block/mirror.c

   1 /*
   2  * Image mirroring
   3  *
   4  * Copyright Red Hat, Inc. 2012
   5  *
   6  * Authors:
   7  *  Paolo Bonzini  <pbonzini@redhat.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 "blockjob.h"
  16 #include "block_int.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     BLOCK_SIZE = 512 * BDRV_SECTORS_PER_DIRTY_CHUNK, /* in bytes */
  26 };
  27
  28 #define SLICE_TIME 100000000ULL /* ns */
  29
  30 typedef struct MirrorBlockJob {
  31     BlockJob common;
  32     RateLimit limit;
  33     BlockDriverState *target;
  34     MirrorSyncMode mode;
  35     BlockdevOnError on_source_error, on_target_error;
  36     bool synced;
  37     bool should_complete;
  38     int64_t sector_num;
  39     uint8_t *buf;
  40 } MirrorBlockJob;
  41
  42 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
  43                                             int error)
  44 {
  45     s->synced = false;
  46     if (read) {
  47         return block_job_error_action(&s->common, s->common.bs,
  48                                       s->on_source_error, true, error);
  49     } else {
  50         return block_job_error_action(&s->common, s->target,
  51                                       s->on_target_error, false, error);
  52     }
  53 }
  54
  55 static int coroutine_fn mirror_iteration(MirrorBlockJob *s,
  56                                          BlockErrorAction *p_action)
  57 {
  58     BlockDriverState *source = s->common.bs;
  59     BlockDriverState *target = s->target;
  60     QEMUIOVector qiov;
  61     int ret, nb_sectors;
  62     int64_t end;
  63     struct iovec iov;
  64
  65     end = s->common.len >> BDRV_SECTOR_BITS;
  66     s->sector_num = bdrv_get_next_dirty(source, s->sector_num);
  67     nb_sectors = MIN(BDRV_SECTORS_PER_DIRTY_CHUNK, end - s->sector_num);
  68     bdrv_reset_dirty(source, s->sector_num, nb_sectors);
  69
  70     /* Copy the dirty cluster.  */
  71     iov.iov_base = s->buf;
  72     iov.iov_len  = nb_sectors * 512;
  73     qemu_iovec_init_external(&qiov, &iov, 1);
  74
  75     trace_mirror_one_iteration(s, s->sector_num, nb_sectors);
  76     ret = bdrv_co_readv(source, s->sector_num, nb_sectors, &qiov);
  77     if (ret < 0) {
  78         *p_action = mirror_error_action(s, true, -ret);
  79         goto fail;
  80     }
  81     ret = bdrv_co_writev(target, s->sector_num, nb_sectors, &qiov);
  82     if (ret < 0) {
  83         *p_action = mirror_error_action(s, false, -ret);
  84         s->synced = false;
  85         goto fail;
  86     }
  87     return 0;
  88
  89 fail:
  90     /* Try again later.  */
  91     bdrv_set_dirty(source, s->sector_num, nb_sectors);
  92     return ret;
  93 }
  94
  95 static void coroutine_fn mirror_run(void *opaque)
  96 {
  97     MirrorBlockJob *s = opaque;
  98     BlockDriverState *bs = s->common.bs;
  99     int64_t sector_num, end;
 100     int ret = 0;
 101     int n;
 102
 103     if (block_job_is_cancelled(&s->common)) {
 104         goto immediate_exit;
 105     }
 106
 107     s->common.len = bdrv_getlength(bs);
 108     if (s->common.len < 0) {
 109         block_job_completed(&s->common, s->common.len);
 110         return;
 111     }
 112
 113     end = s->common.len >> BDRV_SECTOR_BITS;
 114     s->buf = qemu_blockalign(bs, BLOCK_SIZE);
 115
 116     if (s->mode != MIRROR_SYNC_MODE_NONE) {
 117         /* First part, loop on the sectors and initialize the dirty bitmap.  */
 118         BlockDriverState *base;
 119         base = s->mode == MIRROR_SYNC_MODE_FULL ? NULL : bs->backing_hd;
 120         for (sector_num = 0; sector_num < end; ) {
 121             int64_t next = (sector_num | (BDRV_SECTORS_PER_DIRTY_CHUNK - 1)) + 1;
 122             ret = bdrv_co_is_allocated_above(bs, base,
 123                                              sector_num, next - sector_num, &n);
