block: Byte-based bdrv_co_do_copy_on_readv()
[qemu.git] / block / mirror.c
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 DEFAULT_MIRROR_BUF_SIZE (10 << 20)
27
28 /* The mirroring buffer is a list of granularity-sized chunks.
29 * Free chunks are organized in a list.
30 */
31 typedef struct MirrorBuffer {
32 QSIMPLEQ_ENTRY(MirrorBuffer) next;
33 } MirrorBuffer;
34
35 typedef struct MirrorBlockJob {
36 BlockJob common;
37 RateLimit limit;
38 BlockBackend *target;
39 BlockDriverState *base;
40 /* The name of the graph node to replace */
41 char *replaces;
42 /* The BDS to replace */
43 BlockDriverState *to_replace;
44 /* Used to block operations on the drive-mirror-replace target */
45 Error *replace_blocker;
46 bool is_none_mode;
47 BlockdevOnError on_source_error, on_target_error;
48 bool synced;
49 bool should_complete;
50 int64_t granularity;
51 size_t buf_size;
52 int64_t bdev_length;
53 unsigned long *cow_bitmap;
54 BdrvDirtyBitmap *dirty_bitmap;
55 HBitmapIter hbi;
56 uint8_t *buf;
57 QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
58 int buf_free_count;
59
60 unsigned long *in_flight_bitmap;
61 int in_flight;
62 int sectors_in_flight;
63 int ret;
64 bool unmap;
65 bool waiting_for_io;
66 int target_cluster_sectors;
67 int max_iov;
68 } MirrorBlockJob;
69
70 typedef struct MirrorOp {
71 MirrorBlockJob *s;
72 QEMUIOVector qiov;
73 int64_t sector_num;
74 int nb_sectors;
75 } MirrorOp;
76
77 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
78 int error)
79 {
80 s->synced = false;
81 if (read) {
82 return block_job_error_action(&s->common, s->on_source_error,
83 true, error);
84 } else {
85 return block_job_error_action(&s->common, s->on_target_error,
86 false, error);
87 }
88 }
89
90 static void mirror_iteration_done(MirrorOp *op, int ret)
91 {
92 MirrorBlockJob *s = op->s;
93 struct iovec *iov;
94 int64_t chunk_num;
95 int i, nb_chunks, sectors_per_chunk;
96
97 trace_mirror_iteration_done(s, op->sector_num, op->nb_sectors, ret);
98
99 s->in_flight--;
100 s->sectors_in_flight -= op->nb_sectors;
101 iov = op->qiov.iov;
102 for (i = 0; i < op->qiov.niov; i++) {
103 MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
104 QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
105 s->buf_free_count++;
106 }
107
108 sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
109 chunk_num = op->sector_num / sectors_per_chunk;
110 nb_chunks = DIV_ROUND_UP(op->nb_sectors, sectors_per_chunk);
111 bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
112 if (ret >= 0) {
113 if (s->cow_bitmap) {
114 bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
115 }
116 s->common.offset += (uint64_t)op->nb_sectors * BDRV_SECTOR_SIZE;
117 }
118
119 qemu_iovec_destroy(&op->qiov);
120 g_free(op);
121
122 if (s->waiting_for_io) {
123 qemu_coroutine_enter(s->common.co, NULL);
124 }
125 }
126
127 static void mirror_write_complete(void *opaque, int ret)
128 {
129 MirrorOp *op = opaque;
130 MirrorBlockJob *s = op->s;
131 if (ret < 0) {
132 BlockErrorAction action;
133
134 bdrv_set_dirty_bitmap(s->dirty_bitmap, op->sector_num, op->nb_sectors);
135 action = mirror_error_action(s, false, -ret);
136 if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
137 s->ret = ret;
138 }
139 }
140 mirror_iteration_done(op, ret);
141 }
142
143 static void mirror_read_complete(void *opaque, int ret)
144 {
145 MirrorOp *op = opaque;
146 MirrorBlockJob *s = op->s;
147 if (ret < 0) {
148 BlockErrorAction action;
149
150 bdrv_set_dirty_bitmap(s->dirty_bitmap, op->sector_num, op->nb_sectors);
151 action = mirror_error_action(s, true, -ret);
152 if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
153 s->ret = ret;
154 }
155
