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) 2012, 2018 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/zio.h>
37 #include <sys/abd.h>
38 #include <sys/fs/zfs.h>
39 
40 /*
41  * Virtual device vector for mirroring.
42  */
43 
44 typedef struct mirror_child {
45 	vdev_t		*mc_vd;
46 	uint64_t	mc_offset;
47 	int		mc_error;
48 	uint8_t		mc_tried;
49 	uint8_t		mc_skipped;
50 	uint8_t		mc_speculative;
51 } mirror_child_t;
52 
53 typedef struct mirror_map {
54 	int		mm_children;
55 	int		mm_resilvering;
56 	int		mm_preferred;
57 	int		mm_root;
58 	mirror_child_t	mm_child[1];
59 } mirror_map_t;
60 
61 int vdev_mirror_shift = 21;
62 
63 static void
64 vdev_mirror_map_free(zio_t *zio)
65 {
66 	mirror_map_t *mm = zio->io_vsd;
67 
68 	kmem_free(mm, offsetof(mirror_map_t, mm_child[mm->mm_children]));
69 }
70 
71 static const zio_vsd_ops_t vdev_mirror_vsd_ops = {
72 	vdev_mirror_map_free,
73 	zio_vsd_default_cksum_report
74 };
75 
76 static mirror_map_t *
77 vdev_mirror_map_alloc(zio_t *zio)
78 {
79 	mirror_map_t *mm = NULL;
80 	mirror_child_t *mc;
81 	vdev_t *vd = zio->io_vd;
82 	int c, d;
83 
84 	if (vd == NULL) {
85 		dva_t *dva = zio->io_bp->blk_dva;
86 		spa_t *spa = zio->io_spa;
87 		dva_t dva_copy[SPA_DVAS_PER_BP];
88 
89 		c = BP_GET_NDVAS(zio->io_bp);
90 
91 		/*
92 		 * If we do not trust the pool config, some DVAs might be
93 		 * invalid or point to vdevs that do not exist. We skip them.
94 		 */
95 		if (!spa_trust_config(spa)) {
96 			ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ);
97 			int j = 0;
98 			for (int i = 0; i < c; i++) {
99 				if (zfs_dva_valid(spa, &dva[i], zio->io_bp))
100 					dva_copy[j++] = dva[i];
101 			}
102 			if (j == 0) {
103 				zio->io_vsd = NULL;
104 				zio->io_error = ENXIO;
105 				return (NULL);
106 			}
107 			if (j < c) {
108 				dva = dva_copy;
109 				c = j;
110 			}
111 		}
112 
113 		mm = kmem_zalloc(offsetof(mirror_map_t, mm_child[c]), KM_SLEEP);
114 		mm->mm_children = c;
115 		mm->mm_resilvering = B_FALSE;
116 		mm->mm_preferred = spa_get_random(c);
117 		mm->mm_root = B_TRUE;
118 
119 		/*
120 		 * Check the other, lower-index DVAs to see if they're on
121 		 * the same vdev as the child we picked.  If they are, use
122 		 * them since they are likely to have been allocated from
123 		 * the primary metaslab in use at the time, and hence are
124 		 * more likely to have locality with single-copy data.
125 		 */
126 		for (c = mm->mm_preferred, d = c - 1; d >= 0; d--) {
127 			if (DVA_GET_VDEV(&dva[d]) == DVA_GET_VDEV(&dva[c]))
128 				mm->mm_preferred = d;
129 		}
130 
131 		for (c = 0; c < mm->mm_children; c++) {
132 			mc = &mm->mm_child[c];
133 
134 			mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c]));
135 			mc->mc_offset = DVA_GET_OFFSET(&dva[c]);
136 		}
137 	} else {
138 		int replacing;
139 
140 		c = vd->vdev_children;
141 
142 		mm = kmem_zalloc(offsetof(mirror_map_t, mm_child[c]), KM_SLEEP);
143 		mm->mm_children = c;
144 		/*
145 		 * If we are resilvering, then we should handle scrub reads
146 		 * differently; we shouldn't issue them to the resilvering
147 		 * device because it might not have those blocks.
