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 http://www.opensolaris.org/os/licensing.
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;
931 * In the general case of reconstruction, we must solve the system of linear
932 * equations defined by the coefficients used to generate parity as well as
933 * the contents of the data and parity disks. This can be expressed with
934 * vectors for the original data (D) and the actual data (d) and parity (p)
935 * and a matrix composed of the identity matrix (I) and a dispersal matrix (V):
939 * | V | | D_0 | | p_m-1 |
940 * | | x | : | = | d_0 |
941 * | I | | D_n-1 | | : |
942 * | | ~~ ~~ | d_n-1 |
945 * I is simply a square identity matrix of size n, and V is a vandermonde
946 * matrix defined by the coefficients we chose for the various parity columns
947 * (1, 2, 4). Note that these values were chosen both for simplicity, speedy
948 * computation as well as linear separability.
951 * | 1 .. 1 1 1 | | p_0 |
952 * | 2^n-1 .. 4 2 1 | __ __ | : |
953 * | 4^n-1 .. 16 4 1 | | D_0 | | p_m-1 |
954 * | 1 .. 0 0 0 | | D_1 | | d_0 |
955 * | 0 .. 0 0 0 | x | D_2 | = | d_1 |
956 * | : : : : | | : | | d_2 |
957 * | 0 .. 1 0 0 | | D_n-1 | | : |
958 * | 0 .. 0 1 0 | ~~ ~~ | : |
959 * | 0 .. 0 0 1 | | d_n-1 |
962 * Note that I, V, d, and p are known. To compute D, we must invert the
963 * matrix and use the known data and parity values to reconstruct the unknown
964 * data values. We begin by removing the rows in V|I and d|p that correspond
965 * to failed or missing columns; we then make V|I square (n x n) and d|p
966 * sized n by removing rows corresponding to unused parity from the bottom up
967 * to generate (V|I)' and (d|p)'. We can then generate the inverse of (V|I)'
968 * using Gauss-Jordan elimination. In the example below we use m=3 parity
969 * columns, n=8 data columns, with errors in d_1, d_2, and p_1:
971 * | 1 1 1 1 1 1 1 1 |
972 * | 128 64 32 16 8 4 2 1 | <-----+-+-- missing disks
973 * | 19 205 116 29 64 16 4 1 | / /
974 * | 1 0 0 0 0 0 0 0 | / /
975 * | 0 1 0 0 0 0 0 0 | <--' /
976 * (V|I) = | 0 0 1 0 0 0 0 0 | <---'
977 * | 0 0 0 1 0 0 0 0 |
978 * | 0 0 0 0 1 0 0 0 |
979 * | 0 0 0 0 0 1 0 0 |
980 * | 0 0 0 0 0 0 1 0 |
981 * | 0 0 0 0 0 0 0 1 |
984 * | 1 1 1 1 1 1 1 1 |
985 * | 128 64 32 16 8 4 2 1 |
986 * | 19 205 116 29 64 16 4 1 |
987 * | 1 0 0 0 0 0 0 0 |
988 * | 0 1 0 0 0 0 0 0 |
989 * (V|I)' = | 0 0 1 0 0 0 0 0 |
990 * | 0 0 0 1 0 0 0 0 |
991 * | 0 0 0 0 1 0 0 0 |
992 * | 0 0 0 0 0 1 0 0 |
993 * | 0 0 0 0 0 0 1 0 |
994 * | 0 0 0 0 0 0 0 1 |
997 * Here we employ Gauss-Jordan elimination to find the inverse of (V|I)'. We
998 * have carefully chosen the seed values 1, 2, and 4 to ensure that this
999 * matrix is not singular.
1001 * | 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 |
1002 * | 19 205 116 29 64 16 4 1 0 1 0 0 0 0 0 0 |
1003 * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 |
1004 * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 |
1005 * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 |
1006 * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 |
1007 * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 |
1008 * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 |
1011 * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 |
1012 * | 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 |
1013 * | 19 205 116 29 64 16 4 1 0 1 0 0 0 0 0 0 |
1014 * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 |
1015 * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 |
1016 * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 |
1017 * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 |
1018 * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 |
1021 * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 |
1022 * | 0 1 1 0 0 0 0 0 1 0 1 1 1 1 1 1 |
1023 * | 0 205 116 0 0 0 0 0 0 1 19 29 64 16 4 1 |
1024 * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 |
1025 * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 |
1026 * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 |
1027 * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 |
1028 * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 |
1031 * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 |
1032 * | 0 1 1 0 0 0 0 0 1 0 1 1 1 1 1 1 |
1033 * | 0 0 185 0 0 0 0 0 205 1 222 208 141 221 201 204 |
1034 * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 |
1035 * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 |
1036 * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 |
1037 * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 |
1038 * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 |
1041 * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 |
1042 * | 0 1 1 0 0 0 0 0 1 0 1 1 1 1 1 1 |
1043 * | 0 0 1 0 0 0 0 0 166 100 4 40 158 168 216 209 |
1044 * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 |
1045 * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 |
1046 * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 |
1047 * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 |
1048 * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 |
1051 * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 |
1052 * | 0 1 0 0 0 0 0 0 167 100 5 41 159 169 217 208 |
1053 * | 0 0 1 0 0 0 0 0 166 100 4 40 158 168 216 209 |
1054 * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 |
1055 * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 |
1056 * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 |
1057 * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 |
1058 * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 |
1061 * | 0 0 1 0 0 0 0 0 |
1062 * | 167 100 5 41 159 169 217 208 |
1063 * | 166 100 4 40 158 168 216 209 |
1064 * (V|I)'^-1 = | 0 0 0 1 0 0 0 0 |
1065 * | 0 0 0 0 1 0 0 0 |
1066 * | 0 0 0 0 0 1 0 0 |
1067 * | 0 0 0 0 0 0 1 0 |
1068 * | 0 0 0 0 0 0 0 1 |
1071 * We can then simply compute D = (V|I)'^-1 x (d|p)' to discover the values
1072 * of the missing data.
1074 * As is apparent from the example above, the only non-trivial rows in the
1075 * inverse matrix correspond to the data disks that we're trying to
1076 * reconstruct. Indeed, those are the only rows we need as the others would
1077 * only be useful for reconstructing data known or assumed to be valid. For
1078 * that reason, we only build the coefficients in the rows that correspond to
1084 vdev_raidz_matrix_init(raidz_row_t *rr, int n, int nmap, int *map,
1090 ASSERT(n == rr->rr_cols - rr->rr_firstdatacol);
1093 * Fill in the missing rows of interest.
1095 for (i = 0; i < nmap; i++) {
1096 ASSERT3S(0, <=, map[i]);
1097 ASSERT3S(map[i], <=, 2);
1104 for (j = 0; j < n; j++) {
1108 rows[i][j] = vdev_raidz_pow2[pow];
1114 vdev_raidz_matrix_invert(raidz_row_t *rr, int n, int nmissing, int *missing,
1115 uint8_t **rows, uint8_t **invrows, const uint8_t *used)
1121 * Assert that the first nmissing entries from the array of used
1122 * columns correspond to parity columns and that subsequent entries
1123 * correspond to data columns.
