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  * Copyright (c) 2013 Steven Hartland. All rights reserved.
  29  */
  30 
  31 #include <sys/zfs_context.h>
  32 #include <sys/spa.h>
  33 #include <sys/vdev_impl.h>
  34 #include <sys/zio.h>
  35 #include <sys/fs/zfs.h>
  36 
  37 /*
  38  * Virtual device vector for mirroring.
  39  */
  40 
  41 typedef struct mirror_child {
  42         vdev_t          *mc_vd;
  43         uint64_t        mc_offset;
  44         int             mc_error;
  45         int             mc_load;
  46         uint8_t         mc_tried;
  47         uint8_t         mc_skipped;
  48         uint8_t         mc_speculative;
  49 } mirror_child_t;
  50 
  51 typedef struct mirror_map {
  52         int             *mm_preferred;
  53         int             mm_preferred_cnt;
  54         int             mm_children;
  55         boolean_t       mm_replacing;
  56         boolean_t       mm_root;
  57         mirror_child_t  mm_child[];
  58 } mirror_map_t;
  59 
  60 static int vdev_mirror_shift = 21;
  61 
  62 /*
  63  * The load configuration settings below are tuned by default for
  64  * the case where all devices are of the same rotational type.
  65  *
  66  * If there is a mixture of rotating and non-rotating media, setting
  67  * non_rotating_seek_inc to 0 may well provide better results as it
  68  * will direct more reads to the non-rotating vdevs which are more
  69  * likely to have a higher performance.
  70  */
  71 
  72 /* Rotating media load calculation configuration. */
  73 /* Rotating media load increment for non-seeking I/O's. */
  74 static int rotating_inc = 0;
  75 
  76 /* Rotating media load increment for seeking I/O's. */
  77 static int rotating_seek_inc = 5;
  78 
  79 /*
  80  * Offset in bytes from the last I/O which triggers a reduced rotating media
  81  * seek increment.
  82  */
  83 static int rotating_seek_offset = 1 * 1024 * 1024;
  84 
  85 /* Non-rotating media load calculation configuration. */
  86 /* Non-rotating media load increment for non-seeking I/O's. */
  87 static int non_rotating_inc = 0;
  88 
  89 /* Non-rotating media load increment for seeking I/O's. */
  90 static int non_rotating_seek_inc = 1;
  91 
  92 static inline size_t
  93 vdev_mirror_map_size(int children)
  94 {
  95         return (offsetof(mirror_map_t, mm_child[children]) +
  96             sizeof (int) * children);
  97 }
  98 
  99 static inline mirror_map_t *
 100 vdev_mirror_map_alloc(int children, boolean_t replacing, boolean_t root)
 101 {
 102         mirror_map_t *mm;
 103 
 104         mm = kmem_zalloc(vdev_mirror_map_size(children), KM_SLEEP);
 105         mm->mm_children = children;
 106         mm->mm_replacing = replacing;
 107         mm->mm_root = root;
 108         mm->mm_preferred = (int *)((uintptr_t)mm +
 109             offsetof(mirror_map_t, mm_child[children]));
 110 
 111         return (mm);
 112 }
 113 
 114 static void
 115 vdev_mirror_map_free(zio_t *zio)
 116 {
 117         mirror_map_t *mm = zio->io_vsd;
 118 
 119         kmem_free(mm, vdev_mirror_map_size(mm->mm_children));
 120 }
 121 
 122 static const zio_vsd_ops_t vdev_mirror_vsd_ops = {
 123         vdev_mirror_map_free,
 124         zio_vsd_default_cksum_report
 125 };
 126 
 127 static int
 128 vdev_mirror_load(mirror_map_t *mm, vdev_t *vd, uint64_t zio_offset)
 129 {
 130         uint64_t lastoffset;
 131         int load;
 132 
 133         /* All DVAs have equal weight at the root. */
 134         if (mm->mm_root)
 135                 return (INT_MAX);
 136 
 137         /*
 138          * We don't return INT_MAX if the device is resilvering i.e.
