sys/contrib/openzfs/module/zfs/vdev_mirror.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or https://opensource.org/licenses/CDDL-1.0.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
  23  * Use is subject to license terms.
  24  */
  25
  26 /*
  27  * Copyright (c) 2012, 2015 by Delphix. All rights reserved.
  28  */
  29
  30 #include <sys/zfs_context.h>
  31 #include <sys/spa.h>
  32 #include <sys/spa_impl.h>
  33 #include <sys/dsl_pool.h>
  34 #include <sys/dsl_scan.h>
  35 #include <sys/vdev_impl.h>
  36 #include <sys/vdev_draid.h>
  37 #include <sys/zio.h>
  38 #include <sys/zio_checksum.h>
  39 #include <sys/abd.h>
  40 #include <sys/fs/zfs.h>
  41
  42 /*
  43  * Vdev mirror kstats
  44  */
  45 static kstat_t *mirror_ksp = NULL;
  46
  47 typedef struct mirror_stats {
  48         kstat_named_t vdev_mirror_stat_rotating_linear;
  49         kstat_named_t vdev_mirror_stat_rotating_offset;
  50         kstat_named_t vdev_mirror_stat_rotating_seek;
  51         kstat_named_t vdev_mirror_stat_non_rotating_linear;
  52         kstat_named_t vdev_mirror_stat_non_rotating_seek;
  53
  54         kstat_named_t vdev_mirror_stat_preferred_found;
  55         kstat_named_t vdev_mirror_stat_preferred_not_found;
  56 } mirror_stats_t;
  57
  58 static mirror_stats_t mirror_stats = {
  59         /* New I/O follows directly the last I/O */
  60         { "rotating_linear",                    KSTAT_DATA_UINT64 },
  61         /* New I/O is within zfs_vdev_mirror_rotating_seek_offset of the last */
  62         { "rotating_offset",                    KSTAT_DATA_UINT64 },
  63         /* New I/O requires random seek */
  64         { "rotating_seek",                      KSTAT_DATA_UINT64 },
  65         /* New I/O follows directly the last I/O  (nonrot) */
  66         { "non_rotating_linear",                KSTAT_DATA_UINT64 },
  67         /* New I/O requires random seek (nonrot) */
  68         { "non_rotating_seek",                  KSTAT_DATA_UINT64 },
  69         /* Preferred child vdev found */
  70         { "preferred_found",                    KSTAT_DATA_UINT64 },
  71         /* Preferred child vdev not found or equal load  */
  72         { "preferred_not_found",                KSTAT_DATA_UINT64 },
  73
  74 };
  75
  76 #define MIRROR_STAT(stat)               (mirror_stats.stat.value.ui64)
  77 #define MIRROR_INCR(stat, val)          atomic_add_64(&MIRROR_STAT(stat), val)
  78 #define MIRROR_BUMP(stat)               MIRROR_INCR(stat, 1)
  79
  80 void
  81 vdev_mirror_stat_init(void)
  82 {
  83         mirror_ksp = kstat_create("zfs", 0, "vdev_mirror_stats",
  84             "misc", KSTAT_TYPE_NAMED,
  85             sizeof (mirror_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
  86         if (mirror_ksp != NULL) {
  87                 mirror_ksp->ks_data = &mirror_stats;
  88                 kstat_install(mirror_ksp);
  89         }
  90 }
  91
  92 void
  93 vdev_mirror_stat_fini(void)
  94 {
  95         if (mirror_ksp != NULL) {
  96                 kstat_delete(mirror_ksp);
  97                 mirror_ksp = NULL;
  98         }
  99 }
 100
 101 /*
 102  * Virtual device vector for mirroring.
 103  */
 104 typedef struct mirror_child {
 105         vdev_t          *mc_vd;
 106         abd_t           *mc_abd;
 107         uint64_t        mc_offset;
 108         int             mc_error;
 109         int             mc_load;
 110         uint8_t         mc_tried;
 111         uint8_t         mc_skipped;
 112         uint8_t         mc_speculative;
 113         uint8_t         mc_rebuilding;
 114 } mirror_child_t;
 115
 116 typedef struct mirror_map {
 117         int             *mm_preferred;
 118         int             mm_preferred_cnt;
 119         int             mm_children;
 120         boolean_t       mm_resilvering;
 121         boolean_t       mm_rebuilding;
 122         boolean_t       mm_root;
 123         mirror_child_t  mm_child[];
 124 } mirror_map_t;
 125
 126 static const int vdev_mirror_shift = 21;
 127
 128 /*
 129  * The load configuration settings below are tuned by default for
 130  * the case where all devices are of the same rotational type.
 131  *
 132  * If there is a mixture of rotating and non-rotating media, setting
 133  * zfs_vdev_mirror_non_rotating_seek_inc to 0 may well provide better results
 134  * as it will direct more reads to the non-rotating vdevs which are more likely
 135  * to have a higher performance.