1125 for (i = 0; i < nmissing; i++) {
1126 ASSERT3S(used[i], <, rr->rr_firstdatacol);
1128 for (; i < n; i++) {
1129 ASSERT3S(used[i], >=, rr->rr_firstdatacol);
1133 * First initialize the storage where we'll compute the inverse rows.
1135 for (i = 0; i < nmissing; i++) {
1136 for (j = 0; j < n; j++) {
1137 invrows[i][j] = (i == j) ? 1 : 0;
1142 * Subtract all trivial rows from the rows of consequence.
1144 for (i = 0; i < nmissing; i++) {
1145 for (j = nmissing; j < n; j++) {
1146 ASSERT3U(used[j], >=, rr->rr_firstdatacol);
1147 jj = used[j] - rr->rr_firstdatacol;
1149 invrows[i][j] = rows[i][jj];
1155 * For each of the rows of interest, we must normalize it and subtract
1156 * a multiple of it from the other rows.
1158 for (i = 0; i < nmissing; i++) {
1159 for (j = 0; j < missing[i]; j++) {
1160 ASSERT0(rows[i][j]);
1162 ASSERT3U(rows[i][missing[i]], !=, 0);
1165 * Compute the inverse of the first element and multiply each
1166 * element in the row by that value.
1168 log = 255 - vdev_raidz_log2[rows[i][missing[i]]];
1170 for (j = 0; j < n; j++) {
1171 rows[i][j] = vdev_raidz_exp2(rows[i][j], log);
1172 invrows[i][j] = vdev_raidz_exp2(invrows[i][j], log);
1175 for (ii = 0; ii < nmissing; ii++) {
1179 ASSERT3U(rows[ii][missing[i]], !=, 0);
1181 log = vdev_raidz_log2[rows[ii][missing[i]]];
1183 for (j = 0; j < n; j++) {
1185 vdev_raidz_exp2(rows[i][j], log);
1187 vdev_raidz_exp2(invrows[i][j], log);
1193 * Verify that the data that is left in the rows are properly part of
1194 * an identity matrix.
1196 for (i = 0; i < nmissing; i++) {
1197 for (j = 0; j < n; j++) {
1198 if (j == missing[i]) {
1199 ASSERT3U(rows[i][j], ==, 1);
1201 ASSERT0(rows[i][j]);
1208 vdev_raidz_matrix_reconstruct(raidz_row_t *rr, int n, int nmissing,
1209 int *missing, uint8_t **invrows, const uint8_t *used)
1214 uint8_t *dst[VDEV_RAIDZ_MAXPARITY] = { NULL };
1215 uint64_t dcount[VDEV_RAIDZ_MAXPARITY] = { 0 };
1219 uint8_t *invlog[VDEV_RAIDZ_MAXPARITY];
1223 psize = sizeof (invlog[0][0]) * n * nmissing;
1224 p = kmem_alloc(psize, KM_SLEEP);
1226 for (pp = p, i = 0; i < nmissing; i++) {
1231 for (i = 0; i < nmissing; i++) {
1232 for (j = 0; j < n; j++) {
1233 ASSERT3U(invrows[i][j], !=, 0);
1234 invlog[i][j] = vdev_raidz_log2[invrows[i][j]];
1238 for (i = 0; i < n; i++) {
1240 ASSERT3U(c, <, rr->rr_cols);
1242 ccount = rr->rr_col[c].rc_size;
1243 ASSERT(ccount >= rr->rr_col[missing[0]].rc_size || i > 0);
1246 src = abd_to_buf(rr->rr_col[c].rc_abd);
1247 for (j = 0; j < nmissing; j++) {
1248 cc = missing[j] + rr->rr_firstdatacol;
1249 ASSERT3U(cc, >=, rr->rr_firstdatacol);
1250 ASSERT3U(cc, <, rr->rr_cols);
1251 ASSERT3U(cc, !=, c);
1253 dcount[j] = rr->rr_col[cc].rc_size;
1255 dst[j] = abd_to_buf(rr->rr_col[cc].rc_abd);
1258 for (x = 0; x < ccount; x++, src++) {
1260 log = vdev_raidz_log2[*src];
1262 for (cc = 0; cc < nmissing; cc++) {
1263 if (x >= dcount[cc])
1269 if ((ll = log + invlog[cc][i]) >= 255)
1271 val = vdev_raidz_pow2[ll];
1282 kmem_free(p, psize);
1286 vdev_raidz_reconstruct_general(raidz_row_t *rr, int *tgts, int ntgts)
1290 int missing_rows[VDEV_RAIDZ_MAXPARITY];
1291 int parity_map[VDEV_RAIDZ_MAXPARITY];
1294 uint8_t *rows[VDEV_RAIDZ_MAXPARITY];
1295 uint8_t *invrows[VDEV_RAIDZ_MAXPARITY];
1298 abd_t **bufs = NULL;
1301 * Matrix reconstruction can't use scatter ABDs yet, so we allocate
1302 * temporary linear ABDs if any non-linear ABDs are found.
1304 for (i = rr->rr_firstdatacol; i < rr->rr_cols; i++) {
1305 if (!abd_is_linear(rr->rr_col[i].rc_abd)) {
1306 bufs = kmem_alloc(rr->rr_cols * sizeof (abd_t *),
1309 for (c = rr->rr_firstdatacol; c < rr->rr_cols; c++) {
1310 raidz_col_t *col = &rr->rr_col[c];
1312 bufs[c] = col->rc_abd;
1313 if (bufs[c] != NULL) {
1314 col->rc_abd = abd_alloc_linear(
1315 col->rc_size, B_TRUE);
1316 abd_copy(col->rc_abd, bufs[c],
1325 n = rr->rr_cols - rr->rr_firstdatacol;
1328 * Figure out which data columns are missing.
1331 for (t = 0; t < ntgts; t++) {
1332 if (tgts[t] >= rr->rr_firstdatacol) {
1333 missing_rows[nmissing_rows++] =
1334 tgts[t] - rr->rr_firstdatacol;
1339 * Figure out which parity columns to use to help generate the missing
1342 for (tt = 0, c = 0, i = 0; i < nmissing_rows; c++) {
1344 ASSERT(c < rr->rr_firstdatacol);
1347 * Skip any targeted parity columns.
1349 if (c == tgts[tt]) {
1358 psize = (sizeof (rows[0][0]) + sizeof (invrows[0][0])) *
1359 nmissing_rows * n + sizeof (used[0]) * n;
1360 p = kmem_alloc(psize, KM_SLEEP);
1362 for (pp = p, i = 0; i < nmissing_rows; i++) {
1370 for (i = 0; i < nmissing_rows; i++) {
1371 used[i] = parity_map[i];
1374 for (tt = 0, c = rr->rr_firstdatacol; c < rr->rr_cols; c++) {
1375 if (tt < nmissing_rows &&
1376 c == missing_rows[tt] + rr->rr_firstdatacol) {
1387 * Initialize the interesting rows of the matrix.
