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     /*
  68      * bdrv_set_backing_hd() requires that unfiltered_bs is drained. Drain
  69      * already here and use bdrv_set_backing_hd_drained() instead because
  70      * the polling during drained_begin() might change the graph, and if we do
  71      * this only later, we may end up working with the wrong base node (or it
  72      * might even have gone away by the time we want to use it).
  73      */
  74     bdrv_drained_begin(unfiltered_bs);
  75
  76     base = bdrv_filter_or_cow_bs(s->above_base);
  77     unfiltered_base = bdrv_skip_filters(base);
  78
  79     if (bdrv_cow_child(unfiltered_bs)) {
  80         const char *base_id = NULL, *base_fmt = NULL;
  81         if (unfiltered_base) {
  82             base_id = s->backing_file_str ?: unfiltered_base->filename;
  83             if (unfiltered_base->drv) {
  84                 base_fmt = unfiltered_base->drv->format_name;
  85             }
  86         }
  87
  88         bdrv_set_backing_hd_drained(unfiltered_bs, base, &local_err);
  89
  90         /*
  91          * This call will do I/O, so the graph can change again from here on.
  92          * We have already completed the graph change, so we are not in danger
  93          * of operating on the wrong node any more if this happens.
  94          */
  95         ret = bdrv_change_backing_file(unfiltered_bs, base_id, base_fmt, false);
  96         if (local_err) {
  97             error_report_err(local_err);
  98             ret = -EPERM;
  99             goto out;
 100         }
 101     }
 102
 103 out:
 104     bdrv_drained_end(unfiltered_bs);
 105     return ret;
 106 }
 107
 108 static void stream_clean(Job *job)
 109 {
 110     StreamBlockJob *s = container_of(job, StreamBlockJob, common.job);
 111
 112     if (s->cor_filter_bs) {
 113         bdrv_cor_filter_drop(s->cor_filter_bs);
 114         s->cor_filter_bs = NULL;
 115     }
 116
 117     blk_unref(s->blk);
 118     s->blk = NULL;
 119
 120     /* Reopen the image back in read-only mode if necessary */
 121     if (s->bs_read_only) {
 122         /* Give up write permissions before making it read-only */
 123         bdrv_reopen_set_read_only(s->target_bs, true, NULL);
 124     }
 125
 126     g_free(s->backing_file_str);
 127 }
 128
 129 static int coroutine_fn stream_run(Job *job, Error **errp)
 130 {
 131     StreamBlockJob *s = container_of(job, StreamBlockJob, common.job);
 132     BlockDriverState *unfiltered_bs = bdrv_skip_filters(s->target_bs);
 133     int64_t len;
 134     int64_t offset = 0;
 135     uint64_t delay_ns = 0;
 136     int error = 0;
 137     int64_t n = 0; /* bytes */
 138
 139     if (unfiltered_bs == s->base_overlay) {
 140         /* Nothing to stream */
 141         return 0;
 142     }
 143
 144     len = bdrv_getlength(s->target_bs);
 145     if (len < 0) {
 146         return len;
 147     }
 148     job_progress_set_remaining(&s->common.job, len);
 149
 150     for ( ; offset < len; offset += n) {
 151         bool copy;
 152         int ret;
 153
 154         /* Note that even when no rate limit is applied we need to yield
 155          * with no pending I/O here so that bdrv_drain_all() returns.
