xen: fix ioreq handling
[qemu/kevin.git] / block / backup.c
blobea387338495f145ddac3cc2899ed4c701ea60d11
1 /*
2 * QEMU backup
4 * Copyright (C) 2013 Proxmox Server Solutions
6 * Authors:
7 * Dietmar Maurer (dietmar@proxmox.com)
9 * This work is licensed under the terms of the GNU GPL, version 2 or later.
10 * See the COPYING file in the top-level directory.
14 #include "qemu/osdep.h"
16 #include "trace.h"
17 #include "block/block.h"
18 #include "block/block_int.h"
19 #include "block/blockjob_int.h"
20 #include "block/block_backup.h"
21 #include "qapi/error.h"
22 #include "qapi/qmp/qerror.h"
23 #include "qemu/ratelimit.h"
24 #include "qemu/cutils.h"
25 #include "sysemu/block-backend.h"
26 #include "qemu/bitmap.h"
28 #define BACKUP_CLUSTER_SIZE_DEFAULT (1 << 16)
29 #define SLICE_TIME 100000000ULL /* ns */
31 typedef struct BackupBlockJob {
32 BlockJob common;
33 BlockBackend *target;
34 /* bitmap for sync=incremental */
35 BdrvDirtyBitmap *sync_bitmap;
36 MirrorSyncMode sync_mode;
37 RateLimit limit;
38 BlockdevOnError on_source_error;
39 BlockdevOnError on_target_error;
40 CoRwlock flush_rwlock;
41 uint64_t sectors_read;
42 unsigned long *done_bitmap;
43 int64_t cluster_size;
44 bool compress;
45 NotifierWithReturn before_write;
46 QLIST_HEAD(, CowRequest) inflight_reqs;
47 } BackupBlockJob;
49 /* Size of a cluster in sectors, instead of bytes. */
50 static inline int64_t cluster_size_sectors(BackupBlockJob *job)
52 return job->cluster_size / BDRV_SECTOR_SIZE;
55 /* See if in-flight requests overlap and wait for them to complete */
56 static void coroutine_fn wait_for_overlapping_requests(BackupBlockJob *job,
57 int64_t start,
58 int64_t end)
60 CowRequest *req;
61 bool retry;
63 do {
64 retry = false;
65 QLIST_FOREACH(req, &job->inflight_reqs, list) {
66 if (end > req->start && start < req->end) {
67 qemu_co_queue_wait(&req->wait_queue);
68 retry = true;
69 break;
72 } while (retry);
75 /* Keep track of an in-flight request */
76 static void cow_request_begin(CowRequest *req, BackupBlockJob *job,
77 int64_t start, int64_t end)
79 req->start = start;
80 req->end = end;
81 qemu_co_queue_init(&req->wait_queue);
82 QLIST_INSERT_HEAD(&job->inflight_reqs, req, list);
85 /* Forget about a completed request */
86 static void cow_request_end(CowRequest *req)
88 QLIST_REMOVE(req, list);
89 qemu_co_queue_restart_all(&req->wait_queue);
92 static int coroutine_fn backup_do_cow(BackupBlockJob *job,
93 int64_t sector_num, int nb_sectors,
94 bool *error_is_read,
95 bool is_write_notifier)
97 BlockBackend *blk = job->common.blk;
98 CowRequest cow_request;
99 struct iovec iov;
100 QEMUIOVector bounce_qiov;
101 void *bounce_buffer = NULL;
102 int ret = 0;
103 int64_t sectors_per_cluster = cluster_size_sectors(job);
104 int64_t start, end;
105 int n;
107 qemu_co_rwlock_rdlock(&job->flush_rwlock);
109 start = sector_num / sectors_per_cluster;
110 end = DIV_ROUND_UP(sector_num + nb_sectors, sectors_per_cluster);
112 trace_backup_do_cow_enter(job, start, sector_num, nb_sectors);
114 wait_for_overlapping_requests(job, start, end);
115 cow_request_begin(&cow_request, job, start, end);
117 for (; start < end; start++) {
118 if (test_bit(start, job->done_bitmap)) {
119 trace_backup_do_cow_skip(job, start);
120 continue; /* already copied */
123 trace_backup_do_cow_process(job, start);
125 n = MIN(sectors_per_cluster,
126 job->common.len / BDRV_SECTOR_SIZE -
127 start * sectors_per_cluster);
129 if (!bounce_buffer) {
130 bounce_buffer = blk_blockalign(blk, job->cluster_size);
132 iov.iov_base = bounce_buffer;
133 iov.iov_len = n * BDRV_SECTOR_SIZE;
134 qemu_iovec_init_external(&bounce_qiov, &iov, 1);
136 ret = blk_co_preadv(blk, start * job->cluster_size,