1389 vdev_raidz_matrix_init(rr, n, nmissing_rows, parity_map, rows);
1392 * Invert the matrix.
1394 vdev_raidz_matrix_invert(rr, n, nmissing_rows, missing_rows, rows,
1398 * Reconstruct the missing data using the generated matrix.
1400 vdev_raidz_matrix_reconstruct(rr, n, nmissing_rows, missing_rows,
1403 kmem_free(p, psize);
1406 * copy back from temporary linear abds and free them
1409 for (c = rr->rr_firstdatacol; c < rr->rr_cols; c++) {
1410 raidz_col_t *col = &rr->rr_col[c];
1412 if (bufs[c] != NULL) {
1413 abd_copy(bufs[c], col->rc_abd, col->rc_size);
1414 abd_free(col->rc_abd);
1416 col->rc_abd = bufs[c];
1418 kmem_free(bufs, rr->rr_cols * sizeof (abd_t *));
1423 vdev_raidz_reconstruct_row(raidz_map_t *rm, raidz_row_t *rr,
1424 const int *t, int nt)
1426 int tgts[VDEV_RAIDZ_MAXPARITY], *dt;
1429 int nbadparity, nbaddata;
1430 int parity_valid[VDEV_RAIDZ_MAXPARITY];
1432 nbadparity = rr->rr_firstdatacol;
1433 nbaddata = rr->rr_cols - nbadparity;
1435 for (i = 0, c = 0; c < rr->rr_cols; c++) {
1436 if (c < rr->rr_firstdatacol)
1437 parity_valid[c] = B_FALSE;
1439 if (i < nt && c == t[i]) {
1442 } else if (rr->rr_col[c].rc_error != 0) {
1444 } else if (c >= rr->rr_firstdatacol) {
1447 parity_valid[c] = B_TRUE;
1452 ASSERT(ntgts >= nt);
1453 ASSERT(nbaddata >= 0);
1454 ASSERT(nbaddata + nbadparity == ntgts);
1456 dt = &tgts[nbadparity];
1458 /* Reconstruct using the new math implementation */
1459 ret = vdev_raidz_math_reconstruct(rm, rr, parity_valid, dt, nbaddata);
1460 if (ret != RAIDZ_ORIGINAL_IMPL)
1464 * See if we can use any of our optimized reconstruction routines.
1468 if (parity_valid[VDEV_RAIDZ_P]) {
1469 vdev_raidz_reconstruct_p(rr, dt, 1);
1473 ASSERT(rr->rr_firstdatacol > 1);
1475 if (parity_valid[VDEV_RAIDZ_Q]) {
1476 vdev_raidz_reconstruct_q(rr, dt, 1);
1480 ASSERT(rr->rr_firstdatacol > 2);
1484 ASSERT(rr->rr_firstdatacol > 1);
1486 if (parity_valid[VDEV_RAIDZ_P] &&
1487 parity_valid[VDEV_RAIDZ_Q]) {
1488 vdev_raidz_reconstruct_pq(rr, dt, 2);
1492 ASSERT(rr->rr_firstdatacol > 2);
1497 vdev_raidz_reconstruct_general(rr, tgts, ntgts);
1501 vdev_raidz_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize,
1502 uint64_t *logical_ashift, uint64_t *physical_ashift)
1504 vdev_raidz_t *vdrz = vd->vdev_tsd;
1505 uint64_t nparity = vdrz->vd_nparity;
1510 ASSERT(nparity > 0);
1512 if (nparity > VDEV_RAIDZ_MAXPARITY ||
1513 vd->vdev_children < nparity + 1) {
1514 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
1515 return (SET_ERROR(EINVAL));
1518 vdev_open_children(vd);
1520 for (c = 0; c < vd->vdev_children; c++) {
1521 vdev_t *cvd = vd->vdev_child[c];
1523 if (cvd->vdev_open_error != 0) {
1524 lasterror = cvd->vdev_open_error;
1529 *asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1;
1530 *max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1;
1531 *logical_ashift = MAX(*logical_ashift, cvd->vdev_ashift);
1532 *physical_ashift = MAX(*physical_ashift,
1533 cvd->vdev_physical_ashift);
1536 *asize *= vd->vdev_children;
1537 *max_asize *= vd->vdev_children;
1539 if (numerrors > nparity) {
1540 vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS;
1548 vdev_raidz_close(vdev_t *vd)
1550 for (int c = 0; c < vd->vdev_children; c++) {
1551 if (vd->vdev_child[c] != NULL)
1552 vdev_close(vd->vdev_child[c]);
1557 vdev_raidz_asize(vdev_t *vd, uint64_t psize)
1559 vdev_raidz_t *vdrz = vd->vdev_tsd;
1561 uint64_t ashift = vd->vdev_top->vdev_ashift;
1562 uint64_t cols = vdrz->vd_logical_width;
1563 uint64_t nparity = vdrz->vd_nparity;
1565 asize = ((psize - 1) >> ashift) + 1;
1566 asize += nparity * ((asize + cols - nparity - 1) / (cols - nparity));
1567 asize = roundup(asize, nparity + 1) << ashift;
1573 * The allocatable space for a raidz vdev is N * sizeof(smallest child)
1574 * so each child must provide at least 1/Nth of its asize.
1577 vdev_raidz_min_asize(vdev_t *vd)
1579 return ((vd->vdev_min_asize + vd->vdev_children - 1) /
1584 vdev_raidz_child_done(zio_t *zio)
1586 raidz_col_t *rc = zio->io_private;
1588 ASSERT3P(rc->rc_abd, !=, NULL);
1589 rc->rc_error = zio->io_error;
1595 vdev_raidz_io_verify(vdev_t *vd, raidz_row_t *rr, int col)
1598 vdev_t *tvd = vd->vdev_top;
1600 range_seg64_t logical_rs, physical_rs, remain_rs;
1601 logical_rs.rs_start = rr->rr_offset;
1602 logical_rs.rs_end = logical_rs.rs_start +
1603 vdev_raidz_asize(vd, rr->rr_size);
1605 raidz_col_t *rc = &rr->rr_col[col];
1606 vdev_t *cvd = vd->vdev_child[rc->rc_devidx];
1608 vdev_xlate(cvd, &logical_rs, &physical_rs, &remain_rs);
1609 ASSERT(vdev_xlate_is_empty(&remain_rs));
1610 ASSERT3U(rc->rc_offset, ==, physical_rs.rs_start);
1611 ASSERT3U(rc->rc_offset, <, physical_rs.rs_end);
1613 * It would be nice to assert that rs_end is equal
1614 * to rc_offset + rc_size but there might be an
1615 * optional I/O at the end that is not accounted in
1618 if (physical_rs.rs_end > rc->rc_offset + rc->rc_size) {
1619 ASSERT3U(physical_rs.rs_end, ==, rc->rc_offset +
1620 rc->rc_size + (1 << tvd->vdev_ashift));
1622 ASSERT3U(physical_rs.rs_end, ==, rc->rc_offset + rc->rc_size);
1628 vdev_raidz_io_start_write(zio_t *zio, raidz_row_t *rr, uint64_t ashift)
1630 vdev_t *vd = zio->io_vd;
1631 raidz_map_t *rm = zio->io_vsd;
1633 vdev_raidz_generate_parity_row(rm, rr);
1635 for (int c = 0; c < rr->rr_scols; c++) {
1636 raidz_col_t *rc = &rr->rr_col[c];
1637 vdev_t *cvd = vd->vdev_child[rc->rc_devidx];
1639 /* Verify physical to logical translation */
1640 vdev_raidz_io_verify(vd, rr, c);
1642 if (rc->rc_size > 0) {
1643 ASSERT3P(rc->rc_abd, !=, NULL);
1644 zio_nowait(zio_vdev_child_io(zio, NULL, cvd,
1645 rc->rc_offset, rc->rc_abd,
1646 abd_get_size(rc->rc_abd), zio->io_type,
1647 zio->io_priority, 0, vdev_raidz_child_done, rc));
1650 * Generate optional write for skip sector to improve
1651 * aggregation contiguity.