 124
 125             if (ret < 0) {
 126                 goto immediate_exit;
 127             }
 128
 129             assert(n > 0);
 130             if (ret == 1) {
 131                 bdrv_set_dirty(bs, sector_num, n);
 132                 sector_num = next;
 133             } else {
 134                 sector_num += n;
 135             }
 136         }
 137     }
 138
 139     s->sector_num = -1;
 140     for (;;) {
 141         uint64_t delay_ns;
 142         int64_t cnt;
 143         bool should_complete;
 144
 145         cnt = bdrv_get_dirty_count(bs);
 146         if (cnt != 0) {
 147             BlockErrorAction action = BDRV_ACTION_REPORT;
 148             ret = mirror_iteration(s, &action);
 149             if (ret < 0 && action == BDRV_ACTION_REPORT) {
 150                 goto immediate_exit;
 151             }
 152             cnt = bdrv_get_dirty_count(bs);
 153         }
 154
 155         should_complete = false;
 156         if (cnt == 0) {
 157             trace_mirror_before_flush(s);
 158             ret = bdrv_flush(s->target);
 159             if (ret < 0) {
 160                 if (mirror_error_action(s, false, -ret) == BDRV_ACTION_REPORT) {
 161                     goto immediate_exit;
 162                 }
 163             } else {
 164                 /* We're out of the streaming phase.  From now on, if the job
 165                  * is cancelled we will actually complete all pending I/O and
 166                  * report completion.  This way, block-job-cancel will leave
 167                  * the target in a consistent state.
 168                  */
 169                 s->common.offset = end * BDRV_SECTOR_SIZE;
 170                 if (!s->synced) {
 171                     block_job_ready(&s->common);
 172                     s->synced = true;
 173                 }
 174
 175                 should_complete = s->should_complete ||
 176                     block_job_is_cancelled(&s->common);
 177                 cnt = bdrv_get_dirty_count(bs);
 178             }
 179         }
 180
 181         if (cnt == 0 && should_complete) {
 182             /* The dirty bitmap is not updated while operations are pending.
 183              * If we're about to exit, wait for pending operations before
 184              * calling bdrv_get_dirty_count(bs), or we may exit while the
 185              * source has dirty data to copy!
 186              *
 187              * Note that I/O can be submitted by the guest while
 188              * mirror_populate runs.
 189              */
 190             trace_mirror_before_drain(s, cnt);
 191             bdrv_drain_all();
 192             cnt = bdrv_get_dirty_count(bs);
 193         }
 194
 195         ret = 0;
 196         trace_mirror_before_sleep(s, cnt, s->synced);
 197         if (!s->synced) {
 198             /* Publish progress */
 199             s->common.offset = end * BDRV_SECTOR_SIZE - cnt * BLOCK_SIZE;
 200
 201             if (s->common.speed) {
 202                 delay_ns = ratelimit_calculate_delay(&s->limit, BDRV_SECTORS_PER_DIRTY_CHUNK);
 203             } else {
 204                 delay_ns = 0;
 205             }
 206
 207             /* Note that even when no rate limit is applied we need to yield
 208              * with no pending I/O here so that qemu_aio_flush() returns.
 209              */
 210             block_job_sleep_ns(&s->common, rt_clock, delay_ns);
 211             if (block_job_is_cancelled(&s->common)) {
 212                 break;
 213             }
 214         } else if (!should_complete) {
 215             delay_ns = (cnt == 0 ? SLICE_TIME : 0);
 216             block_job_sleep_ns(&s->common, rt_clock, delay_ns);
 217         } else if (cnt == 0) {
 218             /* The two disks are in sync.  Exit and report successful
 219              * completion.