156 mirror_iteration_done(op, ret);
157 return;
158 }
159 blk_aio_pwritev(s->target, op->sector_num * BDRV_SECTOR_SIZE, &op->qiov,
160 0, mirror_write_complete, op);
161 }
162
163 static inline void mirror_clip_sectors(MirrorBlockJob *s,
164 int64_t sector_num,
165 int *nb_sectors)
166 {
167 *nb_sectors = MIN(*nb_sectors,
168 s->bdev_length / BDRV_SECTOR_SIZE - sector_num);
169 }
170
171 /* Round sector_num and/or nb_sectors to target cluster if COW is needed, and
172 * return the offset of the adjusted tail sector against original. */
173 static int mirror_cow_align(MirrorBlockJob *s,
174 int64_t *sector_num,
175 int *nb_sectors)
176 {
177 bool need_cow;
178 int ret = 0;
179 int chunk_sectors = s->granularity >> BDRV_SECTOR_BITS;
180 int64_t align_sector_num = *sector_num;
181 int align_nb_sectors = *nb_sectors;
182 int max_sectors = chunk_sectors * s->max_iov;
183
184 need_cow = !test_bit(*sector_num / chunk_sectors, s->cow_bitmap);
185 need_cow |= !test_bit((*sector_num + *nb_sectors - 1) / chunk_sectors,
186 s->cow_bitmap);
187 if (need_cow) {
188 bdrv_round_sectors_to_clusters(blk_bs(s->target), *sector_num,
189 *nb_sectors, &align_sector_num,
190 &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 BlockBackend *source = s->common.blk;
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 blk_aio_preadv(source, sector_num * BDRV_SECTOR_SIZE, &op->qiov, 0,
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 blk_aio_discard(s->target, sector_num, op->nb_sectors,
300 mirror_write_complete, op);
301 } else {
302 blk_aio_pwrite_zeroes(s->target, sector_num * BDRV_SECTOR_SIZE,
303 op->nb_sectors * BDRV_SECTOR_SIZE,
304 s->unmap ? BDRV_REQ_MAY_UNMAP : 0,
305 mirror_write_complete, op);
306 }
307 }
308
309 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
310 {
311 BlockDriverState *source = blk_bs(s->common.blk);
312 int64_t sector_num, first_chunk;
313 uint64_t delay_ns = 0;
314 /* At least the first dirty chunk is mirrored in one iteration. */
315 int nb_chunks = 1;
316 int64_t end = s->bdev_length / BDRV_SECTOR_SIZE;
317 int sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
318
319 sector_num = hbitmap_iter_next(&s->hbi);
320 if (sector_num < 0) {
321 bdrv_dirty_iter_init(s->dirty_bitmap, &s->hbi);
322 sector_num = hbitmap_iter_next(&s->hbi);
323 trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
324 assert(sector_num >= 0);
325 }
326
327 first_chunk = sector_num / sectors_per_chunk;
328 while (test_bit(first_chunk, s->in_flight_bitmap)) {
329 trace_mirror_yield_in_flight(s, first_chunk, s->in_flight);
330 mirror_wait_for_io(s);
331 }
332
333 /* Find the number of consective dirty chunks following the first dirty
334 * one, and wait for in flight requests in them. */
335 while (nb_chunks * sectors_per_chunk < (s->buf_size >> BDRV_SECTOR_BITS)) {
336 int64_t hbitmap_next;
337 int64_t next_sector = sector_num + nb_chunks * sectors_per_chunk;
338 int64_t next_chunk = next_sector / sectors_per_chunk;
339 if (next_sector >= end ||
340 !bdrv_get_dirty(source, s->dirty_bitmap, next_sector)) {
341 break;
342 }
343 if (test_bit(next_chunk, s->in_flight_bitmap)) {
344 break;
345 }
346
347 hbitmap_next = hbitmap_iter_next(&s->hbi);
348 if (hbitmap_next > next_sector || hbitmap_next < 0) {
349 /* The bitmap iterator's cache is stale, refresh it */
350 bdrv_set_dirty_iter(&s->hbi, next_sector);
351 hbitmap_next = hbitmap_iter_next(&s->hbi);
352 }
353 assert(hbitmap_next == next_sector);
354 nb_chunks++;
355 }
356
357 /* Clear dirty bits before querying the block status, because
358 * calling bdrv_get_block_status_above could yield - if some blocks are
359 * marked dirty in this window, we need to know.