 136  */
 137
 138 /* Rotating media load calculation configuration. */
 139 static int zfs_vdev_mirror_rotating_inc = 0;
 140 static int zfs_vdev_mirror_rotating_seek_inc = 5;
 141 static int zfs_vdev_mirror_rotating_seek_offset = 1 * 1024 * 1024;
 142
 143 /* Non-rotating media load calculation configuration. */
 144 static int zfs_vdev_mirror_non_rotating_inc = 0;
 145 static int zfs_vdev_mirror_non_rotating_seek_inc = 1;
 146
 147 static inline size_t
 148 vdev_mirror_map_size(int children)
 149 {
 150         return (offsetof(mirror_map_t, mm_child[children]) +
 151             sizeof (int) * children);
 152 }
 153
 154 static inline mirror_map_t *
 155 vdev_mirror_map_alloc(int children, boolean_t resilvering, boolean_t root)
 156 {
 157         mirror_map_t *mm;
 158
 159         mm = kmem_zalloc(vdev_mirror_map_size(children), KM_SLEEP);
 160         mm->mm_children = children;
 161         mm->mm_resilvering = resilvering;
 162         mm->mm_root = root;
 163         mm->mm_preferred = (int *)((uintptr_t)mm +
 164             offsetof(mirror_map_t, mm_child[children]));
 165
 166         return (mm);
 167 }
 168
 169 static void
 170 vdev_mirror_map_free(zio_t *zio)
 171 {
 172         mirror_map_t *mm = zio->io_vsd;
 173
 174         kmem_free(mm, vdev_mirror_map_size(mm->mm_children));
 175 }
 176
 177 static const zio_vsd_ops_t vdev_mirror_vsd_ops = {
 178         .vsd_free = vdev_mirror_map_free,
 179 };
 180
 181 static int
 182 vdev_mirror_load(mirror_map_t *mm, vdev_t *vd, uint64_t zio_offset)
 183 {
 184         uint64_t last_offset;
 185         int64_t offset_diff;
 186         int load;
 187
 188         /* All DVAs have equal weight at the root. */
 189         if (mm->mm_root)
 190                 return (INT_MAX);
 191
 192         /*
 193          * We don't return INT_MAX if the device is resilvering i.e.
 194          * vdev_resilver_txg != 0 as when tested performance was slightly
 195          * worse overall when resilvering with compared to without.
 196          */
 197
 198         /* Fix zio_offset for leaf vdevs */
 199         if (vd->vdev_ops->vdev_op_leaf)
 200                 zio_offset += VDEV_LABEL_START_SIZE;
 201
 202         /* Standard load based on pending queue length. */
 203         load = vdev_queue_length(vd);
 204         last_offset = vdev_queue_last_offset(vd);
 205
 206         if (vd->vdev_nonrot) {
 207                 /* Non-rotating media. */
 208                 if (last_offset == zio_offset) {
 209                         MIRROR_BUMP(vdev_mirror_stat_non_rotating_linear);
 210                         return (load + zfs_vdev_mirror_non_rotating_inc);
 211                 }
 212
 213                 /*
 214                  * Apply a seek penalty even for non-rotating devices as
 215                  * sequential I/O's can be aggregated into fewer operations on
 216                  * the device, thus avoiding unnecessary per-command overhead
 217                  * and boosting performance.
 218                  */
 219                 MIRROR_BUMP(vdev_mirror_stat_non_rotating_seek);
 220                 return (load + zfs_vdev_mirror_non_rotating_seek_inc);
 221         }
 222
 223         /* Rotating media I/O's which directly follow the last I/O. */
 224         if (last_offset == zio_offset) {
 225                 MIRROR_BUMP(vdev_mirror_stat_rotating_linear);
 226                 return (load + zfs_vdev_mirror_rotating_inc);
 227         }
 228
 229         /*
 230          * Apply half the seek increment to I/O's within seek offset
 231          * of the last I/O issued to this vdev as they should incur less
 232          * of a seek increment.
 233          */
 234         offset_diff = (int64_t)(last_offset - zio_offset);
 235         if (ABS(offset_diff) < zfs_vdev_mirror_rotating_seek_offset) {
 236                 MIRROR_BUMP(vdev_mirror_stat_rotating_offset);
 237                 return (load + (zfs_vdev_mirror_rotating_seek_inc / 2));
 238         }
 239
 240         /* Apply the full seek increment to all other I/O's. */
 241         MIRROR_BUMP(vdev_mirror_stat_rotating_seek);
 242         return (load + zfs_vdev_mirror_rotating_seek_inc);
 243 }
 244
 245 static boolean_t
 246 vdev_mirror_rebuilding(vdev_t *vd)
 247 {
 248         if (vd->vdev_ops->vdev_op_leaf && vd->vdev_rebuild_txg)
 249                 return (B_TRUE);
 250
 251         for (int i = 0; i < vd->vdev_children; i++) {
 252                 if (vdev_mirror_rebuilding(vd->vdev_child[i])) {
 253                         return (B_TRUE);
 254                 }
 255         }
 256
 257         return (B_FALSE);
 258 }
 259
 260 /*
 261  * Avoid inlining the function to keep vdev_mirror_io_start(), which
 262  * is this functions only caller, as small as possible on the stack.
 263  */
 264 noinline static mirror_map_t *
 265 vdev_mirror_map_init(zio_t *zio)
 266 {
 267         mirror_map_t *mm = NULL;
 268         mirror_child_t *mc;
 269         vdev_t *vd = zio->io_vd;
 270         int c;
 271
 272         if (vd == NULL) {
 273                 dva_t *dva = zio->io_bp->blk_dva;
 274                 spa_t *spa = zio->io_spa;
 275                 dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
 276                 dva_t dva_copy[SPA_DVAS_PER_BP];
 277
 278                 /*
 279                  * The sequential scrub code sorts and issues all DVAs
 280                  * of a bp separately. Each of these IOs includes all
 281                  * original DVA copies so that repairs can be performed
 282                  * in the event of an error, but we only actually want
 283                  * to check the first DVA since the others will be
 284                  * checked by their respective sorted IOs. Only if we
 285                  * hit an error will we try all DVAs upon retrying.
