xref: /openbmc/qemu/block/commit.c (revision 0e39bb022b5fa8c11964968885f3263c02ce42b0)
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 "trace.h"
17 #include "block/block_int.h"
18 #include "block/blockjob.h"
19 #include "qapi/error.h"
20 #include "qapi/qmp/qerror.h"
21 #include "qemu/ratelimit.h"
22 #include "sysemu/block-backend.h"
23 
24 enum {
25     /*
26      * Size of data buffer for populating the image file.  This should be large
27      * enough to process multiple clusters in a single call, so that populating
28      * contiguous regions of the image is efficient.
29      */
30     COMMIT_BUFFER_SIZE = 512 * 1024, /* in bytes */
31 };
32 
33 #define SLICE_TIME 100000000ULL /* ns */
34 
35 typedef struct CommitBlockJob {
36     BlockJob common;
37     RateLimit limit;
38     BlockDriverState *active;
39     BlockBackend *top;
40     BlockBackend *base;
41     BlockdevOnError on_error;
42     int base_flags;
43     int orig_overlay_flags;
44     char *backing_file_str;
45 } CommitBlockJob;
46 
47 static int coroutine_fn commit_populate(BlockBackend *bs, BlockBackend *base,
48                                         int64_t sector_num, int nb_sectors,
49                                         void *buf)
50 {
51     int ret = 0;
52     QEMUIOVector qiov;
53     struct iovec iov = {
54         .iov_base = buf,
55         .iov_len = nb_sectors * BDRV_SECTOR_SIZE,
56     };
57 
58     qemu_iovec_init_external(&qiov, &iov, 1);
59 
60     ret = blk_co_preadv(bs, sector_num * BDRV_SECTOR_SIZE,
61                         qiov.size, &qiov, 0);
62     if (ret < 0) {
63         return ret;
64     }
65 
66     ret = blk_co_pwritev(base, sector_num * BDRV_SECTOR_SIZE,
67                          qiov.size, &qiov, 0);
68     if (ret < 0) {
69         return ret;
70     }
71 
72     return 0;
73 }
74 
75 typedef struct {
76     int ret;
77 } CommitCompleteData;
78 
79 static void commit_complete(BlockJob *job, void *opaque)
80 {
81     CommitBlockJob *s = container_of(job, CommitBlockJob, common);
82     CommitCompleteData *data = opaque;
83     BlockDriverState *active = s->active;
84     BlockDriverState *top = blk_bs(s->top);
85     BlockDriverState *base = blk_bs(s->base);
86     BlockDriverState *overlay_bs;
87     int ret = data->ret;
88 
89     if (!block_job_is_cancelled(&s->common) && ret == 0) {
90         /* success */
91         ret = bdrv_drop_intermediate(active, top, base, s->backing_file_str);
92     }
93 
94     /* restore base open flags here if appropriate (e.g., change the base back
95      * to r/o). These reopens do not need to be atomic, since we won't abort
96      * even on failure here */
97     if (s->base_flags != bdrv_get_flags(base)) {
98         bdrv_reopen(base, s->base_flags, NULL);
99     }
100     overlay_bs = bdrv_find_overlay(active, top);
101     if (overlay_bs && s->orig_overlay_flags != bdrv_get_flags(overlay_bs)) {
102         bdrv_reopen(overlay_bs, s->orig_overlay_flags, NULL);
103     }
104     g_free(s->backing_file_str);
105     blk_unref(s->top);
106     blk_unref(s->base);
107     block_job_completed(&s->common, ret);
108     g_free(data);
109 }
110 
111 static void coroutine_fn commit_run(void *opaque)
112 {
113     CommitBlockJob *s = opaque;
114     CommitCompleteData *data;
115     int64_t sector_num, end;
116     uint64_t delay_ns = 0;
117     int ret = 0;
118     int n = 0;
119     void *buf = NULL;
120     int bytes_written = 0;
121     int64_t base_len;
122 
123     ret = s->common.len = blk_getlength(s->top);
124 
125 
126     if (s->common.len < 0) {
127         goto out;
128     }
129 
130     ret = base_len = blk_getlength(s->base);
131     if (base_len < 0) {
132         goto out;
133     }
134 
135     if (base_len < s->common.len) {
136         ret = blk_truncate(s->base, s->common.len);
137         if (ret) {
138             goto out;
139         }
140     }
141 
142     end = s->common.len >> BDRV_SECTOR_BITS;
143     buf = blk_blockalign(s->top, COMMIT_BUFFER_SIZE);
144 
145     for (sector_num = 0; sector_num < end; sector_num += n) {
146         bool copy;
147 
148         /* Note that even when no rate limit is applied we need to yield
149          * with no pending I/O here so that bdrv_drain_all() returns.
