Make vfio MSI interrupt be non-threaded.
[kvmfornfv.git] / qemu / block / commit.c
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 "trace.h"
16 #include "block/block_int.h"
17 #include "block/blockjob.h"
18 #include "qapi/qmp/qerror.h"
19 #include "qemu/ratelimit.h"
20
21 enum {
22     /*
23      * Size of data buffer for populating the image file.  This should be large
24      * enough to process multiple clusters in a single call, so that populating
25      * contiguous regions of the image is efficient.
26      */
27     COMMIT_BUFFER_SIZE = 512 * 1024, /* in bytes */
28 };
29
30 #define SLICE_TIME 100000000ULL /* ns */
31
32 typedef struct CommitBlockJob {
33     BlockJob common;
34     RateLimit limit;
35     BlockDriverState *active;
36     BlockDriverState *top;
37     BlockDriverState *base;
38     BlockdevOnError on_error;
39     int base_flags;
40     int orig_overlay_flags;
41     char *backing_file_str;
42 } CommitBlockJob;
43
44 static int coroutine_fn commit_populate(BlockDriverState *bs,
45                                         BlockDriverState *base,
46                                         int64_t sector_num, int nb_sectors,
47                                         void *buf)
48 {
49     int ret = 0;
50
51     ret = bdrv_read(bs, sector_num, buf, nb_sectors);
52     if (ret) {
53         return ret;
54     }
55
56     ret = bdrv_write(base, sector_num, buf, nb_sectors);
57     if (ret) {
58         return ret;
59     }
60
61     return 0;
62 }
63
64 typedef struct {
65     int ret;
66 } CommitCompleteData;
67
68 static void commit_complete(BlockJob *job, void *opaque)
69 {
70     CommitBlockJob *s = container_of(job, CommitBlockJob, common);
71     CommitCompleteData *data = opaque;
72     BlockDriverState *active = s->active;
73     BlockDriverState *top = s->top;
74     BlockDriverState *base = s->base;
75     BlockDriverState *overlay_bs;
76     int ret = data->ret;
77
78     if (!block_job_is_cancelled(&s->common) && ret == 0) {
79         /* success */
80         ret = bdrv_drop_intermediate(active, top, base, s->backing_file_str);
81     }
82
83     /* restore base open flags here if appropriate (e.g., change the base back
84      * to r/o). These reopens do not need to be atomic, since we won't abort
85      * even on failure here */
86     if (s->base_flags != bdrv_get_flags(base)) {
87         bdrv_reopen(base, s->base_flags, NULL);
88     }
89     overlay_bs = bdrv_find_overlay(active, top);
90     if (overlay_bs && s->orig_overlay_flags != bdrv_get_flags(overlay_bs)) {
91         bdrv_reopen(overlay_bs, s->orig_overlay_flags, NULL);
92     }
93     g_free(s->backing_file_str);
94     block_job_completed(&s->common, ret);
95     g_free(data);
96 }
97
98 static void coroutine_fn commit_run(void *opaque)
99 {
100     CommitBlockJob *s = opaque;
101     CommitCompleteData *data;
102     BlockDriverState *top = s->top;
103     BlockDriverState *base = s->base;
104     int64_t sector_num, end;
105     int ret = 0;
106     int n = 0;
107     void *buf = NULL;
108     int bytes_written = 0;
109     int64_t base_len;
110
111     ret = s->common.len = bdrv_getlength(top);
112
113
114     if (s->common.len < 0) {
115         goto out;
116     }
117
118     ret = base_len = bdrv_getlength(base);
119     if (base_len < 0) {
120         goto out;
121     }
122
123     if (base_len < s->common.len) {
124         ret = bdrv_truncate(base, s->common.len);
125         if (ret) {
126             goto out;
127         }
128     }
129
130     end = s->common.len >> BDRV_SECTOR_BITS;
131     buf = qemu_blockalign(top, COMMIT_BUFFER_SIZE);
132
133     for (sector_num = 0; sector_num < end; sector_num += n) {
134         uint64_t delay_ns = 0;
135         bool copy;
136
137 wait:
138         /* Note that even when no rate limit is applied we need to yield
139          * with no pending I/O here so that bdrv_drain_all() returns.
