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 "trace.h"
  15 #include "block/block_int.h"
  16 #include "block/blockjob.h"
  17 #include "qemu/ratelimit.h"
  18
  19 enum {
  20     /*
  21      * Size of data buffer for populating the image file.  This should be large
  22      * enough to process multiple clusters in a single call, so that populating
  23      * contiguous regions of the image is efficient.
  24      */
  25     STREAM_BUFFER_SIZE = 512 * 1024, /* in bytes */
  26 };
  27
  28 #define SLICE_TIME 100000000ULL /* ns */
  29
  30 typedef struct StreamBlockJob {
  31     BlockJob common;
  32     RateLimit limit;
  33     BlockDriverState *base;
  34     BlockdevOnError on_error;
  35     char *backing_file_str;
  36 } StreamBlockJob;
  37
  38 static int coroutine_fn stream_populate(BlockDriverState *bs,
  39                                         int64_t sector_num, int nb_sectors,
  40                                         void *buf)
  41 {
  42     struct iovec iov = {
  43         .iov_base = buf,
  44         .iov_len  = nb_sectors * BDRV_SECTOR_SIZE,
  45     };
  46     QEMUIOVector qiov;
  47
  48     qemu_iovec_init_external(&qiov, &iov, 1);
  49
  50     /* Copy-on-read the unallocated clusters */
  51     return bdrv_co_copy_on_readv(bs, sector_num, nb_sectors, &qiov);
  52 }
  53
  54 static void close_unused_images(BlockDriverState *top, BlockDriverState *base,
  55                                 const char *base_id)
  56 {
  57     BlockDriverState *intermediate;
  58     intermediate = top->backing_hd;
  59
  60     /* Must assign before bdrv_delete() to prevent traversing dangling pointer
  61      * while we delete backing image instances.
  62      */
  63     bdrv_set_backing_hd(top, base);
  64
  65     while (intermediate) {
  66         BlockDriverState *unused;
  67
  68         /* reached base */
  69         if (intermediate == base) {
  70             break;
  71         }
  72
  73         unused = intermediate;
  74         intermediate = intermediate->backing_hd;
  75         bdrv_set_backing_hd(unused, NULL);
  76         bdrv_unref(unused);
  77     }
  78
  79     bdrv_refresh_limits(top, NULL);
  80 }
  81
  82 typedef struct {
  83     int ret;
  84     bool reached_end;
  85 } StreamCompleteData;
  86
  87 static void stream_complete(BlockJob *job, void *opaque)
  88 {
  89     StreamBlockJob *s = container_of(job, StreamBlockJob, common);
  90     StreamCompleteData *data = opaque;
  91     BlockDriverState *base = s->base;
  92
  93     if (!block_job_is_cancelled(&s->common) && data->reached_end &&
  94         data->ret == 0) {
  95         const char *base_id = NULL, *base_fmt = NULL;
  96         if (base) {
  97             base_id = s->backing_file_str;
  98             if (base->drv) {
  99                 base_fmt = base->drv->format_name;
 100             }
 101         }
 102         data->ret = bdrv_change_backing_file(job->bs, base_id, base_fmt);
 103         close_unused_images(job->bs, base, base_id);
 104     }
 105
 106     g_free(s->backing_file_str);
 107     block_job_completed(&s->common, data->ret);
 108     g_free(data);
 109 }
 110
 111 static void coroutine_fn stream_run(void *opaque)
 112 {
 113     StreamBlockJob *s = opaque;
 114     StreamCompleteData *data;
 115     BlockDriverState *bs = s->common.bs;
 116     BlockDriverState *base = s->base;
 117     int64_t sector_num, end;
 118     int error = 0;
 119     int ret = 0;
 120     int n = 0;
 121     void *buf;
 122
 123     if (!bs->backing_hd) {
 124         block_job_completed(&s->common, 0);
 125         return;
 126     }
 127
 128     s->common.len = bdrv_getlength(bs);
 129     if (s->common.len < 0) {
 130         block_job_completed(&s->common, s->common.len);
 131         return;
 132     }
 133
 134     end = s->common.len >> BDRV_SECTOR_BITS;
 135     buf = qemu_blockalign(bs, STREAM_BUFFER_SIZE);
 136
 137     /* Turn on copy-on-read for the whole block device so that guest read
 138      * requests help us make progress.  Only do this when copying the entire
 139      * backing chain since the copy-on-read operation does not take base into
 140      * account.
