);
523         bdrv_unref(s->to_replace);
524     }
525     if (replace_aio_context) {
526         aio_context_release(replace_aio_context);
527     }
528     g_free(s->replaces);
529     bdrv_op_unblock_all(target_bs, s->common.blocker);
530     blk_unref(s->target);
531     block_job_completed(&s->common, data->ret);
532     g_free(data);
533     bdrv_drained_end(src);
534     bdrv_unref(src);
535 }
536 
537 static void mirror_throttle(MirrorBlockJob *s)
538 {
539     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
540 
541     if (now - s->last_pause_ns > SLICE_TIME) {
542         s->last_pause_ns = now;
543         block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, 0);
544     } else {
545         block_job_pause_point(&s->common);
546     }
547 }
548 
549 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
550 {
551     int64_t sector_num, end;
552     BlockDriverState *base = s->base;
553     BlockDriverState *bs = blk_bs(s->common.blk);
554     BlockDriverState *target_bs = blk_bs(s->target);
555     int ret, n;
556 
557     end = s->bdev_length / BDRV_SECTOR_SIZE;
558 
559     if (base == NULL && !bdrv_has_zero_init(target_bs)) {
560         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
561             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, end);
562             return 0;
563         }
564 
565         for (sector_num = 0; sector_num < end; ) {
566             int nb_sectors = MIN(end - sector_num,
567                 QEMU_ALIGN_DOWN(INT_MAX, s->granularity) >> BDRV_SECTOR_BITS);
568 
569             mirror_throttle(s);
570 
571             if (block_job_is_cancelled(&s->common)) {
572                 return 0;
573             }
574 
575             if (s->in_flight >= MAX_IN_FLIGHT) {
576                 trace_mirror_yield(s, s->in_flight, s->buf_free_count, -1);
577                 mirror_wait_for_io(s);
578                 continue;
579             }
580 
581             mirror_do_zero_or_discard(s, sector_num, nb_sectors, false);
582             sector_num += nb_sectors;
583         }
584 
585         mirror_drain(s);
586     }
587 
588     /* First part, loop on the sectors and initialize the dirty bitmap.  */
589     for (sector_num = 0; sector_num < end; ) {
590         /* Just to make sure we are not exceeding int limit. */
591         int nb_sectors = MIN(INT_MAX >> BDRV_SECTOR_BITS,
592                              end - sector_num);
593 
594         mirror_throttle(s);
595 
596         if (block_job_is_cancelled(&s->common)) {
597             return 0;
598         }
599 
600         ret = bdrv_is_allocated_above(bs, base, sector_num, nb_sectors, &n);
601         if (ret < 0) {
602             return ret;
603         }
604 
605         assert(n > 0);
606         if (ret == 1) {
607             bdrv_set_dirty_bitmap(s->dirty_bitmap, sector_num, n);
608         }
609         sector_num += n;
610     }
611     return 0;
612 }
613 
614 static void coroutine_fn mirror_run(void *opaque)
615 {
616     MirrorBlockJob *s = opaque;
617     MirrorExitData *data;
618     BlockDriverState *bs = blk_bs(s->common.blk);
619     BlockDriverState *target_bs = blk_bs(s->target);
620     int64_t length;
621     BlockDriverInfo bdi;
622     char backing_filename[2]; /* we only need 2 characters because we are only
623                                  checking for a NULL string */
624     int ret = 0;
625     int target_cluster_size = BDRV_SECTOR_SIZE;
626 
627     if (block_job_is_cancelled(&s->common)) {
628         goto immediate_exit;
629     }
630 
631     s->bdev_length = bdrv_getlength(bs);
632     if (s->bdev_length < 0) {
633         ret = s->bdev_length;
634         goto immediate_exit;
635     } else if (s->bdev_length == 0) {
636         /* Report BLOCK_JOB_READY and wait for complete. */
637         block_job_event_ready(&s->common);
638         s->synced = true;
639         while (!block_job_is_cancelled(&s->common) && !s->should_complete) {
640             block_job_yield(&s->common);
641         }
642         s->common.cancelled = false;
643         goto immediate_exit;
644     }
645 
646     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
647     s->in_flight_bitmap = bitmap_new(length);
648 
649     /* If we have no backing file yet in the destination, we cannot let
650      * the destination do COW.  Instead, we copy sectors around the
651      * dirty data if needed.  We need a bitmap to do that.
