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.
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.
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]
23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2012, 2020 by Delphix. All rights reserved.
25 * Copyright (c) 2016 Gvozden Nešković. All rights reserved.
28 #include <sys/zfs_context.h>
30 #include <sys/vdev_impl.h>
32 #include <sys/zio_checksum.h>
34 #include <sys/fs/zfs.h>
35 #include <sys/fm/fs/zfs.h>
36 #include <sys/vdev_raidz.h>
37 #include <sys/vdev_raidz_impl.h>
38 #include <sys/vdev_draid.h>
41 #include <sys/vdev.h> /* For vdev_xlate() in vdev_raidz_io_verify() */
45 * Virtual device vector for RAID-Z.
47 * This vdev supports single, double, and triple parity. For single parity,
48 * we use a simple XOR of all the data columns. For double or triple parity,
49 * we use a special case of Reed-Solomon coding. This extends the
50 * technique described in "The mathematics of RAID-6" by H. Peter Anvin by
51 * drawing on the system described in "A Tutorial on Reed-Solomon Coding for
52 * Fault-Tolerance in RAID-like Systems" by James S. Plank on which the
53 * former is also based. The latter is designed to provide higher performance
56 * Note that the Plank paper claimed to support arbitrary N+M, but was then
57 * amended six years later identifying a critical flaw that invalidates its
58 * claims. Nevertheless, the technique can be adapted to work for up to
59 * triple parity. For additional parity, the amendment "Note: Correction to
60 * the 1997 Tutorial on Reed-Solomon Coding" by James S. Plank and Ying Ding
61 * is viable, but the additional complexity means that write performance will
64 * All of the methods above operate on a Galois field, defined over the
65 * integers mod 2^N. In our case we choose N=8 for GF(8) so that all elements
66 * can be expressed with a single byte. Briefly, the operations on the
67 * field are defined as follows:
69 * o addition (+) is represented by a bitwise XOR
70 * o subtraction (-) is therefore identical to addition: A + B = A - B
71 * o multiplication of A by 2 is defined by the following bitwise expression:
76 * (A * 2)_4 = A_3 + A_7
77 * (A * 2)_3 = A_2 + A_7
78 * (A * 2)_2 = A_1 + A_7
82 * In C, multiplying by 2 is therefore ((a << 1) ^ ((a & 0x80) ? 0x1d : 0)).
83 * As an aside, this multiplication is derived from the error correcting
84 * primitive polynomial x^8 + x^4 + x^3 + x^2 + 1.
86 * Observe that any number in the field (except for 0) can be expressed as a
87 * power of 2 -- a generator for the field. We store a table of the powers of
88 * 2 and logs base 2 for quick look ups, and exploit the fact that A * B can
89 * be rewritten as 2^(log_2(A) + log_2(B)) (where '+' is normal addition rather
90 * than field addition). The inverse of a field element A (A^-1) is therefore
91 * A ^ (255 - 1) = A^254.
93 * The up-to-three parity columns, P, Q, R over several data columns,
94 * D_0, ... D_n-1, can be expressed by field operations:
96 * P = D_0 + D_1 + ... + D_n-2 + D_n-1
97 * Q = 2^n-1 * D_0 + 2^n-2 * D_1 + ... + 2^1 * D_n-2 + 2^0 * D_n-1
98 * = ((...((D_0) * 2 + D_1) * 2 + ...) * 2 + D_n-2) * 2 + D_n-1
99 * R = 4^n-1 * D_0 + 4^n-2 * D_1 + ... + 4^1 * D_n-2 + 4^0 * D_n-1
100 * = ((...((D_0) * 4 + D_1) * 4 + ...) * 4 + D_n-2) * 4 + D_n-1
102 * We chose 1, 2, and 4 as our generators because 1 corresponds to the trivial
103 * XOR operation, and 2 and 4 can be computed quickly and generate linearly-
104 * independent coefficients. (There are no additional coefficients that have
105 * this property which is why the uncorrected Plank method breaks down.)
107 * See the reconstruction code below for how P, Q and R can used individually
108 * or in concert to recover missing data columns.
111 #define VDEV_RAIDZ_P 0
112 #define VDEV_RAIDZ_Q 1
113 #define VDEV_RAIDZ_R 2
115 #define VDEV_RAIDZ_MUL_2(x) (((x) << 1) ^ (((x) & 0x80) ? 0x1d : 0))
116 #define VDEV_RAIDZ_MUL_4(x) (VDEV_RAIDZ_MUL_2(VDEV_RAIDZ_MUL_2(x)))
119 * We provide a mechanism to perform the field multiplication operation on a
120 * 64-bit value all at once rather than a byte at a time. This works by
121 * creating a mask from the top bit in each byte and using that to
122 * conditionally apply the XOR of 0x1d.
124 #define VDEV_RAIDZ_64MUL_2(x, mask) \
126 (mask) = (x) & 0x8080808080808080ULL; \
127 (mask) = ((mask) << 1) - ((mask) >> 7); \
128 (x) = (((x) << 1) & 0xfefefefefefefefeULL) ^ \
129 ((mask) & 0x1d1d1d1d1d1d1d1dULL); \
132 #define VDEV_RAIDZ_64MUL_4(x, mask) \
134 VDEV_RAIDZ_64MUL_2((x), mask); \
135 VDEV_RAIDZ_64MUL_2((x), mask); \
139 vdev_raidz_row_free(raidz_row_t *rr)
141 for (int c = 0; c < rr->rr_cols; c++) {
142 raidz_col_t *rc = &rr->rr_col[c];
144 if (rc->rc_size != 0)
145 abd_free(rc->rc_abd);
146 if (rc->rc_orig_data != NULL)
147 abd_free(rc->rc_orig_data);
150 if (rr->rr_abd_empty != NULL)
151 abd_free(rr->rr_abd_empty);
153 kmem_free(rr, offsetof(raidz_row_t, rr_col[rr->rr_scols]));
157 vdev_raidz_map_free(raidz_map_t *rm)
159 for (int i = 0; i < rm->rm_nrows; i++)
160 vdev_raidz_row_free(rm->rm_row[i]);
162 kmem_free(rm, offsetof(raidz_map_t, rm_row[rm->rm_nrows]));
166 vdev_raidz_map_free_vsd(zio_t *zio)
168 raidz_map_t *rm = zio->io_vsd;
170 vdev_raidz_map_free(rm);
173 const zio_vsd_ops_t vdev_raidz_vsd_ops = {
174 .vsd_free = vdev_raidz_map_free_vsd,
178 vdev_raidz_map_alloc_write(zio_t *zio, raidz_map_t *rm, uint64_t ashift)
183 raidz_row_t *rr = rm->rm_row[0];
185 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
186 ASSERT3U(rm->rm_nrows, ==, 1);
189 * Pad any parity columns with additional space to account for skip
192 if (rm->rm_skipstart < rr->rr_firstdatacol) {
193 ASSERT0(rm->rm_skipstart);
194 nwrapped = rm->rm_nskip;
195 } else if (rr->rr_scols < (rm->rm_skipstart + rm->rm_nskip)) {
197 (rm->rm_skipstart + rm->rm_nskip) % rr->rr_scols;
201 * Optional single skip sectors (rc_size == 0) will be handled in
202 * vdev_raidz_io_start_write().
204 int skipped = rr->rr_scols - rr->rr_cols;
206 /* Allocate buffers for the parity columns */
207 for (c = 0; c < rr->rr_firstdatacol; c++) {
208 raidz_col_t *rc = &rr->rr_col[c];
211 * Parity columns will pad out a linear ABD to account for
212 * the skip sector. A linear ABD is used here because
213 * parity calculations use the ABD buffer directly to calculate
214 * parity. This avoids doing a memcpy back to the ABD after the
215 * parity has been calculated. By issuing the parity column
216 * with the skip sector we can reduce contention on the child
217 * VDEV queue locks (vq_lock).
220 rc->rc_abd = abd_alloc_linear(
221 rc->rc_size + (1ULL << ashift), B_FALSE);
222 abd_zero_off(rc->rc_abd, rc->rc_size, 1ULL << ashift);
225 rc->rc_abd = abd_alloc_linear(rc->rc_size, B_FALSE);
229 for (off = 0; c < rr->rr_cols; c++) {
230 raidz_col_t *rc = &rr->rr_col[c];
231 abd_t *abd = abd_get_offset_struct(&rc->rc_abdstruct,
232 zio->io_abd, off, rc->rc_size);
235 * Generate I/O for skip sectors to improve aggregation
236 * continuity. We will use gang ABD's to reduce contention
237 * on the child VDEV queue locks (vq_lock) by issuing
238 * a single I/O that contains the data and skip sector.
240 * It is important to make sure that rc_size is not updated
241 * even though we are adding a skip sector to the ABD. When
242 * calculating the parity in vdev_raidz_generate_parity_row()
243 * the rc_size is used to iterate through the ABD's. We can
244 * not have zero'd out skip sectors used for calculating
245 * parity for raidz, because those same sectors are not used
246 * during reconstruction.
248 if (c >= rm->rm_skipstart && skipped < rm->rm_nskip) {
249 rc->rc_abd = abd_alloc_gang();
250 abd_gang_add(rc->rc_abd, abd, B_TRUE);
251 abd_gang_add(rc->rc_abd,
252 abd_get_zeros(1ULL << ashift), B_TRUE);
260 ASSERT3U(off, ==, zio->io_size);
261 ASSERT3S(skipped, ==, rm->rm_nskip);
265 vdev_raidz_map_alloc_read(zio_t *zio, raidz_map_t *rm)
268 raidz_row_t *rr = rm->rm_row[0];
270 ASSERT3U(rm->rm_nrows, ==, 1);
272 /* Allocate buffers for the parity columns */
273 for (c = 0; c < rr->rr_firstdatacol; c++)
274 rr->rr_col[c].rc_abd =
275 abd_alloc_linear(rr->rr_col[c].rc_size, B_FALSE);
277 for (uint64_t off = 0; c < rr->rr_cols; c++) {
278 raidz_col_t *rc = &rr->rr_col[c];
279 rc->rc_abd = abd_get_offset_struct(&rc->rc_abdstruct,
280 zio->io_abd, off, rc->rc_size);
286 * Divides the IO evenly across all child vdevs; usually, dcols is
287 * the number of children in the target vdev.
289 * Avoid inlining the function to keep vdev_raidz_io_start(), which
290 * is this functions only caller, as small as possible on the stack.
