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     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     pstrcpy(top->backing_file, sizeof(top->backing_file), "");
 101     pstrcpy(top->backing_format, sizeof(top->backing_format), "");
 102     if (base_id) {
 103         pstrcpy(top->backing_file, sizeof(top->backing_file), base_id);
 104         if (base->drv) {
 105             pstrcpy(top->backing_format, sizeof(top->backing_format),
 106                     base->drv->format_name);
 107         }
 108     }
 109
 110 }
 111
 112 /*
 113  * Given an image chain: [BASE] -> [INTER1] -> [INTER2] -> [TOP]
 114  *
 115  * Return true if the given sector is allocated in top.
 116  * Return false if the given sector is allocated in intermediate images.
 117  * Return true otherwise.
 118  *
 119  * 'pnum' is set to the number of sectors (including and immediately following
 120  *  the specified sector) that are known to be in the same
 121  *  allocated/unallocated state.
 122  *
 123  */
 124 static int coroutine_fn is_allocated_base(BlockDriverState *top,
 125                                           BlockDriverState *base,
 126                                           int64_t sector_num,
 127                                           int nb_sectors, int *pnum)
 128 {
 129     BlockDriverState *intermediate;
 130     int ret, n;
 131
 132     ret = bdrv_co_is_allocated(top, sector_num, nb_sectors, &n);
 133     if (ret) {
 134         *pnum = n;
 135         return ret;
 136     }
 137
 138     /*
 139      * Is the unallocated chunk [sector_num, n] also
 140      * unallocated between base and top?
 141      */
 142     intermediate = top->backing_hd;
 143
 144     while (intermediate) {
 145         int pnum_inter;
 146
 147         /* reached base */
 148         if (intermediate == base) {
 149             *pnum = n;
 150             return 1;
 151         }
 152         ret = bdrv_co_is_allocated(intermediate, sector_num, nb_sectors,
 153                                    &pnum_inter);
 154         if (ret < 0) {
 155             return ret;
 156         } else if (ret) {
 157             *pnum = pnum_inter;
 158             return 0;
 159         }
 160
 161         /*
 162          * [sector_num, nb_sectors] is unallocated on top but intermediate
 163          * might have
 164          *
 165          * [sector_num+x, nr_sectors] allocated.
 166          */
 167         if (n > pnum_inter) {
 168             n = pnum_inter;
 169         }
 170
 171         intermediate = intermediate->backing_hd;
 172     }
 173
 174     return 1;
 175 }
 176
 177 static void coroutine_fn stream_run(void *opaque)
 178 {
 179     StreamBlockJob *s = opaque;
 180     BlockDriverState *bs = s->common.bs;
 181     BlockDriverState *base = s->base;
 182     int64_t sector_num, end;
 183     int ret = 0;
 184     int n;
 185     void *buf;
 186
 187     s->common.len = bdrv_getlength(bs);
 188     if (s->common.len < 0) {
 189         block_job_complete(&s->common, s->common.len);
 190         return;
 191     }
 192
 193     end = s->common.len >> BDRV_SECTOR_BITS;
 194     buf = qemu_blockalign(bs, STREAM_BUFFER_SIZE);
 195
 196     /* Turn on copy-on-read for the whole block device so that guest read
 197      * requests help us make progress.  Only do this when copying the entire
 198      * backing chain since the copy-on-read operation does not take base into
 199      * account.
 200      */
 201     if (!base) {
 202         bdrv_enable_copy_on_read(bs);
 203     }
 204
 205     for (sector_num = 0; sector_num < end; sector_num += n) {
 206 retry:
 207         if (block_job_is_cancelled(&s->common)) {
 208             break;
 209         }
 210
 211         s->common.busy = true;
 212         if (base) {
 213             ret = is_allocated_base(bs, base, sector_num,
 214                                     STREAM_BUFFER_SIZE / BDRV_SECTOR_SIZE, &n);
 215         } else {
 216             ret = bdrv_co_is_allocated(bs, sector_num,
 217                                        STREAM_BUFFER_SIZE / BDRV_SECTOR_SIZE,
 218                                        &n);
 219         }
 220         trace_stream_one_iteration(s, sector_num, n, ret);
 221         if (ret == 0) {
 222             if (s->common.speed) {
 223                 uint64_t delay_ns = ratelimit_calculate_delay(&s->limit, n);
 224                 if (delay_ns > 0) {
 225                     s->common.busy = false;
 226                     co_sleep_ns(rt_clock, delay_ns);
 227
 228                     /* Recheck cancellation and that sectors are unallocated */
 229                     goto retry;
 230                 }
 231             }
 232             ret = stream_populate(bs, sector_num, n, buf);
 233         }
 234         if (ret < 0) {
 235             break;
 236         }
 237         ret = 0;
 238
 239         /* Publish progress */
 240         s->common.offset += n * BDRV_SECTOR_SIZE;
 241
 242         /* Note that even when no rate limit is applied we need to yield
 243          * with no pending I/O here so that qemu_aio_flush() returns.
 244          */
 245         s->common.busy = false;
 246         co_sleep_ns(rt_clock, 0);
 247     }
 248
 249     if (!base) {
 250         bdrv_disable_copy_on_read(bs);
 251     }
 252
 253     if (!block_job_is_cancelled(&s->common) && sector_num == end && ret == 0) {
 254         const char *base_id = NULL;
 255         if (base) {
 256             base_id = s->backing_file_id;
 257         }
 258         ret = bdrv_change_backing_file(bs, base_id, NULL);
 259         close_unused_images(bs, base, base_id);
 260     }
 261
 262     qemu_vfree(buf);
 263     block_job_complete(&s->common, ret);
 264 }
 265
 266 static void stream_set_speed(BlockJob *job, int64_t speed, Error **errp)
 267 {
 268     StreamBlockJob *s = container_of(job, StreamBlockJob, common);
 269
 270     if (speed < 0) {
 271         error_set(errp, QERR_INVALID_PARAMETER, "speed");
 272         return;
 273     }
 274     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE);
 275 }
 276
 277 static BlockJobType stream_job_type = {
 278     .instance_size = sizeof(StreamBlockJob),
 279     .job_type      = "stream",
 280     .set_speed     = stream_set_speed,
 281 };
 282
 283 void stream_start(BlockDriverState *bs, BlockDriverState *base,
 284                   const char *base_id, BlockDriverCompletionFunc *cb,
 285                   void *opaque, Error **errp)
 286 {
 287     StreamBlockJob *s;
 288     Coroutine *co;
 289
 290     s = block_job_create(&stream_job_type, bs, cb, opaque, errp);
 291     if (!s) {
 292         return;
 293     }
 294
 295     s->base = base;
 296     if (base_id) {
 297         pstrcpy(s->backing_file_id, sizeof(s->backing_file_id), base_id);
 298     }
 299
 300     co = qemu_coroutine_create(stream_run);
 301     trace_stream_start(bs, base, s, co, opaque);
 302     qemu_coroutine_enter(co, s);
 303 }