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/qerror.h"
  20 #include "qapi/qmp/qdict.h"
  21 #include "qemu/ratelimit.h"
  22 #include "sysemu/block-backend.h"
  23 #include "block/copy-on-read.h"
  24
  25 enum {
  26     /*
  27      * Maximum chunk size to feed to copy-on-read.  This should be
  28      * large enough to process multiple clusters in a single call, so
  29      * that populating contiguous regions of the image is efficient.
  30      */
  31     STREAM_CHUNK = 512 * 1024, /* in bytes */
  32 };
  33
  34 typedef struct StreamBlockJob {
  35     BlockJob common;
  36     BlockBackend *blk;
  37     BlockDriverState *base_overlay; /* COW overlay (stream from this) */
  38     BlockDriverState *above_base;   /* Node directly above the base */
  39     BlockDriverState *cor_filter_bs;
  40     BlockDriverState *target_bs;
  41     BlockdevOnError on_error;
  42     char *backing_file_str;
  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 = bdrv_skip_filters(s->target_bs);
  58     BlockDriverState *base;
  59     BlockDriverState *unfiltered_base;
  60     Error *local_err = NULL;
  61     int ret = 0;
  62
  63     /* We should drop filter at this point, as filter hold the backing chain */
  64     bdrv_cor_filter_drop(s->cor_filter_bs);
  65     s->cor_filter_bs = NULL;
  66
  67     bdrv_subtree_drained_begin(s->above_base);
  68
  69     base = bdrv_filter_or_cow_bs(s->above_base);
  70     if (base) {
  71         bdrv_ref(base);
  72     }
  73
  74     unfiltered_base = bdrv_skip_filters(base);
  75
  76     if (bdrv_cow_child(unfiltered_bs)) {
  77         const char *base_id = NULL, *base_fmt = NULL;
  78         if (unfiltered_base) {
  79             base_id = s->backing_file_str ?: unfiltered_base->filename;
  80             if (unfiltered_base->drv) {
  81                 base_fmt = unfiltered_base->drv->format_name;
  82             }
  83         }
  84
  85         bdrv_set_backing_hd(unfiltered_bs, base, &local_err);
  86         ret = bdrv_change_backing_file(unfiltered_bs, base_id, base_fmt, false);
  87         if (local_err) {
  88             error_report_err(local_err);
  89             ret = -EPERM;
  90             goto out;
  91         }
  92     }
  93
  94 out:
  95     if (base) {
  96         bdrv_unref(base);
  97     }
  98     bdrv_subtree_drained_end(s->above_base);
  99     return ret;
 100 }
 101
 102 static void stream_clean(Job *job)
 103 {
 104     StreamBlockJob *s = container_of(job, StreamBlockJob, common.job);
 105
 106     if (s->cor_filter_bs) {
 107         bdrv_cor_filter_drop(s->cor_filter_bs);
 108         s->cor_filter_bs = NULL;
 109     }
 110
 111     blk_unref(s->blk);
 112     s->blk = NULL;
 113
 114     /* Reopen the image back in read-only mode if necessary */
 115     if (s->bs_read_only) {
 116         /* Give up write permissions before making it read-only */
 117         bdrv_reopen_set_read_only(s->target_bs, true, NULL);
 118     }
 119
 120     g_free(s->backing_file_str);
 121 }
 122
 123 static int coroutine_fn stream_run(Job *job, Error **errp)
 124 {
 125     StreamBlockJob *s = container_of(job, StreamBlockJob, common.job);
 126     BlockDriverState *unfiltered_bs = bdrv_skip_filters(s->target_bs);
 127     int64_t len;
 128     int64_t offset = 0;
 129     uint64_t delay_ns = 0;
 130     int error = 0;
 131     int64_t n = 0; /* bytes */
 132
 133     if (unfiltered_bs == s->base_overlay) {
 134         /* Nothing to stream */
 135         return 0;
 136     }
 137
 138     len = bdrv_getlength(s->target_bs);
 139     if (len < 0) {
 140         return len;
 141     }
 142     job_progress_set_remaining(&s->common.job, len);
 143
 144     for ( ; offset < len; offset += n) {
 145         bool copy;
 146         int ret;
 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         job_sleep_ns(&s->common.job, delay_ns);
 152         if (job_is_cancelled(&s->common.job)) {
 153             break;
 154         }
 155
 156         copy = false;
 157
 158         ret = bdrv_is_allocated(unfiltered_bs, offset, STREAM_CHUNK, &n);
 159         if (ret == 1) {
 160             /* Allocated in the top, no need to copy.  */