1653 ASSERT3P(rc->rc_abd, ==, NULL);
1654 zio_nowait(zio_vdev_child_io(zio, NULL, cvd,
1655 rc->rc_offset, NULL, 1ULL << ashift,
1656 zio->io_type, zio->io_priority,
1657 ZIO_FLAG_NODATA | ZIO_FLAG_OPTIONAL, NULL,
1664 vdev_raidz_io_start_read(zio_t *zio, raidz_row_t *rr)
1666 vdev_t *vd = zio->io_vd;
1669 * Iterate over the columns in reverse order so that we hit the parity
1670 * last -- any errors along the way will force us to read the parity.
1672 for (int c = rr->rr_cols - 1; c >= 0; c--) {
1673 raidz_col_t *rc = &rr->rr_col[c];
1674 if (rc->rc_size == 0)
1676 vdev_t *cvd = vd->vdev_child[rc->rc_devidx];
1677 if (!vdev_readable(cvd)) {
1678 if (c >= rr->rr_firstdatacol)
1679 rr->rr_missingdata++;
1681 rr->rr_missingparity++;
1682 rc->rc_error = SET_ERROR(ENXIO);
1683 rc->rc_tried = 1; /* don't even try */
1687 if (vdev_dtl_contains(cvd, DTL_MISSING, zio->io_txg, 1)) {
1688 if (c >= rr->rr_firstdatacol)
1689 rr->rr_missingdata++;
1691 rr->rr_missingparity++;
1692 rc->rc_error = SET_ERROR(ESTALE);
1696 if (c >= rr->rr_firstdatacol || rr->rr_missingdata > 0 ||
1697 (zio->io_flags & (ZIO_FLAG_SCRUB | ZIO_FLAG_RESILVER))) {
1698 zio_nowait(zio_vdev_child_io(zio, NULL, cvd,
1699 rc->rc_offset, rc->rc_abd, rc->rc_size,
1700 zio->io_type, zio->io_priority, 0,
1701 vdev_raidz_child_done, rc));
1707 * Start an IO operation on a RAIDZ VDev
1710 * - For write operations:
1711 * 1. Generate the parity data
1712 * 2. Create child zio write operations to each column's vdev, for both
1714 * 3. If the column skips any sectors for padding, create optional dummy
1715 * write zio children for those areas to improve aggregation continuity.
1716 * - For read operations:
1717 * 1. Create child zio read operations to each data column's vdev to read
1718 * the range of data required for zio.
1719 * 2. If this is a scrub or resilver operation, or if any of the data
1720 * vdevs have had errors, then create zio read operations to the parity
1721 * columns' VDevs as well.
1724 vdev_raidz_io_start(zio_t *zio)
1726 vdev_t *vd = zio->io_vd;
1727 vdev_t *tvd = vd->vdev_top;
1728 vdev_raidz_t *vdrz = vd->vdev_tsd;
1730 raidz_map_t *rm = vdev_raidz_map_alloc(zio, tvd->vdev_ashift,
1731 vdrz->vd_logical_width, vdrz->vd_nparity);
1733 zio->io_vsd_ops = &vdev_raidz_vsd_ops;
1736 * Until raidz expansion is implemented all maps for a raidz vdev
1737 * contain a single row.
1739 ASSERT3U(rm->rm_nrows, ==, 1);
1740 raidz_row_t *rr = rm->rm_row[0];
1742 if (zio->io_type == ZIO_TYPE_WRITE) {
1743 vdev_raidz_io_start_write(zio, rr, tvd->vdev_ashift);
1745 ASSERT(zio->io_type == ZIO_TYPE_READ);
1746 vdev_raidz_io_start_read(zio, rr);
1753 * Report a checksum error for a child of a RAID-Z device.
1756 vdev_raidz_checksum_error(zio_t *zio, raidz_col_t *rc, abd_t *bad_data)
1758 vdev_t *vd = zio->io_vd->vdev_child[rc->rc_devidx];
1760 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE) &&
1761 zio->io_priority != ZIO_PRIORITY_REBUILD) {
1762 zio_bad_cksum_t zbc;
1763 raidz_map_t *rm = zio->io_vsd;
1765 zbc.zbc_has_cksum = 0;
1766 zbc.zbc_injected = rm->rm_ecksuminjected;
1768 (void) zfs_ereport_post_checksum(zio->io_spa, vd,
1769 &zio->io_bookmark, zio, rc->rc_offset, rc->rc_size,
1770 rc->rc_abd, bad_data, &zbc);
1771 mutex_enter(&vd->vdev_stat_lock);
1772 vd->vdev_stat.vs_checksum_errors++;
1773 mutex_exit(&vd->vdev_stat_lock);
1778 * We keep track of whether or not there were any injected errors, so that
1779 * any ereports we generate can note it.
1782 raidz_checksum_verify(zio_t *zio)
1784 zio_bad_cksum_t zbc;
1785 raidz_map_t *rm = zio->io_vsd;
1787 bzero(&zbc, sizeof (zio_bad_cksum_t));
1789 int ret = zio_checksum_error(zio, &zbc);
1790 if (ret != 0 && zbc.zbc_injected != 0)
1791 rm->rm_ecksuminjected = 1;
1797 * Generate the parity from the data columns. If we tried and were able to
1798 * read the parity without error, verify that the generated parity matches the
1799 * data we read. If it doesn't, we fire off a checksum error. Return the
1800 * number of such failures.
