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 http://www.opensolaris.org/os/licensing.
  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) 2013 by Delphix. All rights reserved.
  28  */
  29 
  30 #include <sys/zfs_context.h>
  31 #include <sys/spa.h>
  32 #include <sys/vdev_impl.h>
  33 #include <sys/zio.h>
  34 #include <sys/fs/zfs.h>
  35 
  36 /*
  37  * Virtual device vector for mirroring.
  38  */
  39 
  40 typedef struct mirror_child {
  41         vdev_t          *mc_vd;
  42         uint64_t        mc_offset;
  43         int             mc_error;
  44         uint8_t         mc_tried;
  45         uint8_t         mc_skipped;
  46         uint8_t         mc_speculative;
  47 } mirror_child_t;
  48 
  49 typedef struct mirror_map {
  50         int             mm_children;
  51         int             mm_replacing;
  52         int             mm_preferred;
  53         int             mm_root;
  54         mirror_child_t  mm_child[1];
  55 } mirror_map_t;
  56 
  57 int vdev_mirror_shift = 21;
  58 
  59 static void
  60 vdev_mirror_map_free(zio_t *zio)
  61 {
  62         mirror_map_t *mm = zio->io_vsd;
  63 
  64         kmem_free(mm, offsetof(mirror_map_t, mm_child[mm->mm_children]));
  65 }
  66 
  67 static const zio_vsd_ops_t vdev_mirror_vsd_ops = {
  68         vdev_mirror_map_free,
  69         zio_vsd_default_cksum_report
  70 };
  71 
  72 static mirror_map_t *
  73 vdev_mirror_map_alloc(zio_t *zio)
  74 {
  75         mirror_map_t *mm = NULL;
  76         mirror_child_t *mc;
  77         vdev_t *vd = zio->io_vd;
  78         int c, d;
  79 
  80         if (vd == NULL) {
  81                 dva_t *dva = zio->io_bp->blk_dva;
  82                 spa_t *spa = zio->io_spa;
  83 
  84                 c = BP_GET_NDVAS(zio->io_bp);
  85 
  86                 mm = kmem_zalloc(offsetof(mirror_map_t, mm_child[c]), KM_SLEEP);
  87                 mm->mm_children = c;
  88                 mm->mm_replacing = B_FALSE;
  89                 mm->mm_preferred = spa_get_random(c);
  90                 mm->mm_root = B_TRUE;
  91 
  92                 /*
  93                  * Check the other, lower-index DVAs to see if they're on
  94                  * the same vdev as the child we picked.  If they are, use
  95                  * them since they are likely to have been allocated from
  96                  * the primary metaslab in use at the time, and hence are
  97                  * more likely to have locality with single-copy data.
  98                  */
  99                 for (c = mm->mm_preferred, d = c - 1; d >= 0; d--) {
 100                         if (DVA_GET_VDEV(&dva[d]) == DVA_GET_VDEV(&dva[c]))
 101                                 mm->mm_preferred = d;
 102                 }
 103 
 104                 for (c = 0; c < mm->mm_children; c++) {
 105                         mc = &mm->mm_child[c];
 106 
 107                         mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c]));
 108                         mc->mc_offset = DVA_GET_OFFSET(&dva[c]);
 109                 }
 110         } else {
 111                 c = vd->vdev_children;
 112 
 113                 mm = kmem_zalloc(offsetof(mirror_map_t, mm_child[c]), KM_SLEEP);
 114                 mm->mm_children = c;
 115                 mm->mm_replacing = (vd->vdev_ops == &vdev_replacing_ops ||
 116                     vd->vdev_ops == &vdev_spare_ops);
 117                 mm->mm_preferred = mm->mm_replacing ? 0 :
 118                     (zio->io_offset >> vdev_mirror_shift) % c;
 119                 mm->mm_root = B_FALSE;
 120 
 121                 for (c = 0; c < mm->mm_children; c++) {
 122                         mc = &mm->mm_child[c];
 123                         mc->mc_vd = vd->vdev_child[c];
 124                         mc->mc_offset = zio->io_offset;
 125                 }
 126         }
 127 
 128         zio->io_vsd = mm;
 129         zio->io_vsd_ops = &vdev_mirror_vsd_ops;
 130         return (mm);
 131 }
 132 
 133 static int
 134 vdev_mirror_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize,
 135     uint64_t *ashift)
 136 {
 137         int numerrors = 0;
 138         int lasterror = 0;
 139 
 140         if (vd->vdev_children == 0) {
