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
  17 enum {
  18     /*
  19      * Size of data buffer for populating the image file.  This should be large
  20      * enough to process multiple clusters in a single call, so that populating
  21      * contiguous regions of the image is efficient.
  22      */
  23     STREAM_BUFFER_SIZE = 512 * 1024, /* in bytes */
  24 };
  25
  26 #define SLICE_TIME 100000000ULL /* ns */
  27
  28 typedef struct {
  29     int64_t next_slice_time;
  30     uint64_t slice_quota;
  31     uint64_t dispatched;
  32 } RateLimit;
  33
  34 static int64_t ratelimit_calculate_delay(RateLimit *limit, uint64_t n)
  35 {
  36     int64_t delay_ns = 0;
  37     int64_t now = qemu_get_clock_ns(rt_clock);
  38
  39     if (limit->next_slice_time < now) {
  40         limit->next_slice_time = now + SLICE_TIME;
  41         limit->dispatched = 0;
  42     }
  43     if (limit->dispatched + n > limit->slice_quota) {
  44         delay_ns = limit->next_slice_time - now;
  45     } else {
  46         limit->dispatched += n;
  47     }
  48     return delay_ns;
  49 }
  50
  51 static void ratelimit_set_speed(RateLimit *limit, uint64_t speed)
  52 {
  53     limit->slice_quota = speed / (1000000000ULL / SLICE_TIME);
  54 }
  55
  56 typedef struct StreamBlockJob {
  57     BlockJob common;
  58     RateLimit limit;
  59     BlockDriverState *base;
  60 } StreamBlockJob;
  61
  62 static int coroutine_fn stream_populate(BlockDriverState *bs,
  63                                         int64_t sector_num, int nb_sectors,
  64                                         void *buf)
  65 {
  66     struct iovec iov = {
  67         .iov_base = buf,
  68         .iov_len  = nb_sectors * BDRV_SECTOR_SIZE,
  69     };
  70     QEMUIOVector qiov;
  71
  72     qemu_iovec_init_external(&qiov, &iov, 1);
  73
  74     /* Copy-on-read the unallocated clusters */
  75     return bdrv_co_copy_on_readv(bs, sector_num, nb_sectors, &qiov);
  76 }
  77
  78 static void coroutine_fn stream_run(void *opaque)
  79 {
  80     StreamBlockJob *s = opaque;
  81     BlockDriverState *bs = s->common.bs;
  82     int64_t sector_num, end;
  83     int ret = 0;
  84     int n;
  85     void *buf;
  86
  87     s->common.len = bdrv_getlength(bs);
  88     if (s->common.len < 0) {
  89         block_job_complete(&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 retry:
 107         if (block_job_is_cancelled(&s->common)) {
 108             break;
 109         }
 110
 111         ret = bdrv_co_is_allocated(bs, sector_num,
 112                                    STREAM_BUFFER_SIZE / BDRV_SECTOR_SIZE, &n);
 113         trace_stream_one_iteration(s, sector_num, n, ret);
 114         if (ret == 0) {
 115             if (s->common.speed) {
 116                 uint64_t delay_ns = ratelimit_calculate_delay(&s->limit, n);
 117                 if (delay_ns > 0) {
 118                     co_sleep_ns(rt_clock, delay_ns);
 119
 120                     /* Recheck cancellation and that sectors are unallocated */
 121                     goto retry;
 122                 }
 123             }
 124             ret = stream_populate(bs, sector_num, n, buf);
 125         }
 126         if (ret < 0) {
 127             break;
 128         }
 129
 130         /* Publish progress */
 131         s->common.offset += n * BDRV_SECTOR_SIZE;
 132
 133         /* Note that even when no rate limit is applied we need to yield
 134          * with no pending I/O here so that qemu_aio_flush() returns.
 135          */
 136         co_sleep_ns(rt_clock, 0);
 137     }
 138
 139     if (!base) {
 140         bdrv_disable_copy_on_read(bs);
 141     }
 142
 143     if (sector_num == end && ret == 0) {
 144         ret = bdrv_change_backing_file(bs, NULL, NULL);
 145     }
 146
 147     qemu_vfree(buf);
 148     block_job_complete(&s->common, ret);
 149 }
 150
 151 static int stream_set_speed(BlockJob *job, int64_t value)
 152 {
 153     StreamBlockJob *s = container_of(job, StreamBlockJob, common);
 154
 155     if (value < 0) {
 156         return -EINVAL;
 157     }
 158     job->speed = value;
 159     ratelimit_set_speed(&s->limit, value / BDRV_SECTOR_SIZE);
 160     return 0;
 161 }
 162
 163 static BlockJobType stream_job_type = {
 164     .instance_size = sizeof(StreamBlockJob),
 165     .job_type      = "stream",
 166     .set_speed     = stream_set_speed,
 167 };
 168
 169 int stream_start(BlockDriverState *bs, BlockDriverState *base,
 170                  BlockDriverCompletionFunc *cb, void *opaque)
 171 {
 172     StreamBlockJob *s;
 173     Coroutine *co;
 174
 175     s = block_job_create(&stream_job_type, bs, cb, opaque);
 176     if (!s) {
 177         return -EBUSY; /* bs must already be in use */
 178     }
 179
 180     s->base = base;
 181
 182     co = qemu_coroutine_create(stream_run);
 183     trace_stream_start(bs, base, s, co, opaque);
 184     qemu_coroutine_enter(co, s);
 185     return 0;
 186 }