360 */
361 bdrv_reset_dirty_bitmap(s->dirty_bitmap, sector_num,
362 nb_chunks * sectors_per_chunk);
363 bitmap_set(s->in_flight_bitmap, sector_num / sectors_per_chunk, nb_chunks);
364 while (nb_chunks > 0 && sector_num < end) {
365 int ret;
366 int io_sectors;
367 BlockDriverState *file;
368 enum MirrorMethod {
369 MIRROR_METHOD_COPY,
370 MIRROR_METHOD_ZERO,
371 MIRROR_METHOD_DISCARD
372 } mirror_method = MIRROR_METHOD_COPY;
373
374 assert(!(sector_num % sectors_per_chunk));
375 ret = bdrv_get_block_status_above(source, NULL, sector_num,
376 nb_chunks * sectors_per_chunk,
377 &io_sectors, &file);
378 if (ret < 0) {
379 io_sectors = nb_chunks * sectors_per_chunk;
380 }
381
382 io_sectors -= io_sectors % sectors_per_chunk;
383 if (io_sectors < sectors_per_chunk) {
384 io_sectors = sectors_per_chunk;
385 } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
386 int64_t target_sector_num;
387 int target_nb_sectors;
388 bdrv_round_sectors_to_clusters(blk_bs(s->target), sector_num,
389 io_sectors, &target_sector_num,
390 &target_nb_sectors);
391 if (target_sector_num == sector_num &&
392 target_nb_sectors == io_sectors) {
393 mirror_method = ret & BDRV_BLOCK_ZERO ?
394 MIRROR_METHOD_ZERO :
395 MIRROR_METHOD_DISCARD;
396 }
397 }
398
399 mirror_clip_sectors(s, sector_num, &io_sectors);
400 switch (mirror_method) {
401 case MIRROR_METHOD_COPY:
402 io_sectors = mirror_do_read(s, sector_num, io_sectors);
403 break;
404 case MIRROR_METHOD_ZERO:
405 mirror_do_zero_or_discard(s, sector_num, io_sectors, false);
406 break;
407 case MIRROR_METHOD_DISCARD:
408 mirror_do_zero_or_discard(s, sector_num, io_sectors, true);
409 break;
410 default:
411 abort();
412 }
413 assert(io_sectors);
414 sector_num += io_sectors;
415 nb_chunks -= DIV_ROUND_UP(io_sectors, sectors_per_chunk);
416 delay_ns += ratelimit_calculate_delay(&s->limit, io_sectors);
417 }
418 return delay_ns;
419 }
420
421 static void mirror_free_init(MirrorBlockJob *s)
422 {
423 int granularity = s->granularity;
424 size_t buf_size = s->buf_size;
425 uint8_t *buf = s->buf;
426
427 assert(s->buf_free_count == 0);
428 QSIMPLEQ_INIT(&s->buf_free);
429 while (buf_size != 0) {
430 MirrorBuffer *cur = (MirrorBuffer *)buf;
431 QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
432 s->buf_free_count++;
433 buf_size -= granularity;
434 buf += granularity;
435 }
436 }
437
438 static void mirror_drain(MirrorBlockJob *s)
439 {
440 while (s->in_flight > 0) {
441 mirror_wait_for_io(s);
442 }
443 }
444
445 typedef struct {
446 int ret;
447 } MirrorExitData;
448
449 static void mirror_exit(BlockJob *job, void *opaque)
450 {
451 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
452 MirrorExitData *data = opaque;
453 AioContext *replace_aio_context = NULL;
454 BlockDriverState *src = blk_bs(s->common.blk);
455 BlockDriverState *target_bs = blk_bs(s->target);
456
