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 int zfs_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 int zfs_vdev_mirror_rotating_inc = 0;
75
76 /* Rotating media load increment for seeking I/O's. */
77 int zfs_vdev_mirror_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 int zfs_vdev_mirror_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 int zfs_vdev_mirror_non_rotating_inc = 0;
88
89 /* Non-rotating media load increment for seeking I/O's. */
90 int zfs_vdev_mirror_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 /*
128 * Calculate and return the load of the specified vdev adjusted for a zio at
129 * the given offset.
130 *
131 * The calcuation takes into account the vdev's:
132 * 1. Rotation rate
133 * 2. The distance of zio_offset from the last queued request
134 */
135 static int
136 vdev_mirror_load(mirror_map_t *mm, vdev_t *vd, uint64_t zio_offset)
137 {
138 uint64_t lastoffset;
139 int load;
140
141 /* All DVAs have equal weight at the root. */
142 if (mm->mm_root)
143 return (INT_MAX);
144
145 /*
146 * We don't return INT_MAX if the device is resilvering i.e.
147 * vdev_resilver_txg != 0 as when tested performance was slightly
148 * worse overall when resilvering with compared to without.
149 */
150
151 /* Standard load based on pending queue length. */
152 load = vdev_queue_length(vd);
153 lastoffset = vdev_queue_last_queued_offset(vd);
154
155 if (vd->vdev_rotation_rate == VDEV_RATE_NON_ROTATING) {
156 /* Non-rotating media. */
157 if (lastoffset == zio_offset)
158 return (load + zfs_vdev_mirror_non_rotating_inc);
159
160 /*
161 * Apply a seek penalty even for non-rotating devices as
162 * sequential I/O'a can be aggregated into fewer operations
163 * on the device, thus avoiding unnecessary per-command
164 * overhead and boosting performance.
165 */
166 return (load + zfs_vdev_mirror_non_rotating_seek_inc);
167 }
168
169 /* Rotating media I/O's which directly follow the last I/O. */
170 if (lastoffset == zio_offset)
171 return (load + zfs_vdev_mirror_rotating_inc);
172
173 /*
174 * Apply half the seek increment to I/O's within seek offset
175 * of the last I/O queued to this vdev as they should incure less
176 * of a seek increment.
177 */
178 if (ABS(lastoffset - zio_offset) <
179 zfs_vdev_mirror_rotating_seek_offset)
180 return (load + (zfs_vdev_mirror_rotating_seek_inc / 2));
181
182 /* Apply the full seek increment to all other I/O's. */
183 return (load + zfs_vdev_mirror_rotating_seek_inc);
184 }
185
186
187 static mirror_map_t *
188 vdev_mirror_map_init(zio_t *zio)
189 {
190 mirror_map_t *mm = NULL;
191 mirror_child_t *mc;
192 vdev_t *vd = zio->io_vd;
193 int c;
194
195 if (vd == NULL) {
196 dva_t *dva = zio->io_bp->blk_dva;
197 spa_t *spa = zio->io_spa;
198
199 mm = vdev_mirror_map_alloc(BP_GET_NDVAS(zio->io_bp), B_FALSE,
200 B_TRUE);
201 for (c = 0; c < mm->mm_children; c++) {
202 mc = &mm->mm_child[c];
203 mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c]));
204 mc->mc_offset = DVA_GET_OFFSET(&dva[c]);
205 }
206 } else {
207 mm = vdev_mirror_map_alloc(vd->vdev_children,
208 (vd->vdev_ops == &vdev_replacing_ops ||
209 vd->vdev_ops == &vdev_spare_ops), B_FALSE);
210 for (c = 0; c < mm->mm_children; c++) {
211 mc = &mm->mm_child[c];
212 mc->mc_vd = vd->vdev_child[c];
213 mc->mc_offset = zio->io_offset;
214 }
215 }
216
217 zio->io_vsd = mm;
218 zio->io_vsd_ops = &vdev_mirror_vsd_ops;
219 return (mm);
220 }
221
222 static int
223 vdev_mirror_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize,
224 uint64_t *ashift)
225 {
226 int numerrors = 0;
227 int lasterror = 0;
228
229 if (vd->vdev_children == 0) {
230 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
231 return (SET_ERROR(EINVAL));
232 }
233
234 vdev_open_children(vd);
235
236 for (int c = 0; c < vd->vdev_children; c++) {
237 vdev_t *cvd = vd->vdev_child[c];
238
239 if (cvd->vdev_open_error) {
240 lasterror = cvd->vdev_open_error;
241 numerrors++;
242 continue;
243 }
244
245 *asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1;
246 *max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1;
247 *ashift = MAX(*ashift, cvd->vdev_ashift);