 139          * vdev_resilver_txg != 0 as when tested performance was slightly
 140          * worse overall when resilvering with compared to without.
 141          */
 142 
 143         /* Standard load based on pending queue length. */
 144         load = vdev_queue_length(vd);
 145         lastoffset = vdev_queue_lastoffset(vd);
 146 
 147         if (vd->vdev_rotation_rate == VDEV_RATE_NON_ROTATING) {
 148                 /* Non-rotating media. */
 149                 if (lastoffset == zio_offset)
 150                         return (load + non_rotating_inc);
 151 
 152                 /*
 153                  * Apply a seek penalty even for non-rotating devices as
 154                  * sequential I/O'a can be aggregated into fewer operations
 155                  * on the device, thus avoiding unnecessary per-command
 156                  * overhead and boosting performance.
 157                  */
 158                 return (load + non_rotating_seek_inc);
 159         }
 160 
 161         /* Rotating media I/O's which directly follow the last I/O. */
 162         if (lastoffset == zio_offset)
 163                 return (load + rotating_inc);
 164 
 165         /*
 166          * Apply half the seek increment to I/O's within seek offset
 167          * of the last I/O queued to this vdev as they should incure less
 168          * of a seek increment.
 169          */
 170         if (ABS(lastoffset - zio_offset) < rotating_seek_offset)
 171                 return (load + (rotating_seek_inc / 2));
 172 
 173         /* Apply the full seek increment to all other I/O's. */
 174         return (load + rotating_seek_inc);
 175 }
 176 
 177 
 178 static mirror_map_t *
 179 vdev_mirror_map_init(zio_t *zio)
 180 {
 181         mirror_map_t *mm = NULL;
 182         mirror_child_t *mc;
 183         vdev_t *vd = zio->io_vd;
 184         int c;
 185 
 186         if (vd == NULL) {
 187                 dva_t *dva = zio->io_bp->blk_dva;
 188                 spa_t *spa = zio->io_spa;
 189 
 190                 mm = vdev_mirror_map_alloc(BP_GET_NDVAS(zio->io_bp), B_FALSE,
 191                     B_TRUE);
 192                 for (c = 0; c < mm->mm_children; c++) {
 193                         mc = &mm->mm_child[c];
 194                         mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c]));
 195                         mc->mc_offset = DVA_GET_OFFSET(&dva[c]);
 196                 }
 197         } else {
 198                 mm = vdev_mirror_map_alloc(vd->vdev_children,
 199                     (vd->vdev_ops == &vdev_replacing_ops ||
 200                     vd->vdev_ops == &vdev_spare_ops), B_FALSE);
 201                 for (c = 0; c < mm->mm_children; c++) {
 202                         mc = &mm->mm_child[c];
 203                         mc->mc_vd = vd->vdev_child[c];
 204                         mc->mc_offset = zio->io_offset;
 205                 }
 206         }
 207 
 208         zio->io_vsd = mm;
 209         zio->io_vsd_ops = &vdev_mirror_vsd_ops;
 210         return (mm);
 211 }
 212 
 213 static int
 214 vdev_mirror_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize,
 215     uint64_t *ashift)
 216 {
 217         int numerrors = 0;
 218         int lasterror = 0;
 219 
 220         if (vd->vdev_children == 0) {
 221                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 222                 return (SET_ERROR(EINVAL));
 223         }
 224 
 225         vdev_open_children(vd);
 226 
 227         for (int c = 0; c < vd->vdev_children; c++) {
 228                 vdev_t *cvd = vd->vdev_child[c];
 229 
 230                 if (cvd->vdev_open_error) {