1803 raidz_parity_verify(zio_t *zio, raidz_row_t *rr)
1805 abd_t *orig[VDEV_RAIDZ_MAXPARITY];
1807 raidz_map_t *rm = zio->io_vsd;
1810 blkptr_t *bp = zio->io_bp;
1811 enum zio_checksum checksum = (bp == NULL ? zio->io_prop.zp_checksum :
1812 (BP_IS_GANG(bp) ? ZIO_CHECKSUM_GANG_HEADER : BP_GET_CHECKSUM(bp)));
1814 if (checksum == ZIO_CHECKSUM_NOPARITY)
1817 for (c = 0; c < rr->rr_firstdatacol; c++) {
1818 rc = &rr->rr_col[c];
1819 if (!rc->rc_tried || rc->rc_error != 0)
1822 orig[c] = abd_alloc_sametype(rc->rc_abd, rc->rc_size);
1823 abd_copy(orig[c], rc->rc_abd, rc->rc_size);
1827 * Verify any empty sectors are zero filled to ensure the parity
1828 * is calculated correctly even if these non-data sectors are damaged.
1830 if (rr->rr_nempty && rr->rr_abd_empty != NULL)
1831 ret += vdev_draid_map_verify_empty(zio, rr);
1834 * Regenerates parity even for !tried||rc_error!=0 columns. This
1835 * isn't harmful but it does have the side effect of fixing stuff
1836 * we didn't realize was necessary (i.e. even if we return 0).
1838 vdev_raidz_generate_parity_row(rm, rr);
1840 for (c = 0; c < rr->rr_firstdatacol; c++) {
1841 rc = &rr->rr_col[c];
1843 if (!rc->rc_tried || rc->rc_error != 0)
1846 if (abd_cmp(orig[c], rc->rc_abd) != 0) {
1847 vdev_raidz_checksum_error(zio, rc, orig[c]);
1848 rc->rc_error = SET_ERROR(ECKSUM);
1858 vdev_raidz_worst_error(raidz_row_t *rr)
1862 for (int c = 0; c < rr->rr_cols; c++)
1863 error = zio_worst_error(error, rr->rr_col[c].rc_error);
1869 vdev_raidz_io_done_verified(zio_t *zio, raidz_row_t *rr)
1871 int unexpected_errors = 0;
1872 int parity_errors = 0;
1873 int parity_untried = 0;
1874 int data_errors = 0;
1876 ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ);
1878 for (int c = 0; c < rr->rr_cols; c++) {
1879 raidz_col_t *rc = &rr->rr_col[c];
1882 if (c < rr->rr_firstdatacol)
1887 if (!rc->rc_skipped)
1888 unexpected_errors++;
1889 } else if (c < rr->rr_firstdatacol && !rc->rc_tried) {
1895 * If we read more parity disks than were used for
1896 * reconstruction, confirm that the other parity disks produced
1899 * Note that we also regenerate parity when resilvering so we
1900 * can write it out to failed devices later.
1902 if (parity_errors + parity_untried <
1903 rr->rr_firstdatacol - data_errors ||
1904 (zio->io_flags & ZIO_FLAG_RESILVER)) {
1905 int n = raidz_parity_verify(zio, rr);
1906 unexpected_errors += n;
1909 if (zio->io_error == 0 && spa_writeable(zio->io_spa) &&
1910 (unexpected_errors > 0 || (zio->io_flags & ZIO_FLAG_RESILVER))) {
1912 * Use the good data we have in hand to repair damaged children.
1914 for (int c = 0; c < rr->rr_cols; c++) {
1915 raidz_col_t *rc = &rr->rr_col[c];
1916 vdev_t *vd = zio->io_vd;
1917 vdev_t *cvd = vd->vdev_child[rc->rc_devidx];
1919 if (!rc->rc_allow_repair) {
1921 } else if (!rc->rc_force_repair &&
1922 (rc->rc_error == 0 || rc->rc_size == 0)) {
1926 zio_nowait(zio_vdev_child_io(zio, NULL, cvd,
1927 rc->rc_offset, rc->rc_abd, rc->rc_size,
1929 zio->io_priority == ZIO_PRIORITY_REBUILD ?
1930 ZIO_PRIORITY_REBUILD : ZIO_PRIORITY_ASYNC_WRITE,
1931 ZIO_FLAG_IO_REPAIR | (unexpected_errors ?
1932 ZIO_FLAG_SELF_HEAL : 0), NULL, NULL));
1938 raidz_restore_orig_data(raidz_map_t *rm)
1940 for (int i = 0; i < rm->rm_nrows; i++) {
1941 raidz_row_t *rr = rm->rm_row[i];
1942 for (int c = 0; c < rr->rr_cols; c++) {
1943 raidz_col_t *rc = &rr->rr_col[c];
1944 if (rc->rc_need_orig_restore) {
1945 abd_copy(rc->rc_abd,
1946 rc->rc_orig_data, rc->rc_size);
1947 rc->rc_need_orig_restore = B_FALSE;
1954 * returns EINVAL if reconstruction of the block will not be possible
1955 * returns ECKSUM if this specific reconstruction failed
1956 * returns 0 on successful reconstruction
1959 raidz_reconstruct(zio_t *zio, int *ltgts, int ntgts, int nparity)
1961 raidz_map_t *rm = zio->io_vsd;
1963 /* Reconstruct each row */
1964 for (int r = 0; r < rm->rm_nrows; r++) {
1965 raidz_row_t *rr = rm->rm_row[r];
1966 int my_tgts[VDEV_RAIDZ_MAXPARITY]; /* value is child id */
1971 for (int c = 0; c < rr->rr_cols; c++) {
1972 raidz_col_t *rc = &rr->rr_col[c];
1973 ASSERT0(rc->rc_need_orig_restore);
1974 if (rc->rc_error != 0) {
1980 if (rc->rc_size == 0)
1982 for (int lt = 0; lt < ntgts; lt++) {
1983 if (rc->rc_devidx == ltgts[lt]) {
1984 if (rc->rc_orig_data == NULL) {
1987 rc->rc_size, B_TRUE);
1988 abd_copy(rc->rc_orig_data,
1989 rc->rc_abd, rc->rc_size);
1991 rc->rc_need_orig_restore = B_TRUE;
2001 if (dead > nparity) {
2002 /* reconstruction not possible */
2003 raidz_restore_orig_data(rm);
2007 vdev_raidz_reconstruct_row(rm, rr, my_tgts, t);
2010 /* Check for success */
2011 if (raidz_checksum_verify(zio) == 0) {
2013 /* Reconstruction succeeded - report errors */
2014 for (int i = 0; i < rm->rm_nrows; i++) {
2015 raidz_row_t *rr = rm->rm_row[i];
2017 for (int c = 0; c < rr->rr_cols; c++) {
2018 raidz_col_t *rc = &rr->rr_col[c];
2019 if (rc->rc_need_orig_restore) {
2021 * Note: if this is a parity column,
2022 * we don't really know if it's wrong.
2024 * vdev_raidz_io_done_verified() check
2025 * it, and if we set rc_error, it will
2026 * think that it is a "known" error
2027 * that doesn't need to be checked
2030 if (rc->rc_error == 0 &&
2031 c >= rr->rr_firstdatacol) {
2032 vdev_raidz_checksum_error(zio,
2033 rc, rc->rc_orig_data);
2037 rc->rc_need_orig_restore = B_FALSE;
2041 vdev_raidz_io_done_verified(zio, rr);
2044 zio_checksum_verified(zio);
2049 /* Reconstruction failed - restore original data */
2050 raidz_restore_orig_data(rm);
2055 * Iterate over all combinations of N bad vdevs and attempt a reconstruction.
