qemu

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 backing_mask_protocol;
43
    bool bs_read_only;
44
} StreamBlockJob;
45

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

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
    bdrv_drained_begin(unfiltered_bs);
84
    if (unfiltered_bs_cow) {
85
        bdrv_ref(unfiltered_bs_cow);
86
        bdrv_drained_begin(unfiltered_bs_cow);
87
    }
88

89
    bdrv_graph_rdlock_main_loop();
90
    base = bdrv_filter_or_cow_bs(s->above_base);
91
    unfiltered_base = bdrv_skip_filters(base);
92
    bdrv_graph_rdunlock_main_loop();
93

94
    if (unfiltered_bs_cow) {
95
        const char *base_id = NULL, *base_fmt = NULL;
96
        if (unfiltered_base) {
97
            base_id = s->backing_file_str ?: unfiltered_base->filename;
98
            if (unfiltered_base->drv) {
99
                if (s->backing_mask_protocol &&
100
                    unfiltered_base->drv->protocol_name) {
101
                    base_fmt = "raw";
102
                } else {
103
                    base_fmt = unfiltered_base->drv->format_name;
104
                }
105
            }
106
        }
107

108
        bdrv_graph_wrlock();
109
        bdrv_set_backing_hd_drained(unfiltered_bs, base, &local_err);
110
        bdrv_graph_wrunlock();
111

112
        /*
113
         * This call will do I/O, so the graph can change again from here on.
114
         * We have already completed the graph change, so we are not in danger
115
         * of operating on the wrong node any more if this happens.
116
         */
117
        ret = bdrv_change_backing_file(unfiltered_bs, base_id, base_fmt, false);
118
        if (local_err) {
119
            error_report_err(local_err);
120
            ret = -EPERM;
121
            goto out;
122
        }
123
    }
124

125
out:
126
    if (unfiltered_bs_cow) {
127
        bdrv_drained_end(unfiltered_bs_cow);
128
        bdrv_unref(unfiltered_bs_cow);
129
    }
130
    bdrv_drained_end(unfiltered_bs);
131
    return ret;
132
}
133

134
static void stream_clean(Job *job)
135
{
136
    StreamBlockJob *s = container_of(job, StreamBlockJob, common.job);
137

138
    if (s->cor_filter_bs) {
139
        bdrv_cor_filter_drop(s->cor_filter_bs);
140
        s->cor_filter_bs = NULL;
141
    }
142

143
    blk_unref(s->blk);
144
    s->blk = NULL;
145

146
    /* Reopen the image back in read-only mode if necessary */
147
    if (s->bs_read_only) {
148
        /* Give up write permissions before making it read-only */
149
        bdrv_reopen_set_read_only(s->target_bs, true, NULL);
150
    }
151

152
    g_free(s->backing_file_str);
153
}
154

155
static int coroutine_fn stream_run(Job *job, Error **errp)
156
{
157
    StreamBlockJob *s = container_of(job, StreamBlockJob, common.job);
158
    BlockDriverState *unfiltered_bs;
159
    int64_t len;
160
    int64_t offset = 0;
161
    int error = 0;
162
    int64_t n = 0; /* bytes */
163

164
    WITH_GRAPH_RDLOCK_GUARD() {
165
        unfiltered_bs = bdrv_skip_filters(s->target_bs);
166
        if (unfiltered_bs == s->base_overlay) {
167
            /* Nothing to stream */
168
            return 0;
169
        }
170

171
        len = bdrv_co_getlength(s->target_bs);
172
        if (len < 0) {
173
            return len;
174
        }
175
    }
176
    job_progress_set_remaining(&s->common.job, len);
177

178
    for ( ; offset < len; offset += n) {
179
        bool copy;
180
        int ret;
181

182
        /* Note that even when no rate limit is applied we need to yield
183
         * with no pending I/O here so that bdrv_drain_all() returns.
184
         */
185
        block_job_ratelimit_sleep(&s->common);
186
        if (job_is_cancelled(&s->common.job)) {
187
            break;
188
        }
189