457 /* Make sure that the source BDS doesn't go away before we called
458 * block_job_completed(). */
459 bdrv_ref(src);
460
461 if (s->to_replace) {
462 replace_aio_context = bdrv_get_aio_context(s->to_replace);
463 aio_context_acquire(replace_aio_context);
464 }
465
466 if (s->should_complete && data->ret == 0) {
467 BlockDriverState *to_replace = src;
468 if (s->to_replace) {
469 to_replace = s->to_replace;
470 }
471
472 if (bdrv_get_flags(target_bs) != bdrv_get_flags(to_replace)) {
473 bdrv_reopen(target_bs, bdrv_get_flags(to_replace), NULL);
474 }
475
476 /* The mirror job has no requests in flight any more, but we need to
477 * drain potential other users of the BDS before changing the graph. */
478 bdrv_drained_begin(target_bs);
479 bdrv_replace_in_backing_chain(to_replace, target_bs);
480 bdrv_drained_end(target_bs);
481
482 /* We just changed the BDS the job BB refers to */
483 blk_remove_bs(job->blk);
484 blk_insert_bs(job->blk, src);
485 }
486 if (s->to_replace) {
487 bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
488 error_free(s->replace_blocker);
489 bdrv_unref(s->to_replace);
490 }
491 if (replace_aio_context) {
492 aio_context_release(replace_aio_context);
493 }
494 g_free(s->replaces);
495 bdrv_op_unblock_all(target_bs, s->common.blocker);
496 blk_unref(s->target);
497 block_job_completed(&s->common, data->ret);
498 g_free(data);
499 bdrv_drained_end(src);
500 if (qemu_get_aio_context() == bdrv_get_aio_context(src)) {
501 aio_enable_external(iohandler_get_aio_context());
502 }
503 bdrv_unref(src);
504 }
505
506 static void coroutine_fn mirror_run(void *opaque)
507 {
508 MirrorBlockJob *s = opaque;
509 MirrorExitData *data;
510 BlockDriverState *bs = blk_bs(s->common.blk);
511 BlockDriverState *target_bs = blk_bs(s->target);
512 int64_t sector_num, end, length;
513 uint64_t last_pause_ns;
514 BlockDriverInfo bdi;
515 char backing_filename[2]; /* we only need 2 characters because we are only
516 checking for a NULL string */
517 int ret = 0;
518 int n;
519 int target_cluster_size = BDRV_SECTOR_SIZE;
520
521 if (block_job_is_cancelled(&s->common)) {
522 goto immediate_exit;
523 }
524
525 s->bdev_length = bdrv_getlength(bs);
526 if (s->bdev_length < 0) {
527 ret = s->bdev_length;
528 goto immediate_exit;
529 } else if (s->bdev_length == 0) {
530 /* Report BLOCK_JOB_READY and wait for complete. */
531 block_job_event_ready(&s->common);
532 s->synced = true;
533 while (!block_job_is_cancelled(&s->common) && !s->should_complete) {
534 block_job_yield(&s->common);
535 }
536 s->common.cancelled = false;
537 goto immediate_exit;
538 }
539
540 length = DIV_ROUND_UP(s->bdev_length, s->granularity);
541 s->in_flight_bitmap = bitmap_new(length);
542
543 /* If we have no backing file yet in the destination, we cannot let
544 * the destination do COW. Instead, we copy sectors around the
545 * dirty data if needed. We need a bitmap to do that.