150          */
151         block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
152         if (block_job_is_cancelled(&s->common)) {
153             break;
154         }
155         /* Copy if allocated above the base */
156         ret = bdrv_is_allocated_above(blk_bs(s->top), blk_bs(s->base),
157                                       sector_num,
158                                       COMMIT_BUFFER_SIZE / BDRV_SECTOR_SIZE,
159                                       &n);
160         copy = (ret == 1);
161         trace_commit_one_iteration(s, sector_num, n, ret);
162         if (copy) {
163             ret = commit_populate(s->top, s->base, sector_num, n, buf);
164             bytes_written += n * BDRV_SECTOR_SIZE;
165         }
166         if (ret < 0) {
167             BlockErrorAction action =
168                 block_job_error_action(&s->common, false, s->on_error, -ret);
169             if (action == BLOCK_ERROR_ACTION_REPORT) {
170                 goto out;
171             } else {
172                 n = 0;
173                 continue;
174             }
175         }
176         /* Publish progress */
177         s->common.offset += n * BDRV_SECTOR_SIZE;
178 
179         if (copy && s->common.speed) {
180             delay_ns = ratelimit_calculate_delay(&s->limit, n);
181         }
182     }
183 
184     ret = 0;
185 
186 out:
187     qemu_vfree(buf);
188 
189     data = g_malloc(sizeof(*data));
190     data->ret = ret;
191     block_job_defer_to_main_loop(&s->common, commit_complete, data);
192 }
193 
194 static void commit_set_speed(BlockJob *job, int64_t speed, Error **errp)
195 {
196     CommitBlockJob *s = container_of(job, CommitBlockJob, common);
197 
198     if (speed < 0) {
199         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
200         return;
201     }
202     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
203 }
204 
205 static const BlockJobDriver commit_job_driver = {
206     .instance_size = sizeof(CommitBlockJob),
207     .job_type      = BLOCK_JOB_TYPE_COMMIT,
208     .set_speed     = commit_set_speed,
209 };
210 
211 void commit_start(const char *job_id, BlockDriverState *bs,
212                   BlockDriverState *base, BlockDriverState *top, int64_t speed,
213                   BlockdevOnError on_error, BlockCompletionFunc *cb,
214                   void *opaque, const char *backing_file_str, Error **errp)
215 {
216     CommitBlockJob *s;
217     BlockReopenQueue *reopen_queue = NULL;
218     int orig_overlay_flags;
219     int orig_base_flags;
220     BlockDriverState *overlay_bs;
221     Error *local_err = NULL;
222 
223     assert(top != bs);
224     if (top == base) {
225         error_setg(errp, "Invalid files for merge: top and base are the same");
226         return;
227     }
228 
229     overlay_bs = bdrv_find_overlay(bs, top);
230 
231     if (overlay_bs == NULL) {
232         error_setg(errp, "Could not find overlay image for %s:", top->filename);
233         return;
234     }
235 
236     s = block_job_create(job_id, &commit_job_driver, bs, speed,
237                          cb, opaque, errp);
238     if (!s) {
239         return;
240     }
241 
242     orig_base_flags    = bdrv_get_flags(base);
243     orig_overlay_flags = bdrv_get_flags(overlay_bs);
244 
245     /* convert base & overlay_bs to r/w, if necessary */
246     if (!(orig_overlay_flags & BDRV_O_RDWR)) {
247         reopen_queue = bdrv_reopen_queue(reopen_queue, overlay_bs, NULL,
248                                          orig_overlay_flags | BDRV_O_RDWR);
249     }
250     if (!(orig_base_flags & BDRV_O_RDWR)) {
251         reopen_queue = bdrv_reopen_queue(reopen_queue, base, NULL,
252                                          orig_base_flags | BDRV_O_RDWR);
253     }
254     if (reopen_queue) {
255         bdrv_reopen_multiple(reopen_queue, &local_err);
256         if (local_err != NULL) {
257             error_propagate(errp, local_err);
258             block_job_unref(&s->common);
259             return;
260         }
261     }
262 
263 