2056 * Note that the algorithm below is non-optimal because it doesn't take into
2057 * account how reconstruction is actually performed. For example, with
2058 * triple-parity RAID-Z the reconstruction procedure is the same if column 4
2059 * is targeted as invalid as if columns 1 and 4 are targeted since in both
2060 * cases we'd only use parity information in column 0.
2062 * The order that we find the various possible combinations of failed
2063 * disks is dictated by these rules:
2064 * - Examine each "slot" (the "i" in tgts[i])
2065 * - Try to increment this slot (tgts[i] = tgts[i] + 1)
2066 * - if we can't increment because it runs into the next slot,
2067 * reset our slot to the minimum, and examine the next slot
2069 * For example, with a 6-wide RAIDZ3, and no known errors (so we have to choose
2070 * 3 columns to reconstruct), we will generate the following sequence:
2073 * 0 1 2 special case: skip since these are all parity
2074 * 0 1 3 first slot: reset to 0; middle slot: increment to 2
2075 * 0 2 3 first slot: increment to 1
2076 * 1 2 3 first: reset to 0; middle: reset to 1; last: increment to 4
2077 * 0 1 4 first: reset to 0; middle: increment to 2
2078 * 0 2 4 first: increment to 1
2079 * 1 2 4 first: reset to 0; middle: increment to 3
2080 * 0 3 4 first: increment to 1
2081 * 1 3 4 first: increment to 2
2082 * 2 3 4 first: reset to 0; middle: reset to 1; last: increment to 5
2083 * 0 1 5 first: reset to 0; middle: increment to 2
2084 * 0 2 5 first: increment to 1
2085 * 1 2 5 first: reset to 0; middle: increment to 3
2086 * 0 3 5 first: increment to 1
2087 * 1 3 5 first: increment to 2
2088 * 2 3 5 first: reset to 0; middle: increment to 4
2089 * 0 4 5 first: increment to 1
2090 * 1 4 5 first: increment to 2
2091 * 2 4 5 first: increment to 3
2094 * This strategy works for dRAID but is less efficient when there are a large
2095 * number of child vdevs and therefore permutations to check. Furthermore,
2096 * since the raidz_map_t rows likely do not overlap reconstruction would be
2097 * possible as long as there are no more than nparity data errors per row.
2098 * These additional permutations are not currently checked but could be as
2099 * a future improvement.
2102 vdev_raidz_combrec(zio_t *zio)
2104 int nparity = vdev_get_nparity(zio->io_vd);
2105 raidz_map_t *rm = zio->io_vsd;
2107 /* Check if there's enough data to attempt reconstrution. */
2108 for (int i = 0; i < rm->rm_nrows; i++) {
2109 raidz_row_t *rr = rm->rm_row[i];
2110 int total_errors = 0;
2112 for (int c = 0; c < rr->rr_cols; c++) {
2113 if (rr->rr_col[c].rc_error)
2117 if (total_errors > nparity)
2118 return (vdev_raidz_worst_error(rr));
2121 for (int num_failures = 1; num_failures <= nparity; num_failures++) {
2122 int tstore[VDEV_RAIDZ_MAXPARITY + 2];
2123 int *ltgts = &tstore[1]; /* value is logical child ID */
2125 /* Determine number of logical children, n */
2126 int n = zio->io_vd->vdev_children;
2128 ASSERT3U(num_failures, <=, nparity);
2129 ASSERT3U(num_failures, <=, VDEV_RAIDZ_MAXPARITY);
2131 /* Handle corner cases in combrec logic */
2133 for (int i = 0; i < num_failures; i++) {
2136 ltgts[num_failures] = n;
2139 int err = raidz_reconstruct(zio, ltgts, num_failures,
2141 if (err == EINVAL) {
2143 * Reconstruction not possible with this #
2144 * failures; try more failures.
2147 } else if (err == 0)
2150 /* Compute next targets to try */
2151 for (int t = 0; ; t++) {
2152 ASSERT3U(t, <, num_failures);
2154 if (ltgts[t] == n) {
2155 /* try more failures */
2156 ASSERT3U(t, ==, num_failures - 1);
2160 ASSERT3U(ltgts[t], <, n);
2161 ASSERT3U(ltgts[t], <=, ltgts[t + 1]);
2164 * If that spot is available, we're done here.
2165 * Try the next combination.
2167 if (ltgts[t] != ltgts[t + 1])
2171 * Otherwise, reset this tgt to the minimum,
2172 * and move on to the next tgt.
2174 ltgts[t] = ltgts[t - 1] + 1;
2175 ASSERT3U(ltgts[t], ==, t);
2178 /* Increase the number of failures and keep trying. */
2179 if (ltgts[num_failures - 1] == n)
2188 vdev_raidz_reconstruct(raidz_map_t *rm, const int *t, int nt)
2190 for (uint64_t row = 0; row < rm->rm_nrows; row++) {
2191 raidz_row_t *rr = rm->rm_row[row];
2192 vdev_raidz_reconstruct_row(rm, rr, t, nt);
2197 * Complete a write IO operation on a RAIDZ VDev
2200 * 1. Check for errors on the child IOs.
2201 * 2. Return, setting an error code if too few child VDevs were written
2202 * to reconstruct the data later. Note that partial writes are
2203 * considered successful if they can be reconstructed at all.
2206 vdev_raidz_io_done_write_impl(zio_t *zio, raidz_row_t *rr)
2208 int total_errors = 0;
2210 ASSERT3U(rr->rr_missingparity, <=, rr->rr_firstdatacol);
2211 ASSERT3U(rr->rr_missingdata, <=, rr->rr_cols - rr->rr_firstdatacol);
2212 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
2214 for (int c = 0; c < rr->rr_cols; c++) {
2215 raidz_col_t *rc = &rr->rr_col[c];
2218 ASSERT(rc->rc_error != ECKSUM); /* child has no bp */
2225 * Treat partial writes as a success. If we couldn't write enough
2226 * columns to reconstruct the data, the I/O failed. Otherwise,
2229 * Now that we support write reallocation, it would be better
2230 * to treat partial failure as real failure unless there are
2231 * no non-degraded top-level vdevs left, and not update DTLs
2232 * if we intend to reallocate.
