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