 141                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 142                 return (SET_ERROR(EINVAL));
 143         }
 144 
 145         vdev_open_children(vd);
 146 
 147         for (int c = 0; c < vd->vdev_children; c++) {
 148                 vdev_t *cvd = vd->vdev_child[c];
 149 
 150                 if (cvd->vdev_open_error) {
 151                         lasterror = cvd->vdev_open_error;
 152                         numerrors++;
 153                         continue;
 154                 }
 155 
 156                 *asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1;
 157                 *max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1;
 158                 *ashift = MAX(*ashift, cvd->vdev_ashift);
 159         }
 160 
 161         if (numerrors == vd->vdev_children) {
 162                 vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS;
 163                 return (lasterror);
 164         }
 165 
 166         return (0);
 167 }
 168 
 169 static void
 170 vdev_mirror_close(vdev_t *vd)
 171 {
 172         for (int c = 0; c < vd->vdev_children; c++)
 173                 vdev_close(vd->vdev_child[c]);
 174 }
 175 
 176 static void
 177 vdev_mirror_child_done(zio_t *zio)
 178 {
 179         mirror_child_t *mc = zio->io_private;
 180 
 181         mc->mc_error = zio->io_error;
 182         mc->mc_tried = 1;
 183         mc->mc_skipped = 0;
 184 }
 185 
 186 static void
 187 vdev_mirror_scrub_done(zio_t *zio)
 188 {
 189         mirror_child_t *mc = zio->io_private;
 190 
 191         if (zio->io_error == 0) {
 192                 zio_t *pio;
 193 
 194                 mutex_enter(&zio->io_lock);
 195                 while ((pio = zio_walk_parents(zio)) != NULL) {
 196                         mutex_enter(&pio->io_lock);
 197                         ASSERT3U(zio->io_size, >=, pio->io_size);
 198                         bcopy(zio->io_data, pio->io_data, pio->io_size);
 199                         mutex_exit(&pio->io_lock);
 200                 }
 201                 mutex_exit(&zio->io_lock);
 202         }
 203 
 204         zio_buf_free(zio->io_data, zio->io_size);
 205 
 206         mc->mc_error = zio->io_error;
 207         mc->mc_tried = 1;
 208         mc->mc_skipped = 0;
 209 }
 210 
 211 /*
 212  * Try to find a child whose DTL doesn't contain the block we want to read.
 213  * If we can't, try the read on any vdev we haven't already tried.
 214  */
 215 static int
 216 vdev_mirror_child_select(zio_t *zio)
 217 {
 218         mirror_map_t *mm = zio->io_vsd;
 219         mirror_child_t *mc;
 220         uint64_t txg = zio->io_txg;
 221         int i, c;
 222 
 223         ASSERT(zio->io_bp == NULL || BP_PHYSICAL_BIRTH(zio->io_bp) == txg);
 224 
 225         /*
 226          * Try to find a child whose DTL doesn't contain the block to read.
 227          * If a child is known to be completely inaccessible (indicated by
 228          * vdev_readable() returning B_FALSE), don't even try.
 229          */
 230         for (i = 0, c = mm->mm_preferred; i < mm->mm_children; i++, c++) {
 231                 if (c >= mm->mm_children)
 232                         c = 0;
 233                 mc = &mm->mm_child[c];
 234                 if (mc->mc_tried || mc->mc_skipped)
 235                         continue;
 236                 if (!vdev_readable(mc->mc_vd)) {
 237                         mc->mc_error = SET_ERROR(ENXIO);
 238                         mc->mc_tried = 1;    /* don't even try */
 239                         mc->mc_skipped = 1;
 240                         continue;
 241                 }
 242                 if (!vdev_dtl_contains(mc->mc_vd, DTL_MISSING, txg, 1))
 243                         return (c);
 244                 mc->mc_error = SET_ERROR(ESTALE);
 245                 mc->mc_skipped = 1;
 246                 mc->mc_speculative = 1;
 247         }
 248 
 249         /*
 250          * Every device is either missing or has this txg in its DTL.
 251          * Look for any child we haven't already tried before giving up.
 252          */
 253         for (c = 0; c < mm->mm_children; c++)
 254                 if (!mm->mm_child[c].mc_tried)
 255                         return (c);
 256 
 257         /*
 258          * Every child failed.  There's no place left to look.