292 noinline raidz_map_t *
293 vdev_raidz_map_alloc(zio_t *zio, uint64_t ashift, uint64_t dcols,
297 /* The starting RAIDZ (parent) vdev sector of the block. */
298 uint64_t b = zio->io_offset >> ashift;
299 /* The zio's size in units of the vdev's minimum sector size. */
300 uint64_t s = zio->io_size >> ashift;
301 /* The first column for this stripe. */
302 uint64_t f = b % dcols;
303 /* The starting byte offset on each child vdev. */
304 uint64_t o = (b / dcols) << ashift;
305 uint64_t q, r, c, bc, col, acols, scols, coff, devidx, asize, tot;
308 kmem_zalloc(offsetof(raidz_map_t, rm_row[1]), KM_SLEEP);
312 * "Quotient": The number of data sectors for this stripe on all but
313 * the "big column" child vdevs that also contain "remainder" data.
315 q = s / (dcols - nparity);
318 * "Remainder": The number of partial stripe data sectors in this I/O.
319 * This will add a sector to some, but not all, child vdevs.
321 r = s - q * (dcols - nparity);
323 /* The number of "big columns" - those which contain remainder data. */
324 bc = (r == 0 ? 0 : r + nparity);
327 * The total number of data and parity sectors associated with
330 tot = s + nparity * (q + (r == 0 ? 0 : 1));
333 * acols: The columns that will be accessed.
334 * scols: The columns that will be accessed or skipped.
337 /* Our I/O request doesn't span all child vdevs. */
339 scols = MIN(dcols, roundup(bc, nparity + 1));
345 ASSERT3U(acols, <=, scols);
347 rr = kmem_alloc(offsetof(raidz_row_t, rr_col[scols]), KM_SLEEP);
351 rr->rr_scols = scols;
353 rr->rr_missingdata = 0;
354 rr->rr_missingparity = 0;
355 rr->rr_firstdatacol = nparity;
356 rr->rr_abd_empty = NULL;
359 rr->rr_offset = zio->io_offset;
360 rr->rr_size = zio->io_size;
365 for (c = 0; c < scols; c++) {
366 raidz_col_t *rc = &rr->rr_col[c];
371 coff += 1ULL << ashift;
374 rc->rc_offset = coff;
376 rc->rc_orig_data = NULL;
380 rc->rc_force_repair = 0;
381 rc->rc_allow_repair = 1;
382 rc->rc_need_orig_restore = B_FALSE;
387 rc->rc_size = (q + 1) << ashift;
389 rc->rc_size = q << ashift;
391 asize += rc->rc_size;
394 ASSERT3U(asize, ==, tot << ashift);
395 rm->rm_nskip = roundup(tot, nparity + 1) - tot;
396 rm->rm_skipstart = bc;
399 * If all data stored spans all columns, there's a danger that parity
400 * will always be on the same device and, since parity isn't read
401 * during normal operation, that device's I/O bandwidth won't be
402 * used effectively. We therefore switch the parity every 1MB.
404 * ... at least that was, ostensibly, the theory. As a practical
405 * matter unless we juggle the parity between all devices evenly, we
406 * won't see any benefit. Further, occasional writes that aren't a
407 * multiple of the LCM of the number of children and the minimum
408 * stripe width are sufficient to avoid pessimal behavior.
409 * Unfortunately, this decision created an implicit on-disk format
410 * requirement that we need to support for all eternity, but only
411 * for single-parity RAID-Z.
413 * If we intend to skip a sector in the zeroth column for padding
414 * we must make sure to note this swap. We will never intend to
415 * skip the first column since at least one data and one parity
416 * column must appear in each row.
418 ASSERT(rr->rr_cols >= 2);
419 ASSERT(rr->rr_col[0].rc_size == rr->rr_col[1].rc_size);
421 if (rr->rr_firstdatacol == 1 && (zio->io_offset & (1ULL << 20))) {
422 devidx = rr->rr_col[0].rc_devidx;
423 o = rr->rr_col[0].rc_offset;
424 rr->rr_col[0].rc_devidx = rr->rr_col[1].rc_devidx;
425 rr->rr_col[0].rc_offset = rr->rr_col[1].rc_offset;
426 rr->rr_col[1].rc_devidx = devidx;
427 rr->rr_col[1].rc_offset = o;
429 if (rm->rm_skipstart == 0)
430 rm->rm_skipstart = 1;
433 if (zio->io_type == ZIO_TYPE_WRITE) {
434 vdev_raidz_map_alloc_write(zio, rm, ashift);
436 vdev_raidz_map_alloc_read(zio, rm);
439 /* init RAIDZ parity ops */
440 rm->rm_ops = vdev_raidz_math_get_ops();
452 vdev_raidz_p_func(void *buf, size_t size, void *private)
454 struct pqr_struct *pqr = private;
455 const uint64_t *src = buf;
456 int i, cnt = size / sizeof (src[0]);
458 ASSERT(pqr->p && !pqr->q && !pqr->r);
460 for (i = 0; i < cnt; i++, src++, pqr->p++)
467 vdev_raidz_pq_func(void *buf, size_t size, void *private)
469 struct pqr_struct *pqr = private;
470 const uint64_t *src = buf;
472 int i, cnt = size / sizeof (src[0]);
474 ASSERT(pqr->p && pqr->q && !pqr->r);
476 for (i = 0; i < cnt; i++, src++, pqr->p++, pqr->q++) {
478 VDEV_RAIDZ_64MUL_2(*pqr->q, mask);
486 vdev_raidz_pqr_func(void *buf, size_t size, void *private)
488 struct pqr_struct *pqr = private;
489 const uint64_t *src = buf;
491 int i, cnt = size / sizeof (src[0]);
493 ASSERT(pqr->p && pqr->q && pqr->r);
495 for (i = 0; i < cnt; i++, src++, pqr->p++, pqr->q++, pqr->r++) {
497 VDEV_RAIDZ_64MUL_2(*pqr->q, mask);
499 VDEV_RAIDZ_64MUL_4(*pqr->r, mask);
507 vdev_raidz_generate_parity_p(raidz_row_t *rr)
509 uint64_t *p = abd_to_buf(rr->rr_col[VDEV_RAIDZ_P].rc_abd);
511 for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) {
512 abd_t *src = rr->rr_col[c].rc_abd;
514 if (c == rr->rr_firstdatacol) {
515 abd_copy_to_buf(p, src, rr->rr_col[c].rc_size);
517 struct pqr_struct pqr = { p, NULL, NULL };
518 (void) abd_iterate_func(src, 0, rr->rr_col[c].rc_size,
519 vdev_raidz_p_func, &pqr);
525 vdev_raidz_generate_parity_pq(raidz_row_t *rr)
527 uint64_t *p = abd_to_buf(rr->rr_col[VDEV_RAIDZ_P].rc_abd);
528 uint64_t *q = abd_to_buf(rr->rr_col[VDEV_RAIDZ_Q].rc_abd);
529 uint64_t pcnt = rr->rr_col[VDEV_RAIDZ_P].rc_size / sizeof (p[0]);
530 ASSERT(rr->rr_col[VDEV_RAIDZ_P].rc_size ==
531 rr->rr_col[VDEV_RAIDZ_Q].rc_size);
533 for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) {
534 abd_t *src = rr->rr_col[c].rc_abd;
536 uint64_t ccnt = rr->rr_col[c].rc_size / sizeof (p[0]);
538 if (c == rr->rr_firstdatacol) {
539 ASSERT(ccnt == pcnt || ccnt == 0);
540 abd_copy_to_buf(p, src, rr->rr_col[c].rc_size);
541 (void) memcpy(q, p, rr->rr_col[c].rc_size);
543 for (uint64_t i = ccnt; i < pcnt; i++) {
548 struct pqr_struct pqr = { p, q, NULL };
550 ASSERT(ccnt <= pcnt);
551 (void) abd_iterate_func(src, 0, rr->rr_col[c].rc_size,
552 vdev_raidz_pq_func, &pqr);
555 * Treat short columns as though they are full of 0s.
556 * Note that there's therefore nothing needed for P.
559 for (uint64_t i = ccnt; i < pcnt; i++) {
560 VDEV_RAIDZ_64MUL_2(q[i], mask);
567 vdev_raidz_generate_parity_pqr(raidz_row_t *rr)
569 uint64_t *p = abd_to_buf(rr->rr_col[VDEV_RAIDZ_P].rc_abd);
570 uint64_t *q = abd_to_buf(rr->rr_col[VDEV_RAIDZ_Q].rc_abd);
571 uint64_t *r = abd_to_buf(rr->rr_col[VDEV_RAIDZ_R].rc_abd);
572 uint64_t pcnt = rr->rr_col[VDEV_RAIDZ_P].rc_size / sizeof (p[0]);
573 ASSERT(rr->rr_col[VDEV_RAIDZ_P].rc_size ==
574 rr->rr_col[VDEV_RAIDZ_Q].rc_size);
575 ASSERT(rr->rr_col[VDEV_RAIDZ_P].rc_size ==
576 rr->rr_col[VDEV_RAIDZ_R].rc_size);
578 for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) {
579 abd_t *src = rr->rr_col[c].rc_abd;
581 uint64_t ccnt = rr->rr_col[c].rc_size / sizeof (p[0]);
583 if (c == rr->rr_firstdatacol) {
584 ASSERT(ccnt == pcnt || ccnt == 0);
585 abd_copy_to_buf(p, src, rr->rr_col[c].rc_size);
586 (void) memcpy(q, p, rr->rr_col[c].rc_size);
587 (void) memcpy(r, p, rr->rr_col[c].rc_size);
589 for (uint64_t i = ccnt; i < pcnt; i++) {
595 struct pqr_struct pqr = { p, q, r };
597 ASSERT(ccnt <= pcnt);
598 (void) abd_iterate_func(src, 0, rr->rr_col[c].rc_size,
599 vdev_raidz_pqr_func, &pqr);
602 * Treat short columns as though they are full of 0s.
603 * Note that there's therefore nothing needed for P.
606 for (uint64_t i = ccnt; i < pcnt; i++) {
607 VDEV_RAIDZ_64MUL_2(q[i], mask);
608 VDEV_RAIDZ_64MUL_4(r[i], mask);
615 * Generate RAID parity in the first virtual columns according to the number of
616 * parity columns available.
