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