 259          */
 260         return (-1);
 261 }
 262 
 263 static int
 264 vdev_mirror_io_start(zio_t *zio)
 265 {
 266         mirror_map_t *mm;
 267         mirror_child_t *mc;
 268         int c, children;
 269 
 270         mm = vdev_mirror_map_alloc(zio);
 271 
 272         if (zio->io_type == ZIO_TYPE_READ) {
 273                 if ((zio->io_flags & ZIO_FLAG_SCRUB) && !mm->mm_replacing) {
 274                         /*
 275                          * For scrubbing reads we need to allocate a read
 276                          * buffer for each child and issue reads to all
 277                          * children.  If any child succeeds, it will copy its
 278                          * data into zio->io_data in vdev_mirror_scrub_done.
 279                          */
 280                         for (c = 0; c < mm->mm_children; c++) {
 281                                 mc = &mm->mm_child[c];
 282                                 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
 283                                     mc->mc_vd, mc->mc_offset,
 284                                     zio_buf_alloc(zio->io_size), zio->io_size,
 285                                     zio->io_type, zio->io_priority, 0,
 286                                     vdev_mirror_scrub_done, mc));
 287                         }
 288                         return (ZIO_PIPELINE_CONTINUE);
 289                 }
 290                 /*
 291                  * For normal reads just pick one child.
 292                  */
 293                 c = vdev_mirror_child_select(zio);
 294                 children = (c >= 0);
 295         } else {
 296                 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
 297 
 298                 /*
 299                  * Writes go to all children.
 300                  */
 301                 c = 0;
 302                 children = mm->mm_children;
 303         }
 304 
 305         while (children--) {
 306                 mc = &mm->mm_child[c];
 307                 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
 308                     mc->mc_vd, mc->mc_offset, zio->io_data, zio->io_size,
 309                     zio->io_type, zio->io_priority, 0,
 310                     vdev_mirror_child_done, mc));
 311                 c++;
 312         }
 313 
 314         return (ZIO_PIPELINE_CONTINUE);
 315 }
 316 
 317 static int
 318 vdev_mirror_worst_error(mirror_map_t *mm)
 319 {
 320         int error[2] = { 0, 0 };
 321 
 322         for (int c = 0; c < mm->mm_children; c++) {
 323                 mirror_child_t *mc = &mm->mm_child[c];
 324                 int s = mc->mc_speculative;
 325                 error[s] = zio_worst_error(error[s], mc->mc_error);
 326         }
 327 
 328         return (error[0] ? error[0] : error[1]);
 329 }
 330 
 331 static void
 332 vdev_mirror_io_done(zio_t *zio)
 333 {
 334         mirror_map_t *mm = zio->io_vsd;
 335         mirror_child_t *mc;
 336         int c;
 337         int good_copies = 0;
 338         int unexpected_errors = 0;
 339 
 340         for (c = 0; c < mm->mm_children; c++) {
 341                 mc = &mm->mm_child[c];
 342 
 343                 if (mc->mc_error) {
 344                         if (!mc->mc_skipped)
 345                                 unexpected_errors++;
 346                 } else if (mc->mc_tried) {
 347                         good_copies++;
 348                 }
 349         }
 350 
 351         if (zio->io_type == ZIO_TYPE_WRITE) {
 352                 /*
 353                  * XXX -- for now, treat partial writes as success.
 354                  *
 355                  * Now that we support write reallocation, it would be better
 356                  * to treat partial failure as real failure unless there are
 357                  * no non-degraded top-level vdevs left, and not update DTLs
 358                  * if we intend to reallocate.
 359                  */
 360                 /* XXPOLICY */
 361                 if (good_copies != mm->mm_children) {
 362                         /*
 363                          * Always require at least one good copy.
 364                          *
 365                          * For ditto blocks (io_vd == NULL), require
 366                          * all copies to be good.
 367                          *
 368                          * XXX -- for replacing vdevs, there's no great answer.
 369                          * If the old device is really dead, we may not even
 370                          * be able to access it -- so we only want to
 371                          * require good writes to the new device.  But if
 372                          * the new device turns out to be flaky, we want
 373                          * to be able to detach it -- which requires all
 374                          * writes to the old device to have succeeded.
 375                          */
 376                         if (good_copies == 0 || zio->io_vd == NULL)
 377                                 zio->io_error = vdev_mirror_worst_error(mm);
 378                 }
 379                 return;
 380         }
 381 
 382         ASSERT(zio->io_type == ZIO_TYPE_READ);
 383 
 384         /*
 385          * If we don't have a good copy yet, keep trying other children.
