]> git.proxmox.com Git - mirror_qemu.git/blob - block/stream.c
block: Mark bdrv_filter_or_cow_bs() and callers GRAPH_RDLOCK
[mirror_qemu.git] / block / stream.c
1 /*
2 * Image streaming
3 *
4 * Copyright IBM, Corp. 2011
5 *
6 * Authors:
7 * Stefan Hajnoczi <stefanha@linux.vnet.ibm.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/block_int.h"
17 #include "block/blockjob_int.h"
18 #include "qapi/error.h"
19 #include "qapi/qmp/qdict.h"
20 #include "qemu/ratelimit.h"
21 #include "sysemu/block-backend.h"
22 #include "block/copy-on-read.h"
23
24 enum {
25 /*
26 * Maximum chunk size to feed to copy-on-read. This should be
27 * large enough to process multiple clusters in a single call, so
28 * that populating contiguous regions of the image is efficient.
29 */
30 STREAM_CHUNK = 512 * 1024, /* in bytes */
31 };
32
33 typedef struct StreamBlockJob {
34 BlockJob common;
35 BlockBackend *blk;
36 BlockDriverState *base_overlay; /* COW overlay (stream from this) */
37 BlockDriverState *above_base; /* Node directly above the base */
38 BlockDriverState *cor_filter_bs;
39 BlockDriverState *target_bs;
40 BlockdevOnError on_error;
41 char *backing_file_str;
42 bool bs_read_only;
43 } StreamBlockJob;
44
45 static int coroutine_fn stream_populate(BlockBackend *blk,
46 int64_t offset, uint64_t bytes)
47 {
48 assert(bytes < SIZE_MAX);
49
50 return blk_co_preadv(blk, offset, bytes, NULL, BDRV_REQ_PREFETCH);
51 }
52
53 static int stream_prepare(Job *job)
54 {
55 StreamBlockJob *s = container_of(job, StreamBlockJob, common.job);
56 BlockDriverState *unfiltered_bs = bdrv_skip_filters(s->target_bs);
57 BlockDriverState *unfiltered_bs_cow = bdrv_cow_bs(unfiltered_bs);
58 BlockDriverState *base;
59 BlockDriverState *unfiltered_base;
60 Error *local_err = NULL;
61 int ret = 0;
62
63 GLOBAL_STATE_CODE();
64
65 /* We should drop filter at this point, as filter hold the backing chain */
66 bdrv_cor_filter_drop(s->cor_filter_bs);
67 s->cor_filter_bs = NULL;
68
69 /*
70 * bdrv_set_backing_hd() requires that the unfiltered_bs and the COW child
71 * of unfiltered_bs is drained. Drain already here and use
72 * bdrv_set_backing_hd_drained() instead because the polling during
73 * drained_begin() might change the graph, and if we do this only later, we
74 * may end up working with the wrong base node (or it might even have gone
75 * away by the time we want to use it).
76 */
77 bdrv_drained_begin(unfiltered_bs);
78 if (unfiltered_bs_cow) {
79 bdrv_ref(unfiltered_bs_cow);
80 bdrv_drained_begin(unfiltered_bs_cow);
81 }
82
83 bdrv_graph_rdlock_main_loop();
84 base = bdrv_filter_or_cow_bs(s->above_base);
85 unfiltered_base = bdrv_skip_filters(base);
86 bdrv_graph_rdunlock_main_loop();
87
88 if (bdrv_cow_child(unfiltered_bs)) {
89 const char *base_id = NULL, *base_fmt = NULL;
90 if (unfiltered_base) {
91 base_id = s->backing_file_str ?: unfiltered_base->filename;
92 if (unfiltered_base->drv) {
93 base_fmt = unfiltered_base->drv->format_name;
94 }
95 }
96
97 bdrv_set_backing_hd_drained(unfiltered_bs, base, &local_err);
98
99 /*
100 * This call will do I/O, so the graph can change again from here on.
101 * We have already completed the graph change, so we are not in danger
102 * of operating on the wrong node any more if this happens.