546 */
547 bdrv_get_backing_filename(target_bs, backing_filename,
548 sizeof(backing_filename));
549 if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) {
550 target_cluster_size = bdi.cluster_size;
551 }
552 if (backing_filename[0] && !target_bs->backing
553 && s->granularity < target_cluster_size) {
554 s->buf_size = MAX(s->buf_size, target_cluster_size);
555 s->cow_bitmap = bitmap_new(length);
556 }
557 s->target_cluster_sectors = target_cluster_size >> BDRV_SECTOR_BITS;
558 s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
559
560 end = s->bdev_length / BDRV_SECTOR_SIZE;
561 s->buf = qemu_try_blockalign(bs, s->buf_size);
562 if (s->buf == NULL) {
563 ret = -ENOMEM;
564 goto immediate_exit;
565 }
566
567 mirror_free_init(s);
568
569 last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
570 if (!s->is_none_mode) {
571 /* First part, loop on the sectors and initialize the dirty bitmap. */
572 BlockDriverState *base = s->base;
573 bool mark_all_dirty = s->base == NULL && !bdrv_has_zero_init(target_bs);
574
575 for (sector_num = 0; sector_num < end; ) {
576 /* Just to make sure we are not exceeding int limit. */
577 int nb_sectors = MIN(INT_MAX >> BDRV_SECTOR_BITS,
578 end - sector_num);
579 int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
580
581 if (now - last_pause_ns > SLICE_TIME) {
582 last_pause_ns = now;
583 block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, 0);
584 }
585
586 if (block_job_is_cancelled(&s->common)) {
587 goto immediate_exit;
588 }
589
590 ret = bdrv_is_allocated_above(bs, base, sector_num, nb_sectors, &n);
591
592 if (ret < 0) {
593 goto immediate_exit;
594 }
595
596 assert(n > 0);
597 if (ret == 1 || mark_all_dirty) {
598 bdrv_set_dirty_bitmap(s->dirty_bitmap, sector_num, n);
599 }
600 sector_num += n;
601 }
602 }
603
604 bdrv_dirty_iter_init(s->dirty_bitmap, &s->hbi);
605 for (;;) {
606 uint64_t delay_ns = 0;
607 int64_t cnt;
608 bool should_complete;
609
610 if (s->ret < 0) {
611 ret = s->ret;
612 goto immediate_exit;
613 }
614
615 cnt = bdrv_get_dirty_count(s->dirty_bitmap);
616 /* s->common.offset contains the number of bytes already processed so
617 * far, cnt is the number of dirty sectors remaining and
618 * s->sectors_in_flight is the number of sectors currently being
619 * processed; together those are the current total operation length */
620 s->common.len = s->common.offset +
621 (cnt + s->sectors_in_flight) * BDRV_SECTOR_SIZE;
622
623 /* Note that even when no rate limit is applied we need to yield
624 * periodically with no pending I/O so that bdrv_drain_all() returns.
625 * We do so every SLICE_TIME nanoseconds, or when there is an error,
626 * or when the source is clean, whichever comes first.
627 */
628 if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - last_pause_ns < SLICE_TIME &&
629 s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
630 if (s->in_flight == MAX_IN_FLIGHT || s->buf_free_count == 0 ||
631 (cnt == 0 && s->in_flight > 0)) {
632 trace_mirror_yield(s, s->in_flight, s->buf_free_count, cnt);
633 mirror_wait_for_io(s);
634 continue;
635 } else if (cnt != 0) {
636 delay_ns = mirror_iteration(s);
637 }
638 }
639
640 should_complete = false;
641 if (s->in_flight == 0 && cnt == 0) {
642 trace_mirror_before_flush(s);
643 ret = blk_flush(s->target);
644 if (ret < 0) {
645 if (mirror_error_action(s, false, -ret) ==
646 BLOCK_ERROR_ACTION_REPORT) {
647 goto immediate_exit;
648 }
649 } else {
650 /* We're out of the streaming phase. From now on, if the job
651 * is cancelled we will actually complete all pending I/O and
652 * report completion. This way, block-job-cancel will leave
653 * the target in a consistent state.
654 */
655 if (!s->synced) {
656 block_job_event_ready(&s->common);
657 s->synced = true;
658 }
659
660 should_complete = s->should_complete ||
661 block_job_is_cancelled(&s->common);
662 cnt = bdrv_get_dirty_count(s->dirty_bitmap);
663 }
664 }
665
666 if (cnt == 0 && should_complete) {
667 /* The dirty bitmap is not updated while operations are pending.
668 * If we're about to exit, wait for pending operations before
669 * calling bdrv_get_dirty_count(bs), or we may exit while the
670 * source has dirty data to copy!