 286                  *
 287                  * Note: This check is safe even if the user switches
 288                  * from a legacy scrub to a sequential one in the middle
 289                  * of processing, since scn_is_sorted isn't updated until
 290                  * all outstanding IOs from the previous scrub pass
 291                  * complete.
 292                  */
 293                 if ((zio->io_flags & ZIO_FLAG_SCRUB) &&
 294                     !(zio->io_flags & ZIO_FLAG_IO_RETRY) &&
 295                     dsl_scan_scrubbing(spa->spa_dsl_pool) &&
 296                     scn->scn_is_sorted) {
 297                         c = 1;
 298                 } else {
 299                         c = BP_GET_NDVAS(zio->io_bp);
 300                 }
 301
 302                 /*
 303                  * If the pool cannot be written to, then infer that some
 304                  * DVAs might be invalid or point to vdevs that do not exist.
 305                  * We skip them.
 306                  */
 307                 if (!spa_writeable(spa)) {
 308                         ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ);
 309                         int j = 0;
 310                         for (int i = 0; i < c; i++) {
 311                                 if (zfs_dva_valid(spa, &dva[i], zio->io_bp))
 312                                         dva_copy[j++] = dva[i];
 313                         }
 314                         if (j == 0) {
 315                                 zio->io_vsd = NULL;
 316                                 zio->io_error = ENXIO;
 317                                 return (NULL);
 318                         }
 319                         if (j < c) {
 320                                 dva = dva_copy;
 321                                 c = j;
 322                         }
 323                 }
 324
 325                 mm = vdev_mirror_map_alloc(c, B_FALSE, B_TRUE);
 326                 for (c = 0; c < mm->mm_children; c++) {
 327                         mc = &mm->mm_child[c];
 328
 329                         mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c]));
 330                         mc->mc_offset = DVA_GET_OFFSET(&dva[c]);
 331                         if (mc->mc_vd == NULL) {
 332                                 kmem_free(mm, vdev_mirror_map_size(
 333                                     mm->mm_children));
 334                                 zio->io_vsd = NULL;
 335                                 zio->io_error = ENXIO;
 336                                 return (NULL);
 337                         }
 338                 }
 339         } else {
 340                 /*
 341                  * If we are resilvering, then we should handle scrub reads
 342                  * differently; we shouldn't issue them to the resilvering
 343                  * device because it might not have those blocks.
 344                  *
 345                  * We are resilvering iff:
 346                  * 1) We are a replacing vdev (ie our name is "replacing-1" or
 347                  *    "spare-1" or something like that), and
 348                  * 2) The pool is currently being resilvered.
 349                  *
 350                  * We cannot simply check vd->vdev_resilver_txg, because it's
 351                  * not set in this path.
 352                  *
 353                  * Nor can we just check our vdev_ops; there are cases (such as
 354                  * when a user types "zpool replace pool odev spare_dev" and
 355                  * spare_dev is in the spare list, or when a spare device is
 356                  * automatically used to replace a DEGRADED device) when
 357                  * resilvering is complete but both the original vdev and the
 358                  * spare vdev remain in the pool.  That behavior is intentional.
 359                  * It helps implement the policy that a spare should be
 360                  * automatically removed from the pool after the user replaces
 361                  * the device that originally failed.
 362                  *
 363                  * If a spa load is in progress, then spa_dsl_pool may be
 364                  * uninitialized.  But we shouldn't be resilvering during a spa
 365                  * load anyway.