2234 if (total_errors > rr->rr_firstdatacol) {
2235 zio->io_error = zio_worst_error(zio->io_error,
2236 vdev_raidz_worst_error(rr));
2241 vdev_raidz_io_done_reconstruct_known_missing(zio_t *zio, raidz_map_t *rm,
2244 int parity_errors = 0;
2245 int parity_untried = 0;
2246 int data_errors = 0;
2247 int total_errors = 0;
2249 ASSERT3U(rr->rr_missingparity, <=, rr->rr_firstdatacol);
2250 ASSERT3U(rr->rr_missingdata, <=, rr->rr_cols - rr->rr_firstdatacol);
2251 ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ);
2253 for (int c = 0; c < rr->rr_cols; c++) {
2254 raidz_col_t *rc = &rr->rr_col[c];
2257 ASSERT(rc->rc_error != ECKSUM); /* child has no bp */
2259 if (c < rr->rr_firstdatacol)
2265 } else if (c < rr->rr_firstdatacol && !rc->rc_tried) {
2271 * If there were data errors and the number of errors we saw was
2272 * correctable -- less than or equal to the number of parity disks read
2273 * -- reconstruct based on the missing data.
2275 if (data_errors != 0 &&
2276 total_errors <= rr->rr_firstdatacol - parity_untried) {
2278 * We either attempt to read all the parity columns or
2279 * none of them. If we didn't try to read parity, we
2280 * wouldn't be here in the correctable case. There must
2281 * also have been fewer parity errors than parity
2282 * columns or, again, we wouldn't be in this code path.
2284 ASSERT(parity_untried == 0);
2285 ASSERT(parity_errors < rr->rr_firstdatacol);
2288 * Identify the data columns that reported an error.
2291 int tgts[VDEV_RAIDZ_MAXPARITY];
2292 for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) {
2293 raidz_col_t *rc = &rr->rr_col[c];
2294 if (rc->rc_error != 0) {
2295 ASSERT(n < VDEV_RAIDZ_MAXPARITY);
2300 ASSERT(rr->rr_firstdatacol >= n);
2302 vdev_raidz_reconstruct_row(rm, rr, tgts, n);
2307 * Return the number of reads issued.
2310 vdev_raidz_read_all(zio_t *zio, raidz_row_t *rr)
2312 vdev_t *vd = zio->io_vd;
2315 rr->rr_missingdata = 0;
2316 rr->rr_missingparity = 0;
2319 * If this rows contains empty sectors which are not required
2320 * for a normal read then allocate an ABD for them now so they
2321 * may be read, verified, and any needed repairs performed.
2323 if (rr->rr_nempty && rr->rr_abd_empty == NULL)
2324 vdev_draid_map_alloc_empty(zio, rr);
2326 for (int c = 0; c < rr->rr_cols; c++) {
2327 raidz_col_t *rc = &rr->rr_col[c];
2328 if (rc->rc_tried || rc->rc_size == 0)
2331 zio_nowait(zio_vdev_child_io(zio, NULL,
2332 vd->vdev_child[rc->rc_devidx],
2333 rc->rc_offset, rc->rc_abd, rc->rc_size,
2334 zio->io_type, zio->io_priority, 0,
2335 vdev_raidz_child_done, rc));
2342 * We're here because either there were too many errors to even attempt
2343 * reconstruction (total_errors == rm_first_datacol), or vdev_*_combrec()
2344 * failed. In either case, there is enough bad data to prevent reconstruction.
2345 * Start checksum ereports for all children which haven't failed.
2348 vdev_raidz_io_done_unrecoverable(zio_t *zio)
2350 raidz_map_t *rm = zio->io_vsd;
2352 for (int i = 0; i < rm->rm_nrows; i++) {
2353 raidz_row_t *rr = rm->rm_row[i];
2355 for (int c = 0; c < rr->rr_cols; c++) {
2356 raidz_col_t *rc = &rr->rr_col[c];
2357 vdev_t *cvd = zio->io_vd->vdev_child[rc->rc_devidx];
2359 if (rc->rc_error != 0)
2362 zio_bad_cksum_t zbc;
2363 zbc.zbc_has_cksum = 0;
2364 zbc.zbc_injected = rm->rm_ecksuminjected;
2366 (void) zfs_ereport_start_checksum(zio->io_spa,
2367 cvd, &zio->io_bookmark, zio, rc->rc_offset,
2369 mutex_enter(&cvd->vdev_stat_lock);
2370 cvd->vdev_stat.vs_checksum_errors++;
2371 mutex_exit(&cvd->vdev_stat_lock);
2377 vdev_raidz_io_done(zio_t *zio)
2379 raidz_map_t *rm = zio->io_vsd;
2381 if (zio->io_type == ZIO_TYPE_WRITE) {
2382 for (int i = 0; i < rm->rm_nrows; i++) {
2383 vdev_raidz_io_done_write_impl(zio, rm->rm_row[i]);
2386 for (int i = 0; i < rm->rm_nrows; i++) {
2387 raidz_row_t *rr = rm->rm_row[i];
2388 vdev_raidz_io_done_reconstruct_known_missing(zio,
2392 if (raidz_checksum_verify(zio) == 0) {
2393 for (int i = 0; i < rm->rm_nrows; i++) {
2394 raidz_row_t *rr = rm->rm_row[i];
2395 vdev_raidz_io_done_verified(zio, rr);
2397 zio_checksum_verified(zio);
2400 * A sequential resilver has no checksum which makes
2401 * combinatoral reconstruction impossible. This code
2402 * path is unreachable since raidz_checksum_verify()
2403 * has no checksum to verify and must succeed.
2405 ASSERT3U(zio->io_priority, !=, ZIO_PRIORITY_REBUILD);
2408 * This isn't a typical situation -- either we got a
2409 * read error or a child silently returned bad data.
2410 * Read every block so we can try again with as much
2411 * data and parity as we can track down. If we've
2412 * already been through once before, all children will
2413 * be marked as tried so we'll proceed to combinatorial
2417 for (int i = 0; i < rm->rm_nrows; i++) {
2418 nread += vdev_raidz_read_all(zio,
2423 * Normally our stage is VDEV_IO_DONE, but if
2424 * we've already called redone(), it will have
2425 * changed to VDEV_IO_START, in which case we
2426 * don't want to call redone() again.
2428 if (zio->io_stage != ZIO_STAGE_VDEV_IO_START)
2429 zio_vdev_io_redone(zio);
2433 zio->io_error = vdev_raidz_combrec(zio);
2434 if (zio->io_error == ECKSUM &&
2435 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2436 vdev_raidz_io_done_unrecoverable(zio);
2443 vdev_raidz_state_change(vdev_t *vd, int faulted, int degraded)
2445 vdev_raidz_t *vdrz = vd->vdev_tsd;
2446 if (faulted > vdrz->vd_nparity)
2447 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2448 VDEV_AUX_NO_REPLICAS);
2449 else if (degraded + faulted != 0)
2450 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE);
2452 vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE);
2456 * Determine if any portion of the provided block resides on a child vdev
2457 * with a dirty DTL and therefore needs to be resilvered. The function
2458 * assumes that at least one DTL is dirty which implies that full stripe
2459 * width blocks must be resilvered.
