]>
Commit | Line | Data |
---|---|---|
34dc7c2f BB |
1 | /* |
2 | * CDDL HEADER START | |
3 | * | |
4 | * The contents of this file are subject to the terms of the | |
5 | * Common Development and Distribution License (the "License"). | |
6 | * You may not use this file except in compliance with the License. | |
7 | * | |
8 | * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE | |
9 | * or http://www.opensolaris.org/os/licensing. | |
10 | * See the License for the specific language governing permissions | |
11 | * and limitations under the License. | |
12 | * | |
13 | * When distributing Covered Code, include this CDDL HEADER in each | |
14 | * file and include the License file at usr/src/OPENSOLARIS.LICENSE. | |
15 | * If applicable, add the following below this CDDL HEADER, with the | |
16 | * fields enclosed by brackets "[]" replaced with your own identifying | |
17 | * information: Portions Copyright [yyyy] [name of copyright owner] | |
18 | * | |
19 | * CDDL HEADER END | |
20 | */ | |
21 | ||
22 | /* | |
428870ff | 23 | * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. |
4f072827 | 24 | * Copyright (c) 2012, 2020 by Delphix. All rights reserved. |
ab9f4b0b | 25 | * Copyright (c) 2016 Gvozden Nešković. All rights reserved. |
34dc7c2f BB |
26 | */ |
27 | ||
34dc7c2f BB |
28 | #include <sys/zfs_context.h> |
29 | #include <sys/spa.h> | |
30 | #include <sys/vdev_impl.h> | |
31 | #include <sys/zio.h> | |
32 | #include <sys/zio_checksum.h> | |
a6255b7f | 33 | #include <sys/abd.h> |
34dc7c2f BB |
34 | #include <sys/fs/zfs.h> |
35 | #include <sys/fm/fs/zfs.h> | |
ab9f4b0b GN |
36 | #include <sys/vdev_raidz.h> |
37 | #include <sys/vdev_raidz_impl.h> | |
b2255edc | 38 | #include <sys/vdev_draid.h> |
34dc7c2f | 39 | |
619f0976 | 40 | #ifdef ZFS_DEBUG |
1b939560 | 41 | #include <sys/vdev.h> /* For vdev_xlate() in vdev_raidz_io_verify() */ |
619f0976 GW |
42 | #endif |
43 | ||
34dc7c2f BB |
44 | /* |
45 | * Virtual device vector for RAID-Z. | |
46 | * | |
45d1cae3 BB |
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 | |
54 | * for writes. | |
55 | * | |
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 | |
62 | * suffer. | |
63 | * | |
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: | |
34dc7c2f BB |
68 | * |
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: | |
d3cc8b15 | 72 | * |
34dc7c2f BB |
73 | * (A * 2)_7 = A_6 |
74 | * (A * 2)_6 = A_5 | |
75 | * (A * 2)_5 = A_4 | |
76 | * (A * 2)_4 = A_3 + A_7 | |
77 | * (A * 2)_3 = A_2 + A_7 | |
78 | * (A * 2)_2 = A_1 + A_7 | |
79 | * (A * 2)_1 = A_0 | |
80 | * (A * 2)_0 = A_7 | |
81 | * | |
82 | * In C, multiplying by 2 is therefore ((a << 1) ^ ((a & 0x80) ? 0x1d : 0)). | |
45d1cae3 BB |
83 | * As an aside, this multiplication is derived from the error correcting |
84 | * primitive polynomial x^8 + x^4 + x^3 + x^2 + 1. | |
34dc7c2f BB |
85 | * |
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 | |
45d1cae3 BB |
90 | * than field addition). The inverse of a field element A (A^-1) is therefore |
91 | * A ^ (255 - 1) = A^254. | |
34dc7c2f | 92 | * |
45d1cae3 BB |
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: | |
34dc7c2f BB |
95 | * |
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 | |
45d1cae3 BB |
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 | |
34dc7c2f | 101 | * |
e1cfd73f | 102 | * We chose 1, 2, and 4 as our generators because 1 corresponds to the trivial |
45d1cae3 BB |
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.) | |
106 | * | |
107 | * See the reconstruction code below for how P, Q and R can used individually | |
108 | * or in concert to recover missing data columns. | |
34dc7c2f BB |
109 | */ |
110 | ||
34dc7c2f BB |
111 | #define VDEV_RAIDZ_P 0 |
112 | #define VDEV_RAIDZ_Q 1 | |
45d1cae3 | 113 | #define VDEV_RAIDZ_R 2 |
45d1cae3 BB |
114 | |
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))) | |
117 | ||
118 | /* | |
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. | |
123 | */ | |
124 | #define VDEV_RAIDZ_64MUL_2(x, mask) \ | |
125 | { \ | |
126 | (mask) = (x) & 0x8080808080808080ULL; \ | |
127 | (mask) = ((mask) << 1) - ((mask) >> 7); \ | |
128 | (x) = (((x) << 1) & 0xfefefefefefefefeULL) ^ \ | |
c5b3a7bb | 129 | ((mask) & 0x1d1d1d1d1d1d1d1dULL); \ |
45d1cae3 | 130 | } |
34dc7c2f | 131 | |
45d1cae3 BB |
132 | #define VDEV_RAIDZ_64MUL_4(x, mask) \ |
133 | { \ | |
134 | VDEV_RAIDZ_64MUL_2((x), mask); \ | |
135 | VDEV_RAIDZ_64MUL_2((x), mask); \ | |
136 | } | |
34dc7c2f | 137 | |
b2255edc BB |
138 | static void |
139 | vdev_raidz_row_free(raidz_row_t *rr) | |
b128c09f | 140 | { |
e2af2acc MA |
141 | for (int c = 0; c < rr->rr_cols; c++) { |
142 | raidz_col_t *rc = &rr->rr_col[c]; | |
b128c09f | 143 | |
e2af2acc MA |
144 | if (rc->rc_size != 0) |
145 | abd_free(rc->rc_abd); | |
e2af2acc | 146 | if (rc->rc_orig_data != NULL) |
330c6c05 | 147 | abd_free(rc->rc_orig_data); |
b2255edc BB |
148 | } |
149 | ||
b2255edc BB |
150 | if (rr->rr_abd_empty != NULL) |
151 | abd_free(rr->rr_abd_empty); | |
152 | ||
153 | kmem_free(rr, offsetof(raidz_row_t, rr_col[rr->rr_scols])); | |
154 | } | |
428870ff | 155 | |
b2255edc BB |
156 | void |
157 | vdev_raidz_map_free(raidz_map_t *rm) | |
158 | { | |
159 | for (int i = 0; i < rm->rm_nrows; i++) | |