 366                  */
 367                 boolean_t replacing = (vd->vdev_ops == &vdev_replacing_ops ||
 368                     vd->vdev_ops == &vdev_spare_ops) &&
 369                     spa_load_state(vd->vdev_spa) == SPA_LOAD_NONE &&
 370                     dsl_scan_resilvering(vd->vdev_spa->spa_dsl_pool);
 371                 mm = vdev_mirror_map_alloc(vd->vdev_children, replacing,
 372                     B_FALSE);
 373                 for (c = 0; c < mm->mm_children; c++) {
 374                         mc = &mm->mm_child[c];
 375                         mc->mc_vd = vd->vdev_child[c];
 376                         mc->mc_offset = zio->io_offset;
 377
 378                         if (vdev_mirror_rebuilding(mc->mc_vd))
 379                                 mm->mm_rebuilding = mc->mc_rebuilding = B_TRUE;
 380                 }
 381         }
 382
 383         return (mm);
 384 }
 385
 386 static int
 387 vdev_mirror_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize,
 388     uint64_t *logical_ashift, uint64_t *physical_ashift)
 389 {
 390         int numerrors = 0;
 391         int lasterror = 0;
 392
 393         if (vd->vdev_children == 0) {
 394                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 395                 return (SET_ERROR(EINVAL));
 396         }
 397
 398         vdev_open_children(vd);
 399
 400         for (int c = 0; c < vd->vdev_children; c++) {
 401                 vdev_t *cvd = vd->vdev_child[c];
 402
 403                 if (cvd->vdev_open_error) {
 404                         lasterror = cvd->vdev_open_error;
 405                         numerrors++;
 406                         continue;
 407                 }
 408
 409                 *asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1;
 410                 *max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1;
 411                 *logical_ashift = MAX(*logical_ashift, cvd->vdev_ashift);
 412         }
 413         for (int c = 0; c < vd->vdev_children; c++) {
 414                 vdev_t *cvd = vd->vdev_child[c];
 415
 416                 if (cvd->vdev_open_error)
 417                         continue;
 418                 *physical_ashift = vdev_best_ashift(*logical_ashift,
 419                     *physical_ashift, cvd->vdev_physical_ashift);
 420         }
 421
 422         if (numerrors == vd->vdev_children) {
 423                 if (vdev_children_are_offline(vd))
 424                         vd->vdev_stat.vs_aux = VDEV_AUX_CHILDREN_OFFLINE;
 425                 else
 426                         vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS;
 427                 return (lasterror);
 428         }
 429
 430         return (0);
 431 }
 432
 433 static void
 434 vdev_mirror_close(vdev_t *vd)
 435 {
 436         for (int c = 0; c < vd->vdev_children; c++)
 437                 vdev_close(vd->vdev_child[c]);
 438 }
 439
 440 static void
 441 vdev_mirror_child_done(zio_t *zio)
 442 {
 443         mirror_child_t *mc = zio->io_private;
 444
 445         mc->mc_error = zio->io_error;
 446         mc->mc_tried = 1;
 447         mc->mc_skipped = 0;
 448 }
 449
 450 /*
 451  * Check the other, lower-index DVAs to see if they're on the same
 452  * vdev as the child we picked.  If they are, use them since they
 453  * are likely to have been allocated from the primary metaslab in
 454  * use at the time, and hence are more likely to have locality with
 455  * single-copy data.
 456  */
 457 static int
 458 vdev_mirror_dva_select(zio_t *zio, int p)
 459 {
 460         dva_t *dva = zio->io_bp->blk_dva;
 461         mirror_map_t *mm = zio->io_vsd;
 462         int preferred;
 463         int c;
 464
 465         preferred = mm->mm_preferred[p];
 466         for (p--; p >= 0; p--) {
 467                 c = mm->mm_preferred[p];
 468                 if (DVA_GET_VDEV(&dva[c]) == DVA_GET_VDEV(&dva[preferred]))
 469                         preferred = c;
 470         }
 471         return (preferred);
 472 }
 473
 474 static int
 475 vdev_mirror_preferred_child_randomize(zio_t *zio)
 476 {
 477         mirror_map_t *mm = zio->io_vsd;
 478         int p;
 479
 480         if (mm->mm_root) {
 481                 p = random_in_range(mm->mm_preferred_cnt);
 482                 return (vdev_mirror_dva_select(zio, p));
 483         }
 484
 485         /*
 486          * To ensure we don't always favour the first matching vdev,
 487          * which could lead to wear leveling issues on SSD's, we
 488          * use the I/O offset as a pseudo random seed into the vdevs
 489          * which have the lowest load.
 490          */
 491         p = (zio->io_offset >> vdev_mirror_shift) % mm->mm_preferred_cnt;
 492         return (mm->mm_preferred[p]);
 493 }
 494
 495 static boolean_t
 496 vdev_mirror_child_readable(mirror_child_t *mc)
 497 {
 498         vdev_t *vd = mc->mc_vd;
 499
 500         if (vd->vdev_top != NULL && vd->vdev_top->vdev_ops == &vdev_draid_ops)
 501                 return (vdev_draid_readable(vd, mc->mc_offset));
 502         else
 503                 return (vdev_readable(vd));
 504 }
 505
 506 static boolean_t
 507 vdev_mirror_child_missing(mirror_child_t *mc, uint64_t txg, uint64_t size)
 508 {
 509         vdev_t *vd = mc->mc_vd;
 510
 511         if (vd->vdev_top != NULL && vd->vdev_top->vdev_ops == &vdev_draid_ops)
 512                 return (vdev_draid_missing(vd, mc->mc_offset, txg, size));
 513         else
 514                 return (vdev_dtl_contains(vd, DTL_MISSING, txg, size));
 515 }
 516
 517 /*
 518  * Try to find a vdev whose DTL doesn't contain the block we want to read
 519  * preferring vdevs based on determined load. If we can't, try the read on
 520  * any vdev we haven't already tried.
 521  *
 522  * Distributed spares are an exception to the above load rule. They are
 523  * always preferred in order to detect gaps in the distributed spare which
 524  * are created when another disk in the dRAID fails. In order to restore
 525  * redundancy those gaps must be read to trigger the required repair IO.