671 *
672 * Note that I/O can be submitted by the guest while
673 * mirror_populate runs.
674 */
675 trace_mirror_before_drain(s, cnt);
676 bdrv_co_drain(bs);
677 cnt = bdrv_get_dirty_count(s->dirty_bitmap);
678 }
679
680 ret = 0;
681 trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
682 if (!s->synced) {
683 block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
684 if (block_job_is_cancelled(&s->common)) {
685 break;
686 }
687 } else if (!should_complete) {
688 delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0);
689 block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
690 } else if (cnt == 0) {
691 /* The two disks are in sync. Exit and report successful
692 * completion.
693 */
694 assert(QLIST_EMPTY(&bs->tracked_requests));
695 s->common.cancelled = false;
696 break;
697 }
698 last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
699 }
700
701 immediate_exit:
702 if (s->in_flight > 0) {
703 /* We get here only if something went wrong. Either the job failed,
704 * or it was cancelled prematurely so that we do not guarantee that
705 * the target is a copy of the source.
706 */
707 assert(ret < 0 || (!s->synced && block_job_is_cancelled(&s->common)));
708 mirror_drain(s);
709 }
710
711 assert(s->in_flight == 0);
712 qemu_vfree(s->buf);
713 g_free(s->cow_bitmap);
714 g_free(s->in_flight_bitmap);
715 bdrv_release_dirty_bitmap(bs, s->dirty_bitmap);
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(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_complete(BlockJob *job, Error **errp)
743 {
744 MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
745 Error *local_err = NULL;
746 int ret;
747
748 ret = bdrv_open_backing_file(blk_bs(s->target), NULL, "backing",
749 &local_err);
750 if (ret < 0) {
751 error_propagate(errp, local_err);
752 return;
753 }
754 if (!s->synced) {
755 error_setg(errp, QERR_BLOCK_JOB_NOT_READY, job->id);
756 return;
757 }
758
759 /* check the target bs is not blocked and block all operations on it */
760 if (s->replaces) {
761 AioContext *replace_aio_context;
762
763 s->to_replace = bdrv_find_node(s->replaces);
764 if (!s->to_replace) {
765 error_setg(errp, "Node name '%s' not found", s->replaces);
766 return;
767 }
768
769 replace_aio_context = bdrv_get_aio_context(s->to_replace);
770 aio_context_acquire(replace_aio_context);
771
772 error_setg(&s->replace_blocker,
773 "block device is in use by block-job-complete");
774 bdrv_op_block_all(s->to_replace, s->replace_blocker);
775 bdrv_ref(s->to_replace);
776
777 aio_context_release(replace_aio_context);
778 }
779
780 s->should_complete = true;
781 block_job_enter(&s->common);
782 }
783
784 static const BlockJobDriver mirror_job_driver = {
785 .instance_size = sizeof(MirrorBlockJob),
786 .job_type = BLOCK_JOB_TYPE_MIRROR,
787 .set_speed = mirror_set_speed,
788 .complete = mirror_complete,
789 };
790
791 static const BlockJobDriver commit_active_job_driver = {
792 .instance_size = sizeof(MirrorBlockJob),
793 .job_type = BLOCK_JOB_TYPE_COMMIT,
794 .set_speed = mirror_set_speed,
795 .complete = mirror_complete,
796 };
797
798 static void mirror_start_job(BlockDriverState *bs, BlockDriverState *target,
799 const char *replaces,
800 int64_t speed, uint32_t granularity,
801 int64_t buf_size,
802 BlockdevOnError on_source_error,
803 BlockdevOnError on_target_error,
804 bool unmap,
805 BlockCompletionFunc *cb,
806 void *opaque, Error **errp,
807 const BlockJobDriver *driver,
808 bool is_none_mode, BlockDriverState *base)
809 {
810 MirrorBlockJob *s;
811
812 if (granularity == 0) {