160 | vdev_raidz_row_free(rm->rm_row[i]); | |
428870ff | 161 | |
b2255edc | 162 | kmem_free(rm, offsetof(raidz_map_t, rm_row[rm->rm_nrows])); |
b128c09f BB |
163 | } |
164 | ||
428870ff BB |
165 | static void |
166 | vdev_raidz_map_free_vsd(zio_t *zio) | |
167 | { | |
168 | raidz_map_t *rm = zio->io_vsd; | |
169 | ||
330c6c05 | 170 | vdev_raidz_map_free(rm); |
428870ff BB |
171 | } |
172 | ||
330c6c05 | 173 | const zio_vsd_ops_t vdev_raidz_vsd_ops = { |
56d8d8ac | 174 | .vsd_free = vdev_raidz_map_free_vsd, |
428870ff BB |
175 | }; |
176 | ||
345196be BA |
177 | static void |
178 | vdev_raidz_map_alloc_write(zio_t *zio, raidz_map_t *rm, uint64_t ashift) | |
179 | { | |
180 | int c; | |
181 | int nwrapped = 0; | |
182 | uint64_t off = 0; | |
183 | raidz_row_t *rr = rm->rm_row[0]; | |
184 | ||
185 | ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); | |
186 | ASSERT3U(rm->rm_nrows, ==, 1); | |
187 | ||
188 | /* | |
189 | * Pad any parity columns with additional space to account for skip | |
190 | * sectors. | |
191 | */ | |
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)) { | |
196 | nwrapped = | |
197 | (rm->rm_skipstart + rm->rm_nskip) % rr->rr_scols; | |
198 | } | |
199 | ||
200 | /* | |
201 | * Optional single skip sectors (rc_size == 0) will be handled in | |
202 | * vdev_raidz_io_start_write(). | |
203 | */ | |
204 | int skipped = rr->rr_scols - rr->rr_cols; | |
205 | ||
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]; | |
209 | ||
210 | /* | |
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). | |
218 | */ | |
219 | if (c < nwrapped) { | |
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); | |
223 | skipped++; | |
224 | } else { | |
225 | rc->rc_abd = abd_alloc_linear(rc->rc_size, B_FALSE); | |
226 | } | |
227 | } | |
228 | ||
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); | |
233 | ||
234 | /* | |
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. | |
239 | * | |
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. | |
247 | */ | |
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); | |
253 | skipped++; | |
254 | } else { | |
255 | rc->rc_abd = abd; | |
256 | } | |
257 | off += rc->rc_size; | |
258 | } | |
259 | ||
260 | ASSERT3U(off, ==, zio->io_size); | |
261 | ASSERT3S(skipped, ==, rm->rm_nskip); | |
262 | } | |
263 | ||
264 | static void | |
265 | vdev_raidz_map_alloc_read(zio_t *zio, raidz_map_t *rm) | |
266 | { | |
267 | int c; | |
268 | raidz_row_t *rr = rm->rm_row[0]; | |
269 | ||
270 | ASSERT3U(rm->rm_nrows, ==, 1); | |
271 | ||
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); | |
276 | ||
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); | |
281 | off += rc->rc_size; | |
282 | } | |
283 | } | |
284 | ||
e49f1e20 WA |
285 | /* |
286 | * Divides the IO evenly across all child vdevs; usually, dcols is | |
287 | * the number of children in the target vdev. | |
a1687880 BB |
288 | * |
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. | |
e49f1e20 | 291 | */ |
ab9f4b0b | 292 | noinline raidz_map_t * |
3d6da72d | 293 | vdev_raidz_map_alloc(zio_t *zio, uint64_t ashift, uint64_t dcols, |
34dc7c2f BB |
294 | uint64_t nparity) |
295 | { | |
b2255edc | 296 | raidz_row_t *rr; |
e49f1e20 | 297 | /* The starting RAIDZ (parent) vdev sector of the block. */ |
3d6da72d | 298 | uint64_t b = zio->io_offset >> ashift; |
e49f1e20 | 299 | /* The zio's size in units of the vdev's minimum sector size. */ |
3d6da72d | 300 | uint64_t s = zio->io_size >> ashift; |
e49f1e20 | 301 | /* The first column for this stripe. */ |
34dc7c2f | 302 | uint64_t f = b % dcols; |
e49f1e20 | 303 | /* The starting byte offset on each child vdev. */ |
3d6da72d | 304 | uint64_t o = (b / dcols) << ashift; |
45d1cae3 | 305 | uint64_t q, r, c, bc, col, acols, scols, coff, devidx, asize, tot; |
34dc7c2f | 306 | |
b2255edc BB |
307 | raidz_map_t *rm = |
308 | kmem_zalloc(offsetof(raidz_map_t, rm_row[1]), KM_SLEEP); | |
309 | rm->rm_nrows = 1; | |
310 | ||
e49f1e20 WA |
311 | /* |
312 | * "Quotient": The number of data sectors for this stripe on all but | |
313 | * the "big column" child vdevs that also contain "remainder" data. | |
314 | */ | |
34dc7c2f | 315 | q = s / (dcols - nparity); |
e49f1e20 WA |
316 | |
317 | /* | |
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. | |
320 | */ | |
34dc7c2f | 321 | r = s - q * (dcols - nparity); |
e49f1e20 WA |
322 | |
323 | /* The number of "big columns" - those which contain remainder data. */ | |
34dc7c2f | 324 | bc = (r == 0 ? 0 : r + nparity); |
e49f1e20 WA |
325 | |
326 | /* | |
327 | * The total number of data and parity sectors associated with | |
328 | * this I/O. | |
329 | */ | |
45d1cae3 BB |
330 | tot = s + nparity * (q + (r == 0 ? 0 : 1)); |
331 | ||
b2255edc BB |
332 | /* |
333 | * acols: The columns that will be accessed. | |
334 | * scols: The columns that will be accessed or skipped. | |
335 | */ | |
45d1cae3 | 336 | if (q == 0) { |
e49f1e20 | 337 | /* Our I/O request doesn't span all child vdevs. */ |
45d1cae3 BB |
338 | acols = bc; |
339 | scols = MIN(dcols, roundup(bc, nparity + 1)); | |
340 | } else { | |
341 | acols = dcols; | |
342 | scols = dcols; | |
343 | } | |
34dc7c2f | 344 | |
45d1cae3 | 345 | ASSERT3U(acols, <=, scols); |