 526  */
 527 static int
 528 vdev_mirror_child_select(zio_t *zio)
 529 {
 530         mirror_map_t *mm = zio->io_vsd;
 531         uint64_t txg = zio->io_txg;
 532         int c, lowest_load;
 533
 534         ASSERT(zio->io_bp == NULL || BP_PHYSICAL_BIRTH(zio->io_bp) == txg);
 535
 536         lowest_load = INT_MAX;
 537         mm->mm_preferred_cnt = 0;
 538         for (c = 0; c < mm->mm_children; c++) {
 539                 mirror_child_t *mc;
 540
 541                 mc = &mm->mm_child[c];
 542                 if (mc->mc_tried || mc->mc_skipped)
 543                         continue;
 544
 545                 if (mc->mc_vd == NULL ||
 546                     !vdev_mirror_child_readable(mc)) {
 547                         mc->mc_error = SET_ERROR(ENXIO);
 548                         mc->mc_tried = 1;       /* don't even try */
 549                         mc->mc_skipped = 1;
 550                         continue;
 551                 }
 552
 553                 if (vdev_mirror_child_missing(mc, txg, 1)) {
 554                         mc->mc_error = SET_ERROR(ESTALE);
 555                         mc->mc_skipped = 1;
 556                         mc->mc_speculative = 1;
 557                         continue;
 558                 }
 559
 560                 if (mc->mc_vd->vdev_ops == &vdev_draid_spare_ops) {
 561                         mm->mm_preferred[0] = c;
 562                         mm->mm_preferred_cnt = 1;
 563                         break;
 564                 }
 565
 566                 mc->mc_load = vdev_mirror_load(mm, mc->mc_vd, mc->mc_offset);
 567                 if (mc->mc_load > lowest_load)
 568                         continue;
 569
 570                 if (mc->mc_load < lowest_load) {
 571                         lowest_load = mc->mc_load;
 572                         mm->mm_preferred_cnt = 0;
 573                 }
 574                 mm->mm_preferred[mm->mm_preferred_cnt] = c;
 575                 mm->mm_preferred_cnt++;
 576         }
 577
 578         if (mm->mm_preferred_cnt == 1) {
 579                 MIRROR_BUMP(vdev_mirror_stat_preferred_found);
 580                 return (mm->mm_preferred[0]);
 581         }
 582
 583         if (mm->mm_preferred_cnt > 1) {
 584                 MIRROR_BUMP(vdev_mirror_stat_preferred_not_found);
 585                 return (vdev_mirror_preferred_child_randomize(zio));
 586         }
 587
 588         /*
 589          * Every device is either missing or has this txg in its DTL.
 590          * Look for any child we haven't already tried before giving up.
 591          */
 592         for (c = 0; c < mm->mm_children; c++) {
 593                 if (!mm->mm_child[c].mc_tried)
 594                         return (c);
 595         }
 596
 597         /*
 598          * Every child failed.  There's no place left to look.
 599          */
 600         return (-1);
 601 }
 602
 603 static void
 604 vdev_mirror_io_start(zio_t *zio)
 605 {
 606         mirror_map_t *mm;
 607         mirror_child_t *mc;
 608         int c, children;
 609
 610         mm = vdev_mirror_map_init(zio);
 611         zio->io_vsd = mm;
 612         zio->io_vsd_ops = &vdev_mirror_vsd_ops;
 613
 614         if (mm == NULL) {
 615                 ASSERT(!spa_trust_config(zio->io_spa));
 616                 ASSERT(zio->io_type == ZIO_TYPE_READ);
 617                 zio_execute(zio);
 618                 return;
 619         }
 620
 621         if (zio->io_type == ZIO_TYPE_READ) {
 622                 if ((zio->io_flags & ZIO_FLAG_SCRUB) && !mm->mm_resilvering) {
 623                         /*
 624                          * For scrubbing reads we need to issue reads to all
 625                          * children.  One child can reuse parent buffer, but
 626                          * for others we have to allocate separate ones to
 627                          * verify checksums if io_bp is non-NULL, or compare
 628                          * them in vdev_mirror_io_done() otherwise.
 629                          */
 630                         boolean_t first = B_TRUE;
 631                         for (c = 0; c < mm->mm_children; c++) {
 632                                 mc = &mm->mm_child[c];
 633
 634                                 /* Don't issue ZIOs to offline children */
 635                                 if (!vdev_mirror_child_readable(mc)) {
 636                                         mc->mc_error = SET_ERROR(ENXIO);
 637                                         mc->mc_tried = 1;
 638                                         mc->mc_skipped = 1;
 639                                         continue;
 640                                 }
 641
 642                                 mc->mc_abd = first ? zio->io_abd :
 643                                     abd_alloc_sametype(zio->io_abd,
 644                                     zio->io_size);
 645                                 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
 646                                     mc->mc_vd, mc->mc_offset, mc->mc_abd,
 647                                     zio->io_size, zio->io_type,
 648                                     zio->io_priority, 0,
 649                                     vdev_mirror_child_done, mc));
 650                                 first = B_FALSE;
 651                         }
 652                         zio_execute(zio);
 653                         return;
 654                 }
 655                 /*
 656                  * For normal reads just pick one child.
 657                  */
 658                 c = vdev_mirror_child_select(zio);
 659                 children = (c >= 0);
 660         } else {
 661                 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
 662
 663                 /*
 664                  * Writes go to all children.
 665                  */
 666                 c = 0;
 667                 children = mm->mm_children;
 668         }
 669
 670         while (children--) {
 671                 mc = &mm->mm_child[c];
 672                 c++;
 673
 674                 /*
 675                  * When sequentially resilvering only issue write repair
 676                  * IOs to the vdev which is being rebuilt since performance
 677                  * is limited by the slowest child.  This is an issue for
 678                  * faster replacement devices such as distributed spares.
