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 now = qemu_get_clock_ns(rt_clock);
  37
  38     if (limit->next_slice_time < now) {
  39         limit->next_slice_time = now + SLICE_TIME;
  40         limit->dispatched = 0;
  41     }
  42     if (limit->dispatched == 0 || limit->dispatched + n <= limit->slice_quota) {
  43         limit->dispatched += n;
  44         return 0;
  45     } else {
  46         limit->dispatched = n;
  47         return limit->next_slice_time - now;
  48     }
  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     char backing_file_id[1024];
  61 } StreamBlockJob;
  62
  63 static int coroutine_fn stream_populate(BlockDriverState *bs,
  64                                         int64_t sector_num, int nb_sectors,
  65                                         void *buf)
  66 {
  67     struct iovec iov = {
  68         .iov_base = buf,
  69         .iov_len  = nb_sectors * BDRV_SECTOR_SIZE,
  70     };
  71     QEMUIOVector qiov;
  72
  73     qemu_iovec_init_external(&qiov, &iov, 1);
  74
  75     /* Copy-on-read the unallocated clusters */
  76     return bdrv_co_copy_on_readv(bs, sector_num, nb_sectors, &qiov);
  77 }
  78
  79 static void close_unused_images(BlockDriverState *top, BlockDriverState *base,
  80                                 const char *base_id)
  81 {
  82     BlockDriverState *intermediate;
  83     intermediate = top->backing_hd;
  84
  85     while (intermediate) {
  86         BlockDriverState *unused;
  87
  88         /* reached base */
  89         if (intermediate == base) {
  90             break;
  91         }
  92
  93         unused = intermediate;
  94         intermediate = intermediate->backing_hd;
  95         unused->backing_hd = NULL;
  96         bdrv_delete(unused);
  97     }
  98     top->backing_hd = base;
  99 }
 100
 101 /*
 102  * Given an image chain: [BASE] -> [INTER1] -> [INTER2] -> [TOP]
 103  *
 104  * Return true if the given sector is allocated in any image between
 105  * BASE and TOP (inclusive).  BASE can be NULL to check if the given
 106  * sector is allocated in any image of the chain.  Return false otherwise.
 107  *
 108  * 'pnum' is set to the number of sectors (including and immediately following
 109  *  the specified sector) that are known to be in the same
 110  *  allocated/unallocated state.
 111  *
 112  */
 113 static int coroutine_fn is_allocated_above(BlockDriverState *top,
 114                                            BlockDriverState *base,
 115                                            int64_t sector_num,
 116                                            int nb_sectors, int *pnum)
 117 {
 118     BlockDriverState *intermediate;
 119     int ret, n = nb_sectors;
 120
 121     intermediate = top;
 122     while (intermediate != base) {
 123         int pnum_inter;
 124         ret = bdrv_co_is_allocated(intermediate, sector_num, nb_sectors,
 125                                    &pnum_inter);
 126         if (ret < 0) {
 127             return ret;
 128         } else if (ret) {
 129             *pnum = pnum_inter;
 130             return 1;
 131         }
 132
 133         /*
 134          * [sector_num, nb_sectors] is unallocated on top but intermediate
 135          * might have
 136          *
 137          * [sector_num+x, nr_sectors] allocated.
 138          */
 139         if (n > pnum_inter) {
 140             n = pnum_inter;
 141         }
 142
 143         intermediate = intermediate->backing_hd;
 144     }
 145
 146     *pnum = n;
 147     return 0;
 148 }
 149
 150 static void coroutine_fn stream_run(void *opaque)
 151 {
 152     StreamBlockJob *s = opaque;
 153     BlockDriverState *bs = s->common.bs;
 154     BlockDriverState *base = s->base;
 155     int64_t sector_num, end;
 156     int ret = 0;
 157     int n = 0;
 158     void *buf;
 159
 160     s->common.len = bdrv_getlength(bs);
 161     if (s->common.len < 0) {
 162         block_job_complete(&s->common, s->common.len);
 163         return;
 164     }
 165
 166     end = s->common.len >> BDRV_SECTOR_BITS;
 167     buf = qemu_blockalign(bs, STREAM_BUFFER_SIZE);
 168
 169     /* Turn on copy-on-read for the whole block device so that guest read
 170      * requests help us make progress.  Only do this when copying the entire
 171      * backing chain since the copy-on-read operation does not take base into
 172      * account.
