block/commit.c

   1 /*
   2  * Live block commit
   3  *
   4  * Copyright Red Hat, Inc. 2012
   5  *
   6  * Authors:
   7  *  Jeff Cody   <jcody@redhat.com>
   8  *  Based on stream.c by Stefan Hajnoczi
   9  *
  10  * This work is licensed under the terms of the GNU LGPL, version 2 or later.
  11  * See the COPYING.LIB file in the top-level directory.
  12  *
  13  */
  14
  15 #include "qemu/osdep.h"
  16 #include "trace.h"
  17 #include "block/block_int.h"
  18 #include "block/blockjob.h"
  19 #include "qapi/error.h"
  20 #include "qapi/qmp/qerror.h"
  21 #include "qemu/ratelimit.h"
  22 #include "sysemu/block-backend.h"
  23
  24 enum {
  25     /*
  26      * Size of data buffer for populating the image file.  This should be large
  27      * enough to process multiple clusters in a single call, so that populating
  28      * contiguous regions of the image is efficient.
  29      */
  30     COMMIT_BUFFER_SIZE = 512 * 1024, /* in bytes */
  31 };
  32
  33 #define SLICE_TIME 100000000ULL /* ns */
  34
  35 typedef struct CommitBlockJob {
  36     BlockJob common;
  37     RateLimit limit;
  38     BlockDriverState *active;
  39     BlockBackend *top;
  40     BlockBackend *base;
  41     BlockdevOnError on_error;
  42     int base_flags;
  43     int orig_overlay_flags;
  44     char *backing_file_str;
  45 } CommitBlockJob;
  46
  47 static int coroutine_fn commit_populate(BlockBackend *bs, BlockBackend *base,
  48                                         int64_t sector_num, int nb_sectors,
  49                                         void *buf)
  50 {
  51     int ret = 0;
  52     QEMUIOVector qiov;
  53     struct iovec iov = {
  54         .iov_base = buf,
  55         .iov_len = nb_sectors * BDRV_SECTOR_SIZE,
  56     };
  57
  58     qemu_iovec_init_external(&qiov, &iov, 1);
  59
  60     ret = blk_co_preadv(bs, sector_num * BDRV_SECTOR_SIZE,
  61                         qiov.size, &qiov, 0);
  62     if (ret < 0) {
  63         return ret;
  64     }
  65
  66     ret = blk_co_pwritev(base, sector_num * BDRV_SECTOR_SIZE,
  67                          qiov.size, &qiov, 0);
  68     if (ret < 0) {
  69         return ret;
  70     }
  71
  72     return 0;
  73 }
  74
  75 typedef struct {
  76     int ret;
  77 } CommitCompleteData;
  78
  79 static void commit_complete(BlockJob *job, void *opaque)
  80 {
  81     CommitBlockJob *s = container_of(job, CommitBlockJob, common);
  82     CommitCompleteData *data = opaque;
  83     BlockDriverState *active = s->active;
  84     BlockDriverState *top = blk_bs(s->top);
  85     BlockDriverState *base = blk_bs(s->base);
  86     BlockDriverState *overlay_bs;
  87     int ret = data->ret;
  88
  89     if (!block_job_is_cancelled(&s->common) && ret == 0) {
  90         /* success */
  91         ret = bdrv_drop_intermediate(active, top, base, s->backing_file_str);
  92     }
  93
  94     /* restore base open flags here if appropriate (e.g., change the base back
  95      * to r/o). These reopens do not need to be atomic, since we won't abort
  96      * even on failure here */
  97     if (s->base_flags != bdrv_get_flags(base)) {
  98         bdrv_reopen(base, s->base_flags, NULL);
  99     }
 100     overlay_bs = bdrv_find_overlay(active, top);
 101     if (overlay_bs && s->orig_overlay_flags != bdrv_get_flags(overlay_bs)) {
 102         bdrv_reopen(overlay_bs, s->orig_overlay_flags, NULL);
 103     }
 104     g_free(s->backing_file_str);
 105     blk_unref(s->top);
 106     blk_unref(s->base);
 107     block_job_completed(&s->common, ret);
 108     g_free(data);
 109 }
 110
 111 static void coroutine_fn commit_run(void *opaque)
 112 {
 113     CommitBlockJob *s = opaque;
 114     CommitCompleteData *data;
 115     int64_t sector_num, end;
 116     int ret = 0;
 117     int n = 0;
 118     void *buf = NULL;
 119     int bytes_written = 0;
 120     int64_t base_len;
 121
 122     ret = s->common.len = blk_getlength(s->top);
 123
 124
 125     if (s->common.len < 0) {
 126         goto out;
 127     }
 128
 129     ret = base_len = blk_getlength(s->base);
 130     if (base_len < 0) {
 131         goto out;
 132     }
 133
 134     if (base_len < s->common.len) {
 135         ret = blk_truncate(s->base, s->common.len);
 136         if (ret) {
 137             goto out;
 138         }
 139     }
 140
 141     end = s->common.len >> BDRV_SECTOR_BITS;
 142     buf = blk_blockalign(s->top, COMMIT_BUFFER_SIZE);
 143
 144     for (sector_num = 0; sector_num < end; sector_num += n) {
 145         uint64_t delay_ns = 0;
 146         bool copy;
 147
 148 wait:
 149         /* Note that even when no rate limit is applied we need to yield
 150          * with no pending I/O here so that bdrv_drain_all() returns.
