block/mirror.c

   1 /*
   2  * Image mirroring
   3  *
   4  * Copyright Red Hat, Inc. 2012
   5  *
   6  * Authors:
   7  *  Paolo Bonzini  <[email protected]>
   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 "qemu/osdep.h"
  15 #include "trace.h"
  16 #include "block/blockjob_int.h"
  17 #include "block/block_int.h"
  18 #include "sysemu/block-backend.h"
  19 #include "qapi/error.h"
  20 #include "qapi/qmp/qerror.h"
  21 #include "qemu/ratelimit.h"
  22 #include "qemu/bitmap.h"
  23
  24 #define SLICE_TIME    100000000ULL /* ns */
  25 #define MAX_IN_FLIGHT 16
  26 #define MAX_IO_SECTORS ((1 << 20) >> BDRV_SECTOR_BITS) /* 1 Mb */
  27 #define DEFAULT_MIRROR_BUF_SIZE \
  28     (MAX_IN_FLIGHT * MAX_IO_SECTORS * BDRV_SECTOR_SIZE)
  29
  30 /* The mirroring buffer is a list of granularity-sized chunks.
  31  * Free chunks are organized in a list.
  32  */
  33 typedef struct MirrorBuffer {
  34     QSIMPLEQ_ENTRY(MirrorBuffer) next;
  35 } MirrorBuffer;
  36
  37 typedef struct MirrorBlockJob {
  38     BlockJob common;
  39     RateLimit limit;
  40     BlockBackend *target;
  41     BlockDriverState *base;
  42     /* The name of the graph node to replace */
  43     char *replaces;
  44     /* The BDS to replace */
  45     BlockDriverState *to_replace;
  46     /* Used to block operations on the drive-mirror-replace target */
  47     Error *replace_blocker;
  48     bool is_none_mode;
  49     BlockMirrorBackingMode backing_mode;
  50     BlockdevOnError on_source_error, on_target_error;
  51     bool synced;
  52     bool should_complete;
  53     int64_t granularity;
  54     size_t buf_size;
  55     int64_t bdev_length;
  56     unsigned long *cow_bitmap;
  57     BdrvDirtyBitmap *dirty_bitmap;
  58     BdrvDirtyBitmapIter *dbi;
  59     uint8_t *buf;
  60     QSIMPLEQ_HEAD(, MirrorBuffer) buf_free;
  61     int buf_free_count;
  62
  63     uint64_t last_pause_ns;
  64     unsigned long *in_flight_bitmap;
  65     int in_flight;
  66     int64_t sectors_in_flight;
  67     int ret;
  68     bool unmap;
  69     bool waiting_for_io;
  70     int target_cluster_sectors;
  71     int max_iov;
  72     bool initial_zeroing_ongoing;
  73 } MirrorBlockJob;
  74
  75 typedef struct MirrorOp {
  76     MirrorBlockJob *s;
  77     QEMUIOVector qiov;
  78     int64_t sector_num;
  79     int nb_sectors;
  80 } MirrorOp;
  81
  82 static BlockErrorAction mirror_error_action(MirrorBlockJob *s, bool read,
  83                                             int error)
  84 {
  85     s->synced = false;
  86     if (read) {
  87         return block_job_error_action(&s->common, s->on_source_error,
  88                                       true, error);
  89     } else {
  90         return block_job_error_action(&s->common, s->on_target_error,
  91                                       false, error);
  92     }
  93 }
  94
  95 static void mirror_iteration_done(MirrorOp *op, int ret)
  96 {
  97     MirrorBlockJob *s = op->s;
  98     struct iovec *iov;
  99     int64_t chunk_num;
 100     int i, nb_chunks, sectors_per_chunk;
 101
 102     trace_mirror_iteration_done(s, op->sector_num, op->nb_sectors, ret);
 103
 104     s->in_flight--;
 105     s->sectors_in_flight -= op->nb_sectors;
 106     iov = op->qiov.iov;
 107     for (i = 0; i < op->qiov.niov; i++) {
 108         MirrorBuffer *buf = (MirrorBuffer *) iov[i].iov_base;
 109         QSIMPLEQ_INSERT_TAIL(&s->buf_free, buf, next);
 110         s->buf_free_count++;
 111     }
 112
 113     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 114     chunk_num = op->sector_num / sectors_per_chunk;
 115     nb_chunks = DIV_ROUND_UP(op->nb_sectors, sectors_per_chunk);
 116     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
 117     if (ret >= 0) {
 118         if (s->cow_bitmap) {
 119             bitmap_set(s->cow_bitmap, chunk_num, nb_chunks);
 120         }
 121         if (!s->initial_zeroing_ongoing) {
 122             s->common.offset += (uint64_t)op->nb_sectors * BDRV_SECTOR_SIZE;
 123         }
 124     }
 125     qemu_iovec_destroy(&op->qiov);
 126     g_free(op);
 127
 128     if (s->waiting_for_io) {
 129         qemu_coroutine_enter(s->common.co);
 130     }
 131 }
 132
 133 static void mirror_write_complete(void *opaque, int ret)
 134 {
 135     MirrorOp *op = opaque;
 136     MirrorBlockJob *s = op->s;
 137
 138     aio_context_acquire(blk_get_aio_context(s->common.blk));
 139     if (ret < 0) {
 140         BlockErrorAction action;
 141
 142         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->sector_num, op->nb_sectors);
 143         action = mirror_error_action(s, false, -ret);
 144         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
 145             s->ret = ret;
 146         }
 147     }
 148     mirror_iteration_done(op, ret);
 149     aio_context_release(blk_get_aio_context(s->common.blk));
 150 }
 151
 152 static void mirror_read_complete(void *opaque, int ret)
 153 {
 154     MirrorOp *op = opaque;
 155     MirrorBlockJob *s = op->s;
 156