137 bounce_qiov.size, &bounce_qiov,
138 is_write_notifier ? BDRV_REQ_NO_SERIALISING : 0);
139 if (ret < 0) {
140 trace_backup_do_cow_read_fail(job, start, ret);
141 if (error_is_read) {
142 *error_is_read = true;
144 goto out;
147 if (buffer_is_zero(iov.iov_base, iov.iov_len)) {
148 ret = blk_co_pwrite_zeroes(job->target, start * job->cluster_size,
149 bounce_qiov.size, BDRV_REQ_MAY_UNMAP);
150 } else {
151 ret = blk_co_pwritev(job->target, start * job->cluster_size,
152 bounce_qiov.size, &bounce_qiov,
153 job->compress ? BDRV_REQ_WRITE_COMPRESSED : 0);
155 if (ret < 0) {
156 trace_backup_do_cow_write_fail(job, start, ret);
157 if (error_is_read) {
158 *error_is_read = false;
160 goto out;
163 set_bit(start, job->done_bitmap);
165 /* Publish progress, guest I/O counts as progress too. Note that the
166 * offset field is an opaque progress value, it is not a disk offset.
168 job->sectors_read += n;
169 job->common.offset += n * BDRV_SECTOR_SIZE;
172 out:
173 if (bounce_buffer) {
174 qemu_vfree(bounce_buffer);
177 cow_request_end(&cow_request);
179 trace_backup_do_cow_return(job, sector_num, nb_sectors, ret);
181 qemu_co_rwlock_unlock(&job->flush_rwlock);
183 return ret;
186 static int coroutine_fn backup_before_write_notify(
187 NotifierWithReturn *notifier,
188 void *opaque)
190 BackupBlockJob *job = container_of(notifier, BackupBlockJob, before_write);
191 BdrvTrackedRequest *req = opaque;
192 int64_t sector_num = req->offset >> BDRV_SECTOR_BITS;
193 int nb_sectors = req->bytes >> BDRV_SECTOR_BITS;
195 assert(req->bs == blk_bs(job->common.blk));
196 assert((req->offset & (BDRV_SECTOR_SIZE - 1)) == 0);
197 assert((req->bytes & (BDRV_SECTOR_SIZE - 1)) == 0);
199 return backup_do_cow(job, sector_num, nb_sectors, NULL, true);
202 static void backup_set_speed(BlockJob *job, int64_t speed, Error **errp)
204 BackupBlockJob *s = container_of(job, BackupBlockJob, common);
206 if (speed < 0) {
207 error_setg(errp, QERR_INVALID_PARAMETER, "speed");
208 return;
210 ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
213 static void backup_cleanup_sync_bitmap(BackupBlockJob *job, int ret)
215 BdrvDirtyBitmap *bm;
216 BlockDriverState *bs = blk_bs(job->common.blk);
218 if (ret < 0 || block_job_is_cancelled(&job->common)) {
219 /* Merge the successor back into the parent, delete nothing. */
220 bm = bdrv_reclaim_dirty_bitmap(bs, job->sync_bitmap, NULL);
221 assert(bm);
222 } else {
223 /* Everything is fine, delete this bitmap and install the backup. */
224 bm = bdrv_dirty_bitmap_abdicate(bs, job->sync_bitmap, NULL);
225 assert(bm);
229 static void backup_commit(BlockJob *job)
231 BackupBlockJob *s = container_of(job, BackupBlockJob, common);
232 if (s->sync_bitmap) {
233 backup_cleanup_sync_bitmap(s, 0);
237 static void backup_abort(BlockJob *job)
239 BackupBlockJob *s = container_of(job, BackupBlockJob, common);
240 if (s->sync_bitmap) {
241 backup_cleanup_sync_bitmap(s, -1);
245 static void backup_clean(BlockJob *job)
247 BackupBlockJob *s = container_of(job, BackupBlockJob, common);
248 assert(s->target);
249 blk_unref(s->target);
250 s->target = NULL;
253 static void backup_attached_aio_context(BlockJob *job, AioContext *aio_context)
255 BackupBlockJob *s = container_of(job, BackupBlockJob, common);
257 blk_set_aio_context(s->target, aio_context);
260 void backup_do_checkpoint(BlockJob *job, Error **errp)
262 BackupBlockJob *backup_job = container_of(job, BackupBlockJob, common);
263 int64_t len;
265 assert(job->driver->job_type == BLOCK_JOB_TYPE_BACKUP);
267 if (backup_job->sync_mode != MIRROR_SYNC_MODE_NONE) {
268 error_setg(errp, "The backup job only supports block checkpoint in"
269 " sync=none mode");
270 return;