 151          */
 152         block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 153         if (block_job_is_cancelled(&s->common)) {
 154             break;
 155         }
 156         /* Copy if allocated above the base */
 157         ret = bdrv_is_allocated_above(blk_bs(s->top), blk_bs(s->base),
 158                                       sector_num,
 159                                       COMMIT_BUFFER_SIZE / BDRV_SECTOR_SIZE,
 160                                       &n);
 161         copy = (ret == 1);
 162         trace_commit_one_iteration(s, sector_num, n, ret);
 163         if (copy) {
 164             if (s->common.speed) {
 165                 delay_ns = ratelimit_calculate_delay(&s->limit, n);
 166                 if (delay_ns > 0) {
 167                     goto wait;
 168                 }
 169             }
 170             ret = commit_populate(s->top, s->base, sector_num, n, buf);
 171             bytes_written += n * BDRV_SECTOR_SIZE;
 172         }
 173         if (ret < 0) {
 174             if (s->on_error == BLOCKDEV_ON_ERROR_STOP ||
 175                 s->on_error == BLOCKDEV_ON_ERROR_REPORT||
 176                 (s->on_error == BLOCKDEV_ON_ERROR_ENOSPC && ret == -ENOSPC)) {
 177                 goto out;
 178             } else {
 179                 n = 0;
 180                 continue;
 181             }
 182         }
 183         /* Publish progress */
 184         s->common.offset += n * BDRV_SECTOR_SIZE;
 185     }
 186
 187     ret = 0;
 188
 189 out:
 190     qemu_vfree(buf);
 191
 192     data = g_malloc(sizeof(*data));
 193     data->ret = ret;
 194     block_job_defer_to_main_loop(&s->common, commit_complete, data);
 195 }
 196
 197 static void commit_set_speed(BlockJob *job, int64_t speed, Error **errp)
 198 {
 199     CommitBlockJob *s = container_of(job, CommitBlockJob, common);
 200
 201     if (speed < 0) {
 202         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
 203         return;
 204     }
 205     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 206 }
 207
 208 static const BlockJobDriver commit_job_driver = {
 209     .instance_size = sizeof(CommitBlockJob),
 210     .job_type      = BLOCK_JOB_TYPE_COMMIT,
 211     .set_speed     = commit_set_speed,
 212 };
 213
 214 void commit_start(BlockDriverState *bs, BlockDriverState *base,
 215                   BlockDriverState *top, int64_t speed,
 216                   BlockdevOnError on_error, BlockCompletionFunc *cb,
 217                   void *opaque, const char *backing_file_str, Error **errp)
 218 {
 219     CommitBlockJob *s;
 220     BlockReopenQueue *reopen_queue = NULL;
 221     int orig_overlay_flags;
 222     int orig_base_flags;
 223     BlockDriverState *overlay_bs;
 224     Error *local_err = NULL;
 225
 226     assert(top != bs);
 227     if (top == base) {
 228         error_setg(errp, "Invalid files for merge: top and base are the same");
 229         return;
 230     }
 231
 232     overlay_bs = bdrv_find_overlay(bs, top);
 233
 234     if (overlay_bs == NULL) {
 235         error_setg(errp, "Could not find overlay image for %s:", top->filename);
 236         return;
 237     }
 238
 239     s = block_job_create(&commit_job_driver, bs, speed, cb, opaque, errp);
 240     if (!s) {
 241         return;
 242     }
 243
 244     orig_base_flags    = bdrv_get_flags(base);
 245     orig_overlay_flags = bdrv_get_flags(overlay_bs);
 246
 247     /* convert base & overlay_bs to r/w, if necessary */
 248     if (!(orig_overlay_flags & BDRV_O_RDWR)) {
 249         reopen_queue = bdrv_reopen_queue(reopen_queue, overlay_bs, NULL,
 250                                          orig_overlay_flags | BDRV_O_RDWR);
 251     }
 252     if (!(orig_base_flags & BDRV_O_RDWR)) {
 253         reopen_queue = bdrv_reopen_queue(reopen_queue, base, NULL,
 254                                          orig_base_flags | BDRV_O_RDWR);
 255     }
 256     if (reopen_queue) {
 257         bdrv_reopen_multiple(reopen_queue, &local_err);
 258         if (local_err != NULL) {
 259             error_propagate(errp, local_err);
 260             block_job_unref(&s->common);
 261             return;
 262         }
 263     }