652      */
653     bdrv_get_backing_filename(target_bs, backing_filename,
654                               sizeof(backing_filename));
655     if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) {
656         target_cluster_size = bdi.cluster_size;
657     }
658     if (backing_filename[0] && !target_bs->backing
659         && s->granularity < target_cluster_size) {
660         s->buf_size = MAX(s->buf_size, target_cluster_size);
661         s->cow_bitmap = bitmap_new(length);
662     }
663     s->target_cluster_sectors = target_cluster_size >> BDRV_SECTOR_BITS;
664     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
665 
666     s->buf = qemu_try_blockalign(bs, s->buf_size);
667     if (s->buf == NULL) {
668         ret = -ENOMEM;
669         goto immediate_exit;
670     }
671 
672     mirror_free_init(s);
673 
674     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
675     if (!s->is_none_mode) {
676         ret = mirror_dirty_init(s);
677         if (ret < 0 || block_job_is_cancelled(&s->common)) {
678             goto immediate_exit;
679         }
680     }
681 
682     bdrv_dirty_iter_init(s->dirty_bitmap, &s->hbi);
683     for (;;) {
684         uint64_t delay_ns = 0;
685         int64_t cnt, delta;
686         bool should_complete;
687 
688         if (s->ret < 0) {
689             ret = s->ret;
690             goto immediate_exit;
691         }
692 
693         block_job_pause_point(&s->common);
694 
695         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
696         /* s->common.offset contains the number of bytes already processed so
697          * far, cnt is the number of dirty sectors remaining and
698          * s->sectors_in_flight is the number of sectors currently being
699          * processed; together those are the current total operation length */
700         s->common.len = s->common.offset +
701                         (cnt + s->sectors_in_flight) * BDRV_SECTOR_SIZE;
702 
703         /* Note that even when no rate limit is applied we need to yield
704          * periodically with no pending I/O so that bdrv_drain_all() returns.
705          * We do so every SLICE_TIME nanoseconds, or when there is an error,
706          * or when the source is clean, whichever comes first.
707          */
708         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
709         if (delta < SLICE_TIME &&
710             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
711             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
712                 (cnt == 0 && s->in_flight > 0)) {
713                 trace_mirror_yield(s, s->in_flight, s->buf_free_count, cnt);
714                 mirror_wait_for_io(s);
715                 continue;
716             } else if (cnt != 0) {
717                 delay_ns = mirror_iteration(s);
718             }
719         }
720 
721         should_complete = false;
722         if (s->in_flight == 0 && cnt == 0) {
723             trace_mirror_before_flush(s);
724             ret = blk_flush(s->target);
725             if (ret < 0) {
726                 if (mirror_error_action(s, false, -ret) ==
727                     BLOCK_ERROR_ACTION_REPORT) {
728                     goto immediate_exit;
729                 }
730             } else {
731                 /* We're out of the streaming phase.  From now on, if the job
732                  * is cancelled we will actually complete all pending I/O and
733                  * report completion.  This way, block-job-cancel will leave
734                  * the target in a consistent state.
735                  */
736                 if (!s->synced) {
737                     block_job_event_ready(&s->common);
738                     s->synced = true;
739                 }
740 
741                 should_complete = s->should_complete ||
742                     block_job_is_cancelled(&s->common);
743                 cnt = bdrv_get_dirty_count(s->dirty_bitmap);
744             }
745         }
746 
747         if (cnt == 0 && should_complete) {
748             /* The dirty bitmap is not updated while operations are pending.
749              * If we're about to exit, wait for pending operations before
750              * calling bdrv_get_dirty_count(bs), or we may exit while the
751              * source has dirty data to copy!
752              *
753              * Note that I/O can be submitted by the guest while
754              * mirror_populate runs.
755              */
756             trace_mirror_before_drain(s, cnt);
757             bdrv_co_drain(bs);
758             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
759         }
760 
761         ret = 0;
762         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
763         if (!s->synced) {
764             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
765             if (block_job_is_cancelled(&s->common)) {
766                 break;
767             }
768         } else if (!should_complete) {
769             delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0);
770             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
771         } else if (cnt == 0) {
772             /* The two disks are in sync.  Exit and report successful
773              * completion.
774              */
775             assert(QLIST_EMPTY(&bs->tracked_requests));
776             s->common.cancelled = false;
777             break;
778         }
779         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
780     }
781 
782 immediate_exit:
783     if (s->in_flight > 0) {
784         /* We get here only if something went wrong.  Either the job failed,
785          * or it was cancelled prematurely so that we do not guarantee that
786          * the target is a copy of the source.