103 */
104 ret = bdrv_change_backing_file(unfiltered_bs, base_id, base_fmt, false);
105 if (local_err) {
106 error_report_err(local_err);
107 ret = -EPERM;
108 goto out;
109 }
110 }
111
112 out:
113 if (unfiltered_bs_cow) {
114 bdrv_drained_end(unfiltered_bs_cow);
115 bdrv_unref(unfiltered_bs_cow);
116 }
117 bdrv_drained_end(unfiltered_bs);
118 return ret;
119 }
120
121 static void stream_clean(Job *job)
122 {
123 StreamBlockJob *s = container_of(job, StreamBlockJob, common.job);
124
125 if (s->cor_filter_bs) {
126 bdrv_cor_filter_drop(s->cor_filter_bs);
127 s->cor_filter_bs = NULL;
128 }
129
130 blk_unref(s->blk);
131 s->blk = NULL;
132
133 /* Reopen the image back in read-only mode if necessary */
134 if (s->bs_read_only) {
135 /* Give up write permissions before making it read-only */
136 bdrv_reopen_set_read_only(s->target_bs, true, NULL);
137 }
138
139 g_free(s->backing_file_str);
140 }
141
142 static int coroutine_fn stream_run(Job *job, Error **errp)
143 {
144 StreamBlockJob *s = container_of(job, StreamBlockJob, common.job);
145 BlockDriverState *unfiltered_bs = bdrv_skip_filters(s->target_bs);
146 int64_t len;
147 int64_t offset = 0;
148 int error = 0;
149 int64_t n = 0; /* bytes */
150
151 if (unfiltered_bs == s->base_overlay) {
152 /* Nothing to stream */
153 return 0;
154 }
155
156 WITH_GRAPH_RDLOCK_GUARD() {
157 len = bdrv_co_getlength(s->target_bs);
158 if (len < 0) {
159 return len;
160 }
161 }
162 job_progress_set_remaining(&s->common.job, len);
163
164 for ( ; offset < len; offset += n) {
165 bool copy;
166 int ret;
167
168 /* Note that even when no rate limit is applied we need to yield
169 * with no pending I/O here so that bdrv_drain_all() returns.
170 */
171 block_job_ratelimit_sleep(&s->common);
172 if (job_is_cancelled(&s->common.job)) {
173 break;
174 }
175
176 copy = false;
177
178 WITH_GRAPH_RDLOCK_GUARD() {
179 ret = bdrv_co_is_allocated(unfiltered_bs, offset, STREAM_CHUNK, &n);
180 if (ret == 1) {
181 /* Allocated in the top, no need to copy. */
182 } else if (ret >= 0) {
183 /*
184 * Copy if allocated in the intermediate images. Limit to the
185 * known-unallocated area [offset, offset+n*BDRV_SECTOR_SIZE).
186 */
187 ret = bdrv_co_is_allocated_above(bdrv_cow_bs(unfiltered_bs),
188 s->base_overlay, true,
189 offset, n, &n);
190 /* Finish early if end of backing file has been reached */
191 if (ret == 0 && n == 0) {
192 n = len - offset;
193 }
194
195 copy = (ret > 0);
196 }
197 }
198 trace_stream_one_iteration(s, offset, n, ret);
199 if (copy) {
200 ret = stream_populate(s->blk, offset, n);
201 }
202 if (ret < 0) {
203 BlockErrorAction action =
204 block_job_error_action(&s->common, s->on_error, true, -ret);
205 if (action == BLOCK_ERROR_ACTION_STOP) {
206 n = 0;
207 continue;
208 }
209 if (error == 0) {
210 error = ret;
211 }
212 if (action == BLOCK_ERROR_ACTION_REPORT) {
213 break;
214 }
215 }
216
217 /* Publish progress */
218 job_progress_update(&s->common.job, n);
219 if (copy) {
220 block_job_ratelimit_processed_bytes(&s->common, n);
221 }
222 }
223
224 /* Do not remove the backing file if an error was there but ignored. */
225 return error;
226 }
227
228 static const BlockJobDriver stream_job_driver = {
229 .job_driver = {
230 .instance_size = sizeof(StreamBlockJob),
231 .job_type = JOB_TYPE_STREAM,
232 .free = block_job_free,
233 .run = stream_run,
234 .prepare = stream_prepare,
235 .clean = stream_clean,
236 .user_resume = block_job_user_resume,
237 },
238 };
239
240 void stream_start(const char *job_id, BlockDriverState *bs,
241 BlockDriverState *base, const char *backing_file_str,
242 BlockDriverState *bottom,
243 int creation_flags, int64_t speed,
244 BlockdevOnError on_error,
245 const char *filter_node_name,
246 Error **errp)
247 {
248 StreamBlockJob *s = NULL;
249 BlockDriverState *iter;
250 bool bs_read_only;
251 int basic_flags = BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED;
252 BlockDriverState *base_overlay;
253 BlockDriverState *cor_filter_bs = NULL;
254 BlockDriverState *above_base;
255 QDict *opts;
256 int ret;
257
258 GLOBAL_STATE_CODE();
259
260 assert(!(base && bottom));
261 assert(!(backing_file_str && bottom));
262
263 if (bottom) {
264 /*
265 * New simple interface. The code is written in terms of old interface
266 * with @base parameter (still, it doesn't freeze link to base, so in
267 * this mean old code is correct for new interface). So, for now, just
268 * emulate base_overlay and above_base. Still, when old interface
269 * finally removed, we should refactor code to use only "bottom", but
270 * not "*base*" things.