 679                  */
 680                 if ((zio->io_priority == ZIO_PRIORITY_REBUILD) &&
 681                     (zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
 682                     !(zio->io_flags & ZIO_FLAG_SCRUB) &&
 683                     mm->mm_rebuilding && !mc->mc_rebuilding) {
 684                         continue;
 685                 }
 686
 687                 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
 688                     mc->mc_vd, mc->mc_offset, zio->io_abd, zio->io_size,
 689                     zio->io_type, zio->io_priority, 0,
 690                     vdev_mirror_child_done, mc));
 691         }
 692
 693         zio_execute(zio);
 694 }
 695
 696 static int
 697 vdev_mirror_worst_error(mirror_map_t *mm)
 698 {
 699         int error[2] = { 0, 0 };
 700
 701         for (int c = 0; c < mm->mm_children; c++) {
 702                 mirror_child_t *mc = &mm->mm_child[c];
 703                 int s = mc->mc_speculative;
 704                 error[s] = zio_worst_error(error[s], mc->mc_error);
 705         }
 706
 707         return (error[0] ? error[0] : error[1]);
 708 }
 709
 710 static void
 711 vdev_mirror_io_done(zio_t *zio)
 712 {
 713         mirror_map_t *mm = zio->io_vsd;
 714         mirror_child_t *mc;
 715         int c;
 716         int good_copies = 0;
 717         int unexpected_errors = 0;
 718         int last_good_copy = -1;
 719
 720         if (mm == NULL)
 721                 return;
 722
 723         for (c = 0; c < mm->mm_children; c++) {
 724                 mc = &mm->mm_child[c];
 725
 726                 if (mc->mc_error) {
 727                         if (!mc->mc_skipped)
 728                                 unexpected_errors++;
 729                 } else if (mc->mc_tried) {
 730                         last_good_copy = c;
 731                         good_copies++;
 732                 }
 733         }
 734
 735         if (zio->io_type == ZIO_TYPE_WRITE) {
 736                 /*
 737                  * XXX -- for now, treat partial writes as success.
 738                  *
 739                  * Now that we support write reallocation, it would be better
 740                  * to treat partial failure as real failure unless there are
 741                  * no non-degraded top-level vdevs left, and not update DTLs
 742                  * if we intend to reallocate.
 743                  */
 744                 if (good_copies != mm->mm_children) {
 745                         /*
 746                          * Always require at least one good copy.
 747                          *
 748                          * For ditto blocks (io_vd == NULL), require
 749                          * all copies to be good.
 750                          *
 751                          * XXX -- for replacing vdevs, there's no great answer.
 752                          * If the old device is really dead, we may not even
 753                          * be able to access it -- so we only want to
 754                          * require good writes to the new device.  But if
 755                          * the new device turns out to be flaky, we want
 756                          * to be able to detach it -- which requires all
 757                          * writes to the old device to have succeeded.
 758                          */
 759                         if (good_copies == 0 || zio->io_vd == NULL)
 760                                 zio->io_error = vdev_mirror_worst_error(mm);
 761                 }
 762                 return;
 763         }
 764
 765         ASSERT(zio->io_type == ZIO_TYPE_READ);
 766
 767         /*
 768          * If we don't have a good copy yet, keep trying other children.
 769          */
 770         if (good_copies == 0 && (c = vdev_mirror_child_select(zio)) != -1) {
 771                 ASSERT(c >= 0 && c < mm->mm_children);
 772                 mc = &mm->mm_child[c];
 773                 zio_vdev_io_redone(zio);
 774                 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
 775                     mc->mc_vd, mc->mc_offset, zio->io_abd, zio->io_size,
 776                     ZIO_TYPE_READ, zio->io_priority, 0,
 777                     vdev_mirror_child_done, mc));
 778                 return;
 779         }
 780
 781         if (zio->io_flags & ZIO_FLAG_SCRUB && !mm->mm_resilvering) {
 782                 abd_t *best_abd = NULL;
 783                 if (last_good_copy >= 0)
 784                         best_abd = mm->mm_child[last_good_copy].mc_abd;
 785
 786                 /*
 787                  * If we're scrubbing but don't have a BP available (because
 788                  * this vdev is under a raidz or draid vdev) then the best we
 789                  * can do is compare all of the copies read.  If they're not
 790                  * identical then return a checksum error and the most likely
 791                  * correct data.  The raidz code will issue a repair I/O if
 792                  * possible.
 793                  */
 794                 if (zio->io_bp == NULL) {
 795                         ASSERT(zio->io_vd->vdev_ops == &vdev_replacing_ops ||
 796                             zio->io_vd->vdev_ops == &vdev_spare_ops);
 797
 798                         abd_t *pref_abd = NULL;
 799                         for (c = 0; c < last_good_copy; c++) {
 800                                 mc = &mm->mm_child[c];
 801                                 if (mc->mc_error || !mc->mc_tried)
 802                                         continue;
 803
 804                                 if (abd_cmp(mc->mc_abd, best_abd) != 0)
 805                                         zio->io_error = SET_ERROR(ECKSUM);
 806
 807                                 /*
 808                                  * The distributed spare is always prefered
 809                                  * by vdev_mirror_child_select() so it's
 810                                  * considered to be the best candidate.
 811                                  */
 812                                 if (pref_abd == NULL &&
 813                                     mc->mc_vd->vdev_ops ==
 814                                     &vdev_draid_spare_ops)
 815                                         pref_abd = mc->mc_abd;
 816
 817                                 /*
 818                                  * In the absence of a preferred copy, use
 819                                  * the parent pointer to avoid a memory copy.