273 len = DIV_ROUND_UP(backup_job->common.len, backup_job->cluster_size);
274 bitmap_zero(backup_job->done_bitmap, len);
277 void backup_wait_for_overlapping_requests(BlockJob *job, int64_t sector_num,
278 int nb_sectors)
280 BackupBlockJob *backup_job = container_of(job, BackupBlockJob, common);
281 int64_t sectors_per_cluster = cluster_size_sectors(backup_job);
282 int64_t start, end;
284 assert(job->driver->job_type == BLOCK_JOB_TYPE_BACKUP);
286 start = sector_num / sectors_per_cluster;
287 end = DIV_ROUND_UP(sector_num + nb_sectors, sectors_per_cluster);
288 wait_for_overlapping_requests(backup_job, start, end);
291 void backup_cow_request_begin(CowRequest *req, BlockJob *job,
292 int64_t sector_num,
293 int nb_sectors)
295 BackupBlockJob *backup_job = container_of(job, BackupBlockJob, common);
296 int64_t sectors_per_cluster = cluster_size_sectors(backup_job);
297 int64_t start, end;
299 assert(job->driver->job_type == BLOCK_JOB_TYPE_BACKUP);
301 start = sector_num / sectors_per_cluster;
302 end = DIV_ROUND_UP(sector_num + nb_sectors, sectors_per_cluster);
303 cow_request_begin(req, backup_job, start, end);
306 void backup_cow_request_end(CowRequest *req)
308 cow_request_end(req);
311 static void backup_drain(BlockJob *job)
313 BackupBlockJob *s = container_of(job, BackupBlockJob, common);
315 /* Need to keep a reference in case blk_drain triggers execution
316 * of backup_complete...
318 if (s->target) {
319 BlockBackend *target = s->target;
320 blk_ref(target);
321 blk_drain(target);
322 blk_unref(target);
326 static BlockErrorAction backup_error_action(BackupBlockJob *job,
327 bool read, int error)
329 if (read) {
330 return block_job_error_action(&job->common, job->on_source_error,
331 true, error);
332 } else {
333 return block_job_error_action(&job->common, job->on_target_error,
334 false, error);
338 typedef struct {
339 int ret;
340 } BackupCompleteData;
342 static void backup_complete(BlockJob *job, void *opaque)
344 BackupCompleteData *data = opaque;
346 block_job_completed(job, data->ret);
347 g_free(data);
350 static bool coroutine_fn yield_and_check(BackupBlockJob *job)
352 if (block_job_is_cancelled(&job->common)) {
353 return true;
356 /* we need to yield so that bdrv_drain_all() returns.
357 * (without, VM does not reboot)
359 if (job->common.speed) {
360 uint64_t delay_ns = ratelimit_calculate_delay(&job->limit,
361 job->sectors_read);
362 job->sectors_read = 0;
363 block_job_sleep_ns(&job->common, QEMU_CLOCK_REALTIME, delay_ns);
364 } else {
365 block_job_sleep_ns(&job->common, QEMU_CLOCK_REALTIME, 0);
368 if (block_job_is_cancelled(&job->common)) {
369 return true;
372 return false;
375 static int coroutine_fn backup_run_incremental(BackupBlockJob *job)
377 bool error_is_read;
378 int ret = 0;
379 int clusters_per_iter;
380 uint32_t granularity;
381 int64_t sector;
382 int64_t cluster;
383 int64_t end;
384 int64_t last_cluster = -1;
385 int64_t sectors_per_cluster = cluster_size_sectors(job);
386 BdrvDirtyBitmapIter *dbi;
388 granularity = bdrv_dirty_bitmap_granularity(job->sync_bitmap);
389 clusters_per_iter = MAX((granularity / job->cluster_size), 1);
390 dbi = bdrv_dirty_iter_new(job->sync_bitmap, 0);
392 /* Find the next dirty sector(s) */
393 while ((sector = bdrv_dirty_iter_next(dbi)) != -1) {
394 cluster = sector / sectors_per_cluster;
396 /* Fake progress updates for any clusters we skipped */
397 if (cluster != last_cluster + 1) {
398 job->common.offset += ((cluster - last_cluster - 1) *
399 job->cluster_size);
402 for (end = cluster + clusters_per_iter; cluster < end; cluster++) {
403 do {
404 if (yield_and_check(job)) {
405 goto out;
407 ret = backup_do_cow(job, cluster * sectors_per_cluster,