 231                         lasterror = cvd->vdev_open_error;
 232                         numerrors++;
 233                         continue;
 234                 }
 235 
 236                 *asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1;
 237                 *max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1;
 238                 *ashift = MAX(*ashift, cvd->vdev_ashift);
 239         }
 240 
 241         if (numerrors == vd->vdev_children) {
 242                 vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS;
 243                 return (lasterror);
 244         }
 245 
 246         return (0);
 247 }
 248 
 249 static void
 250 vdev_mirror_close(vdev_t *vd)
 251 {
 252         for (int c = 0; c < vd->vdev_children; c++)
 253                 vdev_close(vd->vdev_child[c]);
 254 }
 255 
 256 static void
 257 vdev_mirror_child_done(zio_t *zio)
 258 {
 259         mirror_child_t *mc = zio->io_private;
 260 
 261         mc->mc_error = zio->io_error;
 262         mc->mc_tried = 1;
 263         mc->mc_skipped = 0;
 264 }
 265 
 266 static void
 267 vdev_mirror_scrub_done(zio_t *zio)
 268 {
 269         mirror_child_t *mc = zio->io_private;
 270 
 271         if (zio->io_error == 0) {
 272                 zio_t *pio;
 273 
 274                 mutex_enter(&zio->io_lock);
 275                 while ((pio = zio_walk_parents(zio)) != NULL) {
 276                         mutex_enter(&pio->io_lock);
 277                         ASSERT3U(zio->io_size, >=, pio->io_size);
 278                         bcopy(zio->io_data, pio->io_data, pio->io_size);
 279                         mutex_exit(&pio->io_lock);
 280                 }
 281                 mutex_exit(&zio->io_lock);
 282         }
 283 
 284         zio_buf_free(zio->io_data, zio->io_size);
 285 
 286         mc->mc_error = zio->io_error;
 287         mc->mc_tried = 1;
 288         mc->mc_skipped = 0;
 289 }
 290 
 291 /*
 292  * Check the other, lower-index DVAs to see if they're on the same
 293  * vdev as the child we picked.  If they are, use them since they
 294  * are likely to have been allocated from the primary metaslab in
 295  * use at the time, and hence are more likely to have locality with
 296  * single-copy data.
 297  */
 298 static int
 299 vdev_mirror_dva_select(zio_t *zio, int preferred)
 300 {
 301         dva_t *dva = zio->io_bp->blk_dva;
 302         mirror_map_t *mm = zio->io_vsd;
 303         int c;
 304 
 305         for (c = preferred - 1; c >= 0; c--) {
 306                 if (DVA_GET_VDEV(&dva[c]) == DVA_GET_VDEV(&dva[preferred]))
 307                         preferred = c;
 308         }
 309         return (preferred);
 310 }
 311 
 312 static int
 313 vdev_mirror_preferred_child_randomize(zio_t *zio)
 314 {
 315         mirror_map_t *mm = zio->io_vsd;
 316         int p;
 317 
 318         if (mm->mm_root) {
 319                 p = spa_get_random(mm->mm_preferred_cnt);
 320                 return (vdev_mirror_dva_select(zio, mm->mm_preferred[p]));
 321         }
 322 
 323         /*
 324          * To ensure we don't always favour the first matching vdev,
 325          * which could lead to wear leveling issues on SSD's, we
 326          * use the I/O offset as a pseudo random seed into the vdevs
 327          * which have the lowest load.
 328          */
 329         p = (zio->io_offset >> vdev_mirror_shift) % mm->mm_preferred_cnt;
 330         return (mm->mm_preferred[p]);
 331 }
 332 
 333 /*
 334  * Try to find a vdev whose DTL doesn't contain the block we want to read
 335  * prefering vdevs based on determined load.
 336  *
 337  * If we can't, try the read on any vdev we haven't already tried.