 157     aio_context_acquire(blk_get_aio_context(s->common.blk));
 158     if (ret < 0) {
 159         BlockErrorAction action;
 160
 161         bdrv_set_dirty_bitmap(s->dirty_bitmap, op->sector_num, op->nb_sectors);
 162         action = mirror_error_action(s, true, -ret);
 163         if (action == BLOCK_ERROR_ACTION_REPORT && s->ret >= 0) {
 164             s->ret = ret;
 165         }
 166
 167         mirror_iteration_done(op, ret);
 168     } else {
 169         blk_aio_pwritev(s->target, op->sector_num * BDRV_SECTOR_SIZE, &op->qiov,
 170                         0, mirror_write_complete, op);
 171     }
 172     aio_context_release(blk_get_aio_context(s->common.blk));
 173 }
 174
 175 static inline void mirror_clip_sectors(MirrorBlockJob *s,
 176                                        int64_t sector_num,
 177                                        int *nb_sectors)
 178 {
 179     *nb_sectors = MIN(*nb_sectors,
 180                       s->bdev_length / BDRV_SECTOR_SIZE - sector_num);
 181 }
 182
 183 /* Round sector_num and/or nb_sectors to target cluster if COW is needed, and
 184  * return the offset of the adjusted tail sector against original. */
 185 static int mirror_cow_align(MirrorBlockJob *s,
 186                             int64_t *sector_num,
 187                             int *nb_sectors)
 188 {
 189     bool need_cow;
 190     int ret = 0;
 191     int chunk_sectors = s->granularity >> BDRV_SECTOR_BITS;
 192     int64_t align_sector_num = *sector_num;
 193     int align_nb_sectors = *nb_sectors;
 194     int max_sectors = chunk_sectors * s->max_iov;
 195
 196     need_cow = !test_bit(*sector_num / chunk_sectors, s->cow_bitmap);
 197     need_cow |= !test_bit((*sector_num + *nb_sectors - 1) / chunk_sectors,
 198                           s->cow_bitmap);
 199     if (need_cow) {
 200         bdrv_round_sectors_to_clusters(blk_bs(s->target), *sector_num,
 201                                        *nb_sectors, &align_sector_num,
 202                                        &align_nb_sectors);
 203     }
 204
 205     if (align_nb_sectors > max_sectors) {
 206         align_nb_sectors = max_sectors;
 207         if (need_cow) {
 208             align_nb_sectors = QEMU_ALIGN_DOWN(align_nb_sectors,
 209                                                s->target_cluster_sectors);
 210         }
 211     }
 212     /* Clipping may result in align_nb_sectors unaligned to chunk boundary, but
 213      * that doesn't matter because it's already the end of source image. */
 214     mirror_clip_sectors(s, align_sector_num, &align_nb_sectors);
 215
 216     ret = align_sector_num + align_nb_sectors - (*sector_num + *nb_sectors);
 217     *sector_num = align_sector_num;
 218     *nb_sectors = align_nb_sectors;
 219     assert(ret >= 0);
 220     return ret;
 221 }
 222
 223 static inline void mirror_wait_for_io(MirrorBlockJob *s)
 224 {
 225     assert(!s->waiting_for_io);
 226     s->waiting_for_io = true;
 227     qemu_coroutine_yield();
 228     s->waiting_for_io = false;
 229 }
 230
 231 /* Submit async read while handling COW.
 232  * Returns: The number of sectors copied after and including sector_num,
 233  *          excluding any sectors copied prior to sector_num due to alignment.
 234  *          This will be nb_sectors if no alignment is necessary, or
 235  *          (new_end - sector_num) if tail is rounded up or down due to
 236  *          alignment or buffer limit.
 237  */
 238 static int mirror_do_read(MirrorBlockJob *s, int64_t sector_num,
 239                           int nb_sectors)
 240 {
 241     BlockBackend *source = s->common.blk;
 242     int sectors_per_chunk, nb_chunks;
 243     int ret;
 244     MirrorOp *op;
 245     int max_sectors;
 246
 247     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 248     max_sectors = sectors_per_chunk * s->max_iov;
 249
 250     /* We can only handle as much as buf_size at a time. */
 251     nb_sectors = MIN(s->buf_size >> BDRV_SECTOR_BITS, nb_sectors);
 252     nb_sectors = MIN(max_sectors, nb_sectors);
 253     assert(nb_sectors);
 254     ret = nb_sectors;
 255
 256     if (s->cow_bitmap) {
 257         ret += mirror_cow_align(s, &sector_num, &nb_sectors);
 258     }
 259     assert(nb_sectors << BDRV_SECTOR_BITS <= s->buf_size);
 260     /* The sector range must meet granularity because:
 261      * 1) Caller passes in aligned values;
 262      * 2) mirror_cow_align is used only when target cluster is larger. */
 263     assert(!(sector_num % sectors_per_chunk));
 264     nb_chunks = DIV_ROUND_UP(nb_sectors, sectors_per_chunk);
 265
 266     while (s->buf_free_count < nb_chunks) {
 267         trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
 268         mirror_wait_for_io(s);
 269     }
 270
 271     /* Allocate a MirrorOp that is used as an AIO callback.  */
 272     op = g_new(MirrorOp, 1);
 273     op->s = s;
 274     op->sector_num = sector_num;
 275     op->nb_sectors = nb_sectors;
 276
 277     /* Now make a QEMUIOVector taking enough granularity-sized chunks
 278      * from s->buf_free.