264     s->base = blk_new();
265     blk_insert_bs(s->base, base);
266 
267     s->top = blk_new();
268     blk_insert_bs(s->top, top);
269 
270     s->active = bs;
271 
272     s->base_flags          = orig_base_flags;
273     s->orig_overlay_flags  = orig_overlay_flags;
274 
275     s->backing_file_str = g_strdup(backing_file_str);
276 
277     s->on_error = on_error;
278     s->common.co = qemu_coroutine_create(commit_run, s);
279 
280     trace_commit_start(bs, base, top, s, s->common.co, opaque);
281     qemu_coroutine_enter(s->common.co);
282 }
283 
284 
285 #define COMMIT_BUF_SECTORS 2048
286 
287 /* commit COW file into the raw image */
288 int bdrv_commit(BlockDriverState *bs)
289 {
290     BlockBackend *src, *backing;
291     BlockDriver *drv = bs->drv;
292     int64_t sector, total_sectors, length, backing_length;
293     int n, ro, open_flags;
294     int ret = 0;
295     uint8_t *buf = NULL;
296 
297     if (!drv)
298         return -ENOMEDIUM;
299 
300     if (!bs->backing) {
301         return -ENOTSUP;
302     }
303 
304     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_COMMIT_SOURCE, NULL) ||
305         bdrv_op_is_blocked(bs->backing->bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL)) {
306         return -EBUSY;
307     }
308 
309     ro = bs->backing->bs->read_only;
310     open_flags =  bs->backing->bs->open_flags;
311 
312     if (ro) {
313         if (bdrv_reopen(bs->backing->bs, open_flags | BDRV_O_RDWR, NULL)) {
314             return -EACCES;
315         }
316     }
317 
318     src = blk_new();
319     blk_insert_bs(src, bs);
320 
321     backing = blk_new();
322     blk_insert_bs(backing, bs->backing->bs);
323 
324     length = blk_getlength(src);
325     if (length < 0) {
326         ret = length;
327         goto ro_cleanup;
328     }
329 
330     backing_length = blk_getlength(backing);
331     if (backing_length < 0) {
332         ret = backing_length;
333         goto ro_cleanup;
334     }
335 
336     /* If our top snapshot is larger than the backing file image,
337      * grow the backing file image if possible.  If not possible,
338      * we must return an error */
339     if (length > backing_length) {
340         ret = blk_truncate(backing, length);
341         if (ret < 0) {
342             goto ro_cleanup;
343         }
344     }
345 
346     total_sectors = length >> BDRV_SECTOR_BITS;
347 
348     /* blk_try_blockalign() for src will choose an alignment that works for
349      * backing as well, so no need to compare the alignment manually. */
350     buf = blk_try_blockalign(src, COMMIT_BUF_SECTORS * BDRV_SECTOR_SIZE);
351     if (buf == NULL) {
352         ret = -ENOMEM;
353         goto ro_cleanup;
354     }
355 
356     for (sector = 0; sector < total_sectors; sector += n) {
357         ret = bdrv_is_allocated(bs, sector, COMMIT_BUF_SECTORS, &n);
358         if (ret < 0) {
359             goto ro_cleanup;
360         }
361         if (ret) {
362             ret = blk_pread(src, sector * BDRV_SECTOR_SIZE, buf,
363                             n * BDRV_SECTOR_SIZE);
364             if (ret < 0) {
365                 goto ro_cleanup;
366             }
367 
368             ret = blk_pwrite(backing, sector * BDRV_SECTOR_SIZE, buf,
369                              n * BDRV_SECTOR_SIZE, 0);
370             if (ret < 0) {
371                 goto ro_cleanup;
372             }
373         }
374     }
375 
376     if (drv->bdrv_make_empty) {
377         ret = drv->bdrv_make_empty(bs);
378         if (ret < 0) {
379             goto ro_cleanup;
380         }
381         blk_flush(src);
382     }
383 
384     /*
385      * Make sure all data we wrote to the backing device is actually
386      * stable on disk.
387      */
388     blk_flush(backing);
389 
390     ret = 0;
391 ro_cleanup:
392     qemu_vfree(buf);
393 
394     blk_unref(src);
395     blk_unref(backing);
396 
397     if (ro) {
398         /* ignoring error return here */
399         bdrv_reopen(bs->backing->bs, open_flags & ~BDRV_O_RDWR, NULL);
400     }
401 
402     return ret;
403 }
404