787          */
788         assert(ret < 0 || (!s->synced && block_job_is_cancelled(&s->common)));
789         mirror_drain(s);
790     }
791 
792     assert(s->in_flight == 0);
793     qemu_vfree(s->buf);
794     g_free(s->cow_bitmap);
795     g_free(s->in_flight_bitmap);
796     bdrv_release_dirty_bitmap(bs, s->dirty_bitmap);
797 
798     data = g_malloc(sizeof(*data));
799     data->ret = ret;
800     /* Before we switch to target in mirror_exit, make sure data doesn't
801      * change. */
802     bdrv_drained_begin(bs);
803     block_job_defer_to_main_loop(&s->common, mirror_exit, data);
804 }
805 
806 static void mirror_set_speed(BlockJob *job, int64_t speed, Error **errp)
807 {
808     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
809 
810     if (speed < 0) {
811         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
812         return;
813     }
814     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
815 }
816 
817 static void mirror_complete(BlockJob *job, Error **errp)
818 {
819     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
820     BlockDriverState *src, *target;
821 
822     src = blk_bs(job->blk);
823     target = blk_bs(s->target);
824 
825     if (!s->synced) {
826         error_setg(errp, "The active block job '%s' cannot be completed",
827                    job->id);
828         return;
829     }
830 
831     if (s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
832         int ret;
833 
834         assert(!target->backing);
835         ret = bdrv_open_backing_file(target, NULL, "backing", errp);
836         if (ret < 0) {
837             return;
838         }
839     }
840 
841     /* block all operations on to_replace bs */
842     if (s->replaces) {
843         AioContext *replace_aio_context;
844 
845         s->to_replace = bdrv_find_node(s->replaces);
846         if (!s->to_replace) {
847             error_setg(errp, "Node name '%s' not found", s->replaces);
848             return;
849         }
850 
851         replace_aio_context = bdrv_get_aio_context(s->to_replace);
852         aio_context_acquire(replace_aio_context);
853 
854         error_setg(&s->replace_blocker,
855                    "block device is in use by block-job-complete");
856         bdrv_op_block_all(s->to_replace, s->replace_blocker);
857         bdrv_ref(s->to_replace);
858 
859         aio_context_release(replace_aio_context);
860     }
861 
862     if (s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
863         BlockDriverState *backing = s->is_none_mode ? src : s->base;
864         if (backing_bs(target) != backing) {
865             bdrv_set_backing_hd(target, backing);
866         }
867     }
868 
869     s->should_complete = true;
870     block_job_enter(&s->common);
871 }
872 
873 /* There is no matching mirror_resume() because mirror_run() will begin
874  * iterating again when the job is resumed.
875  */
876 static void coroutine_fn mirror_pause(BlockJob *job)
877 {
878     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
879 
880     mirror_drain(s);
881 }
882 
883 static void mirror_attached_aio_context(BlockJob *job, AioContext *new_context)
884 {
885     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
886 
887     blk_set_aio_context(s->target, new_context);
888 }
889 
890 static const BlockJobDriver mirror_job_driver = {
891     .instance_size          = sizeof(MirrorBlockJob),
892     .job_type               = BLOCK_JOB_TYPE_MIRROR,
893     .set_speed              = mirror_set_speed,
894     .complete               = mirror_complete,
895     .pause                  = mirror_pause,
896     .attached_aio_context   = mirror_attached_aio_context,
897 };
898 
899 static const BlockJobDriver commit_active_job_driver = {
900     .instance_size          = sizeof(MirrorBlockJob),
901     .job_type               = BLOCK_JOB_TYPE_COMMIT,
902     .set_speed              = mirror_set_speed,
903     .complete               = mirror_complete,
904     .pause                  = mirror_pause,
905     .attached_aio_context   = mirror_attached_aio_context,
906 };
907 
908 static void mirror_start_job(const char *job_id, BlockDriverState *bs,
909                              BlockDriverState *target, const char *replaces,
910                              int64_t speed, uint32_t granularity,
911                              int64_t buf_size,
912                              BlockMirrorBackingMode backing_mode,
913                              BlockdevOnError on_source_error,
914                              BlockdevOnError on_target_error,
915                              bool unmap,
916                              BlockCompletionFunc *cb,
917                              void *opaque, Error **errp,
918                              const BlockJobDriver *driver,
919                              bool is_none_mode, BlockDriverState *base)
