xref: /qemu/block/stream.c (revision 83c2201fc47bd0dfa656bde7202bd0e2539d54a0)
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 "qobject/qdict.h"
20 #include "qemu/ratelimit.h"
21 #include "system/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 backing_mask_protocol;
43     bool bs_read_only;
44 } StreamBlockJob;
45 
stream_populate(BlockBackend * blk,int64_t offset,uint64_t bytes)46 static int coroutine_fn stream_populate(BlockBackend *blk,
47                                         int64_t offset, uint64_t bytes)
48 {
49     assert(bytes < SIZE_MAX);
50 
51     return blk_co_preadv(blk, offset, bytes, NULL, BDRV_REQ_PREFETCH);
52 }
53 
stream_prepare(Job * job)54 static int stream_prepare(Job *job)
55 {
56     StreamBlockJob *s = container_of(job, StreamBlockJob, common.job);
57     BlockDriverState *unfiltered_bs;
58     BlockDriverState *unfiltered_bs_cow;
59     BlockDriverState *base;
60     BlockDriverState *unfiltered_base;
61     Error *local_err = NULL;
62     int ret = 0;
63 
64     GLOBAL_STATE_CODE();
65 
66     bdrv_graph_rdlock_main_loop();
67     unfiltered_bs = bdrv_skip_filters(s->target_bs);
68     unfiltered_bs_cow = bdrv_cow_bs(unfiltered_bs);
69     bdrv_graph_rdunlock_main_loop();
70 
71     /* We should drop filter at this point, as filter hold the backing chain */
72     bdrv_cor_filter_drop(s->cor_filter_bs);
73     s->cor_filter_bs = NULL;
74 
75     /*
76      * bdrv_set_backing_hd() requires that the unfiltered_bs and the COW child
77      * of unfiltered_bs is drained. Drain already here and use
78      * bdrv_set_backing_hd_drained() instead because the polling during
79      * drained_begin() might change the graph, and if we do this only later, we
80      * may end up working with the wrong base node (or it might even have gone
81      * away by the time we want to use it).
82      */
83     if (unfiltered_bs_cow) {
84         bdrv_ref(unfiltered_bs_cow);
85     }
86     bdrv_drain_all_begin();
87 
88     bdrv_graph_rdlock_main_loop();
89     base = bdrv_filter_or_cow_bs(s->above_base);
90     unfiltered_base = bdrv_skip_filters(base);
91     bdrv_graph_rdunlock_main_loop();
92 
93     if (unfiltered_bs_cow) {
94         const char *base_id = NULL, *base_fmt = NULL;
95         if (unfiltered_base) {
96             base_id = s->backing_file_str ?: unfiltered_base->filename;
97             if (unfiltered_base->drv) {
98                 if (s->backing_mask_protocol &&
99                     unfiltered_base->drv->protocol_name) {
100                     base_fmt = "raw";
101                 } else {
102                     base_fmt = unfiltered_base->drv->format_name;
103                 }
104             }
105         }
106 
107         bdrv_graph_wrlock();
108         bdrv_set_backing_hd_drained(unfiltered_bs, base, &local_err);
109         bdrv_graph_wrunlock();
110 
111         /*
112          * This call will do I/O, so the graph can change again from here on.
113          * We have already completed the graph change, so we are not in danger
114          * of operating on the wrong node any more if this happens.
115          */
116         ret = bdrv_change_backing_file(unfiltered_bs, base_id, base_fmt, false);
117         if (local_err) {
118             error_report_err(local_err);
119             ret = -EPERM;
120             goto out;
121         }
122     }
123 
124 out:
125     bdrv_drain_all_end();
126     if (unfiltered_bs_cow) {
127         bdrv_unref(unfiltered_bs_cow);
128     }
129     return ret;
130 }
131 
stream_clean(Job * job)132 static void stream_clean(Job *job)
133 {
134     StreamBlockJob *s = container_of(job, StreamBlockJob, common.job);
135 
136     if (s->cor_filter_bs) {
137         bdrv_cor_filter_drop(s->cor_filter_bs);
138         s->cor_filter_bs = NULL;
139     }
140 
141     blk_unref(s->blk);
142     s->blk = NULL;
143 
144     /* Reopen the image back in read-only mode if necessary */
145     if (s->bs_read_only) {
146         /* Give up write permissions before making it read-only */
147         bdrv_reopen_set_read_only(s->target_bs, true, NULL);
148     }
149 
150     g_free(s->backing_file_str);
151 }
152 
stream_run(Job * job,Error ** errp)153 static int coroutine_fn stream_run(Job *job, Error **errp)
154 {
155     StreamBlockJob *s = container_of(job, StreamBlockJob, common.job);
156     BlockDriverState *unfiltered_bs = NULL;
157     int64_t len = -1;
158     int64_t offset = 0;
159     int error = 0;
160     int64_t n = 0; /* bytes */
161 
162     WITH_GRAPH_RDLOCK_GUARD() {
163         unfiltered_bs = bdrv_skip_filters(s->target_bs);
164         if (unfiltered_bs == s->base_overlay) {
165             /* Nothing to stream */
166             return 0;
167         }
168 
169         len = bdrv_co_getlength(s->target_bs);
170         if (len < 0) {
171             return len;
172         }
173     }
174     job_progress_set_remaining(&s->common.job, len);
175 
176     for ( ; offset < len; offset += n) {
177         bool copy;
178         int ret = -1;
179 
180         /* Note that even when no rate limit is applied we need to yield
181          * with no pending I/O here so that bdrv_drain_all() returns.