34dc7c2f | 346 | |
b2255edc BB |
347 | rr = kmem_alloc(offsetof(raidz_row_t, rr_col[scols]), KM_SLEEP); |
348 | rm->rm_row[0] = rr; | |
349 | ||
350 | rr->rr_cols = acols; | |
351 | rr->rr_scols = scols; | |
352 | rr->rr_bigcols = bc; | |
353 | rr->rr_missingdata = 0; | |
354 | rr->rr_missingparity = 0; | |
355 | rr->rr_firstdatacol = nparity; | |
b2255edc BB |
356 | rr->rr_abd_empty = NULL; |
357 | rr->rr_nempty = 0; | |
358 | #ifdef ZFS_DEBUG | |
359 | rr->rr_offset = zio->io_offset; | |
360 | rr->rr_size = zio->io_size; | |
361 | #endif | |
34dc7c2f | 362 | |
45d1cae3 BB |
363 | asize = 0; |
364 | ||
365 | for (c = 0; c < scols; c++) { | |
b2255edc | 366 | raidz_col_t *rc = &rr->rr_col[c]; |
34dc7c2f BB |
367 | col = f + c; |
368 | coff = o; | |
369 | if (col >= dcols) { | |
370 | col -= dcols; | |
3d6da72d | 371 | coff += 1ULL << ashift; |
34dc7c2f | 372 | } |
b2255edc BB |
373 | rc->rc_devidx = col; |
374 | rc->rc_offset = coff; | |
375 | rc->rc_abd = NULL; | |
b2255edc BB |
376 | rc->rc_orig_data = NULL; |
377 | rc->rc_error = 0; | |
378 | rc->rc_tried = 0; | |
379 | rc->rc_skipped = 0; | |
8fb577ae BB |
380 | rc->rc_force_repair = 0; |
381 | rc->rc_allow_repair = 1; | |
b2255edc | 382 | rc->rc_need_orig_restore = B_FALSE; |
45d1cae3 BB |
383 | |
384 | if (c >= acols) | |
b2255edc | 385 | rc->rc_size = 0; |
45d1cae3 | 386 | else if (c < bc) |
b2255edc | 387 | rc->rc_size = (q + 1) << ashift; |
45d1cae3 | 388 | else |
b2255edc | 389 | rc->rc_size = q << ashift; |
45d1cae3 | 390 | |
b2255edc | 391 | asize += rc->rc_size; |
34dc7c2f BB |
392 | } |
393 | ||
3d6da72d | 394 | ASSERT3U(asize, ==, tot << ashift); |
428870ff | 395 | rm->rm_nskip = roundup(tot, nparity + 1) - tot; |
b2255edc | 396 | rm->rm_skipstart = bc; |
34dc7c2f | 397 | |
34dc7c2f BB |
398 | /* |
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 | |
e1cfd73f | 401 | * during normal operation, that device's I/O bandwidth won't be |
34dc7c2f BB |
402 | * used effectively. We therefore switch the parity every 1MB. |
403 | * | |
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. | |
428870ff BB |
412 | * |
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. | |
34dc7c2f | 417 | */ |
b2255edc BB |
418 | ASSERT(rr->rr_cols >= 2); |
419 | ASSERT(rr->rr_col[0].rc_size == rr->rr_col[1].rc_size); | |
34dc7c2f | 420 | |
b2255edc BB |
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; | |
428870ff BB |
428 | |
429 | if (rm->rm_skipstart == 0) | |
430 | rm->rm_skipstart = 1; | |
34dc7c2f BB |
431 | } |
432 | ||
345196be BA |
433 | if (zio->io_type == ZIO_TYPE_WRITE) { |
434 | vdev_raidz_map_alloc_write(zio, rm, ashift); | |
435 | } else { | |
436 | vdev_raidz_map_alloc_read(zio, rm); | |
437 | } | |
438 | ||
c9187d86 GN |
439 | /* init RAIDZ parity ops */ |
440 | rm->rm_ops = vdev_raidz_math_get_ops(); | |
ab9f4b0b | 441 | |
34dc7c2f BB |
442 | return (rm); |
443 | } | |
444 | ||
a6255b7f DQ |
445 | struct pqr_struct { |
446 | uint64_t *p; | |
447 | uint64_t *q; | |
448 | uint64_t *r; | |
449 | }; | |
450 | ||
451 | static int | |
452 | vdev_raidz_p_func(void *buf, size_t size, void *private) | |
453 | { | |
454 | struct pqr_struct *pqr = private; | |
455 | const uint64_t *src = buf; | |
456 | int i, cnt = size / sizeof (src[0]); | |
457 | ||
458 | ASSERT(pqr->p && !pqr->q && !pqr->r); | |
459 | ||
460 | for (i = 0; i < cnt; i++, src++, pqr->p++) | |
461 | *pqr->p ^= *src; | |
462 | ||
463 | return (0); | |
464 | } | |
465 | ||
466 | static int | |
467 | vdev_raidz_pq_func(void *buf, size_t size, void *private) | |
468 | { | |
469 | struct pqr_struct *pqr = private; | |
470 | const uint64_t *src = buf; | |
471 | uint64_t mask; | |
472 | int i, cnt = size / sizeof (src[0]); | |
473 | ||
474 | ASSERT(pqr->p && pqr->q && !pqr->r); | |
475 | ||
476 | for (i = 0; i < cnt; i++, src++, pqr->p++, pqr->q++) { | |
477 | *pqr->p ^= *src; | |
478 | VDEV_RAIDZ_64MUL_2(*pqr->q, mask); | |
479 | *pqr->q ^= *src; | |
480 | } | |
481 | ||
482 | return (0); | |
483 | } | |
484 | ||
485 | static int | |
486 | vdev_raidz_pqr_func(void *buf, size_t size, void *private) | |
487 | { | |
488 | struct pqr_struct *pqr = private; | |
489 | const uint64_t *src = buf; | |
490 | uint64_t mask; | |
491 | int i, cnt = size / sizeof (src[0]); | |
492 | ||
493 | ASSERT(pqr->p && pqr->q && pqr->r); | |
494 | ||
495 | for (i = 0; i < cnt; i++, src++, pqr->p++, pqr->q++, pqr->r++) { | |
496 | *pqr->p ^= *src; | |
497 | VDEV_RAIDZ_64MUL_2(*pqr->q, mask); | |
498 | *pqr->q ^= *src; | |
499 | VDEV_RAIDZ_64MUL_4(*pqr->r, mask); | |
500 | *pqr->r ^= *src; | |
501 | } | |
502 | ||
503 | return (0); | |
504 | } | |
505 | ||
34dc7c2f | 506 | static void |
b2255edc | 507 | vdev_raidz_generate_parity_p(raidz_row_t *rr) |
34dc7c2f | 508 | { |
b2255edc | 509 | uint64_t *p = abd_to_buf(rr->rr_col[VDEV_RAIDZ_P].rc_abd); |
34dc7c2f | 510 | |
b2255edc BB |
511 | for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { |
512 | abd_t *src = rr->rr_col[c].rc_abd; | |
34dc7c2f | 513 | |
b2255edc BB |
514 | if (c == rr->rr_firstdatacol) { |
515 | abd_copy_to_buf(p, src, rr->rr_col[c].rc_size); | |
34dc7c2f | 516 | } else { |
a6255b7f | 517 | struct pqr_struct pqr = { p, NULL, NULL }; |
