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