619 vdev_raidz_generate_parity_row(raidz_map_t *rm, raidz_row_t *rr)
621 ASSERT3U(rr->rr_cols, !=, 0);
623 /* Generate using the new math implementation */
624 if (vdev_raidz_math_generate(rm, rr) != RAIDZ_ORIGINAL_IMPL)
627 switch (rr->rr_firstdatacol) {
629 vdev_raidz_generate_parity_p(rr);
632 vdev_raidz_generate_parity_pq(rr);
635 vdev_raidz_generate_parity_pqr(rr);
638 cmn_err(CE_PANIC, "invalid RAID-Z configuration");
643 vdev_raidz_generate_parity(raidz_map_t *rm)
645 for (int i = 0; i < rm->rm_nrows; i++) {
646 raidz_row_t *rr = rm->rm_row[i];
647 vdev_raidz_generate_parity_row(rm, rr);
652 vdev_raidz_reconst_p_func(void *dbuf, void *sbuf, size_t size, void *private)
655 uint64_t *dst = dbuf;
656 uint64_t *src = sbuf;
657 int cnt = size / sizeof (src[0]);
659 for (int i = 0; i < cnt; i++) {
667 vdev_raidz_reconst_q_pre_func(void *dbuf, void *sbuf, size_t size,
671 uint64_t *dst = dbuf;
672 uint64_t *src = sbuf;
674 int cnt = size / sizeof (dst[0]);
676 for (int i = 0; i < cnt; i++, dst++, src++) {
677 VDEV_RAIDZ_64MUL_2(*dst, mask);
685 vdev_raidz_reconst_q_pre_tail_func(void *buf, size_t size, void *private)
690 int cnt = size / sizeof (dst[0]);
692 for (int i = 0; i < cnt; i++, dst++) {
693 /* same operation as vdev_raidz_reconst_q_pre_func() on dst */
694 VDEV_RAIDZ_64MUL_2(*dst, mask);
700 struct reconst_q_struct {
706 vdev_raidz_reconst_q_post_func(void *buf, size_t size, void *private)
708 struct reconst_q_struct *rq = private;
710 int cnt = size / sizeof (dst[0]);
712 for (int i = 0; i < cnt; i++, dst++, rq->q++) {
717 for (j = 0, b = (uint8_t *)dst; j < 8; j++, b++) {
718 *b = vdev_raidz_exp2(*b, rq->exp);
725 struct reconst_pq_struct {
735 vdev_raidz_reconst_pq_func(void *xbuf, void *ybuf, size_t size, void *private)
737 struct reconst_pq_struct *rpq = private;
741 for (int i = 0; i < size;
742 i++, rpq->p++, rpq->q++, rpq->pxy++, rpq->qxy++, xd++, yd++) {
743 *xd = vdev_raidz_exp2(*rpq->p ^ *rpq->pxy, rpq->aexp) ^
744 vdev_raidz_exp2(*rpq->q ^ *rpq->qxy, rpq->bexp);
745 *yd = *rpq->p ^ *rpq->pxy ^ *xd;
752 vdev_raidz_reconst_pq_tail_func(void *xbuf, size_t size, void *private)
754 struct reconst_pq_struct *rpq = private;
757 for (int i = 0; i < size;
758 i++, rpq->p++, rpq->q++, rpq->pxy++, rpq->qxy++, xd++) {
759 /* same operation as vdev_raidz_reconst_pq_func() on xd */
760 *xd = vdev_raidz_exp2(*rpq->p ^ *rpq->pxy, rpq->aexp) ^
761 vdev_raidz_exp2(*rpq->q ^ *rpq->qxy, rpq->bexp);
768 vdev_raidz_reconstruct_p(raidz_row_t *rr, int *tgts, int ntgts)
773 ASSERT3U(ntgts, ==, 1);
774 ASSERT3U(x, >=, rr->rr_firstdatacol);
775 ASSERT3U(x, <, rr->rr_cols);
777 ASSERT3U(rr->rr_col[x].rc_size, <=, rr->rr_col[VDEV_RAIDZ_P].rc_size);
779 src = rr->rr_col[VDEV_RAIDZ_P].rc_abd;
780 dst = rr->rr_col[x].rc_abd;
782 abd_copy_from_buf(dst, abd_to_buf(src), rr->rr_col[x].rc_size);
784 for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) {
785 uint64_t size = MIN(rr->rr_col[x].rc_size,
786 rr->rr_col[c].rc_size);
788 src = rr->rr_col[c].rc_abd;
793 (void) abd_iterate_func2(dst, src, 0, 0, size,
794 vdev_raidz_reconst_p_func, NULL);
799 vdev_raidz_reconstruct_q(raidz_row_t *rr, int *tgts, int ntgts)
807 ASSERT(rr->rr_col[x].rc_size <= rr->rr_col[VDEV_RAIDZ_Q].rc_size);
809 for (c = rr->rr_firstdatacol; c < rr->rr_cols; c++) {
810 uint64_t size = (c == x) ? 0 : MIN(rr->rr_col[x].rc_size,
811 rr->rr_col[c].rc_size);
813 src = rr->rr_col[c].rc_abd;
814 dst = rr->rr_col[x].rc_abd;
816 if (c == rr->rr_firstdatacol) {
817 abd_copy(dst, src, size);
818 if (rr->rr_col[x].rc_size > size) {
819 abd_zero_off(dst, size,
820 rr->rr_col[x].rc_size - size);
823 ASSERT3U(size, <=, rr->rr_col[x].rc_size);
824 (void) abd_iterate_func2(dst, src, 0, 0, size,
825 vdev_raidz_reconst_q_pre_func, NULL);
826 (void) abd_iterate_func(dst,
827 size, rr->rr_col[x].rc_size - size,
828 vdev_raidz_reconst_q_pre_tail_func, NULL);
832 src = rr->rr_col[VDEV_RAIDZ_Q].rc_abd;
833 dst = rr->rr_col[x].rc_abd;
834 exp = 255 - (rr->rr_cols - 1 - x);
836 struct reconst_q_struct rq = { abd_to_buf(src), exp };
837 (void) abd_iterate_func(dst, 0, rr->rr_col[x].rc_size,
838 vdev_raidz_reconst_q_post_func, &rq);
842 vdev_raidz_reconstruct_pq(raidz_row_t *rr, int *tgts, int ntgts)
844 uint8_t *p, *q, *pxy, *qxy, tmp, a, b, aexp, bexp;
845 abd_t *pdata, *qdata;
846 uint64_t xsize, ysize;
853 ASSERT(x >= rr->rr_firstdatacol);
854 ASSERT(y < rr->rr_cols);
856 ASSERT(rr->rr_col[x].rc_size >= rr->rr_col[y].rc_size);
859 * Move the parity data aside -- we're going to compute parity as
860 * though columns x and y were full of zeros -- Pxy and Qxy. We want to
861 * reuse the parity generation mechanism without trashing the actual
862 * parity so we make those columns appear to be full of zeros by
863 * setting their lengths to zero.
865 pdata = rr->rr_col[VDEV_RAIDZ_P].rc_abd;
866 qdata = rr->rr_col[VDEV_RAIDZ_Q].rc_abd;
867 xsize = rr->rr_col[x].rc_size;
868 ysize = rr->rr_col[y].rc_size;
870 rr->rr_col[VDEV_RAIDZ_P].rc_abd =
871 abd_alloc_linear(rr->rr_col[VDEV_RAIDZ_P].rc_size, B_TRUE);
872 rr->rr_col[VDEV_RAIDZ_Q].rc_abd =
873 abd_alloc_linear(rr->rr_col[VDEV_RAIDZ_Q].rc_size, B_TRUE);
874 rr->rr_col[x].rc_size = 0;
875 rr->rr_col[y].rc_size = 0;
877 vdev_raidz_generate_parity_pq(rr);
879 rr->rr_col[x].rc_size = xsize;
880 rr->rr_col[y].rc_size = ysize;
882 p = abd_to_buf(pdata);
883 q = abd_to_buf(qdata);
884 pxy = abd_to_buf(rr->rr_col[VDEV_RAIDZ_P].rc_abd);
885 qxy = abd_to_buf(rr->rr_col[VDEV_RAIDZ_Q].rc_abd);
886 xd = rr->rr_col[x].rc_abd;
887 yd = rr->rr_col[y].rc_abd;
891 * Pxy = P + D_x + D_y
892 * Qxy = Q + 2^(ndevs - 1 - x) * D_x + 2^(ndevs - 1 - y) * D_y
894 * We can then solve for D_x:
895 * D_x = A * (P + Pxy) + B * (Q + Qxy)
897 * A = 2^(x - y) * (2^(x - y) + 1)^-1
898 * B = 2^(ndevs - 1 - x) * (2^(x - y) + 1)^-1
900 * With D_x in hand, we can easily solve for D_y:
901 * D_y = P + Pxy + D_x
904 a = vdev_raidz_pow2[255 + x - y];
905 b = vdev_raidz_pow2[255 - (rr->rr_cols - 1 - x)];
906 tmp = 255 - vdev_raidz_log2[a ^ 1];
908 aexp = vdev_raidz_log2[vdev_raidz_exp2(a, tmp)];
909 bexp = vdev_raidz_log2[vdev_raidz_exp2(b, tmp)];
911 ASSERT3U(xsize, >=, ysize);
912 struct reconst_pq_struct rpq = { p, q, pxy, qxy, aexp, bexp };
914 (void) abd_iterate_func2(xd, yd, 0, 0, ysize,
915 vdev_raidz_reconst_pq_func, &rpq);
916 (void) abd_iterate_func(xd, ysize, xsize - ysize,
917 vdev_raidz_reconst_pq_tail_func, &rpq);
919 abd_free(rr->rr_col[VDEV_RAIDZ_P].rc_abd);
920 abd_free(rr->rr_col[VDEV_RAIDZ_Q].rc_abd);
923 * Restore the saved parity data.
925 rr->rr_col[VDEV_RAIDZ_P].rc_abd = pdata;
926 rr->rr_col[VDEV_RAIDZ_Q].rc_abd = qdata;
930 * In the general case of reconstruction, we must solve the system of linear
931 * equations defined by the coefficients used to generate parity as well as
932 * the contents of the data and parity disks. This can be expressed with
933 * vectors for the original data (D) and the actual data (d) and parity (p)
934 * and a matrix composed of the identity matrix (I) and a dispersal matrix (V):
938 * | V | | D_0 | | p_m-1 |
939 * | | x | : | = | d_0 |
940 * | I | | D_n-1 | | : |
941 * | | ~~ ~~ | d_n-1 |
944 * I is simply a square identity matrix of size n, and V is a vandermonde
945 * matrix defined by the coefficients we chose for the various parity columns
946 * (1, 2, 4). Note that these values were chosen both for simplicity, speedy
947 * computation as well as linear separability.
