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