 141      */
 142     if (!base) {
 143         bdrv_enable_copy_on_read(bs);
 144     }
 145
 146     for (sector_num = 0; sector_num < end; sector_num += n) {
 147         uint64_t delay_ns = 0;
 148         bool copy;
 149
 150 wait:
 151         /* Note that even when no rate limit is applied we need to yield
 152          * with no pending I/O here so that bdrv_drain_all() returns.
 153          */
 154         block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 155         if (block_job_is_cancelled(&s->common)) {
 156             break;
 157         }
 158
 159         copy = false;
 160
 161         ret = bdrv_is_allocated(bs, sector_num,
 162                                 STREAM_BUFFER_SIZE / BDRV_SECTOR_SIZE, &n);
 163         if (ret == 1) {
 164             /* Allocated in the top, no need to copy.  */
 165         } else if (ret >= 0) {
 166             /* Copy if allocated in the intermediate images.  Limit to the
 167              * known-unallocated area [sector_num, sector_num+n).  */
 168             ret = bdrv_is_allocated_above(bs->backing_hd, base,
 169                                           sector_num, n, &n);
 170
 171             /* Finish early if end of backing file has been reached */
 172             if (ret == 0 && n == 0) {
 173                 n = end - sector_num;
 174             }
 175
 176             copy = (ret == 1);
 177         }
 178         trace_stream_one_iteration(s, sector_num, n, ret);
 179         if (copy) {
 180             if (s->common.speed) {
 181                 delay_ns = ratelimit_calculate_delay(&s->limit, n);
 182                 if (delay_ns > 0) {
 183                     goto wait;
 184                 }
 185             }
 186             ret = stream_populate(bs, sector_num, n, buf);
 187         }
 188         if (ret < 0) {
 189             BlockErrorAction action =
 190                 block_job_error_action(&s->common, s->common.bs, s->on_error,
 191                                        true, -ret);
 192             if (action == BLOCK_ERROR_ACTION_STOP) {
 193                 n = 0;
 194                 continue;
 195             }
 196             if (error == 0) {
 197                 error = ret;
 198             }
 199             if (action == BLOCK_ERROR_ACTION_REPORT) {
 200                 break;
 201             }
 202         }
 203         ret = 0;
 204
 205         /* Publish progress */
 206         s->common.offset += n * BDRV_SECTOR_SIZE;
 207     }
 208
 209     if (!base) {
 210         bdrv_disable_copy_on_read(bs);
 211     }
 212
 213     /* Do not remove the backing file if an error was there but ignored.  */
 214     ret = error;
 215
 216     qemu_vfree(buf);
 217
 218     /* Modify backing chain and close BDSes in main loop */
 219     data = g_malloc(sizeof(*data));
 220     data->ret = ret;
 221     data->reached_end = sector_num == end;
 222     block_job_defer_to_main_loop(&s->common, stream_complete, data);
 223 }
 224
 225 static void stream_set_speed(BlockJob *job, int64_t speed, Error **errp)
 226 {
 227     StreamBlockJob *s = container_of(job, StreamBlockJob, common);
 228
 229     if (speed < 0) {
 230         error_set(errp, QERR_INVALID_PARAMETER, "speed");
 231         return;
 232     }
 233     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 234 }
 235
 236 static const BlockJobDriver stream_job_driver = {
 237     .instance_size = sizeof(StreamBlockJob),
 238     .job_type      = BLOCK_JOB_TYPE_STREAM,
 239     .set_speed     = stream_set_speed,
 240 };
 241
 242 void stream_start(BlockDriverState *bs, BlockDriverState *base,
 243                   const char *backing_file_str, int64_t speed,
 244                   BlockdevOnError on_error,
 245                   BlockCompletionFunc *cb,
 246                   void *opaque, Error **errp)
 247 {
 248     StreamBlockJob *s;
 249
 250     if ((on_error == BLOCKDEV_ON_ERROR_STOP ||
 251          on_error == BLOCKDEV_ON_ERROR_ENOSPC) &&
 252         !bdrv_iostatus_is_enabled(bs)) {
 253         error_set(errp, QERR_INVALID_PARAMETER, "on-error");
 254         return;
 255     }
 256
 257     s = block_job_create(&stream_job_driver, bs, speed, cb, opaque, errp);
 258     if (!s) {
 259         return;
 260     }
 261
 262     s->base = base;
 263     s->backing_file_str = g_strdup(backing_file_str);
 264
 265     s->on_error = on_error;
 266     s->common.co = qemu_coroutine_create(stream_run);
 267     trace_stream_start(bs, base, s, s->common.co, opaque);
 268     qemu_coroutine_enter(s->common.co, s);
 269 }