b2255edc | 518 | (void) abd_iterate_func(src, 0, rr->rr_col[c].rc_size, |
a6255b7f | 519 | vdev_raidz_p_func, &pqr); |
34dc7c2f BB |
520 | } |
521 | } | |
522 | } | |
523 | ||
524 | static void | |
b2255edc | 525 | vdev_raidz_generate_parity_pq(raidz_row_t *rr) |
34dc7c2f | 526 | { |
b2255edc BB |
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); | |
34dc7c2f | 532 | |
b2255edc BB |
533 | for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { |
534 | abd_t *src = rr->rr_col[c].rc_abd; | |
45d1cae3 | 535 | |
b2255edc | 536 | uint64_t ccnt = rr->rr_col[c].rc_size / sizeof (p[0]); |
34dc7c2f | 537 | |
b2255edc | 538 | if (c == rr->rr_firstdatacol) { |
f7e76821 | 539 | ASSERT(ccnt == pcnt || ccnt == 0); |
b2255edc BB |
540 | abd_copy_to_buf(p, src, rr->rr_col[c].rc_size); |
541 | (void) memcpy(q, p, rr->rr_col[c].rc_size); | |
45d1cae3 | 542 | |
b2255edc | 543 | for (uint64_t i = ccnt; i < pcnt; i++) { |
a6255b7f DQ |
544 | p[i] = 0; |
545 | q[i] = 0; | |
45d1cae3 | 546 | } |
a6255b7f | 547 | } else { |
f7e76821 IH |
548 | struct pqr_struct pqr = { p, q, NULL }; |
549 | ||
550 | ASSERT(ccnt <= pcnt); | |
b2255edc | 551 | (void) abd_iterate_func(src, 0, rr->rr_col[c].rc_size, |
f7e76821 | 552 | vdev_raidz_pq_func, &pqr); |
45d1cae3 BB |
553 | |
554 | /* | |
555 | * Treat short columns as though they are full of 0s. | |
556 | * Note that there's therefore nothing needed for P. | |
557 | */ | |
b2255edc BB |
558 | uint64_t mask; |
559 | for (uint64_t i = ccnt; i < pcnt; i++) { | |
a6255b7f | 560 | VDEV_RAIDZ_64MUL_2(q[i], mask); |
45d1cae3 BB |
561 | } |
562 | } | |
563 | } | |
564 | } | |
565 | ||
566 | static void | |
b2255edc | 567 | vdev_raidz_generate_parity_pqr(raidz_row_t *rr) |
45d1cae3 | 568 | { |
b2255edc BB |
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); | |
45d1cae3 | 577 | |
b2255edc BB |
578 | for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { |
579 | abd_t *src = rr->rr_col[c].rc_abd; | |
45d1cae3 | 580 | |
b2255edc | 581 | uint64_t ccnt = rr->rr_col[c].rc_size / sizeof (p[0]); |
45d1cae3 | 582 | |
b2255edc | 583 | if (c == rr->rr_firstdatacol) { |
f7e76821 | 584 | ASSERT(ccnt == pcnt || ccnt == 0); |
b2255edc BB |
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); | |
45d1cae3 | 588 | |
b2255edc | 589 | for (uint64_t i = ccnt; i < pcnt; i++) { |
a6255b7f DQ |
590 | p[i] = 0; |
591 | q[i] = 0; | |
592 | r[i] = 0; | |
34dc7c2f | 593 | } |
a6255b7f | 594 | } else { |
f7e76821 IH |
595 | struct pqr_struct pqr = { p, q, r }; |
596 | ||
597 | ASSERT(ccnt <= pcnt); | |
b2255edc | 598 | (void) abd_iterate_func(src, 0, rr->rr_col[c].rc_size, |
f7e76821 IH |
599 | vdev_raidz_pqr_func, &pqr); |
600 | ||
34dc7c2f BB |
601 | /* |
602 | * Treat short columns as though they are full of 0s. | |
45d1cae3 | 603 | * Note that there's therefore nothing needed for P. |
34dc7c2f | 604 | */ |
b2255edc BB |
605 | uint64_t mask; |
606 | for (uint64_t i = ccnt; i < pcnt; i++) { | |
a6255b7f DQ |
607 | VDEV_RAIDZ_64MUL_2(q[i], mask); |
608 | VDEV_RAIDZ_64MUL_4(r[i], mask); | |
34dc7c2f BB |
609 | } |
610 | } | |
611 | } | |
612 | } | |
613 | ||
45d1cae3 BB |
614 | /* |
615 | * Generate RAID parity in the first virtual columns according to the number of | |
616 | * parity columns available. | |
617 | */ | |
ab9f4b0b | 618 | void |
b2255edc | 619 | vdev_raidz_generate_parity_row(raidz_map_t *rm, raidz_row_t *rr) |
45d1cae3 | 620 | { |
b2255edc BB |
621 | ASSERT3U(rr->rr_cols, !=, 0); |
622 | ||
c9187d86 | 623 | /* Generate using the new math implementation */ |
b2255edc | 624 | if (vdev_raidz_math_generate(rm, rr) != RAIDZ_ORIGINAL_IMPL) |
ab9f4b0b | 625 | return; |
ab9f4b0b | 626 | |
b2255edc | 627 | switch (rr->rr_firstdatacol) { |
45d1cae3 | 628 | case 1: |
b2255edc | 629 | vdev_raidz_generate_parity_p(rr); |
45d1cae3 BB |
630 | break; |
631 | case 2: | |
b2255edc | 632 | vdev_raidz_generate_parity_pq(rr); |
45d1cae3 BB |
633 | break; |
634 | case 3: | |
b2255edc | 635 | vdev_raidz_generate_parity_pqr(rr); |
45d1cae3 BB |
636 | break; |
637 | default: | |
638 | cmn_err(CE_PANIC, "invalid RAID-Z configuration"); | |
639 | } | |
640 | } | |
641 | ||
b2255edc BB |
642 | void |
643 | vdev_raidz_generate_parity(raidz_map_t *rm) | |
644 | { | |
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); | |
648 | } | |
649 | } | |
650 | ||
a6255b7f DQ |
651 | static int |
652 | vdev_raidz_reconst_p_func(void *dbuf, void *sbuf, size_t size, void *private) | |
653 | { | |
14e4e3cb | 654 | (void) private; |
a6255b7f DQ |
655 | uint64_t *dst = dbuf; |
656 | uint64_t *src = sbuf; | |
657 | int cnt = size / sizeof (src[0]); | |
a6255b7f | 658 | |
1c27024e | 659 | for (int i = 0; i < cnt; i++) { |
a6255b7f DQ |
660 | dst[i] ^= src[i]; |
661 | } | |
662 | ||
663 | return (0); | |
664 | } | |
665 | ||
a6255b7f DQ |
666 | static int |
667 | vdev_raidz_reconst_q_pre_func(void *dbuf, void *sbuf, size_t size, | |
668 | void *private) | |
669 | { | |
14e4e3cb | 670 | (void) private; |
a6255b7f DQ |
671 | uint64_t *dst = dbuf; |
672 | uint64_t *src = sbuf; | |
673 | uint64_t mask; | |
674 | int cnt = size / sizeof (dst[0]); | |
a6255b7f | 675 | |
1c27024e | 676 | for (int i = 0; i < cnt; i++, dst++, src++) { |
a6255b7f DQ |
677 | VDEV_RAIDZ_64MUL_2(*dst, mask); |
