1/*
2 * Live block commit
3 *
4 * Copyright Red Hat, Inc. 2012
5 *
6 * Authors:
7 * Jeff Cody <jcody@redhat.com>
8 * Based on stream.c by Stefan Hajnoczi
9 *
10 * This work is licensed under the terms of the GNU LGPL, version 2 or later.
11 * See the COPYING.LIB file in the top-level directory.
12 *
13 */
14
15#include "qemu/osdep.h"
16#include "qemu/cutils.h"
17#include "trace.h"
18#include "block/block_int.h"
19#include "block/blockjob_int.h"
20#include "qapi/error.h"
21#include "qapi/qmp/qerror.h"
22#include "qemu/ratelimit.h"
23#include "sysemu/block-backend.h"
24
25enum {
26 /*
27 * Size of data buffer for populating the image file. This should be large
28 * enough to process multiple clusters in a single call, so that populating
29 * contiguous regions of the image is efficient.
30 */
31 COMMIT_BUFFER_SIZE = 512 * 1024, /* in bytes */
32};
33
34typedef struct CommitBlockJob {
35 BlockJob common;
36 BlockDriverState *commit_top_bs;
37 BlockBackend *top;
38 BlockBackend *base;
39 BlockDriverState *base_bs;
40 BlockdevOnError on_error;
41 bool base_read_only;
42 bool chain_frozen;
43 char *backing_file_str;
44} CommitBlockJob;
45
46static int coroutine_fn commit_populate(BlockBackend *bs, BlockBackend *base,
47 int64_t offset, uint64_t bytes,
48 void *buf)
49{
50 int ret = 0;
51
52 assert(bytes < SIZE_MAX);
53
54 ret = blk_co_pread(bs, offset, bytes, buf, 0);
55 if (ret < 0) {
56 return ret;
57 }
58
59 ret = blk_co_pwrite(base, offset, bytes, buf, 0);
60 if (ret < 0) {
61 return ret;
62 }
63
64 return 0;
65}
66
67static int commit_prepare(Job *job)
68{
69 CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
70
71 bdrv_unfreeze_backing_chain(s->commit_top_bs, s->base_bs);
72 s->chain_frozen = false;
73
74 /* Remove base node parent that still uses BLK_PERM_WRITE/RESIZE before
75 * the normal backing chain can be restored. */
76 blk_unref(s->base);
77 s->base = NULL;
78
79 /* FIXME: bdrv_drop_intermediate treats total failures and partial failures
80 * identically. Further work is needed to disambiguate these cases. */
81 return bdrv_drop_intermediate(s->commit_top_bs, s->base_bs,
82 s->backing_file_str);
83}
84
85static void commit_abort(Job *job)
86{
87 CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
88 BlockDriverState *top_bs = blk_bs(s->top);
89
90 if (s->chain_frozen) {
91 bdrv_unfreeze_backing_chain(s->commit_top_bs, s->base_bs);
92 }
93
94 /* Make sure commit_top_bs and top stay around until bdrv_replace_node() */
95 bdrv_ref(top_bs);
96 bdrv_ref(s->commit_top_bs);
97
98 if (s->base) {
99 blk_unref(s->base);
100 }
101
102 /* free the blockers on the intermediate nodes so that bdrv_replace_nodes
103 * can succeed */
104 block_job_remove_all_bdrv(&s->common);
105
106 /* If bdrv_drop_intermediate() failed (or was not invoked), remove the
107 * commit filter driver from the backing chain now. Do this as the final
108 * step so that the 'consistent read' permission can be granted.