182          */
183         block_job_ratelimit_sleep(&s->common);
184         if (job_is_cancelled(&s->common.job)) {
185             break;
186         }
187 
188         copy = false;
189 
190         WITH_GRAPH_RDLOCK_GUARD() {
191             ret = bdrv_co_is_allocated(unfiltered_bs, offset, STREAM_CHUNK, &n);
192             if (ret == 1) {
193                 /* Allocated in the top, no need to copy.  */
194             } else if (ret >= 0) {
195                 /*
196                  * Copy if allocated in the intermediate images.  Limit to the
197                  * known-unallocated area [offset, offset+n*BDRV_SECTOR_SIZE).
198                  */
199                 ret = bdrv_co_is_allocated_above(bdrv_cow_bs(unfiltered_bs),
200                                                  s->base_overlay, true,
201                                                  offset, n, &n);
202                 /* Finish early if end of backing file has been reached */
203                 if (ret == 0 && n == 0) {
204                     n = len - offset;
205                 }
206 
207                 copy = (ret > 0);
208             }
209         }
210         trace_stream_one_iteration(s, offset, n, ret);
211         if (copy) {
212             ret = stream_populate(s->blk, offset, n);
213         }
214         if (ret < 0) {
215             BlockErrorAction action =
216                 block_job_error_action(&s->common, s->on_error, true, -ret);
217             if (action == BLOCK_ERROR_ACTION_STOP) {
218                 n = 0;
219                 continue;
220             }
221             if (error == 0) {
222                 error = ret;
223             }
224             if (action == BLOCK_ERROR_ACTION_REPORT) {
225                 break;
226             }
227         }
228 
229         /* Publish progress */
230         job_progress_update(&s->common.job, n);
231         if (copy) {
232             block_job_ratelimit_processed_bytes(&s->common, n);
233         }
234     }
235 
236     /* Do not remove the backing file if an error was there but ignored. */
237     return error;
238 }
239 
240 static const BlockJobDriver stream_job_driver = {
241     .job_driver = {
242         .instance_size = sizeof(StreamBlockJob),
243         .job_type      = JOB_TYPE_STREAM,
244         .free          = block_job_free,
245         .run           = stream_run,
246         .prepare       = stream_prepare,
247         .clean         = stream_clean,
248         .user_resume   = block_job_user_resume,
249     },
250 };
251 
stream_start(const char * job_id,BlockDriverState * bs,BlockDriverState * base,const char * backing_file_str,bool backing_mask_protocol,BlockDriverState * bottom,int creation_flags,int64_t speed,BlockdevOnError on_error,const char * filter_node_name,Error ** errp)252 void stream_start(const char *job_id, BlockDriverState *bs,
253                   BlockDriverState *base, const char *backing_file_str,
254                   bool backing_mask_protocol,
255                   BlockDriverState *bottom,
256                   int creation_flags, int64_t speed,
257                   BlockdevOnError on_error,
258                   const char *filter_node_name,
259                   Error **errp)
260 {
261     StreamBlockJob *s = NULL;
262     BlockDriverState *iter;
263     bool bs_read_only;
264     int basic_flags = BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED;
265     BlockDriverState *base_overlay;
266     BlockDriverState *cor_filter_bs = NULL;
267     BlockDriverState *above_base;
268     QDict *opts;
269     int ret;
270 
271     GLOBAL_STATE_CODE();
272 
273     assert(!(base && bottom));
274     assert(!(backing_file_str && bottom));
275 
276     bdrv_graph_rdlock_main_loop();
277 
278     if (bottom) {
279         /*
280          * New simple interface. The code is written in terms of old interface
281          * with @base parameter (still, it doesn't freeze link to base, so in
282          * this mean old code is correct for new interface). So, for now, just
283          * emulate base_overlay and above_base. Still, when old interface
284          * finally removed, we should refactor code to use only "bottom", but
285          * not "*base*" things.
286          */
287         assert(!bottom->drv->is_filter);
288         base_overlay = above_base = bottom;
289     } else {
290         base_overlay = bdrv_find_overlay(bs, base);
291         if (!base_overlay) {
292             error_setg(errp, "'%s' is not in the backing chain of '%s'",
293                        base->node_name, bs->node_name);
294             goto out_rdlock;
295         }
296 
297         /*
298          * Find the node directly above @base.  @base_overlay is a COW overlay,
299          * so it must have a bdrv_cow_child(), but it is the immediate overlay
300          * of @base, so between the two there can only be filters.