148 		 *
149 		 * We are resilvering iff:
150 		 * 1) We are a replacing vdev (ie our name is "replacing-1" or
151 		 *    "spare-1" or something like that), and
152 		 * 2) The pool is currently being resilvered.
153 		 *
154 		 * We cannot simply check vd->vdev_resilver_txg, because it's
155 		 * not set in this path.
156 		 *
157 		 * Nor can we just check our vdev_ops; there are cases (such as
158 		 * when a user types "zpool replace pool odev spare_dev" and
159 		 * spare_dev is in the spare list, or when a spare device is
160 		 * automatically used to replace a DEGRADED device) when
161 		 * resilvering is complete but both the original vdev and the
162 		 * spare vdev remain in the pool.  That behavior is intentional.
163 		 * It helps implement the policy that a spare should be
164 		 * automatically removed from the pool after the user replaces
165 		 * the device that originally failed.
166 		 */
167 		replacing = (vd->vdev_ops == &vdev_replacing_ops ||
168 		    vd->vdev_ops == &vdev_spare_ops);
169 		/*
170 		 * If a spa load is in progress, then spa_dsl_pool may be
171 		 * uninitialized.  But we shouldn't be resilvering during a spa
172 		 * load anyway.
173 		 */
174 		if (replacing &&
175 		    (spa_load_state(vd->vdev_spa) == SPA_LOAD_NONE) &&
176 		    dsl_scan_resilvering(vd->vdev_spa->spa_dsl_pool)) {
177 			mm->mm_resilvering = B_TRUE;
178 		} else {
179 			mm->mm_resilvering = B_FALSE;
180 		}
181 
182 		mm->mm_preferred = mm->mm_resilvering ? 0 :
183 		    (zio->io_offset >> vdev_mirror_shift) % c;
184 		mm->mm_root = B_FALSE;
185 
186 		for (c = 0; c < mm->mm_children; c++) {
187 			mc = &mm->mm_child[c];
188 			mc->mc_vd = vd->vdev_child[c];
189 			mc->mc_offset = zio->io_offset;
190 		}
191 	}
192 
193 	zio->io_vsd = mm;
194 	zio->io_vsd_ops = &vdev_mirror_vsd_ops;
195 	return (mm);
196 }
197 
198 static int
199 vdev_mirror_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize,
200     uint64_t *ashift)
201 {
202 	int numerrors = 0;
203 	int lasterror = 0;
204 
205 	if (vd->vdev_children == 0) {
206 		vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
207 		return (SET_ERROR(EINVAL));
208 	}
209 
210 	vdev_open_children(vd);
211 
212 	for (int c = 0; c < vd->vdev_children; c++) {
213 		vdev_t *cvd = vd->vdev_child[c];
214 
215 		if (cvd->vdev_open_error) {
216 			lasterror = cvd->vdev_open_error;
217 			numerrors++;
218 			continue;
219 		}
220 
221 		*asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1;
222 		*max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1;
223 		*ashift = MAX(*ashift, cvd->vdev_ashift);
224 	}
225 
226 	if (numerrors == vd->vdev_children) {
227 		if (vdev_children_are_offline(vd))
228 			vd->vdev_stat.vs_aux = VDEV_AUX_CHILDREN_OFFLINE;
229 		else
230 			vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS;
231 		return (lasterror);
232 	}
233 
234 	return (0);
235 }
236 
237 static void
238 vdev_mirror_close(vdev_t *vd)
239 {
240 	for (int c = 0; c < vd->vdev_children; c++)
241 		vdev_close(vd->vdev_child[c]);
242 }
243 
244 static void
245 vdev_mirror_child_done(zio_t *zio)
246 {
247 	mirror_child_t *mc = zio->io_private;
248 
249 	mc->mc_error = zio->io_error;
250 	mc->mc_tried = 1;
251 	mc->mc_skipped = 0;
252 }
253 
254 static void
255 vdev_mirror_scrub_done(zio_t *zio)
256 {
257 	mirror_child_t *mc = zio->io_private;
258 
259 	if (zio->io_error == 0) {
260 		zio_t *pio;
261 		zio_link_t *zl = NULL;
262 
263 		mutex_enter(&zio->io_lock);
264 		while ((pio = zio_walk_parents(zio, &zl)) != NULL) {
265 			mutex_enter(&pio->io_lock);
266 			ASSERT3U(zio->io_size, >=, pio->io_size);
267 			abd_copy(pio->io_abd, zio->io_abd, pio->io_size);
268 			mutex_exit(&pio->io_lock);
269 		}
270 		mutex_exit(&zio->io_lock);
271 	}
272 	abd_free(zio->io_abd);
273 
274 	mc->mc_error = zio->io_error;
275 	mc->mc_tried = 1;
276 	mc->mc_skipped = 0;
277 }
278 
279 /*
280  * Try to find a child whose DTL doesn't contain the block we want to read.