190
        copy = false;
191

192
        WITH_GRAPH_RDLOCK_GUARD() {
193
            ret = bdrv_co_is_allocated(unfiltered_bs, offset, STREAM_CHUNK, &n);
194
            if (ret == 1) {
195
                /* Allocated in the top, no need to copy.  */
196
            } else if (ret >= 0) {
197
                /*
198
                 * Copy if allocated in the intermediate images.  Limit to the
199
                 * known-unallocated area [offset, offset+n*BDRV_SECTOR_SIZE).
200
                 */
201
                ret = bdrv_co_is_allocated_above(bdrv_cow_bs(unfiltered_bs),
202
                                                 s->base_overlay, true,
203
                                                 offset, n, &n);
204
                /* Finish early if end of backing file has been reached */
205
                if (ret == 0 && n == 0) {
206
                    n = len - offset;
207
                }
208

209
                copy = (ret > 0);
210
            }
211
        }
212
        trace_stream_one_iteration(s, offset, n, ret);
213
        if (copy) {
214
            ret = stream_populate(s->blk, offset, n);
215
        }
216
        if (ret < 0) {
217
            BlockErrorAction action =
218
                block_job_error_action(&s->common, s->on_error, true, -ret);
219
            if (action == BLOCK_ERROR_ACTION_STOP) {
220
                n = 0;
221
                continue;
222
            }
223
            if (error == 0) {
224
                error = ret;
225
            }
226
            if (action == BLOCK_ERROR_ACTION_REPORT) {
227
                break;
228
            }
229
        }
230

231
        /* Publish progress */
232
        job_progress_update(&s->common.job, n);
233
        if (copy) {
234
            block_job_ratelimit_processed_bytes(&s->common, n);
235
        }
236
    }
237

238
    /* Do not remove the backing file if an error was there but ignored. */
239
    return error;
240
}
241

242
static const BlockJobDriver stream_job_driver = {
243
    .job_driver = {
244
        .instance_size = sizeof(StreamBlockJob),
245
        .job_type      = JOB_TYPE_STREAM,
246
        .free          = block_job_free,
247
        .run           = stream_run,
248
        .prepare       = stream_prepare,
249
        .clean         = stream_clean,
250
        .user_resume   = block_job_user_resume,
251
    },
252
};
253

254
void stream_start(const char *job_id, BlockDriverState *bs,
255
                  BlockDriverState *base, const char *backing_file_str,
256
                  bool backing_mask_protocol,
257
                  BlockDriverState *bottom,
258
                  int creation_flags, int64_t speed,
259
                  BlockdevOnError on_error,
260
                  const char *filter_node_name,
261
                  Error **errp)
262
{
263
    StreamBlockJob *s = NULL;
264
    BlockDriverState *iter;
265
    bool bs_read_only;
266
    int basic_flags = BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED;
267
    BlockDriverState *base_overlay;
268
    BlockDriverState *cor_filter_bs = NULL;
269
    BlockDriverState *above_base;
270
    QDict *opts;
271
    int ret;
272

273
    GLOBAL_STATE_CODE();
274

275
    assert(!(base && bottom));
276
    assert(!(backing_file_str && bottom));
277

278
    bdrv_graph_rdlock_main_loop();
279

280
    if (bottom) {
281
        /*
282
         * New simple interface. The code is written in terms of old interface
283
         * with @base parameter (still, it doesn't freeze link to base, so in
284
         * this mean old code is correct for new interface). So, for now, just
285
         * emulate base_overlay and above_base. Still, when old interface
286
         * finally removed, we should refactor code to use only "bottom", but
287
         * not "*base*" things.
288
         */
289
        assert(!bottom->drv->is_filter);
290
        base_overlay = above_base = bottom;
291
    } else {
292
        base_overlay = bdrv_find_overlay(bs, base);
293
        if (!base_overlay) {
294
            error_setg(errp, "'%s' is not in the backing chain of '%s'",
295
                       base->node_name, bs->node_name);
296
            goto out_rdlock;
297
        }
298