301          */
302         above_base = base_overlay;
303         if (bdrv_cow_bs(above_base) != base) {
304             above_base = bdrv_cow_bs(above_base);
305             while (bdrv_filter_bs(above_base) != base) {
306                 above_base = bdrv_filter_bs(above_base);
307             }
308         }
309     }
310 
311     /* Make sure that the image is opened in read-write mode */
312     bs_read_only = bdrv_is_read_only(bs);
313     if (bs_read_only) {
314         /* Hold the chain during reopen */
315         if (bdrv_freeze_backing_chain(bs, above_base, errp) < 0) {
316             goto out_rdlock;
317         }
318 
319         ret = bdrv_reopen_set_read_only(bs, false, errp);
320 
321         /* failure, or cor-filter will hold the chain */
322         bdrv_unfreeze_backing_chain(bs, above_base);
323 
324         if (ret < 0) {
325             goto out_rdlock;
326         }
327     }
328 
329     bdrv_graph_rdunlock_main_loop();
330 
331     opts = qdict_new();
332 
333     qdict_put_str(opts, "driver", "copy-on-read");
334     qdict_put_str(opts, "file", bdrv_get_node_name(bs));
335     /* Pass the base_overlay node name as 'bottom' to COR driver */
336     qdict_put_str(opts, "bottom", base_overlay->node_name);
337     if (filter_node_name) {
338         qdict_put_str(opts, "node-name", filter_node_name);
339     }
340 
341     cor_filter_bs = bdrv_insert_node(bs, opts, BDRV_O_RDWR, errp);
342     if (!cor_filter_bs) {
343         goto fail;
344     }
345 
346     if (!filter_node_name) {
347         cor_filter_bs->implicit = true;
348     }
349 
350     s = block_job_create(job_id, &stream_job_driver, NULL, cor_filter_bs,
351                          0, BLK_PERM_ALL,
352                          speed, creation_flags, NULL, NULL, errp);
353     if (!s) {
354         goto fail;
355     }
356 
357     s->blk = blk_new_with_bs(cor_filter_bs, BLK_PERM_CONSISTENT_READ,
358                              basic_flags | BLK_PERM_WRITE, errp);
359     if (!s->blk) {
360         goto fail;
361     }
362     /*
363      * Disable request queuing in the BlockBackend to avoid deadlocks on drain:
364      * The job reports that it's busy until it reaches a pause point.
365      */
366     blk_set_disable_request_queuing(s->blk, true);
367     blk_set_allow_aio_context_change(s->blk, true);
368 
369     /*
370      * Prevent concurrent jobs trying to modify the graph structure here, we
371      * already have our own plans. Also don't allow resize as the image size is
372      * queried only at the job start and then cached.
373      */
374     bdrv_drain_all_begin();
375     bdrv_graph_wrlock();
376     if (block_job_add_bdrv(&s->common, "active node", bs, 0,
377                            basic_flags | BLK_PERM_WRITE, errp)) {
378         bdrv_graph_wrunlock();
379         bdrv_drain_all_end();
380         goto fail;
381     }
382 
383     /* Block all intermediate nodes between bs and base, because they will
384      * disappear from the chain after this operation. The streaming job reads
385      * every block only once, assuming that it doesn't change, so forbid writes
386      * and resizes. Reassign the base node pointer because the backing BS of the
387      * bottom node might change after the call to bdrv_reopen_set_read_only()
388      * due to parallel block jobs running.
389      * above_base node might change after the call to
390      * bdrv_reopen_set_read_only() due to parallel block jobs running.
391      */
392     base = bdrv_filter_or_cow_bs(above_base);
393     for (iter = bdrv_filter_or_cow_bs(bs); iter != base;
394          iter = bdrv_filter_or_cow_bs(iter))
395     {
396         ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
397                                  basic_flags, errp);
398         if (ret < 0) {
399             bdrv_graph_wrunlock();
400             bdrv_drain_all_end();
401             goto fail;
402         }
403     }
404     bdrv_graph_wrunlock();
405     bdrv_drain_all_end();
406 
407     s->base_overlay = base_overlay;
408     s->above_base = above_base;
409     s->backing_file_str = g_strdup(backing_file_str);
410     s->backing_mask_protocol = backing_mask_protocol;
411     s->cor_filter_bs = cor_filter_bs;
412     s->target_bs = bs;
413     s->bs_read_only = bs_read_only;
414 
415     s->on_error = on_error;
416     trace_stream_start(bs, base, s);
417     job_start(&s->common.job);
418     return;
419 
420 fail:
421     if (s) {
422         job_early_fail(&s->common.job);
423     }
424     if (cor_filter_bs) {
425         bdrv_cor_filter_drop(cor_filter_bs);
426     }
427     if (bs_read_only) {
428         bdrv_reopen_set_read_only(bs, true, NULL);
429     }
430     return;
431 
432 out_rdlock:
433     bdrv_graph_rdunlock_main_loop();
434 }
435