 279      */
 280     qemu_iovec_init(&op->qiov, nb_chunks);
 281     while (nb_chunks-- > 0) {
 282         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
 283         size_t remaining = nb_sectors * BDRV_SECTOR_SIZE - op->qiov.size;
 284
 285         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
 286         s->buf_free_count--;
 287         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
 288     }
 289
 290     /* Copy the dirty cluster.  */
 291     s->in_flight++;
 292     s->sectors_in_flight += nb_sectors;
 293     trace_mirror_one_iteration(s, sector_num, nb_sectors);
 294
 295     blk_aio_preadv(source, sector_num * BDRV_SECTOR_SIZE, &op->qiov, 0,
 296                    mirror_read_complete, op);
 297     return ret;
 298 }
 299
 300 static void mirror_do_zero_or_discard(MirrorBlockJob *s,
 301                                       int64_t sector_num,
 302                                       int nb_sectors,
 303                                       bool is_discard)
 304 {
 305     MirrorOp *op;
 306
 307     /* Allocate a MirrorOp that is used as an AIO callback. The qiov is zeroed
 308      * so the freeing in mirror_iteration_done is nop. */
 309     op = g_new0(MirrorOp, 1);
 310     op->s = s;
 311     op->sector_num = sector_num;
 312     op->nb_sectors = nb_sectors;
 313
 314     s->in_flight++;
 315     s->sectors_in_flight += nb_sectors;
 316     if (is_discard) {
 317         blk_aio_pdiscard(s->target, sector_num << BDRV_SECTOR_BITS,
 318                          op->nb_sectors << BDRV_SECTOR_BITS,
 319                          mirror_write_complete, op);
 320     } else {
 321         blk_aio_pwrite_zeroes(s->target, sector_num * BDRV_SECTOR_SIZE,
 322                               op->nb_sectors * BDRV_SECTOR_SIZE,
 323                               s->unmap ? BDRV_REQ_MAY_UNMAP : 0,
 324                               mirror_write_complete, op);
 325     }
 326 }
 327
 328 static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
 329 {
 330     BlockDriverState *source = blk_bs(s->common.blk);
 331     int64_t sector_num, first_chunk;
 332     uint64_t delay_ns = 0;
 333     /* At least the first dirty chunk is mirrored in one iteration. */
 334     int nb_chunks = 1;
 335     int64_t end = s->bdev_length / BDRV_SECTOR_SIZE;
 336     int sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 337     bool write_zeroes_ok = bdrv_can_write_zeroes_with_unmap(blk_bs(s->target));
 338     int max_io_sectors = MAX((s->buf_size >> BDRV_SECTOR_BITS) / MAX_IN_FLIGHT,
 339                              MAX_IO_SECTORS);
 340
 341     sector_num = bdrv_dirty_iter_next(s->dbi);
 342     if (sector_num < 0) {
 343         bdrv_set_dirty_iter(s->dbi, 0);
 344         sector_num = bdrv_dirty_iter_next(s->dbi);
 345         trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
 346         assert(sector_num >= 0);
 347     }
 348
 349     first_chunk = sector_num / sectors_per_chunk;
 350     while (test_bit(first_chunk, s->in_flight_bitmap)) {
 351         trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
 352         mirror_wait_for_io(s);
 353     }
 354
 355     block_job_pause_point(&s->common);
 356
 357     /* Find the number of consective dirty chunks following the first dirty
 358      * one, and wait for in flight requests in them. */
 359     while (nb_chunks * sectors_per_chunk < (s->buf_size >> BDRV_SECTOR_BITS)) {
 360         int64_t next_dirty;
 361         int64_t next_sector = sector_num + nb_chunks * sectors_per_chunk;
 362         int64_t next_chunk = next_sector / sectors_per_chunk;
 363         if (next_sector >= end ||
 364             !bdrv_get_dirty(source, s->dirty_bitmap, next_sector)) {
 365             break;
 366         }
 367         if (test_bit(next_chunk, s->in_flight_bitmap)) {
 368             break;
 369         }
 370
 371         next_dirty = bdrv_dirty_iter_next(s->dbi);
 372         if (next_dirty > next_sector || next_dirty < 0) {
 373             /* The bitmap iterator's cache is stale, refresh it */
 374             bdrv_set_dirty_iter(s->dbi, next_sector);
 375             next_dirty = bdrv_dirty_iter_next(s->dbi);
 376         }
 377         assert(next_dirty == next_sector);
 378         nb_chunks++;
 379     }
 380
 381     /* Clear dirty bits before querying the block status, because
 382      * calling bdrv_get_block_status_above could yield - if some blocks are
 383      * marked dirty in this window, we need to know.