248 }
249
250 if (numerrors == vd->vdev_children) {
251 vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS;
252 return (lasterror);
253 }
254
255 return (0);
256 }
257
258 static void
259 vdev_mirror_close(vdev_t *vd)
260 {
261 for (int c = 0; c < vd->vdev_children; c++)
262 vdev_close(vd->vdev_child[c]);
263 }
264
265 static void
266 vdev_mirror_child_done(zio_t *zio)
267 {
268 mirror_child_t *mc = zio->io_private;
269
270 mc->mc_error = zio->io_error;
271 mc->mc_tried = 1;
272 mc->mc_skipped = 0;
273 }
274
275 static void
276 vdev_mirror_scrub_done(zio_t *zio)
277 {
278 mirror_child_t *mc = zio->io_private;
279
280 if (zio->io_error == 0) {
281 zio_t *pio;
282
283 mutex_enter(&zio->io_lock);
284 while ((pio = zio_walk_parents(zio)) != NULL) {
285 mutex_enter(&pio->io_lock);
286 ASSERT3U(zio->io_size, >=, pio->io_size);
287 bcopy(zio->io_data, pio->io_data, pio->io_size);
288 mutex_exit(&pio->io_lock);
289 }
290 mutex_exit(&zio->io_lock);
291 }
292
293 zio_buf_free(zio->io_data, zio->io_size);
294
295 mc->mc_error = zio->io_error;
296 mc->mc_tried = 1;
297 mc->mc_skipped = 0;
298 }
299
300 /*
301 * Check the other, lower-index DVAs to see if they're on the same
302 * vdev as the child we picked. If they are, use them since they
303 * are likely to have been allocated from the primary metaslab in
304 * use at the time, and hence are more likely to have locality with
305 * single-copy data.
306 */
307 static int
308 vdev_mirror_dva_select(zio_t *zio, int preferred)
309 {
310 dva_t *dva = zio->io_bp->blk_dva;
311 int c;
312
313 for (c = preferred - 1; c >= 0; c--) {
314 if (DVA_GET_VDEV(&dva[c]) == DVA_GET_VDEV(&dva[preferred]))
315 preferred = c;
316 }
317 return (preferred);
318 }
319
320 static int
321 vdev_mirror_preferred_child_randomize(zio_t *zio)
322 {
323 mirror_map_t *mm = zio->io_vsd;
324 int p;
325
326 if (mm->mm_root) {
327 p = spa_get_random(mm->mm_preferred_cnt);
328 return (vdev_mirror_dva_select(zio, mm->mm_preferred[p]));
329 }
330
331 /*
332 * To ensure we don't always favour the first matching vdev,
333 * which could lead to wear leveling issues on SSD's, we
334 * use the I/O offset as a pseudo random seed into the vdevs
335 * which have the lowest load.
336 */
337 p = (zio->io_offset >> zfs_vdev_mirror_shift) % mm->mm_preferred_cnt;
338 return (mm->mm_preferred[p]);
339 }
340
341 /*
342 * Try to find a vdev whose DTL doesn't contain the block we want to read
343 * prefering vdevs based on determined load.
344 *
345 * If we can't, try the read on any vdev we haven't already tried.
346 */
347 static int
348 vdev_mirror_child_select(zio_t *zio)
349 {
350 mirror_map_t *mm = zio->io_vsd;
351 uint64_t txg = zio->io_txg;
352 int c, lowest_load;
353
354 ASSERT(zio->io_bp == NULL || BP_PHYSICAL_BIRTH(zio->io_bp) == txg);
355
356 lowest_load = INT_MAX;
357 mm->mm_preferred_cnt = 0;
358 for (c = 0; c < mm->mm_children; c++) {
359 mirror_child_t *mc;
360
361 mc = &mm->mm_child[c];
362 if (mc->mc_tried || mc->mc_skipped)
363 continue;
364
365 if (!vdev_readable(mc->mc_vd)) {
366 mc->mc_error = SET_ERROR(ENXIO);
367 mc->mc_tried = 1; /* don't even try */
368 mc->mc_skipped = 1;
369 continue;
370 }
371
372 if (vdev_dtl_contains(mc->mc_vd, DTL_MISSING, txg, 1)) {
373 mc->mc_error = SET_ERROR(ESTALE);
374 mc->mc_skipped = 1;
375 mc->mc_speculative = 1;
376 continue;
377 }
378
379 mc->mc_load = vdev_mirror_load(mm, mc->mc_vd, mc->mc_offset);
380 if (mc->mc_load > lowest_load)
381 continue;
382
383 if (mc->mc_load < lowest_load) {
384 lowest_load = mc->mc_load;
385 mm->mm_preferred_cnt = 0;
386 }
387 mm->mm_preferred[mm->mm_preferred_cnt] = c;
388 mm->mm_preferred_cnt++;
389 }
390
391 if (mm->mm_preferred_cnt == 1) {
392 vdev_queue_register_last_queued_offset(
393 mm->mm_child[mm->mm_preferred[0]].mc_vd, zio);
394 return (mm->mm_preferred[0]);
395 }
396
397 if (mm->mm_preferred_cnt > 1) {
398 int c = vdev_mirror_preferred_child_randomize(zio);
399
400 vdev_queue_register_last_queued_offset(mm->mm_child[c].mc_vd,
401 zio);
402 return (c);
403 }
404
405 /*
406 * Every device is either missing or has this txg in its DTL.