281  * If we can't, try the read on any vdev we haven't already tried.
282  */
283 static int
284 vdev_mirror_child_select(zio_t *zio)
285 {
286 	mirror_map_t *mm = zio->io_vsd;
287 	mirror_child_t *mc;
288 	uint64_t txg = zio->io_txg;
289 	int i, c;
290 
291 	ASSERT(zio->io_bp == NULL || BP_PHYSICAL_BIRTH(zio->io_bp) == txg);
292 
293 	/*
294 	 * Try to find a child whose DTL doesn't contain the block to read.
295 	 * If a child is known to be completely inaccessible (indicated by
296 	 * vdev_readable() returning B_FALSE), don't even try.
297 	 */
298 	for (i = 0, c = mm->mm_preferred; i < mm->mm_children; i++, c++) {
299 		if (c >= mm->mm_children)
300 			c = 0;
301 		mc = &mm->mm_child[c];
302 		if (mc->mc_tried || mc->mc_skipped)
303 			continue;
304 		if (!vdev_readable(mc->mc_vd)) {
305 			mc->mc_error = SET_ERROR(ENXIO);
306 			mc->mc_tried = 1;	/* don't even try */
307 			mc->mc_skipped = 1;
308 			continue;
309 		}
310 		if (!vdev_dtl_contains(mc->mc_vd, DTL_MISSING, txg, 1))
311 			return (c);
312 		mc->mc_error = SET_ERROR(ESTALE);
313 		mc->mc_skipped = 1;
314 		mc->mc_speculative = 1;
315 	}
316 
317 	/*
318 	 * Every device is either missing or has this txg in its DTL.
319 	 * Look for any child we haven't already tried before giving up.
320 	 */
321 	for (c = 0; c < mm->mm_children; c++)
322 		if (!mm->mm_child[c].mc_tried)
323 			return (c);
324 
325 	/*
326 	 * Every child failed.  There's no place left to look.
327 	 */
328 	return (-1);
329 }
330 
331 static void
332 vdev_mirror_io_start(zio_t *zio)
333 {
334 	mirror_map_t *mm;
335 	mirror_child_t *mc;
336 	int c, children;
337 
338 	mm = vdev_mirror_map_alloc(zio);
339 
340 	if (mm == NULL) {
341 		ASSERT(!spa_trust_config(zio->io_spa));
342 		ASSERT(zio->io_type == ZIO_TYPE_READ);
343 		zio_execute(zio);
344 		return;
345 	}
346 
347 	if (zio->io_type == ZIO_TYPE_READ) {
348 		if (zio->io_bp != NULL &&
349 		    (zio->io_flags & ZIO_FLAG_SCRUB) && !mm->mm_resilvering) {
350 			/*
351 			 * For scrubbing reads (if we can verify the
352 			 * checksum here, as indicated by io_bp being
353 			 * non-NULL) we need to allocate a read buffer for
354 			 * each child and issue reads to all children.  If
355 			 * any child succeeds, it will copy its data into
356 			 * zio->io_data in vdev_mirror_scrub_done.
357 			 */
358 			for (c = 0; c < mm->mm_children; c++) {
359 				mc = &mm->mm_child[c];
360 				zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
361 				    mc->mc_vd, mc->mc_offset,
362 				    abd_alloc_sametype(zio->io_abd,
363 				    zio->io_size), zio->io_size,
364 				    zio->io_type, zio->io_priority, 0,
365 				    vdev_mirror_scrub_done, mc));
366 			}
367 			zio_execute(zio);
368 			return;
369 		}
370 		/*
371 		 * For normal reads just pick one child.
