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_int.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 = bdrv_find_overlay(active, top);
  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     if (overlay_bs && s->orig_overlay_flags != bdrv_get_flags(overlay_bs)) {
 101         bdrv_reopen(overlay_bs, s->orig_overlay_flags, NULL);
 102     }
 103     g_free(s->backing_file_str);
 104     blk_unref(s->top);
 105     blk_unref(s->base);
 106     block_job_completed(&s->common, ret);
 107     g_free(data);
 108 }
 109
 110 static void coroutine_fn commit_run(void *opaque)
 111 {
 112     CommitBlockJob *s = opaque;
 113     CommitCompleteData *data;
 114     int64_t sector_num, end;
 115     uint64_t delay_ns = 0;
 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         bool copy;
 146
 147         /* Note that even when no rate limit is applied we need to yield
 148          * with no pending I/O here so that bdrv_drain_all() returns.
 149          */
 150         block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 151         if (block_job_is_cancelled(&s->common)) {
 152             break;
 153         }
 154         /* Copy if allocated above the base */
 155         ret = bdrv_is_allocated_above(blk_bs(s->top), blk_bs(s->base),
 156                                       sector_num,
 157                                       COMMIT_BUFFER_SIZE / BDRV_SECTOR_SIZE,
 158                                       &n);
 159         copy = (ret == 1);
 160         trace_commit_one_iteration(s, sector_num, n, ret);
 161         if (copy) {
 162             ret = commit_populate(s->top, s->base, sector_num, n, buf);
 163             bytes_written += n * BDRV_SECTOR_SIZE;
 164         }
 165         if (ret < 0) {
 166             BlockErrorAction action =
 167                 block_job_error_action(&s->common, false, s->on_error, -ret);
 168             if (action == BLOCK_ERROR_ACTION_REPORT) {
 169                 goto out;
 170             } else {
 171                 n = 0;
 172                 continue;
 173             }
 174         }
 175         /* Publish progress */
 176         s->common.offset += n * BDRV_SECTOR_SIZE;
 177
 178         if (copy && s->common.speed) {
 179             delay_ns = ratelimit_calculate_delay(&s->limit, n);
 180         }
 181     }
 182
 183     ret = 0;
 184
 185 out:
 186     qemu_vfree(buf);
 187
 188     data = g_malloc(sizeof(*data));
 189     data->ret = ret;
 190     block_job_defer_to_main_loop(&s->common, commit_complete, data);
 191 }
 192
 193 static void commit_set_speed(BlockJob *job, int64_t speed, Error **errp)
 194 {
 195     CommitBlockJob *s = container_of(job, CommitBlockJob, common);
 196
 197     if (speed < 0) {
 198         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
 199         return;
 200     }
 201     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 202 }
 203
 204 static const BlockJobDriver commit_job_driver = {
 205     .instance_size = sizeof(CommitBlockJob),
 206     .job_type      = BLOCK_JOB_TYPE_COMMIT,
 207     .set_speed     = commit_set_speed,
 208     .start         = commit_run,
 209 };
 210
 211 void commit_start(const char *job_id, BlockDriverState *bs,
 212                   BlockDriverState *base, BlockDriverState *top, int64_t speed,
 213                   BlockdevOnError on_error, const char *backing_file_str,
 214                   Error **errp)
 215 {
 216     CommitBlockJob *s;
 217     BlockReopenQueue *reopen_queue = NULL;
 218     int orig_overlay_flags;
 219     int orig_base_flags;
 220     BlockDriverState *iter;
 221     BlockDriverState *overlay_bs;
 222     Error *local_err = NULL;
 223     int ret;
 224
 225     assert(top != bs);
 226     if (top == base) {
 227         error_setg(errp, "Invalid files for merge: top and base are the same");
 228         return;
 229     }
 230
 231     overlay_bs = bdrv_find_overlay(bs, top);
 232
 233     if (overlay_bs == NULL) {
 234         error_setg(errp, "Could not find overlay image for %s:", top->filename);
 235         return;
 236     }
 237
 238     s = block_job_create(job_id, &commit_job_driver, bs, speed,
 239                          BLOCK_JOB_DEFAULT, NULL, NULL, 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_base_flags & BDRV_O_RDWR)) {
 249         reopen_queue = bdrv_reopen_queue(reopen_queue, base, NULL,
 250                                          orig_base_flags | BDRV_O_RDWR);
 251     }
 252     if (!(orig_overlay_flags & BDRV_O_RDWR)) {
 253         reopen_queue = bdrv_reopen_queue(reopen_queue, overlay_bs, NULL,
 254                                          orig_overlay_flags | BDRV_O_RDWR);
 255     }
 256     if (reopen_queue) {
 257         bdrv_reopen_multiple(bdrv_get_aio_context(bs), reopen_queue, &local_err);
 258         if (local_err != NULL) {
 259             error_propagate(errp, local_err);
 260             goto fail;
 261         }
 262     }
 263
 264
 265     /* Block all nodes between top and base, because they will
 266      * disappear from the chain after this operation. */
 267     assert(bdrv_chain_contains(top, base));