 820                                  */
 821                                 if (mc->mc_abd == zio->io_abd)
 822                                         best_abd = mc->mc_abd;
 823                         }
 824                         if (pref_abd)
 825                                 best_abd = pref_abd;
 826                 } else {
 827
 828                         /*
 829                          * If we have a BP available, then checksums are
 830                          * already verified and we just need a buffer
 831                          * with valid data, preferring parent one to
 832                          * avoid a memory copy.
 833                          */
 834                         for (c = 0; c < last_good_copy; c++) {
 835                                 mc = &mm->mm_child[c];
 836                                 if (mc->mc_error || !mc->mc_tried)
 837                                         continue;
 838                                 if (mc->mc_abd == zio->io_abd) {
 839                                         best_abd = mc->mc_abd;
 840                                         break;
 841                                 }
 842                         }
 843                 }
 844
 845                 if (best_abd && best_abd != zio->io_abd)
 846                         abd_copy(zio->io_abd, best_abd, zio->io_size);
 847                 for (c = 0; c < mm->mm_children; c++) {
 848                         mc = &mm->mm_child[c];
 849                         if (mc->mc_abd != zio->io_abd)
 850                                 abd_free(mc->mc_abd);
 851                         mc->mc_abd = NULL;
 852                 }
 853         }
 854
 855         if (good_copies == 0) {
 856                 zio->io_error = vdev_mirror_worst_error(mm);
 857                 ASSERT(zio->io_error != 0);
 858         }
 859
 860         if (good_copies && spa_writeable(zio->io_spa) &&
 861             (unexpected_errors ||
 862             (zio->io_flags & ZIO_FLAG_RESILVER) ||
 863             ((zio->io_flags & ZIO_FLAG_SCRUB) && mm->mm_resilvering))) {
 864                 /*
 865                  * Use the good data we have in hand to repair damaged children.
 866                  */
 867                 for (c = 0; c < mm->mm_children; c++) {
 868                         /*
 869                          * Don't rewrite known good children.
 870                          * Not only is it unnecessary, it could
 871                          * actually be harmful: if the system lost
 872                          * power while rewriting the only good copy,
 873                          * there would be no good copies left!
 874                          */
 875                         mc = &mm->mm_child[c];
 876
 877                         if (mc->mc_error == 0) {
 878                                 vdev_ops_t *ops = mc->mc_vd->vdev_ops;
 879
 880                                 if (mc->mc_tried)
 881                                         continue;
 882                                 /*
 883                                  * We didn't try this child.  We need to
 884                                  * repair it if:
 885                                  * 1. it's a scrub (in which case we have
 886                                  * tried everything that was healthy)
 887                                  *  - or -
 888                                  * 2. it's an indirect or distributed spare
 889                                  * vdev (in which case it could point to any
 890                                  * other vdev, which might have a bad DTL)
 891                                  *  - or -
 892                                  * 3. the DTL indicates that this data is
 893                                  * missing from this vdev
 894                                  */
 895                                 if (!(zio->io_flags & ZIO_FLAG_SCRUB) &&
 896                                     ops != &vdev_indirect_ops &&
 897                                     ops != &vdev_draid_spare_ops &&
 898                                     !vdev_dtl_contains(mc->mc_vd, DTL_PARTIAL,
 899                                     zio->io_txg, 1))
 900                                         continue;
 901                                 mc->mc_error = SET_ERROR(ESTALE);
 902                         }
 903
 904                         zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
 905                             mc->mc_vd, mc->mc_offset,
 906                             zio->io_abd, zio->io_size, ZIO_TYPE_WRITE,
 907                             zio->io_priority == ZIO_PRIORITY_REBUILD ?
 908                             ZIO_PRIORITY_REBUILD : ZIO_PRIORITY_ASYNC_WRITE,
 909                             ZIO_FLAG_IO_REPAIR | (unexpected_errors ?
 910                             ZIO_FLAG_SELF_HEAL : 0), NULL, NULL));
 911                 }
 912         }
 913 }
 914
 915 static void
 916 vdev_mirror_state_change(vdev_t *vd, int faulted, int degraded)
 917 {
 918         if (faulted == vd->vdev_children) {
 919                 if (vdev_children_are_offline(vd)) {
 920                         vdev_set_state(vd, B_FALSE, VDEV_STATE_OFFLINE,
 921                             VDEV_AUX_CHILDREN_OFFLINE);
 922                 } else {
 923                         vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
 924                             VDEV_AUX_NO_REPLICAS);
 925                 }
 926         } else if (degraded + faulted != 0) {
 927                 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE);
 928         } else {
 929                 vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE);
 930         }
 931 }
 932
 933 /*
 934  * Return the maximum asize for a rebuild zio in the provided range.