407 * Look for any child we haven't already tried before giving up.
408 */
409 for (c = 0; c < mm->mm_children; c++) {
410 if (!mm->mm_child[c].mc_tried) {
411 vdev_queue_register_last_queued_offset(
412 mm->mm_child[c].mc_vd, zio);
413 return (c);
414 }
415 }
416
417 /*
418 * Every child failed. There's no place left to look.
419 */
420 return (-1);
421 }
422
423 static int
424 vdev_mirror_io_start(zio_t *zio)
425 {
426 mirror_map_t *mm;
427 mirror_child_t *mc;
428 int c, children;
429
430 mm = vdev_mirror_map_init(zio);
431
432 if (zio->io_type == ZIO_TYPE_READ) {
433 if ((zio->io_flags & ZIO_FLAG_SCRUB) && !mm->mm_replacing) {
434 /*
435 * For scrubbing reads we need to allocate a read
436 * buffer for each child and issue reads to all
437 * children. If any child succeeds, it will copy its
438 * data into zio->io_data in vdev_mirror_scrub_done.
439 */
440 for (c = 0; c < mm->mm_children; c++) {
441 mc = &mm->mm_child[c];
442 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
443 mc->mc_vd, mc->mc_offset,
444 zio_buf_alloc(zio->io_size), zio->io_size,
445 zio->io_type, zio->io_priority, 0,
446 vdev_mirror_scrub_done, mc));
447 }
448 return (ZIO_PIPELINE_CONTINUE);
449 }
450 /*
451 * For normal reads just pick one child.
452 */
453 c = vdev_mirror_child_select(zio);
454 children = (c >= 0);
455 } else {
456 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
457
458 /*
459 * Writes go to all children.
460 */
461 c = 0;
462 children = mm->mm_children;
463 }
464
465 while (children--) {
466 mc = &mm->mm_child[c];
467 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
468 mc->mc_vd, mc->mc_offset, zio->io_data, zio->io_size,
469 zio->io_type, zio->io_priority, 0,
470 vdev_mirror_child_done, mc));
471 c++;
472 }
473
474 return (ZIO_PIPELINE_CONTINUE);
475 }
476
477 static int
478 vdev_mirror_worst_error(mirror_map_t *mm)
479 {
480 int error[2] = { 0, 0 };
481
482 for (int c = 0; c < mm->mm_children; c++) {
483 mirror_child_t *mc = &mm->mm_child[c];
484 int s = mc->mc_speculative;
485 error[s] = zio_worst_error(error[s], mc->mc_error);
486 }
487
488 return (error[0] ? error[0] : error[1]);
489 }
490
491 static void
492 vdev_mirror_io_done(zio_t *zio)
493 {
494 mirror_map_t *mm = zio->io_vsd;
495 mirror_child_t *mc;
496 int c;
497 int good_copies = 0;
498 int unexpected_errors = 0;
499
500 for (c = 0; c < mm->mm_children; c++) {
501 mc = &mm->mm_child[c];
502
503 if (mc->mc_error) {
504 if (!mc->mc_skipped)
505 unexpected_errors++;
506 } else if (mc->mc_tried) {
507 good_copies++;
508 }
509 }
510
511 if (zio->io_type == ZIO_TYPE_WRITE) {
512 /*
513 * XXX -- for now, treat partial writes as success.
514 *
515 * Now that we support write reallocation, it would be better
516 * to treat partial failure as real failure unless there are
517 * no non-degraded top-level vdevs left, and not update DTLs
518 * if we intend to reallocate.
519 */
520 /* XXPOLICY */
521 if (good_copies != mm->mm_children) {
522 /*
523 * Always require at least one good copy.
