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