372 		 */
373 		c = vdev_mirror_child_select(zio);
374 		children = (c >= 0);
375 	} else {
376 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
377 
378 		/*
379 		 * Writes go to all children.
380 		 */
381 		c = 0;
382 		children = mm->mm_children;
383 	}
384 
385 	while (children--) {
386 		mc = &mm->mm_child[c];
387 		zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
388 		    mc->mc_vd, mc->mc_offset, zio->io_abd, zio->io_size,
389 		    zio->io_type, zio->io_priority, 0,
390 		    vdev_mirror_child_done, mc));
391 		c++;
392 	}
393 
394 	zio_execute(zio);
395 }
396 
397 static int
398 vdev_mirror_worst_error(mirror_map_t *mm)
399 {
400 	int error[2] = { 0, 0 };
401 
402 	for (int c = 0; c < mm->mm_children; c++) {
403 		mirror_child_t *mc = &mm->mm_child[c];
404 		int s = mc->mc_speculative;
405 		error[s] = zio_worst_error(error[s], mc->mc_error);
406 	}
407 
408 	return (error[0] ? error[0] : error[1]);
409 }
410 
411 static void
412 vdev_mirror_io_done(zio_t *zio)
413 {
414 	mirror_map_t *mm = zio->io_vsd;
415 	mirror_child_t *mc;
416 	int c;
417 	int good_copies = 0;
418 	int unexpected_errors = 0;
419 
420 	if (mm == NULL)
421 		return;
422 
423 	for (c = 0; c < mm->mm_children; c++) {
424 		mc = &mm->mm_child[c];
425 
426 		if (mc->mc_error) {
427 			if (!mc->mc_skipped)
428 				unexpected_errors++;
429 		} else if (mc->mc_tried) {
430 			good_copies++;
431 		}
432 	}
433 
434 	if (zio->io_type == ZIO_TYPE_WRITE) {
435 		/*
436 		 * XXX -- for now, treat partial writes as success.
437 		 *
438 		 * Now that we support write reallocation, it would be better
439 		 * to treat partial failure as real failure unless there are
440 		 * no non-degraded top-level vdevs left, and not update DTLs
441 		 * if we intend to reallocate.
442 		 */
443 		/* XXPOLICY */
444 		if (good_copies != mm->mm_children) {
445 			/*
446 			 * Always require at least one good copy.
447 			 *
448 			 * For ditto blocks (io_vd == NULL), require
449 			 * all copies to be good.
450 			 *
451 			 * XXX -- for replacing vdevs, there's no great answer.
452 			 * If the old device is really dead, we may not even
453 			 * be able to access it -- so we only want to
454 			 * require good writes to the new device.  But if
455 			 * the new device turns out to be flaky, we want
456 			 * to be able to detach it -- which requires all
457 			 * writes to the old device to have succeeded.
458 			 */
459 			if (good_copies == 0 || zio->io_vd == NULL)
460 				zio->io_error = vdev_mirror_worst_error(mm);
461 		}
462 		return;
463 	}
464 
465 	ASSERT(zio->io_type == ZIO_TYPE_READ);
466 
467 	/*
468 	 * If we don't have a good copy yet, keep trying other children.
469 	 */
470 	/* XXPOLICY */
471 	if (good_copies == 0 && (c = vdev_mirror_child_select(zio)) != -1) {
472 		ASSERT(c >= 0 && c < mm->mm_children);
473 		mc = &mm->mm_child[c];
474 		zio_vdev_io_redone(zio);
475 		zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
476 		    mc->mc_vd, mc->mc_offset, zio->io_abd, zio->io_size,
477 		    ZIO_TYPE_READ, zio->io_priority, 0,
478 		    vdev_mirror_child_done, mc));
479 		return;
480 	}
481 
482 	/* XXPOLICY */
483 	if (good_copies == 0) {
484 		zio->io_error = vdev_mirror_worst_error(mm);
485 		ASSERT(zio->io_error != 0);
486 	}
487 
488 	if (good_copies && spa_writeable(zio->io_spa) &&
489 	    (unexpected_errors ||
490 	    (zio->io_flags & ZIO_FLAG_RESILVER) ||
491 	    ((zio->io_flags & ZIO_FLAG_SCRUB) && mm->mm_resilvering))) {
492 		/*
493 		 * Use the good data we have in hand to repair damaged children.