299
        /*
300
         * Find the node directly above @base.  @base_overlay is a COW overlay,
301
         * so it must have a bdrv_cow_child(), but it is the immediate overlay
302
         * of @base, so between the two there can only be filters.
303
         */
304
        above_base = base_overlay;
305
        if (bdrv_cow_bs(above_base) != base) {
306
            above_base = bdrv_cow_bs(above_base);
307
            while (bdrv_filter_bs(above_base) != base) {
308
                above_base = bdrv_filter_bs(above_base);
309
            }
310
        }
311
    }
312

313
    /* Make sure that the image is opened in read-write mode */
314
    bs_read_only = bdrv_is_read_only(bs);
315
    if (bs_read_only) {
316
        /* Hold the chain during reopen */
317
        if (bdrv_freeze_backing_chain(bs, above_base, errp) < 0) {
318
            goto out_rdlock;
319
        }
320

321
        ret = bdrv_reopen_set_read_only(bs, false, errp);
322

323
        /* failure, or cor-filter will hold the chain */
324
        bdrv_unfreeze_backing_chain(bs, above_base);
325

326
        if (ret < 0) {
327
            goto out_rdlock;
328
        }
329
    }
330

331
    bdrv_graph_rdunlock_main_loop();
332

333
    opts = qdict_new();
334

335
    qdict_put_str(opts, "driver", "copy-on-read");
336
    qdict_put_str(opts, "file", bdrv_get_node_name(bs));
337
    /* Pass the base_overlay node name as 'bottom' to COR driver */
338
    qdict_put_str(opts, "bottom", base_overlay->node_name);
339
    if (filter_node_name) {
340
        qdict_put_str(opts, "node-name", filter_node_name);
341
    }
342

343
    cor_filter_bs = bdrv_insert_node(bs, opts, BDRV_O_RDWR, errp);
344
    if (!cor_filter_bs) {
345
        goto fail;
346
    }
347

348
    if (!filter_node_name) {
349
        cor_filter_bs->implicit = true;
350
    }
351

352
    s = block_job_create(job_id, &stream_job_driver, NULL, cor_filter_bs,
353
                         0, BLK_PERM_ALL,
354
                         speed, creation_flags, NULL, NULL, errp);
355
    if (!s) {
356
        goto fail;
357
    }
358

359
    s->blk = blk_new_with_bs(cor_filter_bs, BLK_PERM_CONSISTENT_READ,
360
                             basic_flags | BLK_PERM_WRITE, errp);
361
    if (!s->blk) {
362
        goto fail;
363
    }
364
    /*
365
     * Disable request queuing in the BlockBackend to avoid deadlocks on drain:
366
     * The job reports that it's busy until it reaches a pause point.
367
     */
368
    blk_set_disable_request_queuing(s->blk, true);
369
    blk_set_allow_aio_context_change(s->blk, true);
370

371
    /*
372
     * Prevent concurrent jobs trying to modify the graph structure here, we
373
     * already have our own plans. Also don't allow resize as the image size is
374
     * queried only at the job start and then cached.
375
     */
376
    bdrv_graph_wrlock();
377
    if (block_job_add_bdrv(&s->common, "active node", bs, 0,
378
                           basic_flags | BLK_PERM_WRITE, errp)) {
379
        bdrv_graph_wrunlock();
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
            goto fail;
401
        }
402
    }
403
    bdrv_graph_wrunlock();
404

405
    s->base_overlay = base_overlay;
406
    s->above_base = above_base;
407
    s->backing_file_str = g_strdup(backing_file_str);
408
    s->backing_mask_protocol = backing_mask_protocol;
409
    s->cor_filter_bs = cor_filter_bs;
410
    s->target_bs = bs;
411
    s->bs_read_only = bs_read_only;
412

413
    s->on_error = on_error;
414
    trace_stream_start(bs, base, s);
415
    job_start(&s->common.job);
416
    return;
417

418
fail:
419
    if (s) {
420
        job_early_fail(&s->common.job);
421
    }
422
    if (cor_filter_bs) {
423
        bdrv_cor_filter_drop(cor_filter_bs);
424
    }
425
    if (bs_read_only) {
426
        bdrv_reopen_set_read_only(bs, true, NULL);
427
    }
428
    return;
429

430
out_rdlock:
431
    bdrv_graph_rdunlock_main_loop();
432
}
433