 384      */
 385     bdrv_reset_dirty_bitmap(s->dirty_bitmap, sector_num,
 386                             nb_chunks * sectors_per_chunk);
 387     bitmap_set(s->in_flight_bitmap, sector_num / sectors_per_chunk, nb_chunks);
 388     while (nb_chunks > 0 && sector_num < end) {
 389         int ret;
 390         int io_sectors, io_sectors_acct;
 391         BlockDriverState *file;
 392         enum MirrorMethod {
 393             MIRROR_METHOD_COPY,
 394             MIRROR_METHOD_ZERO,
 395             MIRROR_METHOD_DISCARD
 396         } mirror_method = MIRROR_METHOD_COPY;
 397
 398         assert(!(sector_num % sectors_per_chunk));
 399         ret = bdrv_get_block_status_above(source, NULL, sector_num,
 400                                           nb_chunks * sectors_per_chunk,
 401                                           &io_sectors, &file);
 402         if (ret < 0) {
 403             io_sectors = MIN(nb_chunks * sectors_per_chunk, max_io_sectors);
 404         } else if (ret & BDRV_BLOCK_DATA) {
 405             io_sectors = MIN(io_sectors, max_io_sectors);
 406         }
 407
 408         io_sectors -= io_sectors % sectors_per_chunk;
 409         if (io_sectors < sectors_per_chunk) {
 410             io_sectors = sectors_per_chunk;
 411         } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
 412             int64_t target_sector_num;
 413             int target_nb_sectors;
 414             bdrv_round_sectors_to_clusters(blk_bs(s->target), sector_num,
 415                                            io_sectors,  &target_sector_num,
 416                                            &target_nb_sectors);
 417             if (target_sector_num == sector_num &&
 418                 target_nb_sectors == io_sectors) {
 419                 mirror_method = ret & BDRV_BLOCK_ZERO ?
 420                                     MIRROR_METHOD_ZERO :
 421                                     MIRROR_METHOD_DISCARD;
 422             }
 423         }
 424
 425         while (s->in_flight >= MAX_IN_FLIGHT) {
 426             trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
 427             mirror_wait_for_io(s);
 428         }
 429
 430         if (s->ret < 0) {
 431             return 0;
 432         }
 433
 434         mirror_clip_sectors(s, sector_num, &io_sectors);
 435         switch (mirror_method) {
 436         case MIRROR_METHOD_COPY:
 437             io_sectors = mirror_do_read(s, sector_num, io_sectors);
 438             io_sectors_acct = io_sectors;
 439             break;
 440         case MIRROR_METHOD_ZERO:
 441         case MIRROR_METHOD_DISCARD:
 442             mirror_do_zero_or_discard(s, sector_num, io_sectors,
 443                                       mirror_method == MIRROR_METHOD_DISCARD);
 444             if (write_zeroes_ok) {
 445                 io_sectors_acct = 0;
 446             } else {
 447                 io_sectors_acct = io_sectors;
 448             }
 449             break;
 450         default:
 451             abort();
 452         }
 453         assert(io_sectors);
 454         sector_num += io_sectors;
 455         nb_chunks -= DIV_ROUND_UP(io_sectors, sectors_per_chunk);
 456         if (s->common.speed) {
 457             delay_ns = ratelimit_calculate_delay(&s->limit, io_sectors_acct);
 458         }
 459     }
 460     return delay_ns;
 461 }
 462
 463 static void mirror_free_init(MirrorBlockJob *s)
 464 {
 465     int granularity = s->granularity;
 466     size_t buf_size = s->buf_size;
 467     uint8_t *buf = s->buf;
 468
 469     assert(s->buf_free_count == 0);
 470     QSIMPLEQ_INIT(&s->buf_free);
 471     while (buf_size != 0) {
 472         MirrorBuffer *cur = (MirrorBuffer *)buf;
 473         QSIMPLEQ_INSERT_TAIL(&s->buf_free, cur, next);
 474         s->buf_free_count++;
 475         buf_size -= granularity;
 476         buf += granularity;
 477     }
 478 }
 479
 480 /* This is also used for the .pause callback. There is no matching
 481  * mirror_resume() because mirror_run() will begin iterating again
 482  * when the job is resumed.