678 | *dst ^= *src; | |
679 | } | |
680 | ||
681 | return (0); | |
682 | } | |
683 | ||
a6255b7f DQ |
684 | static int |
685 | vdev_raidz_reconst_q_pre_tail_func(void *buf, size_t size, void *private) | |
686 | { | |
14e4e3cb | 687 | (void) private; |
a6255b7f DQ |
688 | uint64_t *dst = buf; |
689 | uint64_t mask; | |
690 | int cnt = size / sizeof (dst[0]); | |
a6255b7f | 691 | |
1c27024e | 692 | for (int i = 0; i < cnt; i++, dst++) { |
a6255b7f DQ |
693 | /* same operation as vdev_raidz_reconst_q_pre_func() on dst */ |
694 | VDEV_RAIDZ_64MUL_2(*dst, mask); | |
695 | } | |
696 | ||
697 | return (0); | |
698 | } | |
699 | ||
700 | struct reconst_q_struct { | |
701 | uint64_t *q; | |
702 | int exp; | |
703 | }; | |
704 | ||
705 | static int | |
706 | vdev_raidz_reconst_q_post_func(void *buf, size_t size, void *private) | |
707 | { | |
708 | struct reconst_q_struct *rq = private; | |
709 | uint64_t *dst = buf; | |
710 | int cnt = size / sizeof (dst[0]); | |
a6255b7f | 711 | |
1c27024e | 712 | for (int i = 0; i < cnt; i++, dst++, rq->q++) { |
a6255b7f DQ |
713 | int j; |
714 | uint8_t *b; | |
715 | ||
716 | *dst ^= *rq->q; | |
717 | for (j = 0, b = (uint8_t *)dst; j < 8; j++, b++) { | |
718 | *b = vdev_raidz_exp2(*b, rq->exp); | |
719 | } | |
720 | } | |
721 | ||
722 | return (0); | |
723 | } | |
724 | ||
725 | struct reconst_pq_struct { | |
726 | uint8_t *p; | |
727 | uint8_t *q; | |
728 | uint8_t *pxy; | |
729 | uint8_t *qxy; | |
730 | int aexp; | |
731 | int bexp; | |
732 | }; | |
733 | ||
734 | static int | |
735 | vdev_raidz_reconst_pq_func(void *xbuf, void *ybuf, size_t size, void *private) | |
736 | { | |
737 | struct reconst_pq_struct *rpq = private; | |
738 | uint8_t *xd = xbuf; | |
739 | uint8_t *yd = ybuf; | |
a6255b7f | 740 | |
1c27024e | 741 | for (int i = 0; i < size; |
a6255b7f DQ |
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; | |
746 | } | |
747 | ||
748 | return (0); | |
749 | } | |
750 | ||
751 | static int | |
752 | vdev_raidz_reconst_pq_tail_func(void *xbuf, size_t size, void *private) | |
753 | { | |
754 | struct reconst_pq_struct *rpq = private; | |
755 | uint8_t *xd = xbuf; | |
a6255b7f | 756 | |
1c27024e | 757 | for (int i = 0; i < size; |
a6255b7f DQ |
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); | |
762 | } | |
763 | ||
764 | return (0); | |
765 | } | |
766 | ||
46df6e98 | 767 | static void |
b2255edc | 768 | vdev_raidz_reconstruct_p(raidz_row_t *rr, int *tgts, int ntgts) |
34dc7c2f | 769 | { |
45d1cae3 | 770 | int x = tgts[0]; |
a6255b7f | 771 | abd_t *dst, *src; |
34dc7c2f | 772 | |
b2255edc BB |
773 | ASSERT3U(ntgts, ==, 1); |
774 | ASSERT3U(x, >=, rr->rr_firstdatacol); | |
775 | ASSERT3U(x, <, rr->rr_cols); | |
45d1cae3 | 776 | |
b2255edc | 777 | ASSERT3U(rr->rr_col[x].rc_size, <=, rr->rr_col[VDEV_RAIDZ_P].rc_size); |
34dc7c2f | 778 | |
b2255edc BB |
779 | src = rr->rr_col[VDEV_RAIDZ_P].rc_abd; |
780 | dst = rr->rr_col[x].rc_abd; | |
a6255b7f | 781 | |
b2255edc | 782 | abd_copy_from_buf(dst, abd_to_buf(src), rr->rr_col[x].rc_size); |
34dc7c2f | 783 | |
b2255edc BB |
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); | |
a6255b7f | 787 | |
b2255edc | 788 | src = rr->rr_col[c].rc_abd; |
34dc7c2f BB |
789 | |
790 | if (c == x) | |
791 | continue; | |
792 | ||
a6255b7f DQ |
793 | (void) abd_iterate_func2(dst, src, 0, 0, size, |
794 | vdev_raidz_reconst_p_func, NULL); | |
34dc7c2f BB |
795 | } |
796 | } | |
797 | ||
46df6e98 | 798 | static void |
b2255edc | 799 | vdev_raidz_reconstruct_q(raidz_row_t *rr, int *tgts, int ntgts) |
34dc7c2f | 800 | { |
45d1cae3 | 801 | int x = tgts[0]; |
a6255b7f DQ |
802 | int c, exp; |
803 | abd_t *dst, *src; | |
34dc7c2f | 804 | |
45d1cae3 BB |
805 | ASSERT(ntgts == 1); |
806 | ||
b2255edc | 807 | ASSERT(rr->rr_col[x].rc_size <= rr->rr_col[VDEV_RAIDZ_Q].rc_size); |
34dc7c2f | 808 | |
b2255edc BB |
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); | |
34dc7c2f | 812 | |
b2255edc BB |
813 | src = rr->rr_col[c].rc_abd; |
814 | dst = rr->rr_col[x].rc_abd; | |
34dc7c2f | 815 | |
b2255edc | 816 | if (c == rr->rr_firstdatacol) { |
a6255b7f | 817 | abd_copy(dst, src, size); |
b2255edc | 818 | if (rr->rr_col[x].rc_size > size) { |
a6255b7f | 819 | abd_zero_off(dst, size, |
b2255edc BB |
820 | rr->rr_col[x].rc_size - size); |
821 | } | |
34dc7c2f | 822 | } else { |
b2255edc | 823 | ASSERT3U(size, <=, rr->rr_col[x].rc_size); |
a6255b7f DQ |
824 | (void) abd_iterate_func2(dst, src, 0, 0, size, |
825 | vdev_raidz_reconst_q_pre_func, NULL); | |
826 | (void) abd_iterate_func(dst, | |
b2255edc | 827 | size, rr->rr_col[x].rc_size - size, |
a6255b7f | 828 | vdev_raidz_reconst_q_pre_tail_func, NULL); |
34dc7c2f BB |
829 | } |
830 | } | |
831 | ||
b2255edc BB |
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); | |
34dc7c2f | 835 | |
1c27024e | 836 | struct reconst_q_struct rq = { abd_to_buf(src), exp }; |
b2255edc | 837 | (void) abd_iterate_func(dst, 0, rr->rr_col[x].rc_size, |
a6255b7f | 838 | vdev_raidz_reconst_q_post_func, &rq); |
34dc7c2f BB |
839 | } |
840 | ||
46df6e98 | 841 | static void |
b2255edc | 842 | vdev_raidz_reconstruct_pq(raidz_row_t *rr, int *tgts, int ntgts) |
34dc7c2f | 843 | { |
a6255b7f DQ |
844 | uint8_t *p, *q, *pxy, *qxy, tmp, a, b, aexp, bexp; |