 386          */
 387         /* XXPOLICY */
 388         if (good_copies == 0 && (c = vdev_mirror_child_select(zio)) != -1) {
 389                 ASSERT(c >= 0 && c < mm->mm_children);
 390                 mc = &mm->mm_child[c];
 391                 zio_vdev_io_redone(zio);
 392                 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
 393                     mc->mc_vd, mc->mc_offset, zio->io_data, zio->io_size,
 394                     ZIO_TYPE_READ, zio->io_priority, 0,
 395                     vdev_mirror_child_done, mc));
 396                 return;
 397         }
 398 
 399         /* XXPOLICY */
 400         if (good_copies == 0) {
 401                 zio->io_error = vdev_mirror_worst_error(mm);
 402                 ASSERT(zio->io_error != 0);
 403         }
 404 
 405         if (good_copies && spa_writeable(zio->io_spa) &&
 406             (unexpected_errors ||
 407             (zio->io_flags & ZIO_FLAG_RESILVER) ||
 408             ((zio->io_flags & ZIO_FLAG_SCRUB) && mm->mm_replacing))) {
 409                 /*
 410                  * Use the good data we have in hand to repair damaged children.
 411                  */
 412                 for (c = 0; c < mm->mm_children; c++) {
 413                         /*
 414                          * Don't rewrite known good children.
 415                          * Not only is it unnecessary, it could
 416                          * actually be harmful: if the system lost
 417                          * power while rewriting the only good copy,
 418                          * there would be no good copies left!
 419                          */
 420                         mc = &mm->mm_child[c];
 421 
 422                         if (mc->mc_error == 0) {
 423                                 if (mc->mc_tried)
 424                                         continue;
 425                                 if (!(zio->io_flags & ZIO_FLAG_SCRUB) &&
 426                                     !vdev_dtl_contains(mc->mc_vd, DTL_PARTIAL,
 427                                     zio->io_txg, 1))
 428                                         continue;
 429                                 mc->mc_error = SET_ERROR(ESTALE);
 430                         }
 431 
 432                         zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
 433                             mc->mc_vd, mc->mc_offset,
 434                             zio->io_data, zio->io_size,
 435                             ZIO_TYPE_WRITE, ZIO_PRIORITY_ASYNC_WRITE,
 436                             ZIO_FLAG_IO_REPAIR | (unexpected_errors ?
 437                             ZIO_FLAG_SELF_HEAL : 0), NULL, NULL));
 438                 }
 439         }
 440 }
 441 
 442 static void
 443 vdev_mirror_state_change(vdev_t *vd, int faulted, int degraded)
 444 {
 445         if (faulted == vd->vdev_children)
 446                 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
 447                     VDEV_AUX_NO_REPLICAS);
 448         else if (degraded + faulted != 0)
 449                 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE);
 450         else
 451                 vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE);
 452 }
 453 
 454 vdev_ops_t vdev_mirror_ops = {
 455         vdev_mirror_open,
 456         vdev_mirror_close,
 457         vdev_default_asize,
 458         vdev_mirror_io_start,
 459         vdev_mirror_io_done,
 460         vdev_mirror_state_change,
 461         NULL,
 462         NULL,
 463         VDEV_TYPE_MIRROR,       /* name of this vdev type */
 464         B_FALSE                 /* not a leaf vdev */
 465 };
 466 
 467 vdev_ops_t vdev_replacing_ops = {
 468         vdev_mirror_open,
 469         vdev_mirror_close,
 470         vdev_default_asize,
 471         vdev_mirror_io_start,
 472         vdev_mirror_io_done,
 473         vdev_mirror_state_change,
 474         NULL,
 475         NULL,
 476         VDEV_TYPE_REPLACING,    /* name of this vdev type */
 477         B_FALSE                 /* not a leaf vdev */
 478 };
 479 
 480 vdev_ops_t vdev_spare_ops = {
 481         vdev_mirror_open,
 482         vdev_mirror_close,
 483         vdev_default_asize,
 484         vdev_mirror_io_start,
 485         vdev_mirror_io_done,
 486         vdev_mirror_state_change,
 487         NULL,
 488         NULL,
 489         VDEV_TYPE_SPARE,        /* name of this vdev type */
 490         B_FALSE                 /* not a leaf vdev */
 491 };