 483  */
 484 static void mirror_wait_for_all_io(MirrorBlockJob *s)
 485 {
 486     while (s->in_flight > 0) {
 487         mirror_wait_for_io(s);
 488     }
 489 }
 490
 491 typedef struct {
 492     int ret;
 493 } MirrorExitData;
 494
 495 static void mirror_exit(BlockJob *job, void *opaque)
 496 {
 497     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 498     MirrorExitData *data = opaque;
 499     AioContext *replace_aio_context = NULL;
 500     BlockDriverState *src = blk_bs(s->common.blk);
 501     BlockDriverState *target_bs = blk_bs(s->target);
 502
 503     /* Make sure that the source BDS doesn't go away before we called
 504      * block_job_completed(). */
 505     bdrv_ref(src);
 506
 507     if (s->to_replace) {
 508         replace_aio_context = bdrv_get_aio_context(s->to_replace);
 509         aio_context_acquire(replace_aio_context);
 510     }
 511
 512     if (s->should_complete && data->ret == 0) {
 513         BlockDriverState *to_replace = src;
 514         if (s->to_replace) {
 515             to_replace = s->to_replace;
 516         }
 517
 518         if (bdrv_get_flags(target_bs) != bdrv_get_flags(to_replace)) {
 519             bdrv_reopen(target_bs, bdrv_get_flags(to_replace), NULL);
 520         }
 521
 522         /* The mirror job has no requests in flight any more, but we need to
 523          * drain potential other users of the BDS before changing the graph. */
 524         bdrv_drained_begin(target_bs);
 525         bdrv_replace_in_backing_chain(to_replace, target_bs);
 526         bdrv_drained_end(target_bs);
 527
 528         /* We just changed the BDS the job BB refers to */
 529         blk_remove_bs(job->blk);
 530         blk_insert_bs(job->blk, src);
 531     }
 532     if (s->to_replace) {
 533         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
 534         error_free(s->replace_blocker);
 535         bdrv_unref(s->to_replace);
 536     }
 537     if (replace_aio_context) {
 538         aio_context_release(replace_aio_context);
 539     }
 540     g_free(s->replaces);
 541     blk_unref(s->target);
 542     s->target = NULL;
 543     block_job_completed(&s->common, data->ret);
 544     g_free(data);
 545     bdrv_drained_end(src);
 546     bdrv_unref(src);
 547 }
 548
 549 static void mirror_throttle(MirrorBlockJob *s)
 550 {
 551     int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 552
 553     if (now - s->last_pause_ns > SLICE_TIME) {
 554         s->last_pause_ns = now;
 555         block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, 0);
 556     } else {
 557         block_job_pause_point(&s->common);
 558     }
 559 }
 560
 561 static int coroutine_fn mirror_dirty_init(MirrorBlockJob *s)
 562 {
 563     int64_t sector_num, end;
 564     BlockDriverState *base = s->base;
 565     BlockDriverState *bs = blk_bs(s->common.blk);
 566     BlockDriverState *target_bs = blk_bs(s->target);
 567     int ret, n;
 568
 569     end = s->bdev_length / BDRV_SECTOR_SIZE;
 570
 571     if (base == NULL && !bdrv_has_zero_init(target_bs)) {
 572         if (!bdrv_can_write_zeroes_with_unmap(target_bs)) {
 573             bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, end);
 574             return 0;
 575         }
 576
 577         s->initial_zeroing_ongoing = true;
 578         for (sector_num = 0; sector_num < end; ) {
 579             int nb_sectors = MIN(end - sector_num,
 580                 QEMU_ALIGN_DOWN(INT_MAX, s->granularity) >> BDRV_SECTOR_BITS);
 581
 582             mirror_throttle(s);
 583
 584             if (block_job_is_cancelled(&s->common)) {
 585                 s->initial_zeroing_ongoing = false;
 586                 return 0;
 587             }
 588
 589             if (s->in_flight >= MAX_IN_FLIGHT) {
 590                 trace_mirror_yield(s, s->in_flight, s->buf_free_count, -1);
 591                 mirror_wait_for_io(s);
 592                 continue;
 593             }
 594
 595             mirror_do_zero_or_discard(s, sector_num, nb_sectors, false);
 596             sector_num += nb_sectors;
 597         }
 598
 599         mirror_wait_for_all_io(s);
 600         s->initial_zeroing_ongoing = false;
 601     }
 602
 603     /* First part, loop on the sectors and initialize the dirty bitmap.  */
 604     for (sector_num = 0; sector_num < end; ) {
 605         /* Just to make sure we are not exceeding int limit. */
 606         int nb_sectors = MIN(INT_MAX >> BDRV_SECTOR_BITS,
 607                              end - sector_num);
 608
 609         mirror_throttle(s);
 610
 611         if (block_job_is_cancelled(&s->common)) {
 612             return 0;
 613         }
 614
 615         ret = bdrv_is_allocated_above(bs, base, sector_num, nb_sectors, &n);
 616         if (ret < 0) {
 617             return ret;
 618         }
 619
 620         assert(n > 0);
 621         if (ret == 1) {
 622             bdrv_set_dirty_bitmap(s->dirty_bitmap, sector_num, n);
 623         }
 624         sector_num += n;
 625     }
 626     return 0;
 627 }
 628
 629 /* Called when going out of the streaming phase to flush the bulk of the
 630  * data to the medium, or just before completing.
 631  */
 632 static int mirror_flush(MirrorBlockJob *s)
 633 {
 634     int ret = blk_flush(s->target);
 635     if (ret < 0) {
 636         if (mirror_error_action(s, false, -ret) == BLOCK_ERROR_ACTION_REPORT) {
 637             s->ret = ret;
 638         }
 639     }
 640     return ret;
 641 }
 642
 643 static void coroutine_fn mirror_run(void *opaque)
 644 {
 645     MirrorBlockJob *s = opaque;
 646     MirrorExitData *data;
 647     BlockDriverState *bs = blk_bs(s->common.blk);
 648     BlockDriverState *target_bs = blk_bs(s->target);
 649     bool need_drain = true;
 650     int64_t length;
 651     BlockDriverInfo bdi;
 652     char backing_filename[2]; /* we only need 2 characters because we are only
 653                                  checking for a NULL string */
 654     int ret = 0;
 655     int target_cluster_size = BDRV_SECTOR_SIZE;
 656
 657     if (block_job_is_cancelled(&s->common)) {
 658         goto immediate_exit;
 659     }
 660
 661     s->bdev_length = bdrv_getlength(bs);
 662     if (s->bdev_length < 0) {
 663         ret = s->bdev_length;
 664         goto immediate_exit;
 665     }
 666
 667     /* Active commit must resize the base image if its size differs from the
 668      * active layer. */
 669     if (s->base == blk_bs(s->target)) {
 670         int64_t base_length;
 671
 672         base_length = blk_getlength(s->target);
 673         if (base_length < 0) {
 674             ret = base_length;
 675             goto immediate_exit;
 676         }
 677
 678         if (s->bdev_length > base_length) {
 679             ret = blk_truncate(s->target, s->bdev_length);
 680             if (ret < 0) {
 681                 goto immediate_exit;
 682             }
 683         }
 684     }
 685
 686     if (s->bdev_length == 0) {
 687         /* Report BLOCK_JOB_READY and wait for complete. */
 688         block_job_event_ready(&s->common);
 689         s->synced = true;
 690         while (!block_job_is_cancelled(&s->common) && !s->should_complete) {
 691             block_job_yield(&s->common);
 692         }
 693         s->common.cancelled = false;
 694         goto immediate_exit;
 695     }
 696
 697     length = DIV_ROUND_UP(s->bdev_length, s->granularity);
 698     s->in_flight_bitmap = bitmap_new(length);
 699
 700     /* If we have no backing file yet in the destination, we cannot let
 701      * the destination do COW.  Instead, we copy sectors around the
 702      * dirty data if needed.  We need a bitmap to do that.