271 */
272 assert(!bottom->drv->is_filter);
273 base_overlay = above_base = bottom;
274 } else {
275 GRAPH_RDLOCK_GUARD_MAINLOOP();
276
277 base_overlay = bdrv_find_overlay(bs, base);
278 if (!base_overlay) {
279 error_setg(errp, "'%s' is not in the backing chain of '%s'",
280 base->node_name, bs->node_name);
281 return;
282 }
283
284 /*
285 * Find the node directly above @base. @base_overlay is a COW overlay,
286 * so it must have a bdrv_cow_child(), but it is the immediate overlay
287 * of @base, so between the two there can only be filters.
288 */
289 above_base = base_overlay;
290 if (bdrv_cow_bs(above_base) != base) {
291 above_base = bdrv_cow_bs(above_base);
292 while (bdrv_filter_bs(above_base) != base) {
293 above_base = bdrv_filter_bs(above_base);
294 }
295 }
296 }
297
298 /* Make sure that the image is opened in read-write mode */
299 bs_read_only = bdrv_is_read_only(bs);
300 if (bs_read_only) {
301 /* Hold the chain during reopen */
302 if (bdrv_freeze_backing_chain(bs, above_base, errp) < 0) {
303 return;
304 }
305
306 ret = bdrv_reopen_set_read_only(bs, false, errp);
307
308 /* failure, or cor-filter will hold the chain */
309 bdrv_unfreeze_backing_chain(bs, above_base);
310
311 if (ret < 0) {
312 return;
313 }
314 }
315
316 opts = qdict_new();
317
318 qdict_put_str(opts, "driver", "copy-on-read");
319 qdict_put_str(opts, "file", bdrv_get_node_name(bs));
320 /* Pass the base_overlay node name as 'bottom' to COR driver */
321 qdict_put_str(opts, "bottom", base_overlay->node_name);
322 if (filter_node_name) {
323 qdict_put_str(opts, "node-name", filter_node_name);
324 }
325
326 cor_filter_bs = bdrv_insert_node(bs, opts, BDRV_O_RDWR, errp);
327 if (!cor_filter_bs) {
328 goto fail;
329 }
330
331 if (!filter_node_name) {
332 cor_filter_bs->implicit = true;
333 }
334
335 s = block_job_create(job_id, &stream_job_driver, NULL, cor_filter_bs,
336 0, BLK_PERM_ALL,
337 speed, creation_flags, NULL, NULL, errp);
338 if (!s) {
339 goto fail;
340 }
341
342 s->blk = blk_new_with_bs(cor_filter_bs, BLK_PERM_CONSISTENT_READ,
343 basic_flags | BLK_PERM_WRITE, errp);
344 if (!s->blk) {
345 goto fail;
346 }
347 /*
348 * Disable request queuing in the BlockBackend to avoid deadlocks on drain:
349 * The job reports that it's busy until it reaches a pause point.
350 */
351 blk_set_disable_request_queuing(s->blk, true);
352 blk_set_allow_aio_context_change(s->blk, true);
353
354 /*
355 * Prevent concurrent jobs trying to modify the graph structure here, we
356 * already have our own plans. Also don't allow resize as the image size is
357 * queried only at the job start and then cached.
358 */
359 bdrv_graph_wrlock(bs);
360 if (block_job_add_bdrv(&s->common, "active node", bs, 0,
361 basic_flags | BLK_PERM_WRITE, errp)) {
362 bdrv_graph_wrunlock();
363 goto fail;
364 }
365
366 /* Block all intermediate nodes between bs and base, because they will
367 * disappear from the chain after this operation. The streaming job reads
368 * every block only once, assuming that it doesn't change, so forbid writes
369 * and resizes. Reassign the base node pointer because the backing BS of the
370 * bottom node might change after the call to bdrv_reopen_set_read_only()
371 * due to parallel block jobs running.
372 * above_base node might change after the call to
373 * bdrv_reopen_set_read_only() due to parallel block jobs running.
374 */
375 base = bdrv_filter_or_cow_bs(above_base);
376 for (iter = bdrv_filter_or_cow_bs(bs); iter != base;
377 iter = bdrv_filter_or_cow_bs(iter))
378 {
379 ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
380 basic_flags, errp);
381 if (ret < 0) {
382 bdrv_graph_wrunlock();
383 goto fail;
384 }
385 }
386 bdrv_graph_wrunlock();
387
388 s->base_overlay = base_overlay;
389 s->above_base = above_base;
390 s->backing_file_str = g_strdup(backing_file_str);
391 s->cor_filter_bs = cor_filter_bs;
392 s->target_bs = bs;
393 s->bs_read_only = bs_read_only;
394
395 s->on_error = on_error;
396 trace_stream_start(bs, base, s);
397 job_start(&s->common.job);
398 return;
399
400 fail:
401 if (s) {
402 job_early_fail(&s->common.job);
403 }
404 if (cor_filter_bs) {
405 bdrv_cor_filter_drop(cor_filter_bs);
406 }
407 if (bs_read_only) {
408 bdrv_reopen_set_read_only(bs, true, NULL);
409 }
410 }