 264
 265
 266     s->base = blk_new();
 267     blk_insert_bs(s->base, base);
 268
 269     s->top = blk_new();
 270     blk_insert_bs(s->top, top);
 271
 272     s->active = bs;
 273
 274     s->base_flags          = orig_base_flags;
 275     s->orig_overlay_flags  = orig_overlay_flags;
 276
 277     s->backing_file_str = g_strdup(backing_file_str);
 278
 279     s->on_error = on_error;
 280     s->common.co = qemu_coroutine_create(commit_run);
 281
 282     trace_commit_start(bs, base, top, s, s->common.co, opaque);
 283     qemu_coroutine_enter(s->common.co, s);
 284 }
 285
 286
 287 #define COMMIT_BUF_SECTORS 2048
 288
 289 /* commit COW file into the raw image */
 290 int bdrv_commit(BlockDriverState *bs)
 291 {
 292     BlockBackend *src, *backing;
 293     BlockDriver *drv = bs->drv;
 294     int64_t sector, total_sectors, length, backing_length;
 295     int n, ro, open_flags;
 296     int ret = 0;
 297     uint8_t *buf = NULL;
 298
 299     if (!drv)
 300         return -ENOMEDIUM;
 301
 302     if (!bs->backing) {
 303         return -ENOTSUP;
 304     }
 305
 306     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_COMMIT_SOURCE, NULL) ||
 307         bdrv_op_is_blocked(bs->backing->bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL)) {
 308         return -EBUSY;
 309     }
 310
 311     ro = bs->backing->bs->read_only;
 312     open_flags =  bs->backing->bs->open_flags;
 313
 314     if (ro) {
 315         if (bdrv_reopen(bs->backing->bs, open_flags | BDRV_O_RDWR, NULL)) {
 316             return -EACCES;
 317         }
 318     }
 319
 320     src = blk_new();
 321     blk_insert_bs(src, bs);
 322
 323     backing = blk_new();
 324     blk_insert_bs(backing, bs->backing->bs);
 325
 326     length = blk_getlength(src);
 327     if (length < 0) {
 328         ret = length;
 329         goto ro_cleanup;
 330     }
 331
 332     backing_length = blk_getlength(backing);
 333     if (backing_length < 0) {
 334         ret = backing_length;
 335         goto ro_cleanup;
 336     }
 337
 338     /* If our top snapshot is larger than the backing file image,
 339      * grow the backing file image if possible.  If not possible,
 340      * we must return an error */
 341     if (length > backing_length) {
 342         ret = blk_truncate(backing, length);
 343         if (ret < 0) {
 344             goto ro_cleanup;
 345         }
 346     }
 347
 348     total_sectors = length >> BDRV_SECTOR_BITS;
 349
 350     /* blk_try_blockalign() for src will choose an alignment that works for
 351      * backing as well, so no need to compare the alignment manually. */
 352     buf = blk_try_blockalign(src, COMMIT_BUF_SECTORS * BDRV_SECTOR_SIZE);
 353     if (buf == NULL) {
 354         ret = -ENOMEM;
 355         goto ro_cleanup;
 356     }
 357
 358     for (sector = 0; sector < total_sectors; sector += n) {
 359         ret = bdrv_is_allocated(bs, sector, COMMIT_BUF_SECTORS, &n);
 360         if (ret < 0) {
 361             goto ro_cleanup;
 362         }
 363         if (ret) {
 364             ret = blk_pread(src, sector * BDRV_SECTOR_SIZE, buf,
 365                             n * BDRV_SECTOR_SIZE);
 366             if (ret < 0) {
 367                 goto ro_cleanup;
 368             }
 369
 370             ret = blk_pwrite(backing, sector * BDRV_SECTOR_SIZE, buf,
 371                              n * BDRV_SECTOR_SIZE, 0);
 372             if (ret < 0) {
 373                 goto ro_cleanup;
 374             }
 375         }
 376     }
 377
 378     if (drv->bdrv_make_empty) {
 379         ret = drv->bdrv_make_empty(bs);
 380         if (ret < 0) {
 381             goto ro_cleanup;
 382         }
 383         blk_flush(src);
 384     }
 385
 386     /*
 387      * Make sure all data we wrote to the backing device is actually
 388      * stable on disk.
 389      */
 390     blk_flush(backing);
 391
 392     ret = 0;
 393 ro_cleanup:
 394     qemu_vfree(buf);
 395
 396     blk_unref(src);
 397     blk_unref(backing);
 398
 399     if (ro) {
 400         /* ignoring error return here */
 401         bdrv_reopen(bs->backing->bs, open_flags & ~BDRV_O_RDWR, NULL);
 402     }
 403
 404     return ret;
 405 }