109 *
110 * XXX Can (or should) we somehow keep 'consistent read' blocked even
111 * after the failed/cancelled commit job is gone? If we already wrote
112 * something to base, the intermediate images aren't valid any more. */
113 bdrv_replace_node(s->commit_top_bs, backing_bs(s->commit_top_bs),
114 &error_abort);
115
116 bdrv_unref(s->commit_top_bs);
117 bdrv_unref(top_bs);
118}
119
120static void commit_clean(Job *job)
121{
122 CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
123
124 /* restore base open flags here if appropriate (e.g., change the base back
125 * to r/o). These reopens do not need to be atomic, since we won't abort
126 * even on failure here */
127 if (s->base_read_only) {
128 bdrv_reopen_set_read_only(s->base_bs, true, NULL);
129 }
130
131 g_free(s->backing_file_str);
132 blk_unref(s->top);
133}
134
135static int coroutine_fn commit_run(Job *job, Error **errp)
136{
137 CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
138 int64_t offset;
139 uint64_t delay_ns = 0;
140 int ret = 0;
141 int64_t n = 0; /* bytes */
142 void *buf = NULL;
143 int bytes_written = 0;
144 int64_t len, base_len;
145
146 ret = len = blk_getlength(s->top);
147 if (len < 0) {
148 goto out;
149 }
150 job_progress_set_remaining(&s->common.job, len);
151
152 ret = base_len = blk_getlength(s->base);
153 if (base_len < 0) {
154 goto out;
155 }
156
157 if (base_len < len) {
158 ret = blk_truncate(s->base, len, PREALLOC_MODE_OFF, NULL);
159 if (ret) {
160 goto out;
161 }
162 }
163
164 buf = blk_blockalign(s->top, COMMIT_BUFFER_SIZE);
165
166 for (offset = 0; offset < len; offset += n) {
167 bool copy;
168
169 /* Note that even when no rate limit is applied we need to yield
170 * with no pending I/O here so that bdrv_drain_all() returns.
171 */
172 job_sleep_ns(&s->common.job, delay_ns);
173 if (job_is_cancelled(&s->common.job)) {
174 break;
175 }
176 /* Copy if allocated above the base */
177 ret = bdrv_is_allocated_above(blk_bs(s->top), blk_bs(s->base), false,
178 offset, COMMIT_BUFFER_SIZE, &n);
179 copy = (ret == 1);
180 trace_commit_one_iteration(s, offset, n, ret);
181 if (copy) {
182 ret = commit_populate(s->top, s->base, offset, n, buf);
183 bytes_written += n;
184 }
185 if (ret < 0) {
186 BlockErrorAction action =
187 block_job_error_action(&s->common, false, s->on_error, -ret);
188 if (action == BLOCK_ERROR_ACTION_REPORT) {
189 goto out;
190 } else {
191 n = 0;
192 continue;
193 }
194 }
195 /* Publish progress */
196 job_progress_update(&s->common.job, n);
197
198 if (copy) {
199 delay_ns = block_job_ratelimit_get_delay(&s->common, n);
200 } else {
201 delay_ns = 0;
202 }
203 }
204
205 ret = 0;
206
207out:
208 qemu_vfree(buf);
209
210 return ret;
211}
212
213static const BlockJobDriver commit_job_driver = {
214 .job_driver = {
215 .instance_size = sizeof(CommitBlockJob),
216 .job_type = JOB_TYPE_COMMIT,
217 .free = block_job_free,
218 .user_resume = block_job_user_resume,
219 .drain = block_job_drain,
220 .run = commit_run,
221 .prepare = commit_prepare,
222 .abort = commit_abort,
223 .clean = commit_clean
224 },
225};
226
227static int coroutine_fn bdrv_commit_top_preadv(BlockDriverState *bs,
228 uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags)
229{
230 return bdrv_co_preadv(bs->backing, offset, bytes, qiov, flags);
231}
232
233static void bdrv_commit_top_refresh_filename(BlockDriverState *bs)
234{
235 pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
236 bs->backing->bs->filename);
237}
238
239static void bdrv_commit_top_child_perm(BlockDriverState *bs, BdrvChild *c,
240 const BdrvChildRole *role,
241 BlockReopenQueue *reopen_queue,
242 uint64_t perm, uint64_t shared,
243 uint64_t *nperm, uint64_t *nshared)
244{
245 *nperm = 0;
246 *nshared = BLK_PERM_ALL;
247}
248
249/* Dummy node that provides consistent read to its users without requiring it
250 * from its backing file and that allows writes on the backing file chain. */
251static BlockDriver bdrv_commit_top = {