 268     for (iter = top; iter != backing_bs(base); iter = backing_bs(iter)) {
 269         block_job_add_bdrv(&s->common, iter);
 270     }
 271     /* overlay_bs must be blocked because it needs to be modified to
 272      * update the backing image string, but if it's the root node then
 273      * don't block it again */
 274     if (bs != overlay_bs) {
 275         block_job_add_bdrv(&s->common, overlay_bs);
 276     }
 277
 278     /* FIXME Use real permissions */
 279     s->base = blk_new(0, BLK_PERM_ALL);
 280     ret = blk_insert_bs(s->base, base, errp);
 281     if (ret < 0) {
 282         goto fail;
 283     }
 284
 285     /* FIXME Use real permissions */
 286     s->top = blk_new(0, BLK_PERM_ALL);
 287     ret = blk_insert_bs(s->top, top, errp);
 288     if (ret < 0) {
 289         goto fail;
 290     }
 291
 292     s->active = bs;
 293
 294     s->base_flags          = orig_base_flags;
 295     s->orig_overlay_flags  = orig_overlay_flags;
 296
 297     s->backing_file_str = g_strdup(backing_file_str);
 298
 299     s->on_error = on_error;
 300
 301     trace_commit_start(bs, base, top, s);
 302     block_job_start(&s->common);
 303     return;
 304
 305 fail:
 306     if (s->base) {
 307         blk_unref(s->base);
 308     }
 309     if (s->top) {
 310         blk_unref(s->top);
 311     }
 312     block_job_unref(&s->common);
 313 }
 314
 315
 316 #define COMMIT_BUF_SECTORS 2048
 317
 318 /* commit COW file into the raw image */
 319 int bdrv_commit(BlockDriverState *bs)
 320 {
 321     BlockBackend *src, *backing;
 322     BlockDriver *drv = bs->drv;
 323     int64_t sector, total_sectors, length, backing_length;
 324     int n, ro, open_flags;
 325     int ret = 0;
 326     uint8_t *buf = NULL;
 327
 328     if (!drv)
 329         return -ENOMEDIUM;
 330
 331     if (!bs->backing) {
 332         return -ENOTSUP;
 333     }
 334
 335     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_COMMIT_SOURCE, NULL) ||
 336         bdrv_op_is_blocked(bs->backing->bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL)) {
 337         return -EBUSY;
 338     }
 339
 340     ro = bs->backing->bs->read_only;
 341     open_flags =  bs->backing->bs->open_flags;
 342
 343     if (ro) {
 344         if (bdrv_reopen(bs->backing->bs, open_flags | BDRV_O_RDWR, NULL)) {
 345             return -EACCES;
 346         }
 347     }
 348
 349     /* FIXME Use real permissions */
 350     src = blk_new(0, BLK_PERM_ALL);
 351     backing = blk_new(0, BLK_PERM_ALL);
 352
 353     ret = blk_insert_bs(src, bs, NULL);
 354     if (ret < 0) {
 355         goto ro_cleanup;
 356     }
 357
 358     ret = blk_insert_bs(backing, bs->backing->bs, NULL);
 359     if (ret < 0) {
 360         goto ro_cleanup;
 361     }
 362
 363     length = blk_getlength(src);
 364     if (length < 0) {
 365         ret = length;
 366         goto ro_cleanup;
 367     }
 368
 369     backing_length = blk_getlength(backing);
 370     if (backing_length < 0) {
 371         ret = backing_length;
 372         goto ro_cleanup;
 373     }
 374
 375     /* If our top snapshot is larger than the backing file image,
 376      * grow the backing file image if possible.  If not possible,
 377      * we must return an error */
 378     if (length > backing_length) {
 379         ret = blk_truncate(backing, length);
 380         if (ret < 0) {
 381             goto ro_cleanup;
 382         }
 383     }
 384
 385     total_sectors = length >> BDRV_SECTOR_BITS;
 386
 387     /* blk_try_blockalign() for src will choose an alignment that works for
 388      * backing as well, so no need to compare the alignment manually. */
 389     buf = blk_try_blockalign(src, COMMIT_BUF_SECTORS * BDRV_SECTOR_SIZE);
 390     if (buf == NULL) {
 391         ret = -ENOMEM;
 392         goto ro_cleanup;
 393     }
 394
 395     for (sector = 0; sector < total_sectors; sector += n) {
 396         ret = bdrv_is_allocated(bs, sector, COMMIT_BUF_SECTORS, &n);
 397         if (ret < 0) {
 398             goto ro_cleanup;
 399         }
 400         if (ret) {
 401             ret = blk_pread(src, sector * BDRV_SECTOR_SIZE, buf,
 402                             n * BDRV_SECTOR_SIZE);
 403             if (ret < 0) {
 404                 goto ro_cleanup;
 405             }
 406
 407             ret = blk_pwrite(backing, sector * BDRV_SECTOR_SIZE, buf,
 408                              n * BDRV_SECTOR_SIZE, 0);
 409             if (ret < 0) {
 410                 goto ro_cleanup;
 411             }
 412         }
 413     }
 414
 415     if (drv->bdrv_make_empty) {
 416         ret = drv->bdrv_make_empty(bs);
 417         if (ret < 0) {
 418             goto ro_cleanup;
 419         }
 420         blk_flush(src);
 421     }
 422
 423     /*
 424      * Make sure all data we wrote to the backing device is actually
 425      * stable on disk.
 426      */
 427     blk_flush(backing);
 428
 429     ret = 0;
 430 ro_cleanup:
 431     qemu_vfree(buf);
 432
 433     blk_unref(src);
 434     blk_unref(backing);
 435
 436     if (ro) {
 437         /* ignoring error return here */
 438         bdrv_reopen(bs->backing->bs, open_flags & ~BDRV_O_RDWR, NULL);
 439     }
 440
 441     return ret;
 442 }