 935  */
 936 static uint64_t
 937 vdev_mirror_rebuild_asize(vdev_t *vd, uint64_t start, uint64_t asize,
 938     uint64_t max_segment)
 939 {
 940         (void) start;
 941
 942         uint64_t psize = MIN(P2ROUNDUP(max_segment, 1 << vd->vdev_ashift),
 943             SPA_MAXBLOCKSIZE);
 944
 945         return (MIN(asize, vdev_psize_to_asize(vd, psize)));
 946 }
 947
 948 vdev_ops_t vdev_mirror_ops = {
 949         .vdev_op_init = NULL,
 950         .vdev_op_fini = NULL,
 951         .vdev_op_open = vdev_mirror_open,
 952         .vdev_op_close = vdev_mirror_close,
 953         .vdev_op_asize = vdev_default_asize,
 954         .vdev_op_min_asize = vdev_default_min_asize,
 955         .vdev_op_min_alloc = NULL,
 956         .vdev_op_io_start = vdev_mirror_io_start,
 957         .vdev_op_io_done = vdev_mirror_io_done,
 958         .vdev_op_state_change = vdev_mirror_state_change,
 959         .vdev_op_need_resilver = vdev_default_need_resilver,
 960         .vdev_op_hold = NULL,
 961         .vdev_op_rele = NULL,
 962         .vdev_op_remap = NULL,
 963         .vdev_op_xlate = vdev_default_xlate,
 964         .vdev_op_rebuild_asize = vdev_mirror_rebuild_asize,
 965         .vdev_op_metaslab_init = NULL,
 966         .vdev_op_config_generate = NULL,
 967         .vdev_op_nparity = NULL,
 968         .vdev_op_ndisks = NULL,
 969         .vdev_op_type = VDEV_TYPE_MIRROR,       /* name of this vdev type */
 970         .vdev_op_leaf = B_FALSE                 /* not a leaf vdev */
 971 };
 972
 973 vdev_ops_t vdev_replacing_ops = {
 974         .vdev_op_init = NULL,
 975         .vdev_op_fini = NULL,
 976         .vdev_op_open = vdev_mirror_open,
 977         .vdev_op_close = vdev_mirror_close,
 978         .vdev_op_asize = vdev_default_asize,
 979         .vdev_op_min_asize = vdev_default_min_asize,
 980         .vdev_op_min_alloc = NULL,
 981         .vdev_op_io_start = vdev_mirror_io_start,
 982         .vdev_op_io_done = vdev_mirror_io_done,
 983         .vdev_op_state_change = vdev_mirror_state_change,
 984         .vdev_op_need_resilver = vdev_default_need_resilver,
 985         .vdev_op_hold = NULL,
 986         .vdev_op_rele = NULL,
 987         .vdev_op_remap = NULL,
 988         .vdev_op_xlate = vdev_default_xlate,
 989         .vdev_op_rebuild_asize = vdev_mirror_rebuild_asize,
 990         .vdev_op_metaslab_init = NULL,
 991         .vdev_op_config_generate = NULL,
 992         .vdev_op_nparity = NULL,
 993         .vdev_op_ndisks = NULL,
 994         .vdev_op_type = VDEV_TYPE_REPLACING,    /* name of this vdev type */
 995         .vdev_op_leaf = B_FALSE                 /* not a leaf vdev */
 996 };
 997
 998 vdev_ops_t vdev_spare_ops = {
 999         .vdev_op_init = NULL,
1000         .vdev_op_fini = NULL,
1001         .vdev_op_open = vdev_mirror_open,
1002         .vdev_op_close = vdev_mirror_close,
1003         .vdev_op_asize = vdev_default_asize,
1004         .vdev_op_min_asize = vdev_default_min_asize,
1005         .vdev_op_min_alloc = NULL,
1006         .vdev_op_io_start = vdev_mirror_io_start,
1007         .vdev_op_io_done = vdev_mirror_io_done,
1008         .vdev_op_state_change = vdev_mirror_state_change,
1009         .vdev_op_need_resilver = vdev_default_need_resilver,
1010         .vdev_op_hold = NULL,
1011         .vdev_op_rele = NULL,
1012         .vdev_op_remap = NULL,
1013         .vdev_op_xlate = vdev_default_xlate,
1014         .vdev_op_rebuild_asize = vdev_mirror_rebuild_asize,
1015         .vdev_op_metaslab_init = NULL,
1016         .vdev_op_config_generate = NULL,
1017         .vdev_op_nparity = NULL,
1018         .vdev_op_ndisks = NULL,
1019         .vdev_op_type = VDEV_TYPE_SPARE,        /* name of this vdev type */
1020         .vdev_op_leaf = B_FALSE                 /* not a leaf vdev */
1021 };
1022
1023 ZFS_MODULE_PARAM(zfs_vdev_mirror, zfs_vdev_mirror_, rotating_inc, INT, ZMOD_RW,
1024         "Rotating media load increment for non-seeking I/Os");
1025
1026 ZFS_MODULE_PARAM(zfs_vdev_mirror, zfs_vdev_mirror_, rotating_seek_inc, INT,
1027         ZMOD_RW, "Rotating media load increment for seeking I/Os");
1028
1029 /* BEGIN CSTYLED */
1030 ZFS_MODULE_PARAM(zfs_vdev_mirror, zfs_vdev_mirror_, rotating_seek_offset, INT,
1031         ZMOD_RW,
1032         "Offset in bytes from the last I/O which triggers "
1033         "a reduced rotating media seek increment");
1034 /* END CSTYLED */
1035
1036 ZFS_MODULE_PARAM(zfs_vdev_mirror, zfs_vdev_mirror_, non_rotating_inc, INT,
1037         ZMOD_RW, "Non-rotating media load increment for non-seeking I/Os");
1038
1039 ZFS_MODULE_PARAM(zfs_vdev_mirror, zfs_vdev_mirror_, non_rotating_seek_inc, INT,
1040         ZMOD_RW, "Non-rotating media load increment for seeking I/Os");