 161         } else if (ret >= 0) {
 162             /* Copy if allocated in the intermediate images.  Limit to the
 163              * known-unallocated area [offset, offset+n*BDRV_SECTOR_SIZE).  */
 164             ret = bdrv_is_allocated_above(bdrv_cow_bs(unfiltered_bs),
 165                                           s->base_overlay, true,
 166                                           offset, n, &n);
 167             /* Finish early if end of backing file has been reached */
 168             if (ret == 0 && n == 0) {
 169                 n = len - offset;
 170             }
 171
 172             copy = (ret > 0);
 173         }
 174         trace_stream_one_iteration(s, offset, n, ret);
 175         if (copy) {
 176             ret = stream_populate(s->blk, offset, n);
 177         }
 178         if (ret < 0) {
 179             BlockErrorAction action =
 180                 block_job_error_action(&s->common, s->on_error, true, -ret);
 181             if (action == BLOCK_ERROR_ACTION_STOP) {
 182                 n = 0;
 183                 continue;
 184             }
 185             if (error == 0) {
 186                 error = ret;
 187             }
 188             if (action == BLOCK_ERROR_ACTION_REPORT) {
 189                 break;
 190             }
 191         }
 192
 193         /* Publish progress */
 194         job_progress_update(&s->common.job, n);
 195         if (copy) {
 196             delay_ns = block_job_ratelimit_get_delay(&s->common, n);
 197         } else {
 198             delay_ns = 0;
 199         }
 200     }
 201
 202     /* Do not remove the backing file if an error was there but ignored. */
 203     return error;
 204 }
 205
 206 static const BlockJobDriver stream_job_driver = {
 207     .job_driver = {
 208         .instance_size = sizeof(StreamBlockJob),
 209         .job_type      = JOB_TYPE_STREAM,
 210         .free          = block_job_free,
 211         .run           = stream_run,
 212         .prepare       = stream_prepare,
 213         .clean         = stream_clean,
 214         .user_resume   = block_job_user_resume,
 215     },
 216 };
 217
 218 void stream_start(const char *job_id, BlockDriverState *bs,
 219                   BlockDriverState *base, const char *backing_file_str,
 220                   BlockDriverState *bottom,
 221                   int creation_flags, int64_t speed,
 222                   BlockdevOnError on_error,
 223                   const char *filter_node_name,
 224                   Error **errp)
 225 {
 226     StreamBlockJob *s = NULL;
 227     BlockDriverState *iter;
 228     bool bs_read_only;
 229     int basic_flags = BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED;
 230     BlockDriverState *base_overlay;
 231     BlockDriverState *cor_filter_bs = NULL;
 232     BlockDriverState *above_base;
 233     QDict *opts;
 234     int ret;
 235
 236     GLOBAL_STATE_CODE();
 237
 238     assert(!(base && bottom));
 239     assert(!(backing_file_str && bottom));
 240
 241     if (bottom) {
 242         /*
 243          * New simple interface. The code is written in terms of old interface
 244          * with @base parameter (still, it doesn't freeze link to base, so in
 245          * this mean old code is correct for new interface). So, for now, just
 246          * emulate base_overlay and above_base. Still, when old interface
 247          * finally removed, we should refactor code to use only "bottom", but
 248          * not "*base*" things.
 249          */
 250         assert(!bottom->drv->is_filter);
 251         base_overlay = above_base = bottom;
 252     } else {
 253         base_overlay = bdrv_find_overlay(bs, base);
 254         if (!base_overlay) {
 255             error_setg(errp, "'%s' is not in the backing chain of '%s'",
 256                        base->node_name, bs->node_name);
 257             return;
 258         }
 259
 260         /*
 261          * Find the node directly above @base.  @base_overlay is a COW overlay,
 262          * so it must have a bdrv_cow_child(), but it is the immediate overlay
 263          * of @base, so between the two there can only be filters.