950 * | 1 .. 1 1 1 | | p_0 |
951 * | 2^n-1 .. 4 2 1 | __ __ | : |
952 * | 4^n-1 .. 16 4 1 | | D_0 | | p_m-1 |
953 * | 1 .. 0 0 0 | | D_1 | | d_0 |
954 * | 0 .. 0 0 0 | x | D_2 | = | d_1 |
955 * | : : : : | | : | | d_2 |
956 * | 0 .. 1 0 0 | | D_n-1 | | : |
957 * | 0 .. 0 1 0 | ~~ ~~ | : |
958 * | 0 .. 0 0 1 | | d_n-1 |
961 * Note that I, V, d, and p are known. To compute D, we must invert the
962 * matrix and use the known data and parity values to reconstruct the unknown
963 * data values. We begin by removing the rows in V|I and d|p that correspond
964 * to failed or missing columns; we then make V|I square (n x n) and d|p
965 * sized n by removing rows corresponding to unused parity from the bottom up
966 * to generate (V|I)' and (d|p)'. We can then generate the inverse of (V|I)'
967 * using Gauss-Jordan elimination. In the example below we use m=3 parity
968 * columns, n=8 data columns, with errors in d_1, d_2, and p_1:
970 * | 1 1 1 1 1 1 1 1 |
971 * | 128 64 32 16 8 4 2 1 | <-----+-+-- missing disks
972 * | 19 205 116 29 64 16 4 1 | / /
973 * | 1 0 0 0 0 0 0 0 | / /
974 * | 0 1 0 0 0 0 0 0 | <--' /
975 * (V|I) = | 0 0 1 0 0 0 0 0 | <---'
976 * | 0 0 0 1 0 0 0 0 |
977 * | 0 0 0 0 1 0 0 0 |
978 * | 0 0 0 0 0 1 0 0 |
979 * | 0 0 0 0 0 0 1 0 |
980 * | 0 0 0 0 0 0 0 1 |
983 * | 1 1 1 1 1 1 1 1 |
984 * | 128 64 32 16 8 4 2 1 |
985 * | 19 205 116 29 64 16 4 1 |
986 * | 1 0 0 0 0 0 0 0 |
987 * | 0 1 0 0 0 0 0 0 |
988 * (V|I)' = | 0 0 1 0 0 0 0 0 |
989 * | 0 0 0 1 0 0 0 0 |
990 * | 0 0 0 0 1 0 0 0 |
991 * | 0 0 0 0 0 1 0 0 |
992 * | 0 0 0 0 0 0 1 0 |
993 * | 0 0 0 0 0 0 0 1 |
996 * Here we employ Gauss-Jordan elimination to find the inverse of (V|I)'. We
997 * have carefully chosen the seed values 1, 2, and 4 to ensure that this
998 * matrix is not singular.
1000 * | 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 |
1001 * | 19 205 116 29 64 16 4 1 0 1 0 0 0 0 0 0 |
1002 * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 |
1003 * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 |
1004 * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 |
1005 * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 |
1006 * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 |
1007 * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 |
1010 * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 |
1011 * | 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 |
1012 * | 19 205 116 29 64 16 4 1 0 1 0 0 0 0 0 0 |
1013 * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 |
1014 * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 |
1015 * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 |
1016 * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 |
1017 * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 |
1020 * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 |
1021 * | 0 1 1 0 0 0 0 0 1 0 1 1 1 1 1 1 |
1022 * | 0 205 116 0 0 0 0 0 0 1 19 29 64 16 4 1 |
1023 * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 |
1024 * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 |
1025 * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 |
1026 * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 |
1027 * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 |
1030 * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 |
1031 * | 0 1 1 0 0 0 0 0 1 0 1 1 1 1 1 1 |
1032 * | 0 0 185 0 0 0 0 0 205 1 222 208 141 221 201 204 |
1033 * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 |
1034 * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 |
1035 * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 |
1036 * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 |
1037 * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 |
1040 * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 |
1041 * | 0 1 1 0 0 0 0 0 1 0 1 1 1 1 1 1 |
1042 * | 0 0 1 0 0 0 0 0 166 100 4 40 158 168 216 209 |
1043 * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 |
1044 * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 |
1045 * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 |
1046 * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 |
1047 * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 |
1050 * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 |
1051 * | 0 1 0 0 0 0 0 0 167 100 5 41 159 169 217 208 |
1052 * | 0 0 1 0 0 0 0 0 166 100 4 40 158 168 216 209 |
1053 * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 |
1054 * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 |
1055 * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 |
1056 * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 |
1057 * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 |
1060 * | 0 0 1 0 0 0 0 0 |
1061 * | 167 100 5 41 159 169 217 208 |
1062 * | 166 100 4 40 158 168 216 209 |
1063 * (V|I)'^-1 = | 0 0 0 1 0 0 0 0 |
1064 * | 0 0 0 0 1 0 0 0 |
1065 * | 0 0 0 0 0 1 0 0 |
1066 * | 0 0 0 0 0 0 1 0 |
1067 * | 0 0 0 0 0 0 0 1 |
1070 * We can then simply compute D = (V|I)'^-1 x (d|p)' to discover the values
1071 * of the missing data.
1073 * As is apparent from the example above, the only non-trivial rows in the
1074 * inverse matrix correspond to the data disks that we're trying to
1075 * reconstruct. Indeed, those are the only rows we need as the others would
1076 * only be useful for reconstructing data known or assumed to be valid. For
1077 * that reason, we only build the coefficients in the rows that correspond to
1082 vdev_raidz_matrix_init(raidz_row_t *rr, int n, int nmap, int *map,
1088 ASSERT(n == rr->rr_cols - rr->rr_firstdatacol);
1091 * Fill in the missing rows of interest.
1093 for (i = 0; i < nmap; i++) {
1094 ASSERT3S(0, <=, map[i]);
1095 ASSERT3S(map[i], <=, 2);
1102 for (j = 0; j < n; j++) {
1106 rows[i][j] = vdev_raidz_pow2[pow];
1112 vdev_raidz_matrix_invert(raidz_row_t *rr, int n, int nmissing, int *missing,
1113 uint8_t **rows, uint8_t **invrows, const uint8_t *used)
1119 * Assert that the first nmissing entries from the array of used
1120 * columns correspond to parity columns and that subsequent entries
1121 * correspond to data columns.
1123 for (i = 0; i < nmissing; i++) {
1124 ASSERT3S(used[i], <, rr->rr_firstdatacol);
1126 for (; i < n; i++) {
1127 ASSERT3S(used[i], >=, rr->rr_firstdatacol);
1131 * First initialize the storage where we'll compute the inverse rows.
1133 for (i = 0; i < nmissing; i++) {
1134 for (j = 0; j < n; j++) {
1135 invrows[i][j] = (i == j) ? 1 : 0;
1140 * Subtract all trivial rows from the rows of consequence.
1142 for (i = 0; i < nmissing; i++) {
1143 for (j = nmissing; j < n; j++) {
1144 ASSERT3U(used[j], >=, rr->rr_firstdatacol);
1145 jj = used[j] - rr->rr_firstdatacol;
1147 invrows[i][j] = rows[i][jj];
1153 * For each of the rows of interest, we must normalize it and subtract
1154 * a multiple of it from the other rows.
1156 for (i = 0; i < nmissing; i++) {
1157 for (j = 0; j < missing[i]; j++) {
1158 ASSERT0(rows[i][j]);
1160 ASSERT3U(rows[i][missing[i]], !=, 0);
1163 * Compute the inverse of the first element and multiply each
1164 * element in the row by that value.
1166 log = 255 - vdev_raidz_log2[rows[i][missing[i]]];
1168 for (j = 0; j < n; j++) {
1169 rows[i][j] = vdev_raidz_exp2(rows[i][j], log);
1170 invrows[i][j] = vdev_raidz_exp2(invrows[i][j], log);
1173 for (ii = 0; ii < nmissing; ii++) {
1177 ASSERT3U(rows[ii][missing[i]], !=, 0);
1179 log = vdev_raidz_log2[rows[ii][missing[i]]];
1181 for (j = 0; j < n; j++) {
1183 vdev_raidz_exp2(rows[i][j], log);
1185 vdev_raidz_exp2(invrows[i][j], log);
1191 * Verify that the data that is left in the rows are properly part of
1192 * an identity matrix.
1194 for (i = 0; i < nmissing; i++) {
1195 for (j = 0; j < n; j++) {
1196 if (j == missing[i]) {
1197 ASSERT3U(rows[i][j], ==, 1);
1199 ASSERT0(rows[i][j]);
1206 vdev_raidz_matrix_reconstruct(raidz_row_t *rr, int n, int nmissing,
1207 int *missing, uint8_t **invrows, const uint8_t *used)
1212 uint8_t *dst[VDEV_RAIDZ_MAXPARITY] = { NULL };
1213 uint64_t dcount[VDEV_RAIDZ_MAXPARITY] = { 0 };
1217 uint8_t *invlog[VDEV_RAIDZ_MAXPARITY];
1221 psize = sizeof (invlog[0][0]) * n * nmissing;
1222 p = kmem_alloc(psize, KM_SLEEP);
1224 for (pp = p, i = 0; i < nmissing; i++) {
1229 for (i = 0; i < nmissing; i++) {
1230 for (j = 0; j < n; j++) {
1231 ASSERT3U(invrows[i][j], !=, 0);
1232 invlog[i][j] = vdev_raidz_log2[invrows[i][j]];
1236 for (i = 0; i < n; i++) {
1238 ASSERT3U(c, <, rr->rr_cols);
1240 ccount = rr->rr_col[c].rc_size;
1241 ASSERT(ccount >= rr->rr_col[missing[0]].rc_size || i > 0);
1244 src = abd_to_buf(rr->rr_col[c].rc_abd);
1245 for (j = 0; j < nmissing; j++) {
1246 cc = missing[j] + rr->rr_firstdatacol;
1247 ASSERT3U(cc, >=, rr->rr_firstdatacol);
1248 ASSERT3U(cc, <, rr->rr_cols);
1249 ASSERT3U(cc, !=, c);
1251 dcount[j] = rr->rr_col[cc].rc_size;
1253 dst[j] = abd_to_buf(rr->rr_col[cc].rc_abd);
1256 for (x = 0; x < ccount; x++, src++) {
1258 log = vdev_raidz_log2[*src];
1260 for (cc = 0; cc < nmissing; cc++) {
1261 if (x >= dcount[cc])
1267 if ((ll = log + invlog[cc][i]) >= 255)
1269 val = vdev_raidz_pow2[ll];
1280 kmem_free(p, psize);
1284 vdev_raidz_reconstruct_general(raidz_row_t *rr, int *tgts, int ntgts)
1288 int missing_rows[VDEV_RAIDZ_MAXPARITY];
1289 int parity_map[VDEV_RAIDZ_MAXPARITY];
1292 uint8_t *rows[VDEV_RAIDZ_MAXPARITY];
1293 uint8_t *invrows[VDEV_RAIDZ_MAXPARITY];
1296 abd_t **bufs = NULL;
1299 * Matrix reconstruction can't use scatter ABDs yet, so we allocate
1300 * temporary linear ABDs if any non-linear ABDs are found.
1302 for (i = rr->rr_firstdatacol; i < rr->rr_cols; i++) {
1303 if (!abd_is_linear(rr->rr_col[i].rc_abd)) {
1304 bufs = kmem_alloc(rr->rr_cols * sizeof (abd_t *),
1307 for (c = rr->rr_firstdatacol; c < rr->rr_cols; c++) {
1308 raidz_col_t *col = &rr->rr_col[c];
1310 bufs[c] = col->rc_abd;
1311 if (bufs[c] != NULL) {
1312 col->rc_abd = abd_alloc_linear(
1313 col->rc_size, B_TRUE);
1314 abd_copy(col->rc_abd, bufs[c],
1323 n = rr->rr_cols - rr->rr_firstdatacol;
1326 * Figure out which data columns are missing.