920 {
921     MirrorBlockJob *s;
922 
923     if (granularity == 0) {
924         granularity = bdrv_get_default_bitmap_granularity(target);
925     }
926 
927     assert ((granularity & (granularity - 1)) == 0);
928 
929     if (buf_size < 0) {
930         error_setg(errp, "Invalid parameter 'buf-size'");
931         return;
932     }
933 
934     if (buf_size == 0) {
935         buf_size = DEFAULT_MIRROR_BUF_SIZE;
936     }
937 
938     s = block_job_create(job_id, driver, bs, speed, cb, opaque, errp);
939     if (!s) {
940         return;
941     }
942 
943     s->target = blk_new();
944     blk_insert_bs(s->target, target);
945 
946     s->replaces = g_strdup(replaces);
947     s->on_source_error = on_source_error;
948     s->on_target_error = on_target_error;
949     s->is_none_mode = is_none_mode;
950     s->backing_mode = backing_mode;
951     s->base = base;
952     s->granularity = granularity;
953     s->buf_size = ROUND_UP(buf_size, granularity);
954     s->unmap = unmap;
955 
956     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
957     if (!s->dirty_bitmap) {
958         g_free(s->replaces);
959         blk_unref(s->target);
960         block_job_unref(&s->common);
961         return;
962     }
963 
964     bdrv_op_block_all(target, s->common.blocker);
965 
966     s->common.co = qemu_coroutine_create(mirror_run, s);
967     trace_mirror_start(bs, s, s->common.co, opaque);
968     qemu_coroutine_enter(s->common.co);
969 }
970 
971 void mirror_start(const char *job_id, BlockDriverState *bs,
972                   BlockDriverState *target, const char *replaces,
973                   int64_t speed, uint32_t granularity, int64_t buf_size,
974                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
975                   BlockdevOnError on_source_error,
976                   BlockdevOnError on_target_error,
977                   bool unmap,
978                   BlockCompletionFunc *cb,
979                   void *opaque, Error **errp)
980 {
981     bool is_none_mode;
982     BlockDriverState *base;
983 
984     if (mode == MIRROR_SYNC_MODE_INCREMENTAL) {
985         error_setg(errp, "Sync mode 'incremental' not supported");
986         return;
987     }
988     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
989     base = mode == MIRROR_SYNC_MODE_TOP ? backing_bs(bs) : NULL;
990     mirror_start_job(job_id, bs, target, replaces,
991                      speed, granularity, buf_size, backing_mode,
992                      on_source_error, on_target_error, unmap, cb, opaque, errp,
993                      &mirror_job_driver, is_none_mode, base);
994 }
995 
996 void commit_active_start(const char *job_id, BlockDriverState *bs,
997                          BlockDriverState *base, int64_t speed,
998                          BlockdevOnError on_error,
999                          BlockCompletionFunc *cb,
1000                          void *opaque, Error **errp)
1001 {
1002     int64_t length, base_length;
1003     int orig_base_flags;
1004     int ret;
1005     Error *local_err = NULL;
1006 
1007     orig_base_flags = bdrv_get_flags(base);
1008 
1009     if (bdrv_reopen(base, bs->open_flags, errp)) {
1010         return;
1011     }
1012 
1013     length = bdrv_getlength(bs);
1014     if (length < 0) {
1015         error_setg_errno(errp, -length,
1016                          "Unable to determine length of %s", bs->filename);
1017         goto error_restore_flags;
1018     }
1019 
1020     base_length = bdrv_getlength(base);
1021     if (base_length < 0) {
1022         error_setg_errno(errp, -base_length,
1023                          "Unable to determine length of %s", base->filename);
1024         goto error_restore_flags;
1025     }
1026 
1027     if (length > base_length) {
1028         ret = bdrv_truncate(base, length);
1029         if (ret < 0) {
1030             error_setg_errno(errp, -ret,
1031                             "Top image %s is larger than base image %s, and "
1032                              "resize of base image failed",
1033                              bs->filename, base->filename);
1034             goto error_restore_flags;
1035         }
1036     }
1037 
1038     mirror_start_job(job_id, bs, base, NULL, speed, 0, 0,
1039                      MIRROR_LEAVE_BACKING_CHAIN,
1040                      on_error, on_error, false, cb, opaque, &local_err,
1041                      &commit_active_job_driver, false, base);
1042     if (local_err) {
1043         error_propagate(errp, local_err);
1044         goto error_restore_flags;
1045     }
1046 
1047     return;
1048 
1049 error_restore_flags:
1050     /* ignore error and errp for bdrv_reopen, because we want to propagate
1051      * the original error */
1052     bdrv_reopen(base, orig_base_flags, NULL);
1053     return;
1054 }
1055