 338  */
 339 static int
 340 vdev_mirror_child_select(zio_t *zio)
 341 {
 342         mirror_map_t *mm = zio->io_vsd;
 343         uint64_t txg = zio->io_txg;
 344         int c, lowest_load;
 345 
 346         ASSERT(zio->io_bp == NULL || BP_PHYSICAL_BIRTH(zio->io_bp) == txg);
 347 
 348         lowest_load = INT_MAX;
 349         mm->mm_preferred_cnt = 0;
 350         for (c = 0; c < mm->mm_children; c++) {
 351                 mirror_child_t *mc;
 352 
 353                 mc = &mm->mm_child[c];
 354                 if (mc->mc_tried || mc->mc_skipped)
 355                         continue;
 356 
 357                 if (!vdev_readable(mc->mc_vd)) {
 358                         mc->mc_error = SET_ERROR(ENXIO);
 359                         mc->mc_tried = 1;    /* don't even try */
 360                         mc->mc_skipped = 1;
 361                         continue;
 362                 }
 363 
 364                 if (vdev_dtl_contains(mc->mc_vd, DTL_MISSING, txg, 1)) {
 365                         mc->mc_error = SET_ERROR(ESTALE);
 366                         mc->mc_skipped = 1;
 367                         mc->mc_speculative = 1;
 368                         continue;
 369                 }
 370 
 371                 mc->mc_load = vdev_mirror_load(mm, mc->mc_vd, mc->mc_offset);
 372                 if (mc->mc_load > lowest_load)
 373                         continue;
 374 
 375                 if (mc->mc_load < lowest_load) {
 376                         lowest_load = mc->mc_load;
 377                         mm->mm_preferred_cnt = 0;
 378                 }
 379                 mm->mm_preferred[mm->mm_preferred_cnt] = c;
 380                 mm->mm_preferred_cnt++;
 381         }
 382 
 383         if (mm->mm_preferred_cnt == 1) {
 384                 vdev_queue_register_lastoffset(
 385                     mm->mm_child[mm->mm_preferred[0]].mc_vd, zio);
 386                 return (mm->mm_preferred[0]);
 387         }
 388 
 389         if (mm->mm_preferred_cnt > 1) {
 390                 int c = vdev_mirror_preferred_child_randomize(zio);
 391 
 392                 vdev_queue_register_lastoffset(mm->mm_child[c].mc_vd, zio);
 393                 return (c);
 394         }
 395 
 396         /*
 397          * Every device is either missing or has this txg in its DTL.
 398          * Look for any child we haven't already tried before giving up.
 399          */
 400         for (c = 0; c < mm->mm_children; c++) {
 401                 if (!mm->mm_child[c].mc_tried) {
 402                         vdev_queue_register_lastoffset(mm->mm_child[c].mc_vd,
 403                             zio);
 404                         return (c);
 405                 }
 406         }
 407 
 408         /*
 409          * Every child failed.  There's no place left to look.
 410          */
 411         return (-1);
 412 }
 413 
 414 static int
 415 vdev_mirror_io_start(zio_t *zio)
 416 {
 417         mirror_map_t *mm;
 418         mirror_child_t *mc;
 419         int c, children;
 420 
 421         mm = vdev_mirror_map_init(zio);
 422 
 423         if (zio->io_type == ZIO_TYPE_READ) {
 424                 if ((zio->io_flags & ZIO_FLAG_SCRUB) && !mm->mm_replacing) {
 425                         /*
 426                          * For scrubbing reads we need to allocate a read
 427                          * buffer for each child and issue reads to all
 428                          * children.  If any child succeeds, it will copy its
 429                          * data into zio->io_data in vdev_mirror_scrub_done.
 430                          */
 431                         for (c = 0; c < mm->mm_children; c++) {
 432                                 mc = &mm->mm_child[c];
 433                                 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
 434                                     mc->mc_vd, mc->mc_offset,
 435                                     zio_buf_alloc(zio->io_size), zio->io_size,
 436                                     zio->io_type, zio->io_priority, 0,
 437                                     vdev_mirror_scrub_done, mc));
 438                         }
 439                         return (ZIO_PIPELINE_CONTINUE);
 440                 }
 441                 /*
 442                  * For normal reads just pick one child.