408 sectors_per_cluster, &error_is_read,
409 false);
410 if ((ret < 0) &&
411 backup_error_action(job, error_is_read, -ret) ==
412 BLOCK_ERROR_ACTION_REPORT) {
413 goto out;
415 } while (ret < 0);
418 /* If the bitmap granularity is smaller than the backup granularity,
419 * we need to advance the iterator pointer to the next cluster. */
420 if (granularity < job->cluster_size) {
421 bdrv_set_dirty_iter(dbi, cluster * sectors_per_cluster);
424 last_cluster = cluster - 1;
427 /* Play some final catchup with the progress meter */
428 end = DIV_ROUND_UP(job->common.len, job->cluster_size);
429 if (last_cluster + 1 < end) {
430 job->common.offset += ((end - last_cluster - 1) * job->cluster_size);
433 out:
434 bdrv_dirty_iter_free(dbi);
435 return ret;
438 static void coroutine_fn backup_run(void *opaque)
440 BackupBlockJob *job = opaque;
441 BackupCompleteData *data;
442 BlockDriverState *bs = blk_bs(job->common.blk);
443 int64_t start, end;
444 int64_t sectors_per_cluster = cluster_size_sectors(job);
445 int ret = 0;
447 QLIST_INIT(&job->inflight_reqs);
448 qemu_co_rwlock_init(&job->flush_rwlock);
450 start = 0;
451 end = DIV_ROUND_UP(job->common.len, job->cluster_size);
453 job->done_bitmap = bitmap_new(end);
455 job->before_write.notify = backup_before_write_notify;
456 bdrv_add_before_write_notifier(bs, &job->before_write);
458 if (job->sync_mode == MIRROR_SYNC_MODE_NONE) {
459 while (!block_job_is_cancelled(&job->common)) {
460 /* Yield until the job is cancelled. We just let our before_write
461 * notify callback service CoW requests. */
462 block_job_yield(&job->common);
464 } else if (job->sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
465 ret = backup_run_incremental(job);
466 } else {
467 /* Both FULL and TOP SYNC_MODE's require copying.. */
468 for (; start < end; start++) {
469 bool error_is_read;
470 if (yield_and_check(job)) {
471 break;
474 if (job->sync_mode == MIRROR_SYNC_MODE_TOP) {
475 int i, n;
476 int alloced = 0;
478 /* Check to see if these blocks are already in the
479 * backing file. */
481 for (i = 0; i < sectors_per_cluster;) {
482 /* bdrv_is_allocated() only returns true/false based
483 * on the first set of sectors it comes across that
484 * are are all in the same state.
485 * For that reason we must verify each sector in the
486 * backup cluster length. We end up copying more than
487 * needed but at some point that is always the case. */
488 alloced =
489 bdrv_is_allocated(bs,
490 start * sectors_per_cluster + i,
491 sectors_per_cluster - i, &n);
492 i += n;
494 if (alloced == 1 || n == 0) {
495 break;
499 /* If the above loop never found any sectors that are in
500 * the topmost image, skip this backup. */
501 if (alloced == 0) {
502 continue;
505 /* FULL sync mode we copy the whole drive. */
506 ret = backup_do_cow(job, start * sectors_per_cluster,
507 sectors_per_cluster, &error_is_read, false);
508 if (ret < 0) {
509 /* Depending on error action, fail now or retry cluster */
510 BlockErrorAction action =
511 backup_error_action(job, error_is_read, -ret);
512 if (action == BLOCK_ERROR_ACTION_REPORT) {
513 break;
514 } else {
515 start--;
516 continue;
522 notifier_with_return_remove(&job->before_write);
524 /* wait until pending backup_do_cow() calls have completed */
525 qemu_co_rwlock_wrlock(&job->flush_rwlock);
526 qemu_co_rwlock_unlock(&job->flush_rwlock);
527 g_free(job->done_bitmap);
529 data = g_malloc(sizeof(*data));
530 data->ret = ret;
531 block_job_defer_to_main_loop(&job->common, backup_complete, data);
534 static const BlockJobDriver backup_job_driver = {
535 .instance_size = sizeof(BackupBlockJob),
536 .job_type = BLOCK_JOB_TYPE_BACKUP,
537 .start = backup_run,
538 .set_speed = backup_set_speed,