 173      */
 174     if (!base) {
 175         bdrv_enable_copy_on_read(bs);
 176     }
 177
 178     for (sector_num = 0; sector_num < end; sector_num += n) {
 179         uint64_t delay_ns = 0;
 180         bool copy;
 181
 182 wait:
 183         /* Note that even when no rate limit is applied we need to yield
 184          * with no pending I/O here so that qemu_aio_flush() returns.
 185          */
 186         block_job_sleep_ns(&s->common, rt_clock, delay_ns);
 187         if (block_job_is_cancelled(&s->common)) {
 188             break;
 189         }
 190
 191         ret = bdrv_co_is_allocated(bs, sector_num,
 192                                    STREAM_BUFFER_SIZE / BDRV_SECTOR_SIZE, &n);
 193         if (ret == 1) {
 194             /* Allocated in the top, no need to copy.  */
 195             copy = false;
 196         } else {
 197             /* Copy if allocated in the intermediate images.  Limit to the
 198              * known-unallocated area [sector_num, sector_num+n).  */
 199             ret = is_allocated_above(bs->backing_hd, base, sector_num, n, &n);
 200             copy = (ret == 1);
 201         }
 202
 203         trace_stream_one_iteration(s, sector_num, n, ret);
 204         if (ret >= 0 && copy) {
 205             if (s->common.speed) {
 206                 delay_ns = ratelimit_calculate_delay(&s->limit, n);
 207                 if (delay_ns > 0) {
 208                     goto wait;
 209                 }
 210             }
 211             ret = stream_populate(bs, sector_num, n, buf);
 212         }
 213         if (ret < 0) {
 214             break;
 215         }
 216         ret = 0;
 217
 218         /* Publish progress */
 219         s->common.offset += n * BDRV_SECTOR_SIZE;
 220     }
 221
 222     if (!base) {
 223         bdrv_disable_copy_on_read(bs);
 224     }
 225
 226     if (!block_job_is_cancelled(&s->common) && sector_num == end && ret == 0) {
 227         const char *base_id = NULL, *base_fmt = NULL;
 228         if (base) {
 229             base_id = s->backing_file_id;
 230             if (base->drv) {
 231                 base_fmt = base->drv->format_name;
 232             }
 233         }
 234         ret = bdrv_change_backing_file(bs, base_id, base_fmt);
 235         close_unused_images(bs, base, base_id);
 236     }
 237
 238     qemu_vfree(buf);
 239     block_job_complete(&s->common, ret);
 240 }
 241
 242 static void stream_set_speed(BlockJob *job, int64_t speed, Error **errp)
 243 {
 244     StreamBlockJob *s = container_of(job, StreamBlockJob, common);
 245
 246     if (speed < 0) {
 247         error_set(errp, QERR_INVALID_PARAMETER, "speed");
 248         return;
 249     }
 250     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE);
 251 }
 252
 253 static BlockJobType stream_job_type = {
 254     .instance_size = sizeof(StreamBlockJob),
 255     .job_type      = "stream",
 256     .set_speed     = stream_set_speed,
 257 };
 258
 259 void stream_start(BlockDriverState *bs, BlockDriverState *base,
 260                   const char *base_id, int64_t speed,
 261                   BlockDriverCompletionFunc *cb,
 262                   void *opaque, Error **errp)
 263 {
 264     StreamBlockJob *s;
 265
 266     s = block_job_create(&stream_job_type, bs, speed, cb, opaque, errp);
 267     if (!s) {
 268         return;
 269     }
 270
 271     s->base = base;
 272     if (base_id) {
 273         pstrcpy(s->backing_file_id, sizeof(s->backing_file_id), base_id);
 274     }
 275
 276     s->common.co = qemu_coroutine_create(stream_run);
 277     trace_stream_start(bs, base, s, s->common.co, opaque);
 278     qemu_coroutine_enter(s->common.co, s);
 279 }