 156          */
 157         job_sleep_ns(&s->common.job, delay_ns);
 158         if (job_is_cancelled(&s->common.job)) {
 159             break;
 160         }
 161
 162         copy = false;
 163
 164         ret = bdrv_is_allocated(unfiltered_bs, offset, STREAM_CHUNK, &n);
 165         if (ret == 1) {
 166             /* Allocated in the top, no need to copy.  */
 167         } else if (ret >= 0) {
 168             /* Copy if allocated in the intermediate images.  Limit to the
 169              * known-unallocated area [offset, offset+n*BDRV_SECTOR_SIZE).  */
 170             ret = bdrv_is_allocated_above(bdrv_cow_bs(unfiltered_bs),
 171                                           s->base_overlay, true,
 172                                           offset, n, &n);
 173             /* Finish early if end of backing file has been reached */
 174             if (ret == 0 && n == 0) {
 175                 n = len - offset;
 176             }
 177
 178             copy = (ret > 0);
 179         }
 180         trace_stream_one_iteration(s, offset, n, ret);
 181         if (copy) {
 182             ret = stream_populate(s->blk, offset, n);
 183         }
 184         if (ret < 0) {
 185             BlockErrorAction action =
 186                 block_job_error_action(&s->common, s->on_error, true, -ret);
 187             if (action == BLOCK_ERROR_ACTION_STOP) {
 188                 n = 0;
 189                 continue;
 190             }
 191             if (error == 0) {
 192                 error = ret;
 193             }
 194             if (action == BLOCK_ERROR_ACTION_REPORT) {
 195                 break;
 196             }
 197         }
 198
 199         /* Publish progress */
 200         job_progress_update(&s->common.job, n);
 201         if (copy) {
 202             delay_ns = block_job_ratelimit_get_delay(&s->common, n);
 203         } else {
 204             delay_ns = 0;
 205         }
 206     }
 207
 208     /* Do not remove the backing file if an error was there but ignored. */
 209     return error;
 210 }
 211
 212 static const BlockJobDriver stream_job_driver = {
 213     .job_driver = {
 214         .instance_size = sizeof(StreamBlockJob),
 215         .job_type      = JOB_TYPE_STREAM,
 216         .free          = block_job_free,
 217         .run           = stream_run,
 218         .prepare       = stream_prepare,
 219         .clean         = stream_clean,
 220         .user_resume   = block_job_user_resume,
 221     },
 222 };
 223
 224 void stream_start(const char *job_id, BlockDriverState *bs,
 225                   BlockDriverState *base, const char *backing_file_str,
 226                   BlockDriverState *bottom,
 227                   int creation_flags, int64_t speed,
 228                   BlockdevOnError on_error,
 229                   const char *filter_node_name,
 230                   Error **errp)
 231 {
 232     StreamBlockJob *s = NULL;
 233     BlockDriverState *iter;
 234     bool bs_read_only;
 235     int basic_flags = BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED;
 236     BlockDriverState *base_overlay;
 237     BlockDriverState *cor_filter_bs = NULL;
 238     BlockDriverState *above_base;
 239     QDict *opts;
 240     int ret;
 241
 242     GLOBAL_STATE_CODE();
 243
 244     assert(!(base && bottom));
 245     assert(!(backing_file_str && bottom));
 246
 247     if (bottom) {
 248         /*
 249          * New simple interface. The code is written in terms of old interface
 250          * with @base parameter (still, it doesn't freeze link to base, so in
 251          * this mean old code is correct for new interface). So, for now, just
 252          * emulate base_overlay and above_base. Still, when old interface
 253          * finally removed, we should refactor code to use only "bottom", but
 254          * not "*base*" things.
 255          */
 256         assert(!bottom->drv->is_filter);
 257         base_overlay = above_base = bottom;
 258     } else {
 259         base_overlay = bdrv_find_overlay(bs, base);
 260         if (!base_overlay) {
 261             error_setg(errp, "'%s' is not in the backing chain of '%s'",
 262                        base->node_name, bs->node_name);
 263             return;
 264         }
 265
 266         /*
 267          * Find the node directly above @base.  @base_overlay is a COW overlay,
 268          * so it must have a bdrv_cow_child(), but it is the immediate overlay
 269          * of @base, so between the two there can only be filters.