539 .commit = backup_commit,
540 .abort = backup_abort,
541 .clean = backup_clean,
542 .attached_aio_context = backup_attached_aio_context,
543 .drain = backup_drain,
546 BlockJob *backup_job_create(const char *job_id, BlockDriverState *bs,
547 BlockDriverState *target, int64_t speed,
548 MirrorSyncMode sync_mode, BdrvDirtyBitmap *sync_bitmap,
549 bool compress,
550 BlockdevOnError on_source_error,
551 BlockdevOnError on_target_error,
552 int creation_flags,
553 BlockCompletionFunc *cb, void *opaque,
554 BlockJobTxn *txn, Error **errp)
556 int64_t len;
557 BlockDriverInfo bdi;
558 BackupBlockJob *job = NULL;
559 int ret;
561 assert(bs);
562 assert(target);
564 if (bs == target) {
565 error_setg(errp, "Source and target cannot be the same");
566 return NULL;
569 if (!bdrv_is_inserted(bs)) {
570 error_setg(errp, "Device is not inserted: %s",
571 bdrv_get_device_name(bs));
572 return NULL;
575 if (!bdrv_is_inserted(target)) {
576 error_setg(errp, "Device is not inserted: %s",
577 bdrv_get_device_name(target));
578 return NULL;
581 if (compress && target->drv->bdrv_co_pwritev_compressed == NULL) {
582 error_setg(errp, "Compression is not supported for this drive %s",
583 bdrv_get_device_name(target));
584 return NULL;
587 if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_BACKUP_SOURCE, errp)) {
588 return NULL;
591 if (bdrv_op_is_blocked(target, BLOCK_OP_TYPE_BACKUP_TARGET, errp)) {
592 return NULL;
595 if (sync_mode == MIRROR_SYNC_MODE_INCREMENTAL) {
596 if (!sync_bitmap) {
597 error_setg(errp, "must provide a valid bitmap name for "
598 "\"incremental\" sync mode");
599 return NULL;
602 /* Create a new bitmap, and freeze/disable this one. */
603 if (bdrv_dirty_bitmap_create_successor(bs, sync_bitmap, errp) < 0) {
604 return NULL;
606 } else if (sync_bitmap) {
607 error_setg(errp,
608 "a sync_bitmap was provided to backup_run, "
609 "but received an incompatible sync_mode (%s)",
610 MirrorSyncMode_lookup[sync_mode]);
611 return NULL;
614 len = bdrv_getlength(bs);
615 if (len < 0) {
616 error_setg_errno(errp, -len, "unable to get length for '%s'",
617 bdrv_get_device_name(bs));
618 goto error;
621 job = block_job_create(job_id, &backup_job_driver, bs, speed,
622 creation_flags, cb, opaque, errp);
623 if (!job) {
624 goto error;
627 job->target = blk_new();
628 blk_insert_bs(job->target, target);
630 job->on_source_error = on_source_error;
631 job->on_target_error = on_target_error;
632 job->sync_mode = sync_mode;
633 job->sync_bitmap = sync_mode == MIRROR_SYNC_MODE_INCREMENTAL ?
634 sync_bitmap : NULL;
635 job->compress = compress;
637 /* If there is no backing file on the target, we cannot rely on COW if our
638 * backup cluster size is smaller than the target cluster size. Even for
639 * targets with a backing file, try to avoid COW if possible. */
640 ret = bdrv_get_info(target, &bdi);
641 if (ret < 0 && !target->backing) {
642 error_setg_errno(errp, -ret,
643 "Couldn't determine the cluster size of the target image, "
644 "which has no backing file");
645 error_append_hint(errp,
646 "Aborting, since this may create an unusable destination image\n");
647 goto error;
648 } else if (ret < 0 && target->backing) {
649 /* Not fatal; just trudge on ahead. */
650 job->cluster_size = BACKUP_CLUSTER_SIZE_DEFAULT;
651 } else {
652 job->cluster_size = MAX(BACKUP_CLUSTER_SIZE_DEFAULT, bdi.cluster_size);
655 block_job_add_bdrv(&job->common, target);
656 job->common.len = len;
657 block_job_txn_add_job(txn, &job->common);
659 return &job->common;
661 error:
662 if (sync_bitmap) {
663 bdrv_reclaim_dirty_bitmap(bs, sync_bitmap, NULL);
665 if (job) {
666 backup_clean(&job->common);
667 block_job_unref(&job->common);
670 return NULL;