 443                  */
 444                 c = vdev_mirror_child_select(zio);
 445                 children = (c >= 0);
 446         } else {
 447                 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
 448 
 449                 /*
 450                  * Writes go to all children.
 451                  */
 452                 c = 0;
 453                 children = mm->mm_children;
 454         }
 455 
 456         while (children--) {
 457                 mc = &mm->mm_child[c];
 458                 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
 459                     mc->mc_vd, mc->mc_offset, zio->io_data, zio->io_size,
 460                     zio->io_type, zio->io_priority, 0,
 461                     vdev_mirror_child_done, mc));
 462                 c++;
 463         }
 464 
 465         return (ZIO_PIPELINE_CONTINUE);
 466 }
 467 
 468 static int
 469 vdev_mirror_worst_error(mirror_map_t *mm)
 470 {
 471         int error[2] = { 0, 0 };
 472 
 473         for (int c = 0; c < mm->mm_children; c++) {
 474                 mirror_child_t *mc = &mm->mm_child[c];
 475                 int s = mc->mc_speculative;
 476                 error[s] = zio_worst_error(error[s], mc->mc_error);
 477         }
 478 
 479         return (error[0] ? error[0] : error[1]);
 480 }
 481 
 482 static void
 483 vdev_mirror_io_done(zio_t *zio)
 484 {
 485         mirror_map_t *mm = zio->io_vsd;
 486         mirror_child_t *mc;
 487         int c;
 488         int good_copies = 0;
 489         int unexpected_errors = 0;
 490 
 491         for (c = 0; c < mm->mm_children; c++) {
 492                 mc = &mm->mm_child[c];
 493 
 494                 if (mc->mc_error) {
 495                         if (!mc->mc_skipped)
 496                                 unexpected_errors++;
 497                 } else if (mc->mc_tried) {
 498                         good_copies++;
 499                 }
 500         }
 501 
 502         if (zio->io_type == ZIO_TYPE_WRITE) {
 503                 /*
 504                  * XXX -- for now, treat partial writes as success.
 505                  *
 506                  * Now that we support write reallocation, it would be better
 507                  * to treat partial failure as real failure unless there are
 508                  * no non-degraded top-level vdevs left, and not update DTLs
 509                  * if we intend to reallocate.
 510                  */
 511                 /* XXPOLICY */
 512                 if (good_copies != mm->mm_children) {
 513                         /*
 514                          * Always require at least one good copy.
 515                          *
 516                          * For ditto blocks (io_vd == NULL), require
 517                          * all copies to be good.
 518                          *
 519                          * XXX -- for replacing vdevs, there's no great answer.
 520                          * If the old device is really dead, we may not even
 521                          * be able to access it -- so we only want to
 522                          * require good writes to the new device.  But if
 523                          * the new device turns out to be flaky, we want
 524                          * to be able to detach it -- which requires all
 525                          * writes to the old device to have succeeded.
 526                          */
 527                         if (good_copies == 0 || zio->io_vd == NULL)
 528                                 zio->io_error = vdev_mirror_worst_error(mm);
 529                 }
 530                 return;
 531         }
 532 
 533         ASSERT(zio->io_type == ZIO_TYPE_READ);
 534 
 535         /*
 536          * If we don't have a good copy yet, keep trying other children.