 220              */
 221             assert(QLIST_EMPTY(&bs->tracked_requests));
 222             s->common.cancelled = false;
 223             break;
 224         }
 225     }
 226
 227 immediate_exit:
 228     g_free(s->buf);
 229     bdrv_set_dirty_tracking(bs, false);
 230     bdrv_iostatus_disable(s->target);
 231     if (s->should_complete && ret == 0) {
 232         if (bdrv_get_flags(s->target) != bdrv_get_flags(s->common.bs)) {
 233             bdrv_reopen(s->target, bdrv_get_flags(s->common.bs), NULL);
 234         }
 235         bdrv_swap(s->target, s->common.bs);
 236     }
 237     bdrv_close(s->target);
 238     bdrv_delete(s->target);
 239     block_job_completed(&s->common, ret);
 240 }
 241
 242 static void mirror_set_speed(BlockJob *job, int64_t speed, Error **errp)
 243 {
 244     MirrorBlockJob *s = container_of(job, MirrorBlockJob, 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, SLICE_TIME);
 251 }
 252
 253 static void mirror_iostatus_reset(BlockJob *job)
 254 {
 255     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 256
 257     bdrv_iostatus_reset(s->target);
 258 }
 259
 260 static void mirror_complete(BlockJob *job, Error **errp)
 261 {
 262     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 263     int ret;
 264
 265     ret = bdrv_open_backing_file(s->target);
 266     if (ret < 0) {
 267         char backing_filename[PATH_MAX];
 268         bdrv_get_full_backing_filename(s->target, backing_filename,
 269                                        sizeof(backing_filename));
 270         error_set(errp, QERR_OPEN_FILE_FAILED, backing_filename);
 271         return;
 272     }
 273     if (!s->synced) {
 274         error_set(errp, QERR_BLOCK_JOB_NOT_READY, job->bs->device_name);
 275         return;
 276     }
 277
 278     s->should_complete = true;
 279     block_job_resume(job);
 280 }
 281
 282 static BlockJobType mirror_job_type = {
 283     .instance_size = sizeof(MirrorBlockJob),
 284     .job_type      = "mirror",
 285     .set_speed     = mirror_set_speed,
 286     .iostatus_reset= mirror_iostatus_reset,
 287     .complete      = mirror_complete,
 288 };
 289
 290 void mirror_start(BlockDriverState *bs, BlockDriverState *target,
 291                   int64_t speed, MirrorSyncMode mode,
 292                   BlockdevOnError on_source_error,
 293                   BlockdevOnError on_target_error,
 294                   BlockDriverCompletionFunc *cb,
 295                   void *opaque, Error **errp)
 296 {
 297     MirrorBlockJob *s;
 298
 299     if ((on_source_error == BLOCKDEV_ON_ERROR_STOP ||
 300          on_source_error == BLOCKDEV_ON_ERROR_ENOSPC) &&
 301         !bdrv_iostatus_is_enabled(bs)) {
 302         error_set(errp, QERR_INVALID_PARAMETER, "on-source-error");
 303         return;
 304     }
 305
 306     s = block_job_create(&mirror_job_type, bs, speed, cb, opaque, errp);
 307     if (!s) {
 308         return;
 309     }
 310
 311     s->on_source_error = on_source_error;
 312     s->on_target_error = on_target_error;
 313     s->target = target;
 314     s->mode = mode;
 315     bdrv_set_dirty_tracking(bs, true);
 316     bdrv_set_enable_write_cache(s->target, true);
 317     bdrv_set_on_error(s->target, on_target_error, on_target_error);
 318     bdrv_iostatus_enable(s->target);
 319     s->common.co = qemu_coroutine_create(mirror_run);
 320     trace_mirror_start(bs, s, s->common.co, opaque);
 321     qemu_coroutine_enter(s->common.co, s);
 322 }