845 | abd_t *pdata, *qdata; | |
846 | uint64_t xsize, ysize; | |
45d1cae3 BB |
847 | int x = tgts[0]; |
848 | int y = tgts[1]; | |
a6255b7f | 849 | abd_t *xd, *yd; |
34dc7c2f | 850 | |
45d1cae3 | 851 | ASSERT(ntgts == 2); |
34dc7c2f | 852 | ASSERT(x < y); |
b2255edc BB |
853 | ASSERT(x >= rr->rr_firstdatacol); |
854 | ASSERT(y < rr->rr_cols); | |
34dc7c2f | 855 | |
b2255edc | 856 | ASSERT(rr->rr_col[x].rc_size >= rr->rr_col[y].rc_size); |
34dc7c2f BB |
857 | |
858 | /* | |
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. | |
864 | */ | |
b2255edc BB |
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; | |
34dc7c2f | 869 | |
b2255edc BB |
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; | |
34dc7c2f | 876 | |
b2255edc | 877 | vdev_raidz_generate_parity_pq(rr); |
34dc7c2f | 878 | |
b2255edc BB |
879 | rr->rr_col[x].rc_size = xsize; |
880 | rr->rr_col[y].rc_size = ysize; | |
34dc7c2f | 881 | |
a6255b7f DQ |
882 | p = abd_to_buf(pdata); |
883 | q = abd_to_buf(qdata); | |
b2255edc BB |
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; | |
34dc7c2f BB |
888 | |
889 | /* | |
890 | * We now have: | |
891 | * Pxy = P + D_x + D_y | |
892 | * Qxy = Q + 2^(ndevs - 1 - x) * D_x + 2^(ndevs - 1 - y) * D_y | |
893 | * | |
894 | * We can then solve for D_x: | |
895 | * D_x = A * (P + Pxy) + B * (Q + Qxy) | |
896 | * where | |
897 | * A = 2^(x - y) * (2^(x - y) + 1)^-1 | |
898 | * B = 2^(ndevs - 1 - x) * (2^(x - y) + 1)^-1 | |
899 | * | |
900 | * With D_x in hand, we can easily solve for D_y: | |
901 | * D_y = P + Pxy + D_x | |
902 | */ | |
903 | ||
904 | a = vdev_raidz_pow2[255 + x - y]; | |
b2255edc | 905 | b = vdev_raidz_pow2[255 - (rr->rr_cols - 1 - x)]; |
34dc7c2f BB |
906 | tmp = 255 - vdev_raidz_log2[a ^ 1]; |
907 | ||
908 | aexp = vdev_raidz_log2[vdev_raidz_exp2(a, tmp)]; | |
909 | bexp = vdev_raidz_log2[vdev_raidz_exp2(b, tmp)]; | |
910 | ||
a6255b7f | 911 | ASSERT3U(xsize, >=, ysize); |
1c27024e | 912 | struct reconst_pq_struct rpq = { p, q, pxy, qxy, aexp, bexp }; |
34dc7c2f | 913 | |
a6255b7f DQ |
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); | |
34dc7c2f | 918 | |
b2255edc BB |
919 | abd_free(rr->rr_col[VDEV_RAIDZ_P].rc_abd); |
920 | abd_free(rr->rr_col[VDEV_RAIDZ_Q].rc_abd); | |
34dc7c2f BB |
921 | |
922 | /* | |
923 | * Restore the saved parity data. | |
924 | */ | |
b2255edc BB |
925 | rr->rr_col[VDEV_RAIDZ_P].rc_abd = pdata; |
926 | rr->rr_col[VDEV_RAIDZ_Q].rc_abd = qdata; | |
45d1cae3 BB |
927 | } |
928 | ||
45d1cae3 BB |
929 | /* |
930 | * In the general case of reconstruction, we must solve the system of linear | |
dd4bc569 | 931 | * equations defined by the coefficients used to generate parity as well as |
45d1cae3 BB |
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): | |
935 | * | |
936 | * __ __ __ __ | |
937 | * | | __ __ | p_0 | | |
938 | * | V | | D_0 | | p_m-1 | | |
939 | * | | x | : | = | d_0 | | |
940 | * | I | | D_n-1 | | : | | |
941 | * | | ~~ ~~ | d_n-1 | | |
942 | * ~~ ~~ ~~ ~~ | |
943 | * | |
944 | * I is simply a square identity matrix of size n, and V is a vandermonde | |
dd4bc569 | 945 | * matrix defined by the coefficients we chose for the various parity columns |
45d1cae3 BB |
946 | * (1, 2, 4). Note that these values were chosen both for simplicity, speedy |
947 | * computation as well as linear separability. | |
948 | * | |
949 | * __ __ __ __ | |
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 | | |
959 | * ~~ ~~ ~~ ~~ | |
960 | * | |
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: | |
969 | * __ __ | |
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 | | |
981 | * ~~ ~~ | |
982 | * __ __ | |
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 | | |
994 | * ~~ ~~ | |
995 | * | |
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. | |
999 | * __ __ | |
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 | | |
1008 | * ~~ ~~ | |
1009 | * __ __ | |
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 | | |
1018 | * ~~ ~~ | |
1019 | * __ __ | |
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 | | |
1028 | * ~~ ~~ | |
1029 | * __ __ | |
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 | | |
1038 | * ~~ ~~ | |
1039 | * __ __ | |
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 | | |
1048 | * ~~ ~~ | |
1049 | * __ __ | |
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 | | |
1058 | * ~~ ~~ | |
1059 | * __ __ | |
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 | | |
1068 | * ~~ ~~ | |
1069 | * | |
1070 | * We can then simply compute D = (V|I)'^-1 x (d|p)' to discover the values | |
1071 | * of the missing data. | |
1072 | * | |
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 | |
1078 | * targeted columns. | |
1079 | */ | |
45d1cae3 BB |
1080 | |
1081 | static void | |
b2255edc | 1082 | vdev_raidz_matrix_init(raidz_row_t *rr, int n, int nmap, int *map, |
45d1cae3 BB |
1083 | uint8_t **rows) |
1084 | { | |
1085 | int i, j; | |
1086 | int pow; | |
1087 | ||
b2255edc | 1088 | ASSERT(n == rr->rr_cols - rr->rr_firstdatacol); |
45d1cae3 BB |
1089 | |
1090 | /* | |
1091 | * Fill in the missing rows of interest. | |
1092 | */ | |
1093 | for (i = 0; i < nmap; i++) { | |
1094 | ASSERT3S(0, <=, map[i]); | |
1095 | ASSERT3S(map[i], <=, 2); | |
