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