 703      */
 704     bdrv_get_backing_filename(target_bs, backing_filename,
 705                               sizeof(backing_filename));
 706     if (!bdrv_get_info(target_bs, &bdi) && bdi.cluster_size) {
 707         target_cluster_size = bdi.cluster_size;
 708     }
 709     if (backing_filename[0] && !target_bs->backing
 710         && s->granularity < target_cluster_size) {
 711         s->buf_size = MAX(s->buf_size, target_cluster_size);
 712         s->cow_bitmap = bitmap_new(length);
 713     }
 714     s->target_cluster_sectors = target_cluster_size >> BDRV_SECTOR_BITS;
 715     s->max_iov = MIN(bs->bl.max_iov, target_bs->bl.max_iov);
 716
 717     s->buf = qemu_try_blockalign(bs, s->buf_size);
 718     if (s->buf == NULL) {
 719         ret = -ENOMEM;
 720         goto immediate_exit;
 721     }
 722
 723     mirror_free_init(s);
 724
 725     s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 726     if (!s->is_none_mode) {
 727         ret = mirror_dirty_init(s);
 728         if (ret < 0 || block_job_is_cancelled(&s->common)) {
 729             goto immediate_exit;
 730         }
 731     }
 732
 733     assert(!s->dbi);
 734     s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap, 0);
 735     for (;;) {
 736         uint64_t delay_ns = 0;
 737         int64_t cnt, delta;
 738         bool should_complete;
 739
 740         if (s->ret < 0) {
 741             ret = s->ret;
 742             goto immediate_exit;
 743         }
 744
 745         block_job_pause_point(&s->common);
 746
 747         cnt = bdrv_get_dirty_count(s->dirty_bitmap);
 748         /* s->common.offset contains the number of bytes already processed so
 749          * far, cnt is the number of dirty sectors remaining and
 750          * s->sectors_in_flight is the number of sectors currently being
 751          * processed; together those are the current total operation length */
 752         s->common.len = s->common.offset +
 753                         (cnt + s->sectors_in_flight) * BDRV_SECTOR_SIZE;
 754
 755         /* Note that even when no rate limit is applied we need to yield
 756          * periodically with no pending I/O so that bdrv_drain_all() returns.
 757          * We do so every SLICE_TIME nanoseconds, or when there is an error,
 758          * or when the source is clean, whichever comes first.
 759          */
 760         delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns;
 761         if (delta < SLICE_TIME &&
 762             s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {
 763             if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 ||
 764                 (cnt == 0 && s->in_flight > 0)) {
 765                 trace_mirror_yield(s, s->in_flight, s->buf_free_count, cnt);
 766                 mirror_wait_for_io(s);
 767                 continue;
 768             } else if (cnt != 0) {
 769                 delay_ns = mirror_iteration(s);
 770             }
 771         }
 772
 773         should_complete = false;
 774         if (s->in_flight == 0 && cnt == 0) {
 775             trace_mirror_before_flush(s);
 776             if (!s->synced) {
 777                 if (mirror_flush(s) < 0) {
 778                     /* Go check s->ret.  */
 779                     continue;
 780                 }
 781                 /* We're out of the streaming phase.  From now on, if the job
 782                  * is cancelled we will actually complete all pending I/O and
 783                  * report completion.  This way, block-job-cancel will leave
 784                  * the target in a consistent state.
 785                  */
 786                 block_job_event_ready(&s->common);
 787                 s->synced = true;
 788             }
 789
 790             should_complete = s->should_complete ||
 791                 block_job_is_cancelled(&s->common);
 792             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
 793         }
 794
 795         if (cnt == 0 && should_complete) {
 796             /* The dirty bitmap is not updated while operations are pending.
 797              * If we're about to exit, wait for pending operations before
 798              * calling bdrv_get_dirty_count(bs), or we may exit while the
 799              * source has dirty data to copy!
 800              *
 801              * Note that I/O can be submitted by the guest while
 802              * mirror_populate runs, so pause it now.  Before deciding
 803              * whether to switch to target check one last time if I/O has
 804              * come in the meanwhile, and if not flush the data to disk.