1096 | ||
1097 | pow = map[i] * n; | |
1098 | if (pow > 255) | |
1099 | pow -= 255; | |
1100 | ASSERT(pow <= 255); | |
1101 | ||
1102 | for (j = 0; j < n; j++) { | |
1103 | pow -= map[i]; | |
1104 | if (pow < 0) | |
1105 | pow += 255; | |
1106 | rows[i][j] = vdev_raidz_pow2[pow]; | |
1107 | } | |
1108 | } | |
1109 | } | |
1110 | ||
1111 | static void | |
b2255edc | 1112 | vdev_raidz_matrix_invert(raidz_row_t *rr, int n, int nmissing, int *missing, |
45d1cae3 BB |
1113 | uint8_t **rows, uint8_t **invrows, const uint8_t *used) |
1114 | { | |
1115 | int i, j, ii, jj; | |
1116 | uint8_t log; | |
1117 | ||
1118 | /* | |
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. | |
1122 | */ | |
1123 | for (i = 0; i < nmissing; i++) { | |
b2255edc | 1124 | ASSERT3S(used[i], <, rr->rr_firstdatacol); |
45d1cae3 BB |
1125 | } |
1126 | for (; i < n; i++) { | |
b2255edc | 1127 | ASSERT3S(used[i], >=, rr->rr_firstdatacol); |
45d1cae3 BB |
1128 | } |
1129 | ||
1130 | /* | |
1131 | * First initialize the storage where we'll compute the inverse rows. | |
1132 | */ | |
1133 | for (i = 0; i < nmissing; i++) { | |
1134 | for (j = 0; j < n; j++) { | |
1135 | invrows[i][j] = (i == j) ? 1 : 0; | |
1136 | } | |
1137 | } | |
1138 | ||
1139 | /* | |
1140 | * Subtract all trivial rows from the rows of consequence. | |
1141 | */ | |
1142 | for (i = 0; i < nmissing; i++) { | |
1143 | for (j = nmissing; j < n; j++) { | |
b2255edc BB |
1144 | ASSERT3U(used[j], >=, rr->rr_firstdatacol); |
1145 | jj = used[j] - rr->rr_firstdatacol; | |
45d1cae3 BB |
1146 | ASSERT3S(jj, <, n); |
1147 | invrows[i][j] = rows[i][jj]; | |
1148 | rows[i][jj] = 0; | |
1149 | } | |
1150 | } | |
1151 | ||
1152 | /* | |
1153 | * For each of the rows of interest, we must normalize it and subtract | |
1154 | * a multiple of it from the other rows. | |
1155 | */ | |
1156 | for (i = 0; i < nmissing; i++) { | |
1157 | for (j = 0; j < missing[i]; j++) { | |
c99c9001 | 1158 | ASSERT0(rows[i][j]); |
45d1cae3 BB |
1159 | } |
1160 | ASSERT3U(rows[i][missing[i]], !=, 0); | |
1161 | ||
1162 | /* | |
1163 | * Compute the inverse of the first element and multiply each | |
1164 | * element in the row by that value. | |
1165 | */ | |
1166 | log = 255 - vdev_raidz_log2[rows[i][missing[i]]]; | |
1167 | ||
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); | |
1171 | } | |
1172 | ||
1173 | for (ii = 0; ii < nmissing; ii++) { | |
1174 | if (i == ii) | |
1175 | continue; | |
1176 | ||
1177 | ASSERT3U(rows[ii][missing[i]], !=, 0); | |
1178 | ||
1179 | log = vdev_raidz_log2[rows[ii][missing[i]]]; | |
1180 | ||
1181 | for (j = 0; j < n; j++) { | |
1182 | rows[ii][j] ^= | |
1183 | vdev_raidz_exp2(rows[i][j], log); | |
1184 | invrows[ii][j] ^= | |
1185 | vdev_raidz_exp2(invrows[i][j], log); | |
1186 | } | |
1187 | } | |
1188 | } | |
1189 | ||
1190 | /* | |
1191 | * Verify that the data that is left in the rows are properly part of | |
1192 | * an identity matrix. | |
1193 | */ | |
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); | |
1198 | } else { | |
c99c9001 | 1199 | ASSERT0(rows[i][j]); |
45d1cae3 BB |
1200 | } |
1201 | } | |
1202 | } | |
1203 | } | |
1204 | ||
1205 | static void | |
b2255edc | 1206 | vdev_raidz_matrix_reconstruct(raidz_row_t *rr, int n, int nmissing, |
45d1cae3 BB |
1207 | int *missing, uint8_t **invrows, const uint8_t *used) |
1208 | { | |
1209 | int i, j, x, cc, c; | |
1210 | uint8_t *src; | |
1211 | uint64_t ccount; | |
689f093e GN |
1212 | uint8_t *dst[VDEV_RAIDZ_MAXPARITY] = { NULL }; |
1213 | uint64_t dcount[VDEV_RAIDZ_MAXPARITY] = { 0 }; | |
a117a6d6 GW |
1214 | uint8_t log = 0; |
1215 | uint8_t val; | |
45d1cae3 BB |
1216 | int ll; |
1217 | uint8_t *invlog[VDEV_RAIDZ_MAXPARITY]; | |
1218 | uint8_t *p, *pp; | |
1219 | size_t psize; | |
1220 | ||
1221 | psize = sizeof (invlog[0][0]) * n * nmissing; | |
79c76d5b | 1222 | p = kmem_alloc(psize, KM_SLEEP); |
45d1cae3 BB |
1223 | |
1224 | for (pp = p, i = 0; i < nmissing; i++) { | |
1225 | invlog[i] = pp; | |
1226 | pp += n; | |
1227 | } | |
1228 | ||
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]]; | |
1233 | } | |
1234 | } | |
1235 | ||
1236 | for (i = 0; i < n; i++) { | |
1237 | c = used[i]; | |
b2255edc | 1238 | ASSERT3U(c, <, rr->rr_cols); |
45d1cae3 | 1239 | |
b2255edc BB |
1240 | ccount = rr->rr_col[c].rc_size; |
1241 | ASSERT(ccount >= rr->rr_col[missing[0]].rc_size || i > 0); | |
1242 | if (ccount == 0) | |
1243 | continue; | |
1244 | src = abd_to_buf(rr->rr_col[c].rc_abd); | |
45d1cae3 | 1245 | for (j = 0; j < nmissing; j++) { |
b2255edc BB |
1246 | cc = missing[j] + rr->rr_firstdatacol; |
1247 | ASSERT3U(cc, >=, rr->rr_firstdatacol); | |
1248 | ASSERT3U(cc, <, rr->rr_cols); | |
45d1cae3 BB |
1249 | ASSERT3U(cc, !=, c); |
1250 | ||
b2255edc BB |
1251 | dcount[j] = rr->rr_col[cc].rc_size; |
1252 | if (dcount[j] != 0) | |
1253 | dst[j] = abd_to_buf(rr->rr_col[cc].rc_abd); | |
45d1cae3 BB |
1254 | } |
1255 | ||
45d1cae3 BB |
1256 | for (x = 0; x < ccount; x++, src++) { |
1257 | if (*src != 0) | |
1258 | log = vdev_raidz_log2[*src]; | |
1259 | ||
1260 | for (cc = 0; cc < nmissing; cc++) { | |
1261 | if (x >= dcount[cc]) | |
1262 | continue; | |
1263 | ||
1264 | if (*src == 0) { | |
1265 | val = 0; | |