1329 for (t = 0; t < ntgts; t++) {
1330 if (tgts[t] >= rr->rr_firstdatacol) {
1331 missing_rows[nmissing_rows++] =
1332 tgts[t] - rr->rr_firstdatacol;
1337 * Figure out which parity columns to use to help generate the missing
1340 for (tt = 0, c = 0, i = 0; i < nmissing_rows; c++) {
1342 ASSERT(c < rr->rr_firstdatacol);
1345 * Skip any targeted parity columns.
1347 if (c == tgts[tt]) {
1356 psize = (sizeof (rows[0][0]) + sizeof (invrows[0][0])) *
1357 nmissing_rows * n + sizeof (used[0]) * n;
1358 p = kmem_alloc(psize, KM_SLEEP);
1360 for (pp = p, i = 0; i < nmissing_rows; i++) {
1368 for (i = 0; i < nmissing_rows; i++) {
1369 used[i] = parity_map[i];
1372 for (tt = 0, c = rr->rr_firstdatacol; c < rr->rr_cols; c++) {
1373 if (tt < nmissing_rows &&
1374 c == missing_rows[tt] + rr->rr_firstdatacol) {
1385 * Initialize the interesting rows of the matrix.
1387 vdev_raidz_matrix_init(rr, n, nmissing_rows, parity_map, rows);
1390 * Invert the matrix.
1392 vdev_raidz_matrix_invert(rr, n, nmissing_rows, missing_rows, rows,
1396 * Reconstruct the missing data using the generated matrix.
1398 vdev_raidz_matrix_reconstruct(rr, n, nmissing_rows, missing_rows,
1401 kmem_free(p, psize);
1404 * copy back from temporary linear abds and free them
1407 for (c = rr->rr_firstdatacol; c < rr->rr_cols; c++) {
1408 raidz_col_t *col = &rr->rr_col[c];
1410 if (bufs[c] != NULL) {
1411 abd_copy(bufs[c], col->rc_abd, col->rc_size);
1412 abd_free(col->rc_abd);
1414 col->rc_abd = bufs[c];
1416 kmem_free(bufs, rr->rr_cols * sizeof (abd_t *));
1421 vdev_raidz_reconstruct_row(raidz_map_t *rm, raidz_row_t *rr,
1422 const int *t, int nt)
1424 int tgts[VDEV_RAIDZ_MAXPARITY], *dt;
1427 int nbadparity, nbaddata;
1428 int parity_valid[VDEV_RAIDZ_MAXPARITY];
1430 nbadparity = rr->rr_firstdatacol;
1431 nbaddata = rr->rr_cols - nbadparity;
1433 for (i = 0, c = 0; c < rr->rr_cols; c++) {
1434 if (c < rr->rr_firstdatacol)
1435 parity_valid[c] = B_FALSE;
1437 if (i < nt && c == t[i]) {
1440 } else if (rr->rr_col[c].rc_error != 0) {
1442 } else if (c >= rr->rr_firstdatacol) {
1445 parity_valid[c] = B_TRUE;
1450 ASSERT(ntgts >= nt);
1451 ASSERT(nbaddata >= 0);
1452 ASSERT(nbaddata + nbadparity == ntgts);
1454 dt = &tgts[nbadparity];
1456 /* Reconstruct using the new math implementation */
1457 ret = vdev_raidz_math_reconstruct(rm, rr, parity_valid, dt, nbaddata);
1458 if (ret != RAIDZ_ORIGINAL_IMPL)
1462 * See if we can use any of our optimized reconstruction routines.
1466 if (parity_valid[VDEV_RAIDZ_P]) {
1467 vdev_raidz_reconstruct_p(rr, dt, 1);
1471 ASSERT(rr->rr_firstdatacol > 1);
1473 if (parity_valid[VDEV_RAIDZ_Q]) {
1474 vdev_raidz_reconstruct_q(rr, dt, 1);
1478 ASSERT(rr->rr_firstdatacol > 2);
1482 ASSERT(rr->rr_firstdatacol > 1);
1484 if (parity_valid[VDEV_RAIDZ_P] &&
1485 parity_valid[VDEV_RAIDZ_Q]) {
1486 vdev_raidz_reconstruct_pq(rr, dt, 2);
1490 ASSERT(rr->rr_firstdatacol > 2);
1495 vdev_raidz_reconstruct_general(rr, tgts, ntgts);
1499 vdev_raidz_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize,
1500 uint64_t *logical_ashift, uint64_t *physical_ashift)
1502 vdev_raidz_t *vdrz = vd->vdev_tsd;
1503 uint64_t nparity = vdrz->vd_nparity;
1508 ASSERT(nparity > 0);
1510 if (nparity > VDEV_RAIDZ_MAXPARITY ||
1511 vd->vdev_children < nparity + 1) {
1512 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
1513 return (SET_ERROR(EINVAL));
1516 vdev_open_children(vd);
1518 for (c = 0; c < vd->vdev_children; c++) {
1519 vdev_t *cvd = vd->vdev_child[c];
1521 if (cvd->vdev_open_error != 0) {
1522 lasterror = cvd->vdev_open_error;
1527 *asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1;
1528 *max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1;
1529 *logical_ashift = MAX(*logical_ashift, cvd->vdev_ashift);
1531 for (c = 0; c < vd->vdev_children; c++) {
1532 vdev_t *cvd = vd->vdev_child[c];
1534 if (cvd->vdev_open_error != 0)
1536 *physical_ashift = vdev_best_ashift(*logical_ashift,
1537 *physical_ashift, cvd->vdev_physical_ashift);
1540 *asize *= vd->vdev_children;
1541 *max_asize *= vd->vdev_children;
1543 if (numerrors > nparity) {
1544 vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS;
1552 vdev_raidz_close(vdev_t *vd)
1554 for (int c = 0; c < vd->vdev_children; c++) {
1555 if (vd->vdev_child[c] != NULL)
1556 vdev_close(vd->vdev_child[c]);
1561 vdev_raidz_asize(vdev_t *vd, uint64_t psize)
1563 vdev_raidz_t *vdrz = vd->vdev_tsd;
1565 uint64_t ashift = vd->vdev_top->vdev_ashift;
1566 uint64_t cols = vdrz->vd_logical_width;
1567 uint64_t nparity = vdrz->vd_nparity;
1569 asize = ((psize - 1) >> ashift) + 1;
1570 asize += nparity * ((asize + cols - nparity - 1) / (cols - nparity));
1571 asize = roundup(asize, nparity + 1) << ashift;
1577 * The allocatable space for a raidz vdev is N * sizeof(smallest child)
1578 * so each child must provide at least 1/Nth of its asize.
1581 vdev_raidz_min_asize(vdev_t *vd)
1583 return ((vd->vdev_min_asize + vd->vdev_children - 1) /
1588 vdev_raidz_child_done(zio_t *zio)
1590 raidz_col_t *rc = zio->io_private;
1592 ASSERT3P(rc->rc_abd, !=, NULL);
1593 rc->rc_error = zio->io_error;
1599 vdev_raidz_io_verify(vdev_t *vd, raidz_row_t *rr, int col)
1602 vdev_t *tvd = vd->vdev_top;
1604 range_seg64_t logical_rs, physical_rs, remain_rs;
1605 logical_rs.rs_start = rr->rr_offset;
1606 logical_rs.rs_end = logical_rs.rs_start +
1607 vdev_raidz_asize(vd, rr->rr_size);
1609 raidz_col_t *rc = &rr->rr_col[col];
1610 vdev_t *cvd = vd->vdev_child[rc->rc_devidx];
1612 vdev_xlate(cvd, &logical_rs, &physical_rs, &remain_rs);
1613 ASSERT(vdev_xlate_is_empty(&remain_rs));
1614 ASSERT3U(rc->rc_offset, ==, physical_rs.rs_start);
1615 ASSERT3U(rc->rc_offset, <, physical_rs.rs_end);
1617 * It would be nice to assert that rs_end is equal
1618 * to rc_offset + rc_size but there might be an
1619 * optional I/O at the end that is not accounted in
1622 if (physical_rs.rs_end > rc->rc_offset + rc->rc_size) {
1623 ASSERT3U(physical_rs.rs_end, ==, rc->rc_offset +
1624 rc->rc_size + (1 << tvd->vdev_ashift));
1626 ASSERT3U(physical_rs.rs_end, ==, rc->rc_offset + rc->rc_size);
1632 vdev_raidz_io_start_write(zio_t *zio, raidz_row_t *rr, uint64_t ashift)
1634 vdev_t *vd = zio->io_vd;
1635 raidz_map_t *rm = zio->io_vsd;
1637 vdev_raidz_generate_parity_row(rm, rr);
1639 for (int c = 0; c < rr->rr_scols; c++) {
1640 raidz_col_t *rc = &rr->rr_col[c];
1641 vdev_t *cvd = vd->vdev_child[rc->rc_devidx];
1643 /* Verify physical to logical translation */
1644 vdev_raidz_io_verify(vd, rr, c);
1646 if (rc->rc_size > 0) {
1647 ASSERT3P(rc->rc_abd, !=, NULL);
1648 zio_nowait(zio_vdev_child_io(zio, NULL, cvd,
1649 rc->rc_offset, rc->rc_abd,
1650 abd_get_size(rc->rc_abd), zio->io_type,
1651 zio->io_priority, 0, vdev_raidz_child_done, rc));
1654 * Generate optional write for skip sector to improve
1655 * aggregation contiguity.
1657 ASSERT3P(rc->rc_abd, ==, NULL);
1658 zio_nowait(zio_vdev_child_io(zio, NULL, cvd,
1659 rc->rc_offset, NULL, 1ULL << ashift,
1660 zio->io_type, zio->io_priority,
1661 ZIO_FLAG_NODATA | ZIO_FLAG_OPTIONAL, NULL,
1668 vdev_raidz_io_start_read(zio_t *zio, raidz_row_t *rr)
1670 vdev_t *vd = zio->io_vd;
1673 * Iterate over the columns in reverse order so that we hit the parity
1674 * last -- any errors along the way will force us to read the parity.