252 .format_name = "commit_top",
253 .bdrv_co_preadv = bdrv_commit_top_preadv,
254 .bdrv_co_block_status = bdrv_co_block_status_from_backing,
255 .bdrv_refresh_filename = bdrv_commit_top_refresh_filename,
256 .bdrv_child_perm = bdrv_commit_top_child_perm,
257};
258
259void commit_start(const char *job_id, BlockDriverState *bs,
260 BlockDriverState *base, BlockDriverState *top,
261 int creation_flags, int64_t speed,
262 BlockdevOnError on_error, const char *backing_file_str,
263 const char *filter_node_name, Error **errp)
264{
265 CommitBlockJob *s;
266 BlockDriverState *iter;
267 BlockDriverState *commit_top_bs = NULL;
268 Error *local_err = NULL;
269 int ret;
270
271 assert(top != bs);
272 if (top == base) {
273 error_setg(errp, "Invalid files for merge: top and base are the same");
274 return;
275 }
276
277 s = block_job_create(job_id, &commit_job_driver, NULL, bs, 0, BLK_PERM_ALL,
278 speed, creation_flags, NULL, NULL, errp);
279 if (!s) {
280 return;
281 }
282
283 /* convert base to r/w, if necessary */
284 s->base_read_only = bdrv_is_read_only(base);
285 if (s->base_read_only) {
286 if (bdrv_reopen_set_read_only(base, false, errp) != 0) {
287 goto fail;
288 }
289 }
290
291 /* Insert commit_top block node above top, so we can block consistent read
292 * on the backing chain below it */
293 commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, filter_node_name, 0,
294 errp);
295 if (commit_top_bs == NULL) {
296 goto fail;
297 }
298 if (!filter_node_name) {
299 commit_top_bs->implicit = true;
300 }
301
302 /* So that we can always drop this node */
303 commit_top_bs->never_freeze = true;
304
305 commit_top_bs->total_sectors = top->total_sectors;
306
307 bdrv_append(commit_top_bs, top, &local_err);
308 if (local_err) {
309 commit_top_bs = NULL;
310 error_propagate(errp, local_err);
311 goto fail;
312 }
313
314 s->commit_top_bs = commit_top_bs;
315
316 /* Block all nodes between top and base, because they will
317 * disappear from the chain after this operation. */
318 assert(bdrv_chain_contains(top, base));
319 for (iter = top; iter != base; iter = backing_bs(iter)) {
320 /* XXX BLK_PERM_WRITE needs to be allowed so we don't block ourselves
321 * at s->base (if writes are blocked for a node, they are also blocked
322 * for its backing file). The other options would be a second filter
323 * driver above s->base. */
324 ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
325 BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
326 errp);
327 if (ret < 0) {
328 goto fail;
329 }
330 }
331
332 if (bdrv_freeze_backing_chain(commit_top_bs, base, errp) < 0) {
333 goto fail;
334 }
335 s->chain_frozen = true;
336
337 ret = block_job_add_bdrv(&s->common, "base", base, 0, BLK_PERM_ALL, errp);
338 if (ret < 0) {
339 goto fail;
340 }
341
342 s->base = blk_new(s->common.job.aio_context,
343 BLK_PERM_CONSISTENT_READ
344 | BLK_PERM_WRITE
345 | BLK_PERM_RESIZE,
346 BLK_PERM_CONSISTENT_READ
347 | BLK_PERM_GRAPH_MOD
348 | BLK_PERM_WRITE_UNCHANGED);
349 ret = blk_insert_bs(s->base, base, errp);
350 if (ret < 0) {
351 goto fail;
352 }
353 blk_set_disable_request_queuing(s->base, true);
354 s->base_bs = base;
355
356 /* Required permissions are already taken with block_job_add_bdrv() */
357 s->top = blk_new(s->common.job.aio_context, 0, BLK_PERM_ALL);
358 ret = blk_insert_bs(s->top, top, errp);
359 if (ret < 0) {
360 goto fail;
361 }
362 blk_set_disable_request_queuing(s->top, true);
363
364 s->backing_file_str = g_strdup(backing_file_str);
365 s->on_error = on_error;
366
367 trace_commit_start(bs, base, top, s);
368 job_start(&s->common.job);
369 return;
370
371fail:
372 if (s->chain_frozen) {
373 bdrv_unfreeze_backing_chain(commit_top_bs, base);
374 }
375 if (s->base) {
376 blk_unref(s->base);
377 }
378 if (s->top) {
379 blk_unref(s->top);
380 }
381 if (s->base_read_only) {
382 bdrv_reopen_set_read_only(base, true, NULL);
383 }
384 job_early_fail(&s->common.job);