140          */
141         block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
142         if (block_job_is_cancelled(&s->common)) {
143             break;
144         }
145         /* Copy if allocated above the base */
146         ret = bdrv_is_allocated_above(top, base, sector_num,
147                                       COMMIT_BUFFER_SIZE / BDRV_SECTOR_SIZE,
148                                       &n);
149         copy = (ret == 1);
150         trace_commit_one_iteration(s, sector_num, n, ret);
151         if (copy) {
152             if (s->common.speed) {
153                 delay_ns = ratelimit_calculate_delay(&s->limit, n);
154                 if (delay_ns > 0) {
155                     goto wait;
156                 }
157             }
158             ret = commit_populate(top, base, sector_num, n, buf);
159             bytes_written += n * BDRV_SECTOR_SIZE;
160         }
161         if (ret < 0) {
162             if (s->on_error == BLOCKDEV_ON_ERROR_STOP ||
163                 s->on_error == BLOCKDEV_ON_ERROR_REPORT||
164                 (s->on_error == BLOCKDEV_ON_ERROR_ENOSPC && ret == -ENOSPC)) {
165                 goto out;
166             } else {
167                 n = 0;
168                 continue;
169             }
170         }
171         /* Publish progress */
172         s->common.offset += n * BDRV_SECTOR_SIZE;
173     }
174
175     ret = 0;
176
177 out:
178     qemu_vfree(buf);
179
180     data = g_malloc(sizeof(*data));
181     data->ret = ret;
182     block_job_defer_to_main_loop(&s->common, commit_complete, data);
183 }
184
185 static void commit_set_speed(BlockJob *job, int64_t speed, Error **errp)
186 {
187     CommitBlockJob *s = container_of(job, CommitBlockJob, common);
188
189     if (speed < 0) {
190         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
191         return;
192     }
193     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
194 }
195
196 static const BlockJobDriver commit_job_driver = {
197     .instance_size = sizeof(CommitBlockJob),
198     .job_type      = BLOCK_JOB_TYPE_COMMIT,
199     .set_speed     = commit_set_speed,
200 };
201
202 void commit_start(BlockDriverState *bs, BlockDriverState *base,
203                   BlockDriverState *top, int64_t speed,
204                   BlockdevOnError on_error, BlockCompletionFunc *cb,
205                   void *opaque, const char *backing_file_str, Error **errp)
206 {
207     CommitBlockJob *s;
208     BlockReopenQueue *reopen_queue = NULL;
209     int orig_overlay_flags;
210     int orig_base_flags;
211     BlockDriverState *overlay_bs;
212     Error *local_err = NULL;
213
214     if ((on_error == BLOCKDEV_ON_ERROR_STOP ||
215          on_error == BLOCKDEV_ON_ERROR_ENOSPC) &&
216         !bdrv_iostatus_is_enabled(bs)) {
217         error_setg(errp, "Invalid parameter combination");
218         return;
219     }
220
221     assert(top != bs);
222     if (top == base) {
223         error_setg(errp, "Invalid files for merge: top and base are the same");
224         return;
225     }
226
227     overlay_bs = bdrv_find_overlay(bs, top);
228
229     if (overlay_bs == NULL) {
230         error_setg(errp, "Could not find overlay image for %s:", top->filename);
231         return;
232     }
233
234     orig_base_flags    = bdrv_get_flags(base);
235     orig_overlay_flags = bdrv_get_flags(overlay_bs);
236
237     /* convert base & overlay_bs to r/w, if necessary */
238     if (!(orig_base_flags & BDRV_O_RDWR)) {
239         reopen_queue = bdrv_reopen_queue(reopen_queue, base,
240                                          orig_base_flags | BDRV_O_RDWR);
241     }
242     if (!(orig_overlay_flags & BDRV_O_RDWR)) {
243         reopen_queue = bdrv_reopen_queue(reopen_queue, overlay_bs,
244                                          orig_overlay_flags | BDRV_O_RDWR);
245     }
246     if (reopen_queue) {
247         bdrv_reopen_multiple(reopen_queue, &local_err);
248         if (local_err != NULL) {
249             error_propagate(errp, local_err);
250             return;
251         }
252     }
253
254
255     s = block_job_create(&commit_job_driver, bs, speed, cb, opaque, errp);
256     if (!s) {
257         return;
258     }
259
260     s->base   = base;
261     s->top    = top;
262     s->active = bs;
263
264     s->base_flags          = orig_base_flags;
265     s->orig_overlay_flags  = orig_overlay_flags;
266
267     s->backing_file_str = g_strdup(backing_file_str);
268
269     s->on_error = on_error;
270     s->common.co = qemu_coroutine_create(commit_run);
271
272     trace_commit_start(bs, base, top, s, s->common.co, opaque);
273     qemu_coroutine_enter(s->common.co, s);
274 }