1676 for (int c = rr->rr_cols - 1; c >= 0; c--) {
1677 raidz_col_t *rc = &rr->rr_col[c];
1678 if (rc->rc_size == 0)
1680 vdev_t *cvd = vd->vdev_child[rc->rc_devidx];
1681 if (!vdev_readable(cvd)) {
1682 if (c >= rr->rr_firstdatacol)
1683 rr->rr_missingdata++;
1685 rr->rr_missingparity++;
1686 rc->rc_error = SET_ERROR(ENXIO);
1687 rc->rc_tried = 1; /* don't even try */
1691 if (vdev_dtl_contains(cvd, DTL_MISSING, zio->io_txg, 1)) {
1692 if (c >= rr->rr_firstdatacol)
1693 rr->rr_missingdata++;
1695 rr->rr_missingparity++;
1696 rc->rc_error = SET_ERROR(ESTALE);
1700 if (c >= rr->rr_firstdatacol || rr->rr_missingdata > 0 ||
1701 (zio->io_flags & (ZIO_FLAG_SCRUB | ZIO_FLAG_RESILVER))) {
1702 zio_nowait(zio_vdev_child_io(zio, NULL, cvd,
1703 rc->rc_offset, rc->rc_abd, rc->rc_size,
1704 zio->io_type, zio->io_priority, 0,
1705 vdev_raidz_child_done, rc));
1711 * Start an IO operation on a RAIDZ VDev
1714 * - For write operations:
1715 * 1. Generate the parity data
1716 * 2. Create child zio write operations to each column's vdev, for both
1718 * 3. If the column skips any sectors for padding, create optional dummy
1719 * write zio children for those areas to improve aggregation continuity.
1720 * - For read operations:
1721 * 1. Create child zio read operations to each data column's vdev to read
1722 * the range of data required for zio.
1723 * 2. If this is a scrub or resilver operation, or if any of the data
1724 * vdevs have had errors, then create zio read operations to the parity
1725 * columns' VDevs as well.
1728 vdev_raidz_io_start(zio_t *zio)
1730 vdev_t *vd = zio->io_vd;
1731 vdev_t *tvd = vd->vdev_top;
1732 vdev_raidz_t *vdrz = vd->vdev_tsd;
1734 raidz_map_t *rm = vdev_raidz_map_alloc(zio, tvd->vdev_ashift,
1735 vdrz->vd_logical_width, vdrz->vd_nparity);
1737 zio->io_vsd_ops = &vdev_raidz_vsd_ops;
1740 * Until raidz expansion is implemented all maps for a raidz vdev
1741 * contain a single row.
1743 ASSERT3U(rm->rm_nrows, ==, 1);
1744 raidz_row_t *rr = rm->rm_row[0];
1746 if (zio->io_type == ZIO_TYPE_WRITE) {
1747 vdev_raidz_io_start_write(zio, rr, tvd->vdev_ashift);
1749 ASSERT(zio->io_type == ZIO_TYPE_READ);
1750 vdev_raidz_io_start_read(zio, rr);
1757 * Report a checksum error for a child of a RAID-Z device.
1760 vdev_raidz_checksum_error(zio_t *zio, raidz_col_t *rc, abd_t *bad_data)
1762 vdev_t *vd = zio->io_vd->vdev_child[rc->rc_devidx];
1764 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE) &&
1765 zio->io_priority != ZIO_PRIORITY_REBUILD) {
1766 zio_bad_cksum_t zbc;
1767 raidz_map_t *rm = zio->io_vsd;
1769 zbc.zbc_has_cksum = 0;
1770 zbc.zbc_injected = rm->rm_ecksuminjected;
1772 (void) zfs_ereport_post_checksum(zio->io_spa, vd,
1773 &zio->io_bookmark, zio, rc->rc_offset, rc->rc_size,
1774 rc->rc_abd, bad_data, &zbc);
1775 mutex_enter(&vd->vdev_stat_lock);
1776 vd->vdev_stat.vs_checksum_errors++;
1777 mutex_exit(&vd->vdev_stat_lock);
1782 * We keep track of whether or not there were any injected errors, so that
1783 * any ereports we generate can note it.
1786 raidz_checksum_verify(zio_t *zio)
1788 zio_bad_cksum_t zbc = {{{0}}};
1789 raidz_map_t *rm = zio->io_vsd;
1791 int ret = zio_checksum_error(zio, &zbc);
1792 if (ret != 0 && zbc.zbc_injected != 0)
1793 rm->rm_ecksuminjected = 1;
1799 * Generate the parity from the data columns. If we tried and were able to
1800 * read the parity without error, verify that the generated parity matches the
1801 * data we read. If it doesn't, we fire off a checksum error. Return the
1802 * number of such failures.
1805 raidz_parity_verify(zio_t *zio, raidz_row_t *rr)
1807 abd_t *orig[VDEV_RAIDZ_MAXPARITY];
1809 raidz_map_t *rm = zio->io_vsd;
1812 blkptr_t *bp = zio->io_bp;
1813 enum zio_checksum checksum = (bp == NULL ? zio->io_prop.zp_checksum :
1814 (BP_IS_GANG(bp) ? ZIO_CHECKSUM_GANG_HEADER : BP_GET_CHECKSUM(bp)));
1816 if (checksum == ZIO_CHECKSUM_NOPARITY)
1819 for (c = 0; c < rr->rr_firstdatacol; c++) {
1820 rc = &rr->rr_col[c];
1821 if (!rc->rc_tried || rc->rc_error != 0)
1824 orig[c] = rc->rc_abd;
1825 ASSERT3U(abd_get_size(rc->rc_abd), ==, rc->rc_size);
1826 rc->rc_abd = abd_alloc_linear(rc->rc_size, B_FALSE);
1830 * Verify any empty sectors are zero filled to ensure the parity
1831 * is calculated correctly even if these non-data sectors are damaged.
1833 if (rr->rr_nempty && rr->rr_abd_empty != NULL)
1834 ret += vdev_draid_map_verify_empty(zio, rr);
1837 * Regenerates parity even for !tried||rc_error!=0 columns. This
1838 * isn't harmful but it does have the side effect of fixing stuff
1839 * we didn't realize was necessary (i.e. even if we return 0).
1841 vdev_raidz_generate_parity_row(rm, rr);
1843 for (c = 0; c < rr->rr_firstdatacol; c++) {
1844 rc = &rr->rr_col[c];
1846 if (!rc->rc_tried || rc->rc_error != 0)
1849 if (abd_cmp(orig[c], rc->rc_abd) != 0) {
1850 vdev_raidz_checksum_error(zio, rc, orig[c]);
1851 rc->rc_error = SET_ERROR(ECKSUM);
1861 vdev_raidz_worst_error(raidz_row_t *rr)
1865 for (int c = 0; c < rr->rr_cols; c++)
1866 error = zio_worst_error(error, rr->rr_col[c].rc_error);
1872 vdev_raidz_io_done_verified(zio_t *zio, raidz_row_t *rr)
1874 int unexpected_errors = 0;
1875 int parity_errors = 0;
1876 int parity_untried = 0;
1877 int data_errors = 0;
1879 ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ);
1881 for (int c = 0; c < rr->rr_cols; c++) {
1882 raidz_col_t *rc = &rr->rr_col[c];
1885 if (c < rr->rr_firstdatacol)
1890 if (!rc->rc_skipped)
1891 unexpected_errors++;
1892 } else if (c < rr->rr_firstdatacol && !rc->rc_tried) {
1896 if (rc->rc_force_repair)
1897 unexpected_errors++;
1901 * If we read more parity disks than were used for
1902 * reconstruction, confirm that the other parity disks produced
1905 * Note that we also regenerate parity when resilvering so we
1906 * can write it out to failed devices later.
1908 if (parity_errors + parity_untried <
1909 rr->rr_firstdatacol - data_errors ||
1910 (zio->io_flags & ZIO_FLAG_RESILVER)) {
1911 int n = raidz_parity_verify(zio, rr);
1912 unexpected_errors += n;
1915 if (zio->io_error == 0 && spa_writeable(zio->io_spa) &&
1916 (unexpected_errors > 0 || (zio->io_flags & ZIO_FLAG_RESILVER))) {
1918 * Use the good data we have in hand to repair damaged children.
1920 for (int c = 0; c < rr->rr_cols; c++) {
1921 raidz_col_t *rc = &rr->rr_col[c];
1922 vdev_t *vd = zio->io_vd;
1923 vdev_t *cvd = vd->vdev_child[rc->rc_devidx];
1925 if (!rc->rc_allow_repair) {
1927 } else if (!rc->rc_force_repair &&
1928 (rc->rc_error == 0 || rc->rc_size == 0)) {
1932 zio_nowait(zio_vdev_child_io(zio, NULL, cvd,
1933 rc->rc_offset, rc->rc_abd, rc->rc_size,
1935 zio->io_priority == ZIO_PRIORITY_REBUILD ?
1936 ZIO_PRIORITY_REBUILD : ZIO_PRIORITY_ASYNC_WRITE,
1937 ZIO_FLAG_IO_REPAIR | (unexpected_errors ?
1938 ZIO_FLAG_SELF_HEAL : 0), NULL, NULL));
1944 raidz_restore_orig_data(raidz_map_t *rm)
1946 for (int i = 0; i < rm->rm_nrows; i++) {
1947 raidz_row_t *rr = rm->rm_row[i];
1948 for (int c = 0; c < rr->rr_cols; c++) {
1949 raidz_col_t *rc = &rr->rr_col[c];
1950 if (rc->rc_need_orig_restore) {
1951 abd_copy(rc->rc_abd,
1952 rc->rc_orig_data, rc->rc_size);
1953 rc->rc_need_orig_restore = B_FALSE;
1960 * returns EINVAL if reconstruction of the block will not be possible
1961 * returns ECKSUM if this specific reconstruction failed
1962 * returns 0 on successful reconstruction
1965 raidz_reconstruct(zio_t *zio, int *ltgts, int ntgts, int nparity)
1967 raidz_map_t *rm = zio->io_vsd;
1969 /* Reconstruct each row */
1970 for (int r = 0; r < rm->rm_nrows; r++) {
1971 raidz_row_t *rr = rm->rm_row[r];
1972 int my_tgts[VDEV_RAIDZ_MAXPARITY]; /* value is child id */
1977 for (int c = 0; c < rr->rr_cols; c++) {
1978 raidz_col_t *rc = &rr->rr_col[c];
1979 ASSERT0(rc->rc_need_orig_restore);
1980 if (rc->rc_error != 0) {
1986 if (rc->rc_size == 0)
1988 for (int lt = 0; lt < ntgts; lt++) {
1989 if (rc->rc_devidx == ltgts[lt]) {
1990 if (rc->rc_orig_data == NULL) {
1993 rc->rc_size, B_TRUE);
1994 abd_copy(rc->rc_orig_data,
1995 rc->rc_abd, rc->rc_size);
1997 rc->rc_need_orig_restore = B_TRUE;
2007 if (dead > nparity) {
2008 /* reconstruction not possible */
2009 raidz_restore_orig_data(rm);
2013 vdev_raidz_reconstruct_row(rm, rr, my_tgts, t);
2016 /* Check for success */
2017 if (raidz_checksum_verify(zio) == 0) {
2019 /* Reconstruction succeeded - report errors */
2020 for (int i = 0; i < rm->rm_nrows; i++) {
2021 raidz_row_t *rr = rm->rm_row[i];
2023 for (int c = 0; c < rr->rr_cols; c++) {
2024 raidz_col_t *rc = &rr->rr_col[c];
2025 if (rc->rc_need_orig_restore) {
2027 * Note: if this is a parity column,
2028 * we don't really know if it's wrong.