385 /* commit_top_bs has to be replaced after deleting the block job,
386 * otherwise this would fail because of lack of permissions. */
387 if (commit_top_bs) {
388 bdrv_replace_node(commit_top_bs, top, &error_abort);
389 }
390}
391
392
393#define COMMIT_BUF_SIZE (2048 * BDRV_SECTOR_SIZE)
394
395/* commit COW file into the raw image */
396int bdrv_commit(BlockDriverState *bs)
397{
398 BlockBackend *src, *backing;
399 BlockDriverState *backing_file_bs = NULL;
400 BlockDriverState *commit_top_bs = NULL;
401 BlockDriver *drv = bs->drv;
402 AioContext *ctx;
403 int64_t offset, length, backing_length;
404 int ro;
405 int64_t n;
406 int ret = 0;
407 uint8_t *buf = NULL;
408 Error *local_err = NULL;
409
410 if (!drv)
411 return -ENOMEDIUM;
412
413 if (!bs->backing) {
414 return -ENOTSUP;
415 }
416
417 if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_COMMIT_SOURCE, NULL) ||
418 bdrv_op_is_blocked(bs->backing->bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL)) {
419 return -EBUSY;
420 }
421
422 ro = bs->backing->bs->read_only;
423
424 if (ro) {
425 if (bdrv_reopen_set_read_only(bs->backing->bs, false, NULL)) {
426 return -EACCES;
427 }
428 }
429
430 ctx = bdrv_get_aio_context(bs);
431 src = blk_new(ctx, BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
432 backing = blk_new(ctx, BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL);
433
434 ret = blk_insert_bs(src, bs, &local_err);
435 if (ret < 0) {
436 error_report_err(local_err);
437 goto ro_cleanup;
438 }
439
440 /* Insert commit_top block node above backing, so we can write to it */
441 backing_file_bs = backing_bs(bs);
442
443 commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, NULL, BDRV_O_RDWR,
444 &local_err);
445 if (commit_top_bs == NULL) {
446 error_report_err(local_err);
447 goto ro_cleanup;
448 }
449
450 bdrv_set_backing_hd(commit_top_bs, backing_file_bs, &error_abort);
451 bdrv_set_backing_hd(bs, commit_top_bs, &error_abort);
452
453 ret = blk_insert_bs(backing, backing_file_bs, &local_err);
454 if (ret < 0) {
455 error_report_err(local_err);
456 goto ro_cleanup;
457 }
458
459 length = blk_getlength(src);
460 if (length < 0) {
461 ret = length;
462 goto ro_cleanup;
463 }
464
465 backing_length = blk_getlength(backing);
466 if (backing_length < 0) {
467 ret = backing_length;
468 goto ro_cleanup;
469 }
470
471 /* If our top snapshot is larger than the backing file image,
472 * grow the backing file image if possible. If not possible,
473 * we must return an error */
474 if (length > backing_length) {
475 ret = blk_truncate(backing, length, PREALLOC_MODE_OFF, &local_err);
476 if (ret < 0) {
477 error_report_err(local_err);
478 goto ro_cleanup;
479 }
480 }
481
482 /* blk_try_blockalign() for src will choose an alignment that works for
483 * backing as well, so no need to compare the alignment manually. */
484 buf = blk_try_blockalign(src, COMMIT_BUF_SIZE);
485 if (buf == NULL) {
486 ret = -ENOMEM;
487 goto ro_cleanup;
488 }
489
490 for (offset = 0; offset < length; offset += n) {
491 ret = bdrv_is_allocated(bs, offset, COMMIT_BUF_SIZE, &n);
492 if (ret < 0) {
493 goto ro_cleanup;
494 }
495 if (ret) {
496 ret = blk_pread(src, offset, buf, n);
497 if (ret < 0) {
498 goto ro_cleanup;
499 }
500
501 ret = blk_pwrite(backing, offset, buf, n, 0);
502 if (ret < 0) {
503 goto ro_cleanup;
504 }
505 }
506 }
507
508 if (drv->bdrv_make_empty) {
509 ret = drv->bdrv_make_empty(bs);
510 if (ret < 0) {
511 goto ro_cleanup;
512 }
513 blk_flush(src);
514 }
515
516 /*
517 * Make sure all data we wrote to the backing device is actually
518 * stable on disk.
519 */
520 blk_flush(backing);
521
522 ret = 0;
523ro_cleanup:
524 qemu_vfree(buf);
525
526 blk_unref(backing);
527 if (backing_file_bs) {
528 bdrv_set_backing_hd(bs, backing_file_bs, &error_abort);
529 }
530 bdrv_unref(commit_top_bs);
531 blk_unref(src);
532
533 if (ro) {
534 /* ignoring error return here */
535 bdrv_reopen_set_read_only(bs->backing->bs, true, NULL);
536 }
537
538 return ret;
539}
540