494 		 */
495 		for (c = 0; c < mm->mm_children; c++) {
496 			/*
497 			 * Don't rewrite known good children.
498 			 * Not only is it unnecessary, it could
499 			 * actually be harmful: if the system lost
500 			 * power while rewriting the only good copy,
501 			 * there would be no good copies left!
502 			 */
503 			mc = &mm->mm_child[c];
504 
505 			if (mc->mc_error == 0) {
506 				if (mc->mc_tried)
507 					continue;
508 				/*
509 				 * We didn't try this child.  We need to
510 				 * repair it if:
511 				 * 1. it's a scrub (in which case we have
512 				 * tried everything that was healthy)
513 				 *  - or -
514 				 * 2. it's an indirect vdev (in which case
515 				 * it could point to any other vdev, which
516 				 * might have a bad DTL)
517 				 *  - or -
518 				 * 3. the DTL indicates that this data is
519 				 * missing from this vdev
520 				 */
521 				if (!(zio->io_flags & ZIO_FLAG_SCRUB) &&
522 				    mc->mc_vd->vdev_ops != &vdev_indirect_ops &&
523 				    !vdev_dtl_contains(mc->mc_vd, DTL_PARTIAL,
524 				    zio->io_txg, 1))
525 					continue;
526 				mc->mc_error = SET_ERROR(ESTALE);
527 			}
528 
529 			zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
530 			    mc->mc_vd, mc->mc_offset,
531 			    zio->io_abd, zio->io_size,
532 			    ZIO_TYPE_WRITE, ZIO_PRIORITY_ASYNC_WRITE,
533 			    ZIO_FLAG_IO_REPAIR | (unexpected_errors ?
534 			    ZIO_FLAG_SELF_HEAL : 0), NULL, NULL));
535 		}
536 	}
537 }
538 
539 static void
540 vdev_mirror_state_change(vdev_t *vd, int faulted, int degraded)
541 {
542 	if (faulted == vd->vdev_children) {
543 		if (vdev_children_are_offline(vd)) {
544 			vdev_set_state(vd, B_FALSE, VDEV_STATE_OFFLINE,
545 			    VDEV_AUX_CHILDREN_OFFLINE);
546 		} else {
547 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
548 			    VDEV_AUX_NO_REPLICAS);
549 		}
550 	} else if (degraded + faulted != 0) {
551 		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE);
552 	} else {
553 		vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE);
554 	}
555 }
556 
557 vdev_ops_t vdev_mirror_ops = {
558 	vdev_mirror_open,
559 	vdev_mirror_close,
560 	vdev_default_asize,
561 	vdev_mirror_io_start,
562 	vdev_mirror_io_done,
563 	vdev_mirror_state_change,
564 	NULL,
565 	NULL,
566 	NULL,
567 	vdev_default_xlate,
568 	VDEV_TYPE_MIRROR,	/* name of this vdev type */
569 	B_FALSE			/* not a leaf vdev */
570 };
571 
572 vdev_ops_t vdev_replacing_ops = {
573 	vdev_mirror_open,
574 	vdev_mirror_close,
575 	vdev_default_asize,
576 	vdev_mirror_io_start,
577 	vdev_mirror_io_done,
578 	vdev_mirror_state_change,
579 	NULL,
580 	NULL,
581 	NULL,
582 	vdev_default_xlate,
583 	VDEV_TYPE_REPLACING,	/* name of this vdev type */
584 	B_FALSE			/* not a leaf vdev */
585 };
586 
587 vdev_ops_t vdev_spare_ops = {
588 	vdev_mirror_open,
589 	vdev_mirror_close,
590 	vdev_default_asize,
591 	vdev_mirror_io_start,
592 	vdev_mirror_io_done,
593 	vdev_mirror_state_change,
594 	NULL,
595 	NULL,
596 	NULL,
597 	vdev_default_xlate,
598 	VDEV_TYPE_SPARE,	/* name of this vdev type */
599 	B_FALSE			/* not a leaf vdev */
600 };
601