2030 * vdev_raidz_io_done_verified() check
2031 * it, and if we set rc_error, it will
2032 * think that it is a "known" error
2033 * that doesn't need to be checked
2036 if (rc->rc_error == 0 &&
2037 c >= rr->rr_firstdatacol) {
2038 vdev_raidz_checksum_error(zio,
2039 rc, rc->rc_orig_data);
2043 rc->rc_need_orig_restore = B_FALSE;
2047 vdev_raidz_io_done_verified(zio, rr);
2050 zio_checksum_verified(zio);
2055 /* Reconstruction failed - restore original data */
2056 raidz_restore_orig_data(rm);
2061 * Iterate over all combinations of N bad vdevs and attempt a reconstruction.
2062 * Note that the algorithm below is non-optimal because it doesn't take into
2063 * account how reconstruction is actually performed. For example, with
2064 * triple-parity RAID-Z the reconstruction procedure is the same if column 4
2065 * is targeted as invalid as if columns 1 and 4 are targeted since in both
2066 * cases we'd only use parity information in column 0.
2068 * The order that we find the various possible combinations of failed
2069 * disks is dictated by these rules:
2070 * - Examine each "slot" (the "i" in tgts[i])
2071 * - Try to increment this slot (tgts[i] = tgts[i] + 1)
2072 * - if we can't increment because it runs into the next slot,
2073 * reset our slot to the minimum, and examine the next slot
2075 * For example, with a 6-wide RAIDZ3, and no known errors (so we have to choose
2076 * 3 columns to reconstruct), we will generate the following sequence:
2079 * 0 1 2 special case: skip since these are all parity
2080 * 0 1 3 first slot: reset to 0; middle slot: increment to 2
2081 * 0 2 3 first slot: increment to 1
2082 * 1 2 3 first: reset to 0; middle: reset to 1; last: increment to 4
2083 * 0 1 4 first: reset to 0; middle: increment to 2
2084 * 0 2 4 first: increment to 1
2085 * 1 2 4 first: reset to 0; middle: increment to 3
2086 * 0 3 4 first: increment to 1
2087 * 1 3 4 first: increment to 2
2088 * 2 3 4 first: reset to 0; middle: reset to 1; last: increment to 5
2089 * 0 1 5 first: reset to 0; middle: increment to 2
2090 * 0 2 5 first: increment to 1
2091 * 1 2 5 first: reset to 0; middle: increment to 3
2092 * 0 3 5 first: increment to 1
2093 * 1 3 5 first: increment to 2
2094 * 2 3 5 first: reset to 0; middle: increment to 4
2095 * 0 4 5 first: increment to 1
2096 * 1 4 5 first: increment to 2
2097 * 2 4 5 first: increment to 3
2100 * This strategy works for dRAID but is less efficient when there are a large
2101 * number of child vdevs and therefore permutations to check. Furthermore,
2102 * since the raidz_map_t rows likely do not overlap reconstruction would be
2103 * possible as long as there are no more than nparity data errors per row.
2104 * These additional permutations are not currently checked but could be as
2105 * a future improvement.
2108 vdev_raidz_combrec(zio_t *zio)
2110 int nparity = vdev_get_nparity(zio->io_vd);
2111 raidz_map_t *rm = zio->io_vsd;
2113 /* Check if there's enough data to attempt reconstrution. */
2114 for (int i = 0; i < rm->rm_nrows; i++) {
2115 raidz_row_t *rr = rm->rm_row[i];
2116 int total_errors = 0;
2118 for (int c = 0; c < rr->rr_cols; c++) {
2119 if (rr->rr_col[c].rc_error)
2123 if (total_errors > nparity)
2124 return (vdev_raidz_worst_error(rr));
2127 for (int num_failures = 1; num_failures <= nparity; num_failures++) {
2128 int tstore[VDEV_RAIDZ_MAXPARITY + 2];
2129 int *ltgts = &tstore[1]; /* value is logical child ID */
2131 /* Determine number of logical children, n */
2132 int n = zio->io_vd->vdev_children;
2134 ASSERT3U(num_failures, <=, nparity);
2135 ASSERT3U(num_failures, <=, VDEV_RAIDZ_MAXPARITY);
2137 /* Handle corner cases in combrec logic */
2139 for (int i = 0; i < num_failures; i++) {
2142 ltgts[num_failures] = n;
2145 int err = raidz_reconstruct(zio, ltgts, num_failures,
2147 if (err == EINVAL) {
2149 * Reconstruction not possible with this #
2150 * failures; try more failures.
2153 } else if (err == 0)
2156 /* Compute next targets to try */
2157 for (int t = 0; ; t++) {
2158 ASSERT3U(t, <, num_failures);
2160 if (ltgts[t] == n) {
2161 /* try more failures */
2162 ASSERT3U(t, ==, num_failures - 1);
2166 ASSERT3U(ltgts[t], <, n);
2167 ASSERT3U(ltgts[t], <=, ltgts[t + 1]);
2170 * If that spot is available, we're done here.
2171 * Try the next combination.
2173 if (ltgts[t] != ltgts[t + 1])
2177 * Otherwise, reset this tgt to the minimum,
2178 * and move on to the next tgt.
2180 ltgts[t] = ltgts[t - 1] + 1;
2181 ASSERT3U(ltgts[t], ==, t);
2184 /* Increase the number of failures and keep trying. */
2185 if (ltgts[num_failures - 1] == n)
2194 vdev_raidz_reconstruct(raidz_map_t *rm, const int *t, int nt)
2196 for (uint64_t row = 0; row < rm->rm_nrows; row++) {
2197 raidz_row_t *rr = rm->rm_row[row];
2198 vdev_raidz_reconstruct_row(rm, rr, t, nt);
2203 * Complete a write IO operation on a RAIDZ VDev
2206 * 1. Check for errors on the child IOs.
2207 * 2. Return, setting an error code if too few child VDevs were written
2208 * to reconstruct the data later. Note that partial writes are
2209 * considered successful if they can be reconstructed at all.
2212 vdev_raidz_io_done_write_impl(zio_t *zio, raidz_row_t *rr)
2214 int total_errors = 0;
2216 ASSERT3U(rr->rr_missingparity, <=, rr->rr_firstdatacol);
2217 ASSERT3U(rr->rr_missingdata, <=, rr->rr_cols - rr->rr_firstdatacol);
2218 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
2220 for (int c = 0; c < rr->rr_cols; c++) {
2221 raidz_col_t *rc = &rr->rr_col[c];
2224 ASSERT(rc->rc_error != ECKSUM); /* child has no bp */
2231 * Treat partial writes as a success. If we couldn't write enough
2232 * columns to reconstruct the data, the I/O failed. Otherwise,
2235 * Now that we support write reallocation, it would be better
2236 * to treat partial failure as real failure unless there are
2237 * no non-degraded top-level vdevs left, and not update DTLs
2238 * if we intend to reallocate.
2240 if (total_errors > rr->rr_firstdatacol) {
2241 zio->io_error = zio_worst_error(zio->io_error,
2242 vdev_raidz_worst_error(rr));
2247 vdev_raidz_io_done_reconstruct_known_missing(zio_t *zio, raidz_map_t *rm,
2250 int parity_errors = 0;
2251 int parity_untried = 0;
2252 int data_errors = 0;
2253 int total_errors = 0;
2255 ASSERT3U(rr->rr_missingparity, <=, rr->rr_firstdatacol);
2256 ASSERT3U(rr->rr_missingdata, <=, rr->rr_cols - rr->rr_firstdatacol);
2257 ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ);
2259 for (int c = 0; c < rr->rr_cols; c++) {
2260 raidz_col_t *rc = &rr->rr_col[c];
2263 * If scrubbing and a replacing/sparing child vdev determined
2264 * that not all of its children have an identical copy of the
2265 * data, then clear the error so the column is treated like
2266 * any other read and force a repair to correct the damage.
2268 if (rc->rc_error == ECKSUM) {
2269 ASSERT(zio->io_flags & ZIO_FLAG_SCRUB);
2270 vdev_raidz_checksum_error(zio, rc, rc->rc_abd);
2271 rc->rc_force_repair = 1;
2276 if (c < rr->rr_firstdatacol)
2282 } else if (c < rr->rr_firstdatacol && !rc->rc_tried) {
2288 * If there were data errors and the number of errors we saw was
2289 * correctable -- less than or equal to the number of parity disks read
2290 * -- reconstruct based on the missing data.
2292 if (data_errors != 0 &&
2293 total_errors <= rr->rr_firstdatacol - parity_untried) {
2295 * We either attempt to read all the parity columns or
2296 * none of them. If we didn't try to read parity, we
2297 * wouldn't be here in the correctable case. There must
2298 * also have been fewer parity errors than parity
2299 * columns or, again, we wouldn't be in this code path.
2301 ASSERT(parity_untried == 0);
2302 ASSERT(parity_errors < rr->rr_firstdatacol);
2305 * Identify the data columns that reported an error.
2308 int tgts[VDEV_RAIDZ_MAXPARITY];
2309 for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) {
2310 raidz_col_t *rc = &rr->rr_col[c];
2311 if (rc->rc_error != 0) {
2312 ASSERT(n < VDEV_RAIDZ_MAXPARITY);
2317 ASSERT(rr->rr_firstdatacol >= n);
2319 vdev_raidz_reconstruct_row(rm, rr, tgts, n);
2324 * Return the number of reads issued.
2327 vdev_raidz_read_all(zio_t *zio, raidz_row_t *rr)
2329 vdev_t *vd = zio->io_vd;
2332 rr->rr_missingdata = 0;
2333 rr->rr_missingparity = 0;
2336 * If this rows contains empty sectors which are not required
2337 * for a normal read then allocate an ABD for them now so they
2338 * may be read, verified, and any needed repairs performed.
2340 if (rr->rr_nempty && rr->rr_abd_empty == NULL)
2341 vdev_draid_map_alloc_empty(zio, rr);
2343 for (int c = 0; c < rr->rr_cols; c++) {
2344 raidz_col_t *rc = &rr->rr_col[c];
2345 if (rc->rc_tried || rc->rc_size == 0)
2348 zio_nowait(zio_vdev_child_io(zio, NULL,
2349 vd->vdev_child[rc->rc_devidx],
2350 rc->rc_offset, rc->rc_abd, rc->rc_size,
2351 zio->io_type, zio->io_priority, 0,
2352 vdev_raidz_child_done, rc));
2359 * We're here because either there were too many errors to even attempt
2360 * reconstruction (total_errors == rm_first_datacol), or vdev_*_combrec()
2361 * failed. In either case, there is enough bad data to prevent reconstruction.
2362 * Start checksum ereports for all children which haven't failed.