 270          */
 271         above_base = base_overlay;
 272         if (bdrv_cow_bs(above_base) != base) {
 273             above_base = bdrv_cow_bs(above_base);
 274             while (bdrv_filter_bs(above_base) != base) {
 275                 above_base = bdrv_filter_bs(above_base);
 276             }
 277         }
 278     }
 279
 280     /* Make sure that the image is opened in read-write mode */
 281     bs_read_only = bdrv_is_read_only(bs);
 282     if (bs_read_only) {
 283         int ret;
 284         /* Hold the chain during reopen */
 285         if (bdrv_freeze_backing_chain(bs, above_base, errp) < 0) {
 286             return;
 287         }
 288
 289         ret = bdrv_reopen_set_read_only(bs, false, errp);
 290
 291         /* failure, or cor-filter will hold the chain */
 292         bdrv_unfreeze_backing_chain(bs, above_base);
 293
 294         if (ret < 0) {
 295             return;
 296         }
 297     }
 298
 299     opts = qdict_new();
 300
 301     qdict_put_str(opts, "driver", "copy-on-read");
 302     qdict_put_str(opts, "file", bdrv_get_node_name(bs));
 303     /* Pass the base_overlay node name as 'bottom' to COR driver */
 304     qdict_put_str(opts, "bottom", base_overlay->node_name);
 305     if (filter_node_name) {
 306         qdict_put_str(opts, "node-name", filter_node_name);
 307     }
 308
 309     cor_filter_bs = bdrv_insert_node(bs, opts, BDRV_O_RDWR, errp);
 310     if (!cor_filter_bs) {
 311         goto fail;
 312     }
 313
 314     if (!filter_node_name) {
 315         cor_filter_bs->implicit = true;
 316     }
 317
 318     s = block_job_create(job_id, &stream_job_driver, NULL, cor_filter_bs,
 319                          0, BLK_PERM_ALL,
 320                          speed, creation_flags, NULL, NULL, errp);
 321     if (!s) {
 322         goto fail;
 323     }
 324
 325     s->blk = blk_new_with_bs(cor_filter_bs, BLK_PERM_CONSISTENT_READ,
 326                              basic_flags | BLK_PERM_WRITE, errp);
 327     if (!s->blk) {
 328         goto fail;
 329     }
 330     /*
 331      * Disable request queuing in the BlockBackend to avoid deadlocks on drain:
 332      * The job reports that it's busy until it reaches a pause point.
 333      */
 334     blk_set_disable_request_queuing(s->blk, true);
 335     blk_set_allow_aio_context_change(s->blk, true);
 336
 337     /*
 338      * Prevent concurrent jobs trying to modify the graph structure here, we
 339      * already have our own plans. Also don't allow resize as the image size is
 340      * queried only at the job start and then cached.
 341      */
 342     if (block_job_add_bdrv(&s->common, "active node", bs, 0,
 343                            basic_flags | BLK_PERM_WRITE, errp)) {
 344         goto fail;
 345     }
 346
 347     /* Block all intermediate nodes between bs and base, because they will
 348      * disappear from the chain after this operation. The streaming job reads
 349      * every block only once, assuming that it doesn't change, so forbid writes
 350      * and resizes. Reassign the base node pointer because the backing BS of the
 351      * bottom node might change after the call to bdrv_reopen_set_read_only()
 352      * due to parallel block jobs running.
 353      * above_base node might change after the call to
 354      * bdrv_reopen_set_read_only() due to parallel block jobs running.
 355      */
 356     base = bdrv_filter_or_cow_bs(above_base);
 357     for (iter = bdrv_filter_or_cow_bs(bs); iter != base;
 358          iter = bdrv_filter_or_cow_bs(iter))
 359     {
 360         ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
 361                                  basic_flags, errp);
 362         if (ret < 0) {
 363             goto fail;
 364         }
 365     }
 366
 367     s->base_overlay = base_overlay;
 368     s->above_base = above_base;
 369     s->backing_file_str = g_strdup(backing_file_str);
 370     s->cor_filter_bs = cor_filter_bs;
 371     s->target_bs = bs;
 372     s->bs_read_only = bs_read_only;
 373
 374     s->on_error = on_error;
 375     trace_stream_start(bs, base, s);
 376     job_start(&s->common.job);
 377     return;
 378
 379 fail:
 380     if (s) {
 381         job_early_fail(&s->common.job);
 382     }
 383     if (cor_filter_bs) {
 384         bdrv_cor_filter_drop(cor_filter_bs);
 385     }
 386     if (bs_read_only) {
 387         bdrv_reopen_set_read_only(bs, true, NULL);
 388     }
 389 }