524 *
525 * For ditto blocks (io_vd == NULL), require
526 * all copies to be good.
527 *
528 * XXX -- for replacing vdevs, there's no great answer.
529 * If the old device is really dead, we may not even
530 * be able to access it -- so we only want to
531 * require good writes to the new device. But if
532 * the new device turns out to be flaky, we want
533 * to be able to detach it -- which requires all
534 * writes to the old device to have succeeded.
535 */
536 if (good_copies == 0 || zio->io_vd == NULL)
537 zio->io_error = vdev_mirror_worst_error(mm);
538 }
539 return;
540 }
541
542 ASSERT(zio->io_type == ZIO_TYPE_READ);
543
544 /*
545 * If we don't have a good copy yet, keep trying other children.
546 */
547 /* XXPOLICY */
548 if (good_copies == 0 && (c = vdev_mirror_child_select(zio)) != -1) {
549 ASSERT(c >= 0 && c < mm->mm_children);
550 mc = &mm->mm_child[c];
551 zio_vdev_io_redone(zio);
552 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
553 mc->mc_vd, mc->mc_offset, zio->io_data, zio->io_size,
554 ZIO_TYPE_READ, zio->io_priority, 0,
555 vdev_mirror_child_done, mc));
556 return;
557 }
558
559 /* XXPOLICY */
560 if (good_copies == 0) {
561 zio->io_error = vdev_mirror_worst_error(mm);
562 ASSERT(zio->io_error != 0);
563 }
564
565 if (good_copies && spa_writeable(zio->io_spa) &&
566 (unexpected_errors ||
567 (zio->io_flags & ZIO_FLAG_RESILVER) ||
568 ((zio->io_flags & ZIO_FLAG_SCRUB) && mm->mm_replacing))) {
569 /*
570 * Use the good data we have in hand to repair damaged children.
571 */
572 for (c = 0; c < mm->mm_children; c++) {
573 /*
574 * Don't rewrite known good children.
575 * Not only is it unnecessary, it could
576 * actually be harmful: if the system lost
577 * power while rewriting the only good copy,
578 * there would be no good copies left!
579 */
580 mc = &mm->mm_child[c];
581
582 if (mc->mc_error == 0) {
583 if (mc->mc_tried)
584 continue;
585 if (!(zio->io_flags & ZIO_FLAG_SCRUB) &&
586 !vdev_dtl_contains(mc->mc_vd, DTL_PARTIAL,
587 zio->io_txg, 1))
588 continue;
589 mc->mc_error = SET_ERROR(ESTALE);
590 }
591
592 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
593 mc->mc_vd, mc->mc_offset,
594 zio->io_data, zio->io_size,
595 ZIO_TYPE_WRITE, ZIO_PRIORITY_ASYNC_WRITE,
596 ZIO_FLAG_IO_REPAIR | (unexpected_errors ?
597 ZIO_FLAG_SELF_HEAL : 0), NULL, NULL));
598 }
599 }
600 }
601
602 static void
603 vdev_mirror_state_change(vdev_t *vd, int faulted, int degraded)
604 {
605 if (faulted == vd->vdev_children)
606 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
607 VDEV_AUX_NO_REPLICAS);
608 else if (degraded + faulted != 0)
609 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE);
610 else
611 vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE);
612 }
613
614 vdev_ops_t vdev_mirror_ops = {
615 vdev_mirror_open,
616 vdev_mirror_close,
617 vdev_default_asize,
618 vdev_mirror_io_start,
619 vdev_mirror_io_done,
620 vdev_mirror_state_change,
621 NULL,
622 NULL,
623 VDEV_TYPE_MIRROR, /* name of this vdev type */
624 B_FALSE /* not a leaf vdev */
625 };
626
627 vdev_ops_t vdev_replacing_ops = {
628 vdev_mirror_open,
629 vdev_mirror_close,
630 vdev_default_asize,
631 vdev_mirror_io_start,
632 vdev_mirror_io_done,
633 vdev_mirror_state_change,
634 NULL,
635 NULL,
636 VDEV_TYPE_REPLACING, /* name of this vdev type */
637 B_FALSE /* not a leaf vdev */
638 };
639
640 vdev_ops_t vdev_spare_ops = {
641 vdev_mirror_open,
642 vdev_mirror_close,
643 vdev_default_asize,
644 vdev_mirror_io_start,
645 vdev_mirror_io_done,
646 vdev_mirror_state_change,
647 NULL,
648 NULL,
649 VDEV_TYPE_SPARE, /* name of this vdev type */
650 B_FALSE /* not a leaf vdev */
651 };