2365 vdev_raidz_io_done_unrecoverable(zio_t *zio)
2367 raidz_map_t *rm = zio->io_vsd;
2369 for (int i = 0; i < rm->rm_nrows; i++) {
2370 raidz_row_t *rr = rm->rm_row[i];
2372 for (int c = 0; c < rr->rr_cols; c++) {
2373 raidz_col_t *rc = &rr->rr_col[c];
2374 vdev_t *cvd = zio->io_vd->vdev_child[rc->rc_devidx];
2376 if (rc->rc_error != 0)
2379 zio_bad_cksum_t zbc;
2380 zbc.zbc_has_cksum = 0;
2381 zbc.zbc_injected = rm->rm_ecksuminjected;
2383 (void) zfs_ereport_start_checksum(zio->io_spa,
2384 cvd, &zio->io_bookmark, zio, rc->rc_offset,
2386 mutex_enter(&cvd->vdev_stat_lock);
2387 cvd->vdev_stat.vs_checksum_errors++;
2388 mutex_exit(&cvd->vdev_stat_lock);
2394 vdev_raidz_io_done(zio_t *zio)
2396 raidz_map_t *rm = zio->io_vsd;
2398 if (zio->io_type == ZIO_TYPE_WRITE) {
2399 for (int i = 0; i < rm->rm_nrows; i++) {
2400 vdev_raidz_io_done_write_impl(zio, rm->rm_row[i]);
2403 for (int i = 0; i < rm->rm_nrows; i++) {
2404 raidz_row_t *rr = rm->rm_row[i];
2405 vdev_raidz_io_done_reconstruct_known_missing(zio,
2409 if (raidz_checksum_verify(zio) == 0) {
2410 for (int i = 0; i < rm->rm_nrows; i++) {
2411 raidz_row_t *rr = rm->rm_row[i];
2412 vdev_raidz_io_done_verified(zio, rr);
2414 zio_checksum_verified(zio);
2417 * A sequential resilver has no checksum which makes
2418 * combinatoral reconstruction impossible. This code
2419 * path is unreachable since raidz_checksum_verify()
2420 * has no checksum to verify and must succeed.
2422 ASSERT3U(zio->io_priority, !=, ZIO_PRIORITY_REBUILD);
2425 * This isn't a typical situation -- either we got a
2426 * read error or a child silently returned bad data.
2427 * Read every block so we can try again with as much
2428 * data and parity as we can track down. If we've
2429 * already been through once before, all children will
2430 * be marked as tried so we'll proceed to combinatorial
2434 for (int i = 0; i < rm->rm_nrows; i++) {
2435 nread += vdev_raidz_read_all(zio,
2440 * Normally our stage is VDEV_IO_DONE, but if
2441 * we've already called redone(), it will have
2442 * changed to VDEV_IO_START, in which case we
2443 * don't want to call redone() again.
2445 if (zio->io_stage != ZIO_STAGE_VDEV_IO_START)
2446 zio_vdev_io_redone(zio);
2450 zio->io_error = vdev_raidz_combrec(zio);
2451 if (zio->io_error == ECKSUM &&
2452 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2453 vdev_raidz_io_done_unrecoverable(zio);
2460 vdev_raidz_state_change(vdev_t *vd, int faulted, int degraded)
2462 vdev_raidz_t *vdrz = vd->vdev_tsd;
2463 if (faulted > vdrz->vd_nparity)
2464 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2465 VDEV_AUX_NO_REPLICAS);
2466 else if (degraded + faulted != 0)
2467 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE);
2469 vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE);
2473 * Determine if any portion of the provided block resides on a child vdev
2474 * with a dirty DTL and therefore needs to be resilvered. The function
2475 * assumes that at least one DTL is dirty which implies that full stripe
2476 * width blocks must be resilvered.
2479 vdev_raidz_need_resilver(vdev_t *vd, const dva_t *dva, size_t psize,
2480 uint64_t phys_birth)
2482 vdev_raidz_t *vdrz = vd->vdev_tsd;
2483 uint64_t dcols = vd->vdev_children;
2484 uint64_t nparity = vdrz->vd_nparity;
2485 uint64_t ashift = vd->vdev_top->vdev_ashift;
2486 /* The starting RAIDZ (parent) vdev sector of the block. */
2487 uint64_t b = DVA_GET_OFFSET(dva) >> ashift;
2488 /* The zio's size in units of the vdev's minimum sector size. */
2489 uint64_t s = ((psize - 1) >> ashift) + 1;
2490 /* The first column for this stripe. */
2491 uint64_t f = b % dcols;
2493 /* Unreachable by sequential resilver. */
2494 ASSERT3U(phys_birth, !=, TXG_UNKNOWN);
2496 if (!vdev_dtl_contains(vd, DTL_PARTIAL, phys_birth, 1))
2499 if (s + nparity >= dcols)
2502 for (uint64_t c = 0; c < s + nparity; c++) {
2503 uint64_t devidx = (f + c) % dcols;
2504 vdev_t *cvd = vd->vdev_child[devidx];
2507 * dsl_scan_need_resilver() already checked vd with
2508 * vdev_dtl_contains(). So here just check cvd with
2509 * vdev_dtl_empty(), cheaper and a good approximation.
2511 if (!vdev_dtl_empty(cvd, DTL_PARTIAL))
2519 vdev_raidz_xlate(vdev_t *cvd, const range_seg64_t *logical_rs,
2520 range_seg64_t *physical_rs, range_seg64_t *remain_rs)
2524 vdev_t *raidvd = cvd->vdev_parent;
2525 ASSERT(raidvd->vdev_ops == &vdev_raidz_ops);
2527 uint64_t width = raidvd->vdev_children;
2528 uint64_t tgt_col = cvd->vdev_id;
2529 uint64_t ashift = raidvd->vdev_top->vdev_ashift;
2531 /* make sure the offsets are block-aligned */
2532 ASSERT0(logical_rs->rs_start % (1 << ashift));
2533 ASSERT0(logical_rs->rs_end % (1 << ashift));
2534 uint64_t b_start = logical_rs->rs_start >> ashift;
2535 uint64_t b_end = logical_rs->rs_end >> ashift;
2537 uint64_t start_row = 0;
2538 if (b_start > tgt_col) /* avoid underflow */
2539 start_row = ((b_start - tgt_col - 1) / width) + 1;
2541 uint64_t end_row = 0;
2542 if (b_end > tgt_col)
2543 end_row = ((b_end - tgt_col - 1) / width) + 1;
2545 physical_rs->rs_start = start_row << ashift;
2546 physical_rs->rs_end = end_row << ashift;
2548 ASSERT3U(physical_rs->rs_start, <=, logical_rs->rs_start);
2549 ASSERT3U(physical_rs->rs_end - physical_rs->rs_start, <=,
2550 logical_rs->rs_end - logical_rs->rs_start);
2554 * Initialize private RAIDZ specific fields from the nvlist.
2557 vdev_raidz_init(spa_t *spa, nvlist_t *nv, void **tsd)
2564 int error = nvlist_lookup_nvlist_array(nv,
2565 ZPOOL_CONFIG_CHILDREN, &child, &children);
2567 return (SET_ERROR(EINVAL));
2569 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY, &nparity) == 0) {
2570 if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY)
2571 return (SET_ERROR(EINVAL));
2574 * Previous versions could only support 1 or 2 parity
2577 if (nparity > 1 && spa_version(spa) < SPA_VERSION_RAIDZ2)
2578 return (SET_ERROR(EINVAL));
2579 else if (nparity > 2 && spa_version(spa) < SPA_VERSION_RAIDZ3)
2580 return (SET_ERROR(EINVAL));
2583 * We require the parity to be specified for SPAs that
2584 * support multiple parity levels.
2586 if (spa_version(spa) >= SPA_VERSION_RAIDZ2)
2587 return (SET_ERROR(EINVAL));
2590 * Otherwise, we default to 1 parity device for RAID-Z.
2595 vdrz = kmem_zalloc(sizeof (*vdrz), KM_SLEEP);
2596 vdrz->vd_logical_width = children;
2597 vdrz->vd_nparity = nparity;
2605 vdev_raidz_fini(vdev_t *vd)
2607 kmem_free(vd->vdev_tsd, sizeof (vdev_raidz_t));
2611 * Add RAIDZ specific fields to the config nvlist.
2614 vdev_raidz_config_generate(vdev_t *vd, nvlist_t *nv)
2616 ASSERT3P(vd->vdev_ops, ==, &vdev_raidz_ops);
2617 vdev_raidz_t *vdrz = vd->vdev_tsd;
2620 * Make sure someone hasn't managed to sneak a fancy new vdev
2621 * into a crufty old storage pool.
2623 ASSERT(vdrz->vd_nparity == 1 ||
2624 (vdrz->vd_nparity <= 2 &&
2625 spa_version(vd->vdev_spa) >= SPA_VERSION_RAIDZ2) ||
2626 (vdrz->vd_nparity <= 3 &&
2627 spa_version(vd->vdev_spa) >= SPA_VERSION_RAIDZ3));
2630 * Note that we'll add these even on storage pools where they
2631 * aren't strictly required -- older software will just ignore
2634 fnvlist_add_uint64(nv, ZPOOL_CONFIG_NPARITY, vdrz->vd_nparity);
2638 vdev_raidz_nparity(vdev_t *vd)
2640 vdev_raidz_t *vdrz = vd->vdev_tsd;
2641 return (vdrz->vd_nparity);
2645 vdev_raidz_ndisks(vdev_t *vd)
2647 return (vd->vdev_children);
2650 vdev_ops_t vdev_raidz_ops = {
2651 .vdev_op_init = vdev_raidz_init,
2652 .vdev_op_fini = vdev_raidz_fini,
2653 .vdev_op_open = vdev_raidz_open,
2654 .vdev_op_close = vdev_raidz_close,
2655 .vdev_op_asize = vdev_raidz_asize,
2656 .vdev_op_min_asize = vdev_raidz_min_asize,
2657 .vdev_op_min_alloc = NULL,
2658 .vdev_op_io_start = vdev_raidz_io_start,
2659 .vdev_op_io_done = vdev_raidz_io_done,
2660 .vdev_op_state_change = vdev_raidz_state_change,
2661 .vdev_op_need_resilver = vdev_raidz_need_resilver,
2662 .vdev_op_hold = NULL,
2663 .vdev_op_rele = NULL,
2664 .vdev_op_remap = NULL,
2665 .vdev_op_xlate = vdev_raidz_xlate,
2666 .vdev_op_rebuild_asize = NULL,
2667 .vdev_op_metaslab_init = NULL,
2668 .vdev_op_config_generate = vdev_raidz_config_generate,
2669 .vdev_op_nparity = vdev_raidz_nparity,
2670 .vdev_op_ndisks = vdev_raidz_ndisks,
2671 .vdev_op_type = VDEV_TYPE_RAIDZ, /* name of this vdev type */
2672 .vdev_op_leaf = B_FALSE /* not a leaf vdev */