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_int.h"
  16 #include "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_id[1024];
  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     while (intermediate) {
  61         BlockDriverState *unused;
  62
  63         /* reached base */
  64         if (intermediate == base) {
  65             break;
  66         }
  67
  68         unused = intermediate;
  69         intermediate = intermediate->backing_hd;
  70         unused->backing_hd = NULL;
  71         bdrv_delete(unused);
  72     }
  73     top->backing_hd = base;
  74 }
  75
  76 static void coroutine_fn stream_run(void *opaque)
  77 {
  78     StreamBlockJob *s = opaque;
  79     BlockDriverState *bs = s->common.bs;
  80     BlockDriverState *base = s->base;
  81     int64_t sector_num, end;
  82     int error = 0;
  83     int ret = 0;
  84     int n = 0;
  85     void *buf;
  86
  87     s->common.len = bdrv_getlength(bs);
  88     if (s->common.len < 0) {
  89         block_job_completed(&s->common, s->common.len);
  90         return;
  91     }
  92
  93     end = s->common.len >> BDRV_SECTOR_BITS;
  94     buf = qemu_blockalign(bs, STREAM_BUFFER_SIZE);
  95
  96     /* Turn on copy-on-read for the whole block device so that guest read
  97      * requests help us make progress.  Only do this when copying the entire
  98      * backing chain since the copy-on-read operation does not take base into
  99      * account.
 100      */
 101     if (!base) {
 102         bdrv_enable_copy_on_read(bs);
 103     }
 104
 105     for (sector_num = 0; sector_num < end; sector_num += n) {
 106         uint64_t delay_ns = 0;
 107         bool copy;
 108
 109 wait:
 110         /* Note that even when no rate limit is applied we need to yield
 111          * with no pending I/O here so that qemu_aio_flush() returns.
 112          */
 113         block_job_sleep_ns(&s->common, rt_clock, delay_ns);
 114         if (block_job_is_cancelled(&s->common)) {
 115             break;
 116         }
 117
 118         ret = bdrv_co_is_allocated(bs, sector_num,
 119                                    STREAM_BUFFER_SIZE / BDRV_SECTOR_SIZE, &n);
 120         if (ret == 1) {
 121             /* Allocated in the top, no need to copy.  */
 122             copy = false;
 123         } else {
 124             /* Copy if allocated in the intermediate images.  Limit to the
 125              * known-unallocated area [sector_num, sector_num+n).  */
 126             ret = bdrv_co_is_allocated_above(bs->backing_hd, base,
 127                                              sector_num, n, &n);
 128
 129             /* Finish early if end of backing file has been reached */
 130             if (ret == 0 && n == 0) {
 131                 n = end - sector_num;
 132             }
 133
 134             copy = (ret == 1);
 135         }
 136         trace_stream_one_iteration(s, sector_num, n, ret);
 137         if (ret >= 0 && copy) {
 138             if (s->common.speed) {
 139                 delay_ns = ratelimit_calculate_delay(&s->limit, n);
 140                 if (delay_ns > 0) {
 141                     goto wait;
 142                 }
 143             }
 144             ret = stream_populate(bs, sector_num, n, buf);
 145         }
 146         if (ret < 0) {
 147             BlockErrorAction action =
 148                 block_job_error_action(&s->common, s->common.bs, s->on_error,
 149                                        true, -ret);
 150             if (action == BDRV_ACTION_STOP) {
 151                 n = 0;
 152                 continue;
 153             }
 154             if (error == 0) {
 155                 error = ret;
 156             }
 157             if (action == BDRV_ACTION_REPORT) {
 158                 break;
 159             }
 160         }
 161         ret = 0;
 162
 163         /* Publish progress */
 164         s->common.offset += n * BDRV_SECTOR_SIZE;
 165     }
 166
 167     if (!base) {
 168         bdrv_disable_copy_on_read(bs);
 169     }
 170
 171     /* Do not remove the backing file if an error was there but ignored.  */
 172     ret = error;
 173
 174     if (!block_job_is_cancelled(&s->common) && sector_num == end && ret == 0) {
 175         const char *base_id = NULL, *base_fmt = NULL;
 176         if (base) {
 177             base_id = s->backing_file_id;
 178             if (base->drv) {
 179                 base_fmt = base->drv->format_name;
 180             }
 181         }
 182         ret = bdrv_change_backing_file(bs, base_id, base_fmt);
 183         close_unused_images(bs, base, base_id);
 184     }
 185
 186     qemu_vfree(buf);
 187     block_job_completed(&s->common, ret);
 188 }
 189
 190 static void stream_set_speed(BlockJob *job, int64_t speed, Error **errp)
 191 {
 192     StreamBlockJob *s = container_of(job, StreamBlockJob, common);
 193
 194     if (speed < 0) {
 195         error_set(errp, QERR_INVALID_PARAMETER, "speed");
 196         return;
 197     }
 198     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 199 }
 200
 201 static BlockJobType stream_job_type = {
 202     .instance_size = sizeof(StreamBlockJob),
 203     .job_type      = "stream",
 204     .set_speed     = stream_set_speed,
 205 };
 206
 207 void stream_start(BlockDriverState *bs, BlockDriverState *base,
 208                   const char *base_id, int64_t speed,
 209                   BlockdevOnError on_error,
 210                   BlockDriverCompletionFunc *cb,
 211                   void *opaque, Error **errp)
 212 {
 213     StreamBlockJob *s;
 214
 215     if ((on_error == BLOCKDEV_ON_ERROR_STOP ||
 216          on_error == BLOCKDEV_ON_ERROR_ENOSPC) &&
 217         !bdrv_iostatus_is_enabled(bs)) {
 218         error_set(errp, QERR_INVALID_PARAMETER, "on-error");
 219         return;
 220     }
 221
 222     s = block_job_create(&stream_job_type, bs, speed, cb, opaque, errp);
 223     if (!s) {
 224         return;
 225     }
 226
 227     s->base = base;
 228     if (base_id) {
 229         pstrcpy(s->backing_file_id, sizeof(s->backing_file_id), base_id);
 230     }
 231
 232     s->on_error = on_error;
 233     s->common.co = qemu_coroutine_create(stream_run);
 234     trace_stream_start(bs, base, s, s->common.co, opaque);
 235     qemu_coroutine_enter(s->common.co, s);
 236 }