 805              */
 806             trace_mirror_before_drain(s, cnt);
 807
 808             bdrv_drained_begin(bs);
 809             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
 810             if (cnt > 0 || mirror_flush(s) < 0) {
 811                 bdrv_drained_end(bs);
 812                 continue;
 813             }
 814
 815             /* The two disks are in sync.  Exit and report successful
 816              * completion.
 817              */
 818             assert(QLIST_EMPTY(&bs->tracked_requests));
 819             s->common.cancelled = false;
 820             need_drain = false;
 821             break;
 822         }
 823
 824         ret = 0;
 825         trace_mirror_before_sleep(s, cnt, s->synced, delay_ns);
 826         if (!s->synced) {
 827             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 828             if (block_job_is_cancelled(&s->common)) {
 829                 break;
 830             }
 831         } else if (!should_complete) {
 832             delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0);
 833             block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
 834         }
 835         s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 836     }
 837
 838 immediate_exit:
 839     if (s->in_flight > 0) {
 840         /* We get here only if something went wrong.  Either the job failed,
 841          * or it was cancelled prematurely so that we do not guarantee that
 842          * the target is a copy of the source.
 843          */
 844         assert(ret < 0 || (!s->synced && block_job_is_cancelled(&s->common)));
 845         assert(need_drain);
 846         mirror_wait_for_all_io(s);
 847     }
 848
 849     assert(s->in_flight == 0);
 850     qemu_vfree(s->buf);
 851     g_free(s->cow_bitmap);
 852     g_free(s->in_flight_bitmap);
 853     bdrv_dirty_iter_free(s->dbi);
 854     bdrv_release_dirty_bitmap(bs, s->dirty_bitmap);
 855
 856     data = g_malloc(sizeof(*data));
 857     data->ret = ret;
 858
 859     if (need_drain) {
 860         bdrv_drained_begin(bs);
 861     }
 862     block_job_defer_to_main_loop(&s->common, mirror_exit, data);
 863 }
 864
 865 static void mirror_set_speed(BlockJob *job, int64_t speed, Error **errp)
 866 {
 867     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 868
 869     if (speed < 0) {
 870         error_setg(errp, QERR_INVALID_PARAMETER, "speed");
 871         return;
 872     }
 873     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 874 }
 875
 876 static void mirror_complete(BlockJob *job, Error **errp)
 877 {
 878     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 879     BlockDriverState *src, *target;
 880
 881     src = blk_bs(job->blk);
 882     target = blk_bs(s->target);
 883
 884     if (!s->synced) {
 885         error_setg(errp, "The active block job '%s' cannot be completed",
 886                    job->id);
 887         return;
 888     }
 889
 890     if (s->backing_mode == MIRROR_OPEN_BACKING_CHAIN) {
 891         int ret;
 892
 893         assert(!target->backing);
 894         ret = bdrv_open_backing_file(target, NULL, "backing", errp);
 895         if (ret < 0) {
 896             return;
 897         }
 898     }
 899
 900     /* block all operations on to_replace bs */
 901     if (s->replaces) {
 902         AioContext *replace_aio_context;
 903
 904         s->to_replace = bdrv_find_node(s->replaces);
 905         if (!s->to_replace) {
 906             error_setg(errp, "Node name '%s' not found", s->replaces);
 907             return;
 908         }
 909
 910         replace_aio_context = bdrv_get_aio_context(s->to_replace);
 911         aio_context_acquire(replace_aio_context);
 912
 913         error_setg(&s->replace_blocker,
 914                    "block device is in use by block-job-complete");
 915         bdrv_op_block_all(s->to_replace, s->replace_blocker);
 916         bdrv_ref(s->to_replace);
 917
 918         aio_context_release(replace_aio_context);
 919     }
 920
 921     if (s->backing_mode == MIRROR_SOURCE_BACKING_CHAIN) {
 922         BlockDriverState *backing = s->is_none_mode ? src : s->base;
 923         if (backing_bs(target) != backing) {
 924             bdrv_set_backing_hd(target, backing);
 925         }
 926     }
 927
 928     s->should_complete = true;
 929     block_job_enter(&s->common);
 930 }
 931
 932 static void mirror_pause(BlockJob *job)
 933 {
 934     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 935
 936     mirror_wait_for_all_io(s);
 937 }
 938
 939 static void mirror_attached_aio_context(BlockJob *job, AioContext *new_context)
 940 {
 941     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 942
 943     blk_set_aio_context(s->target, new_context);
 944 }
 945
 946 static void mirror_drain(BlockJob *job)
 947 {
 948     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 949
 950     /* Need to keep a reference in case blk_drain triggers execution
 951      * of mirror_complete...