 264          */
 265         above_base = base_overlay;
 266         if (bdrv_cow_bs(above_base) != base) {
 267             above_base = bdrv_cow_bs(above_base);
 268             while (bdrv_filter_bs(above_base) != base) {
 269                 above_base = bdrv_filter_bs(above_base);
 270             }
 271         }
 272     }
 273
 274     /* Make sure that the image is opened in read-write mode */
 275     bs_read_only = bdrv_is_read_only(bs);
 276     if (bs_read_only) {
 277         int ret;
 278         /* Hold the chain during reopen */
 279         if (bdrv_freeze_backing_chain(bs, above_base, errp) < 0) {
 280             return;
 281         }
 282
 283         ret = bdrv_reopen_set_read_only(bs, false, errp);
 284
 285         /* failure, or cor-filter will hold the chain */
 286         bdrv_unfreeze_backing_chain(bs, above_base);
 287
 288         if (ret < 0) {
 289             return;
 290         }
 291     }
 292
 293     opts = qdict_new();
 294
 295     qdict_put_str(opts, "driver", "copy-on-read");
 296     qdict_put_str(opts, "file", bdrv_get_node_name(bs));
 297     /* Pass the base_overlay node name as 'bottom' to COR driver */
 298     qdict_put_str(opts, "bottom", base_overlay->node_name);
 299     if (filter_node_name) {
 300         qdict_put_str(opts, "node-name", filter_node_name);
 301     }
 302
 303     cor_filter_bs = bdrv_insert_node(bs, opts, BDRV_O_RDWR, errp);
 304     if (!cor_filter_bs) {
 305         goto fail;
 306     }
 307
 308     if (!filter_node_name) {
 309         cor_filter_bs->implicit = true;
 310     }
 311
 312     s = block_job_create(job_id, &stream_job_driver, NULL, cor_filter_bs,
 313                          0, BLK_PERM_ALL,
 314                          speed, creation_flags, NULL, NULL, errp);
 315     if (!s) {
 316         goto fail;
 317     }
 318
 319     s->blk = blk_new_with_bs(cor_filter_bs, BLK_PERM_CONSISTENT_READ,
 320                              basic_flags | BLK_PERM_WRITE, errp);
 321     if (!s->blk) {
 322         goto fail;
 323     }
 324     /*
 325      * Disable request queuing in the BlockBackend to avoid deadlocks on drain:
 326      * The job reports that it's busy until it reaches a pause point.
 327      */
 328     blk_set_disable_request_queuing(s->blk, true);
 329     blk_set_allow_aio_context_change(s->blk, true);
 330
 331     /*
 332      * Prevent concurrent jobs trying to modify the graph structure here, we
 333      * already have our own plans. Also don't allow resize as the image size is
 334      * queried only at the job start and then cached.
 335      */
 336     if (block_job_add_bdrv(&s->common, "active node", bs, 0,
 337                            basic_flags | BLK_PERM_WRITE, errp)) {
 338         goto fail;
 339     }
 340
 341     /* Block all intermediate nodes between bs and base, because they will
 342      * disappear from the chain after this operation. The streaming job reads
 343      * every block only once, assuming that it doesn't change, so forbid writes
 344      * and resizes. Reassign the base node pointer because the backing BS of the
 345      * bottom node might change after the call to bdrv_reopen_set_read_only()
 346      * due to parallel block jobs running.
 347      * above_base node might change after the call to
 348      * bdrv_reopen_set_read_only() due to parallel block jobs running.
 349      */
 350     base = bdrv_filter_or_cow_bs(above_base);
 351     for (iter = bdrv_filter_or_cow_bs(bs); iter != base;
 352          iter = bdrv_filter_or_cow_bs(iter))
 353     {
 354         ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
 355                                  basic_flags, errp);
 356         if (ret < 0) {
 357             goto fail;
 358         }
 359     }
 360
 361     s->base_overlay = base_overlay;
 362     s->above_base = above_base;
 363     s->backing_file_str = g_strdup(backing_file_str);
 364     s->cor_filter_bs = cor_filter_bs;
 365     s->target_bs = bs;
 366     s->bs_read_only = bs_read_only;
 367
 368     s->on_error = on_error;
 369     trace_stream_start(bs, base, s);
 370     job_start(&s->common.job);
 371     return;
 372
 373 fail:
 374     if (s) {
 375         job_early_fail(&s->common.job);
 376     }
 377     if (cor_filter_bs) {
 378         bdrv_cor_filter_drop(cor_filter_bs);
 379     }
 380     if (bs_read_only) {
 381         bdrv_reopen_set_read_only(bs, true, NULL);
 382     }
 383 }