 537          */
 538         /* XXPOLICY */
 539         if (good_copies == 0 && (c = vdev_mirror_child_select(zio)) != -1) {
 540                 ASSERT(c >= 0 && c < mm->mm_children);
 541                 mc = &mm->mm_child[c];
 542                 zio_vdev_io_redone(zio);
 543                 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
 544                     mc->mc_vd, mc->mc_offset, zio->io_data, zio->io_size,
 545                     ZIO_TYPE_READ, zio->io_priority, 0,
 546                     vdev_mirror_child_done, mc));
 547                 return;
 548         }
 549 
 550         /* XXPOLICY */
 551         if (good_copies == 0) {
 552                 zio->io_error = vdev_mirror_worst_error(mm);
 553                 ASSERT(zio->io_error != 0);
 554         }
 555 
 556         if (good_copies && spa_writeable(zio->io_spa) &&
 557             (unexpected_errors ||
 558             (zio->io_flags & ZIO_FLAG_RESILVER) ||
 559             ((zio->io_flags & ZIO_FLAG_SCRUB) && mm->mm_replacing))) {
 560                 /*
 561                  * Use the good data we have in hand to repair damaged children.
 562                  */
 563                 for (c = 0; c < mm->mm_children; c++) {
 564                         /*
 565                          * Don't rewrite known good children.
 566                          * Not only is it unnecessary, it could
 567                          * actually be harmful: if the system lost
 568                          * power while rewriting the only good copy,
 569                          * there would be no good copies left!
 570                          */
 571                         mc = &mm->mm_child[c];
 572 
 573                         if (mc->mc_error == 0) {
 574                                 if (mc->mc_tried)
 575                                         continue;
 576                                 if (!(zio->io_flags & ZIO_FLAG_SCRUB) &&
 577                                     !vdev_dtl_contains(mc->mc_vd, DTL_PARTIAL,
 578                                     zio->io_txg, 1))
 579                                         continue;
 580                                 mc->mc_error = SET_ERROR(ESTALE);
 581                         }
 582 
 583                         zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
 584                             mc->mc_vd, mc->mc_offset,
 585                             zio->io_data, zio->io_size,
 586                             ZIO_TYPE_WRITE, ZIO_PRIORITY_ASYNC_WRITE,
 587                             ZIO_FLAG_IO_REPAIR | (unexpected_errors ?
 588                             ZIO_FLAG_SELF_HEAL : 0), NULL, NULL));
 589                 }
 590         }
 591 }
 592 
 593 static void
 594 vdev_mirror_state_change(vdev_t *vd, int faulted, int degraded)
 595 {
 596         if (faulted == vd->vdev_children)
 597                 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
 598                     VDEV_AUX_NO_REPLICAS);
 599         else if (degraded + faulted != 0)
 600                 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE);
 601         else
 602                 vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE);
 603 }
 604 
 605 vdev_ops_t vdev_mirror_ops = {
 606         vdev_mirror_open,
 607         vdev_mirror_close,
 608         vdev_default_asize,
 609         vdev_mirror_io_start,
 610         vdev_mirror_io_done,
 611         vdev_mirror_state_change,
 612         NULL,
 613         NULL,
 614         VDEV_TYPE_MIRROR,       /* name of this vdev type */
 615         B_FALSE                 /* not a leaf vdev */
 616 };
 617 
 618 vdev_ops_t vdev_replacing_ops = {
 619         vdev_mirror_open,
 620         vdev_mirror_close,
 621         vdev_default_asize,
 622         vdev_mirror_io_start,
 623         vdev_mirror_io_done,
 624         vdev_mirror_state_change,
 625         NULL,
 626         NULL,
 627         VDEV_TYPE_REPLACING,    /* name of this vdev type */
 628         B_FALSE                 /* not a leaf vdev */
 629 };
 630 
 631 vdev_ops_t vdev_spare_ops = {
 632         vdev_mirror_open,
 633         vdev_mirror_close,
 634         vdev_default_asize,
 635         vdev_mirror_io_start,
 636         vdev_mirror_io_done,
 637         vdev_mirror_state_change,
 638         NULL,
 639         NULL,
 640         VDEV_TYPE_SPARE,        /* name of this vdev type */
 641         B_FALSE                 /* not a leaf vdev */
 642 };