 952      */
 953     if (s->target) {
 954         BlockBackend *target = s->target;
 955         blk_ref(target);
 956         blk_drain(target);
 957         blk_unref(target);
 958     }
 959 }
 960
 961 static const BlockJobDriver mirror_job_driver = {
 962     .instance_size          = sizeof(MirrorBlockJob),
 963     .job_type               = BLOCK_JOB_TYPE_MIRROR,
 964     .set_speed              = mirror_set_speed,
 965     .start                  = mirror_run,
 966     .complete               = mirror_complete,
 967     .pause                  = mirror_pause,
 968     .attached_aio_context   = mirror_attached_aio_context,
 969     .drain                  = mirror_drain,
 970 };
 971
 972 static const BlockJobDriver commit_active_job_driver = {
 973     .instance_size          = sizeof(MirrorBlockJob),
 974     .job_type               = BLOCK_JOB_TYPE_COMMIT,
 975     .set_speed              = mirror_set_speed,
 976     .start                  = mirror_run,
 977     .complete               = mirror_complete,
 978     .pause                  = mirror_pause,
 979     .attached_aio_context   = mirror_attached_aio_context,
 980     .drain                  = mirror_drain,
 981 };
 982
 983 static void mirror_start_job(const char *job_id, BlockDriverState *bs,
 984                              int creation_flags, BlockDriverState *target,
 985                              const char *replaces, int64_t speed,
 986                              uint32_t granularity, int64_t buf_size,
 987                              BlockMirrorBackingMode backing_mode,
 988                              BlockdevOnError on_source_error,
 989                              BlockdevOnError on_target_error,
 990                              bool unmap,
 991                              BlockCompletionFunc *cb,
 992                              void *opaque, Error **errp,
 993                              const BlockJobDriver *driver,
 994                              bool is_none_mode, BlockDriverState *base,
 995                              bool auto_complete)
 996 {
 997     MirrorBlockJob *s;
 998
 999     if (granularity == 0) {
1000         granularity = bdrv_get_default_bitmap_granularity(target);
1001     }
1002
1003     assert ((granularity & (granularity - 1)) == 0);
1004
1005     if (buf_size < 0) {
1006         error_setg(errp, "Invalid parameter 'buf-size'");
1007         return;
1008     }
1009
1010     if (buf_size == 0) {
1011         buf_size = DEFAULT_MIRROR_BUF_SIZE;
1012     }
1013
1014     s = block_job_create(job_id, driver, bs, speed, creation_flags,
1015                          cb, opaque, errp);
1016     if (!s) {
1017         return;
1018     }
1019
1020     s->target = blk_new();
1021     blk_insert_bs(s->target, target);
1022
1023     s->replaces = g_strdup(replaces);
1024     s->on_source_error = on_source_error;
1025     s->on_target_error = on_target_error;
1026     s->is_none_mode = is_none_mode;
1027     s->backing_mode = backing_mode;
1028     s->base = base;
1029     s->granularity = granularity;
1030     s->buf_size = ROUND_UP(buf_size, granularity);
1031     s->unmap = unmap;
1032     if (auto_complete) {
1033         s->should_complete = true;
1034     }
1035
1036     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
1037     if (!s->dirty_bitmap) {
1038         g_free(s->replaces);
1039         blk_unref(s->target);
1040         block_job_unref(&s->common);
1041         return;
1042     }
1043
1044     block_job_add_bdrv(&s->common, target);
1045     /* In commit_active_start() all intermediate nodes disappear, so
1046      * any jobs in them must be blocked */
1047     if (bdrv_chain_contains(bs, target)) {
1048         BlockDriverState *iter;
1049         for (iter = backing_bs(bs); iter != target; iter = backing_bs(iter)) {
1050             block_job_add_bdrv(&s->common, iter);
1051         }
1052     }
1053
1054     trace_mirror_start(bs, s, opaque);
1055     block_job_start(&s->common);
1056 }
1057
1058 void mirror_start(const char *job_id, BlockDriverState *bs,
1059                   BlockDriverState *target, const char *replaces,
1060                   int64_t speed, uint32_t granularity, int64_t buf_size,
1061                   MirrorSyncMode mode, BlockMirrorBackingMode backing_mode,
1062                   BlockdevOnError on_source_error,
1063                   BlockdevOnError on_target_error,
1064                   bool unmap, Error **errp)
1065 {
1066     bool is_none_mode;
1067     BlockDriverState *base;
1068
1069     if (mode == MIRROR_SYNC_MODE_INCREMENTAL) {
1070         error_setg(errp, "Sync mode 'incremental' not supported");
1071         return;
1072     }
1073     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
1074     base = mode == MIRROR_SYNC_MODE_TOP ? backing_bs(bs) : NULL;
1075     mirror_start_job(job_id, bs, BLOCK_JOB_DEFAULT, target, replaces,
1076                      speed, granularity, buf_size, backing_mode,
1077                      on_source_error, on_target_error, unmap, NULL, NULL, errp,
1078                      &mirror_job_driver, is_none_mode, base, false);
1079 }
1080
1081 void commit_active_start(const char *job_id, BlockDriverState *bs,
1082                          BlockDriverState *base, int creation_flags,
1083                          int64_t speed, BlockdevOnError on_error,
1084                          BlockCompletionFunc *cb, void *opaque, Error **errp,
1085                          bool auto_complete)
1086 {
1087     int orig_base_flags;
1088     Error *local_err = NULL;
1089
1090     orig_base_flags = bdrv_get_flags(base);
1091
1092     if (bdrv_reopen(base, bs->open_flags, errp)) {
1093         return;
1094     }
1095
1096     mirror_start_job(job_id, bs, creation_flags, base, NULL, speed, 0, 0,
1097                      MIRROR_LEAVE_BACKING_CHAIN,
1098                      on_error, on_error, true, cb, opaque, &local_err,
1099                      &commit_active_job_driver, false, base, auto_complete);
1100     if (local_err) {
1101         error_propagate(errp, local_err);
1102         goto error_restore_flags;
1103     }
1104
1105     return;
1106
1107 error_restore_flags:
1108     /* ignore error and errp for bdrv_reopen, because we want to propagate
1109      * the original error */
1110     bdrv_reopen(base, orig_base_flags, NULL);
1111     return;
1112 }