2462 vdev_raidz_need_resilver(vdev_t *vd, const dva_t *dva, size_t psize,
2463 uint64_t phys_birth)
2465 vdev_raidz_t *vdrz = vd->vdev_tsd;
2466 uint64_t dcols = vd->vdev_children;
2467 uint64_t nparity = vdrz->vd_nparity;
2468 uint64_t ashift = vd->vdev_top->vdev_ashift;
2469 /* The starting RAIDZ (parent) vdev sector of the block. */
2470 uint64_t b = DVA_GET_OFFSET(dva) >> ashift;
2471 /* The zio's size in units of the vdev's minimum sector size. */
2472 uint64_t s = ((psize - 1) >> ashift) + 1;
2473 /* The first column for this stripe. */
2474 uint64_t f = b % dcols;
2476 /* Unreachable by sequential resilver. */
2477 ASSERT3U(phys_birth, !=, TXG_UNKNOWN);
2479 if (!vdev_dtl_contains(vd, DTL_PARTIAL, phys_birth, 1))
2482 if (s + nparity >= dcols)
2485 for (uint64_t c = 0; c < s + nparity; c++) {
2486 uint64_t devidx = (f + c) % dcols;
2487 vdev_t *cvd = vd->vdev_child[devidx];
2490 * dsl_scan_need_resilver() already checked vd with
2491 * vdev_dtl_contains(). So here just check cvd with
2492 * vdev_dtl_empty(), cheaper and a good approximation.
2494 if (!vdev_dtl_empty(cvd, DTL_PARTIAL))
2502 vdev_raidz_xlate(vdev_t *cvd, const range_seg64_t *logical_rs,
2503 range_seg64_t *physical_rs, range_seg64_t *remain_rs)
2507 vdev_t *raidvd = cvd->vdev_parent;
2508 ASSERT(raidvd->vdev_ops == &vdev_raidz_ops);
2510 uint64_t width = raidvd->vdev_children;
2511 uint64_t tgt_col = cvd->vdev_id;
2512 uint64_t ashift = raidvd->vdev_top->vdev_ashift;
2514 /* make sure the offsets are block-aligned */
2515 ASSERT0(logical_rs->rs_start % (1 << ashift));
2516 ASSERT0(logical_rs->rs_end % (1 << ashift));
2517 uint64_t b_start = logical_rs->rs_start >> ashift;
2518 uint64_t b_end = logical_rs->rs_end >> ashift;
2520 uint64_t start_row = 0;
2521 if (b_start > tgt_col) /* avoid underflow */
2522 start_row = ((b_start - tgt_col - 1) / width) + 1;
2524 uint64_t end_row = 0;
2525 if (b_end > tgt_col)
2526 end_row = ((b_end - tgt_col - 1) / width) + 1;
2528 physical_rs->rs_start = start_row << ashift;
2529 physical_rs->rs_end = end_row << ashift;
2531 ASSERT3U(physical_rs->rs_start, <=, logical_rs->rs_start);
2532 ASSERT3U(physical_rs->rs_end - physical_rs->rs_start, <=,
2533 logical_rs->rs_end - logical_rs->rs_start);
2537 * Initialize private RAIDZ specific fields from the nvlist.
2540 vdev_raidz_init(spa_t *spa, nvlist_t *nv, void **tsd)
2547 int error = nvlist_lookup_nvlist_array(nv,
2548 ZPOOL_CONFIG_CHILDREN, &child, &children);
2550 return (SET_ERROR(EINVAL));
2552 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY, &nparity) == 0) {
2553 if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY)
2554 return (SET_ERROR(EINVAL));
2557 * Previous versions could only support 1 or 2 parity
2560 if (nparity > 1 && spa_version(spa) < SPA_VERSION_RAIDZ2)
2561 return (SET_ERROR(EINVAL));
2562 else if (nparity > 2 && spa_version(spa) < SPA_VERSION_RAIDZ3)
2563 return (SET_ERROR(EINVAL));
2566 * We require the parity to be specified for SPAs that
2567 * support multiple parity levels.
2569 if (spa_version(spa) >= SPA_VERSION_RAIDZ2)
2570 return (SET_ERROR(EINVAL));
2573 * Otherwise, we default to 1 parity device for RAID-Z.
2578 vdrz = kmem_zalloc(sizeof (*vdrz), KM_SLEEP);
2579 vdrz->vd_logical_width = children;
2580 vdrz->vd_nparity = nparity;
2588 vdev_raidz_fini(vdev_t *vd)
2590 kmem_free(vd->vdev_tsd, sizeof (vdev_raidz_t));
2594 * Add RAIDZ specific fields to the config nvlist.
2597 vdev_raidz_config_generate(vdev_t *vd, nvlist_t *nv)
2599 ASSERT3P(vd->vdev_ops, ==, &vdev_raidz_ops);
2600 vdev_raidz_t *vdrz = vd->vdev_tsd;
2603 * Make sure someone hasn't managed to sneak a fancy new vdev
2604 * into a crufty old storage pool.
2606 ASSERT(vdrz->vd_nparity == 1 ||
2607 (vdrz->vd_nparity <= 2 &&
2608 spa_version(vd->vdev_spa) >= SPA_VERSION_RAIDZ2) ||
2609 (vdrz->vd_nparity <= 3 &&
2610 spa_version(vd->vdev_spa) >= SPA_VERSION_RAIDZ3));
2613 * Note that we'll add these even on storage pools where they
2614 * aren't strictly required -- older software will just ignore
2617 fnvlist_add_uint64(nv, ZPOOL_CONFIG_NPARITY, vdrz->vd_nparity);
2621 vdev_raidz_nparity(vdev_t *vd)
2623 vdev_raidz_t *vdrz = vd->vdev_tsd;
2624 return (vdrz->vd_nparity);
2628 vdev_raidz_ndisks(vdev_t *vd)
2630 return (vd->vdev_children);
2633 vdev_ops_t vdev_raidz_ops = {
2634 .vdev_op_init = vdev_raidz_init,
2635 .vdev_op_fini = vdev_raidz_fini,
2636 .vdev_op_open = vdev_raidz_open,
2637 .vdev_op_close = vdev_raidz_close,
2638 .vdev_op_asize = vdev_raidz_asize,
2639 .vdev_op_min_asize = vdev_raidz_min_asize,
2640 .vdev_op_min_alloc = NULL,
2641 .vdev_op_io_start = vdev_raidz_io_start,
2642 .vdev_op_io_done = vdev_raidz_io_done,
2643 .vdev_op_state_change = vdev_raidz_state_change,
2644 .vdev_op_need_resilver = vdev_raidz_need_resilver,
2645 .vdev_op_hold = NULL,
2646 .vdev_op_rele = NULL,
2647 .vdev_op_remap = NULL,
2648 .vdev_op_xlate = vdev_raidz_xlate,
2649 .vdev_op_rebuild_asize = NULL,
2650 .vdev_op_metaslab_init = NULL,
2651 .vdev_op_config_generate = vdev_raidz_config_generate,
2652 .vdev_op_nparity = vdev_raidz_nparity,
2653 .vdev_op_ndisks = vdev_raidz_ndisks,
2654 .vdev_op_type = VDEV_TYPE_RAIDZ, /* name of this vdev type */
2655 .vdev_op_leaf = B_FALSE /* not a leaf vdev */