813 granularity = bdrv_get_default_bitmap_granularity(target);
814 }
815
816 assert ((granularity & (granularity - 1)) == 0);
817
818 if (buf_size < 0) {
819 error_setg(errp, "Invalid parameter 'buf-size'");
820 return;
821 }
822
823 if (buf_size == 0) {
824 buf_size = DEFAULT_MIRROR_BUF_SIZE;
825 }
826
827 s = block_job_create(driver, bs, speed, cb, opaque, errp);
828 if (!s) {
829 return;
830 }
831
832 s->target = blk_new();
833 blk_insert_bs(s->target, target);
834
835 s->replaces = g_strdup(replaces);
836 s->on_source_error = on_source_error;
837 s->on_target_error = on_target_error;
838 s->is_none_mode = is_none_mode;
839 s->base = base;
840 s->granularity = granularity;
841 s->buf_size = ROUND_UP(buf_size, granularity);
842 s->unmap = unmap;
843
844 s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
845 if (!s->dirty_bitmap) {
846 g_free(s->replaces);
847 blk_unref(s->target);
848 block_job_unref(&s->common);
849 return;
850 }
851
852 bdrv_op_block_all(target, s->common.blocker);
853
854 s->common.co = qemu_coroutine_create(mirror_run);
855 trace_mirror_start(bs, s, s->common.co, opaque);
856 qemu_coroutine_enter(s->common.co, s);
857 }
858
859 void mirror_start(BlockDriverState *bs, BlockDriverState *target,
860 const char *replaces,
861 int64_t speed, uint32_t granularity, int64_t buf_size,
862 MirrorSyncMode mode, BlockdevOnError on_source_error,
863 BlockdevOnError on_target_error,
864 bool unmap,
865 BlockCompletionFunc *cb,
866 void *opaque, Error **errp)
867 {
868 bool is_none_mode;
869 BlockDriverState *base;
870
871 if (mode == MIRROR_SYNC_MODE_INCREMENTAL) {
872 error_setg(errp, "Sync mode 'incremental' not supported");
873 return;
874 }
875 is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
876 base = mode == MIRROR_SYNC_MODE_TOP ? backing_bs(bs) : NULL;
877 mirror_start_job(bs, target, replaces,
878 speed, granularity, buf_size,
879 on_source_error, on_target_error, unmap, cb, opaque, errp,
880 &mirror_job_driver, is_none_mode, base);
881 }
882
883 void commit_active_start(BlockDriverState *bs, BlockDriverState *base,
884 int64_t speed,
885 BlockdevOnError on_error,
886 BlockCompletionFunc *cb,
887 void *opaque, Error **errp)
888 {
889 int64_t length, base_length;
890 int orig_base_flags;
891 int ret;
892 Error *local_err = NULL;
893
894 orig_base_flags = bdrv_get_flags(base);
895
896 if (bdrv_reopen(base, bs->open_flags, errp)) {
897 return;
898 }
899
900 length = bdrv_getlength(bs);
901 if (length < 0) {
902 error_setg_errno(errp, -length,
903 "Unable to determine length of %s", bs->filename);
904 goto error_restore_flags;
905 }
906
907 base_length = bdrv_getlength(base);
908 if (base_length < 0) {
909 error_setg_errno(errp, -base_length,
910 "Unable to determine length of %s", base->filename);
911 goto error_restore_flags;
912 }
913
914 if (length > base_length) {
915 ret = bdrv_truncate(base, length);
916 if (ret < 0) {
917 error_setg_errno(errp, -ret,
918 "Top image %s is larger than base image %s, and "
919 "resize of base image failed",
920 bs->filename, base->filename);
921 goto error_restore_flags;
922 }
923 }
924
925 mirror_start_job(bs, base, NULL, speed, 0, 0,
926 on_error, on_error, false, cb, opaque, &local_err,
927 &commit_active_job_driver, false, base);
928 if (local_err) {
929 error_propagate(errp, local_err);
930 goto error_restore_flags;
931 }
932
933 return;
934
935 error_restore_flags:
936 /* ignore error and errp for bdrv_reopen, because we want to propagate
937 * the original error */
938 bdrv_reopen(base, orig_base_flags, NULL);
939 return;
940 }