1266 | } else { | |
1267 | if ((ll = log + invlog[cc][i]) >= 255) | |
1268 | ll -= 255; | |
1269 | val = vdev_raidz_pow2[ll]; | |
1270 | } | |
1271 | ||
1272 | if (i == 0) | |
1273 | dst[cc][x] = val; | |
1274 | else | |
1275 | dst[cc][x] ^= val; | |
1276 | } | |
1277 | } | |
1278 | } | |
1279 | ||
1280 | kmem_free(p, psize); | |
1281 | } | |
1282 | ||
46df6e98 | 1283 | static void |
b2255edc | 1284 | vdev_raidz_reconstruct_general(raidz_row_t *rr, int *tgts, int ntgts) |
45d1cae3 BB |
1285 | { |
1286 | int n, i, c, t, tt; | |
1287 | int nmissing_rows; | |
1288 | int missing_rows[VDEV_RAIDZ_MAXPARITY]; | |
1289 | int parity_map[VDEV_RAIDZ_MAXPARITY]; | |
45d1cae3 BB |
1290 | uint8_t *p, *pp; |
1291 | size_t psize; | |
45d1cae3 BB |
1292 | uint8_t *rows[VDEV_RAIDZ_MAXPARITY]; |
1293 | uint8_t *invrows[VDEV_RAIDZ_MAXPARITY]; | |
1294 | uint8_t *used; | |
1295 | ||
a6255b7f DQ |
1296 | abd_t **bufs = NULL; |
1297 | ||
a6255b7f DQ |
1298 | /* |
1299 | * Matrix reconstruction can't use scatter ABDs yet, so we allocate | |
b2255edc | 1300 | * temporary linear ABDs if any non-linear ABDs are found. |
a6255b7f | 1301 | */ |
b2255edc BB |
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 *), | |
1305 | KM_PUSHPAGE); | |
1306 | ||
1307 | for (c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { | |
1308 | raidz_col_t *col = &rr->rr_col[c]; | |
1309 | ||
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], | |
1315 | col->rc_size); | |
1316 | } | |
1317 | } | |
a6255b7f | 1318 | |
b2255edc | 1319 | break; |
a6255b7f DQ |
1320 | } |
1321 | } | |
45d1cae3 | 1322 | |
b2255edc | 1323 | n = rr->rr_cols - rr->rr_firstdatacol; |
45d1cae3 BB |
1324 | |
1325 | /* | |
1326 | * Figure out which data columns are missing. | |
1327 | */ | |
1328 | nmissing_rows = 0; | |
1329 | for (t = 0; t < ntgts; t++) { | |
b2255edc | 1330 | if (tgts[t] >= rr->rr_firstdatacol) { |
45d1cae3 | 1331 | missing_rows[nmissing_rows++] = |
b2255edc | 1332 | tgts[t] - rr->rr_firstdatacol; |
45d1cae3 BB |
1333 | } |
1334 | } | |
1335 | ||
1336 | /* | |
1337 | * Figure out which parity columns to use to help generate the missing | |
1338 | * data columns. | |
1339 | */ | |
1340 | for (tt = 0, c = 0, i = 0; i < nmissing_rows; c++) { | |
1341 | ASSERT(tt < ntgts); | |
b2255edc | 1342 | ASSERT(c < rr->rr_firstdatacol); |
45d1cae3 BB |
1343 | |
1344 | /* | |
1345 | * Skip any targeted parity columns. | |
1346 | */ | |
1347 | if (c == tgts[tt]) { | |
1348 | tt++; | |
1349 | continue; | |
1350 | } | |
1351 | ||
45d1cae3 BB |
1352 | parity_map[i] = c; |
1353 | i++; | |
1354 | } | |
1355 | ||
45d1cae3 BB |
1356 | psize = (sizeof (rows[0][0]) + sizeof (invrows[0][0])) * |
1357 | nmissing_rows * n + sizeof (used[0]) * n; | |
79c76d5b | 1358 | p = kmem_alloc(psize, KM_SLEEP); |
45d1cae3 BB |
1359 | |
1360 | for (pp = p, i = 0; i < nmissing_rows; i++) { | |
1361 | rows[i] = pp; | |
1362 | pp += n; | |
1363 | invrows[i] = pp; | |
1364 | pp += n; | |
1365 | } | |
1366 | used = pp; | |
1367 | ||
1368 | for (i = 0; i < nmissing_rows; i++) { | |
1369 | used[i] = parity_map[i]; | |
1370 | } | |
1371 | ||
b2255edc | 1372 | for (tt = 0, c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { |
45d1cae3 | 1373 | if (tt < nmissing_rows && |
b2255edc | 1374 | c == missing_rows[tt] + rr->rr_firstdatacol) { |
45d1cae3 BB |
1375 | tt++; |
1376 | continue; | |
1377 | } | |
1378 | ||
1379 | ASSERT3S(i, <, n); | |
1380 | used[i] = c; | |
1381 | i++; | |
1382 | } | |
1383 | ||
1384 | /* | |
1385 | * Initialize the interesting rows of the matrix. | |
1386 | */ | |
b2255edc | 1387 | vdev_raidz_matrix_init(rr, n, nmissing_rows, parity_map, rows); |
45d1cae3 BB |
1388 | |
1389 | /* | |
1390 | * Invert the matrix. | |
1391 | */ | |
b2255edc | 1392 | vdev_raidz_matrix_invert(rr, n, nmissing_rows, missing_rows, rows, |
45d1cae3 BB |
1393 | invrows, used); |
1394 | ||
1395 | /* | |
1396 | * Reconstruct the missing data using the generated matrix. | |
1397 | */ | |
b2255edc | 1398 | vdev_raidz_matrix_reconstruct(rr, n, nmissing_rows, missing_rows, |
45d1cae3 BB |
1399 | invrows, used); |
1400 | ||
1401 | kmem_free(p, psize); | |
1402 | ||
a6255b7f DQ |
1403 | /* |
1404 | * copy back from temporary linear abds and free them | |
1405 | */ | |
1406 | if (bufs) { | |
b2255edc BB |
1407 | for (c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { |
1408 | raidz_col_t *col = &rr->rr_col[c]; | |
a6255b7f | 1409 | |
b2255edc BB |
1410 | if (bufs[c] != NULL) { |
1411 | abd_copy(bufs[c], col->rc_abd, col->rc_size); | |
1412 | abd_free(col->rc_abd); | |
1413 | } | |
a6255b7f DQ |
1414 | col->rc_abd = bufs[c]; |
1415 | } | |
b2255edc | 1416 | kmem_free(bufs, rr->rr_cols * sizeof (abd_t *)); |
a6255b7f | 1417 | } |
34dc7c2f BB |
1418 | } |
1419 | ||
46df6e98 | 1420 | static void |
b2255edc BB |
1421 | vdev_raidz_reconstruct_row(raidz_map_t *rm, raidz_row_t *rr, |
1422 | const int *t, int nt) | |
45d1cae3 BB |
1423 | { |
1424 | int tgts[VDEV_RAIDZ_MAXPARITY], *dt; | |
1425 | int ntgts; | |
c9187d86 | 1426 | int i, c, ret; |
45d1cae3 BB |
1427 | int nbadparity, nbaddata; |
1428 | int parity_valid[VDEV_RAIDZ_MAXPARITY]; | |
1429 | ||
b2255edc BB |
1430 | nbadparity = rr->rr_firstdatacol; |
1431 | nbaddata = rr->rr_cols - nbadparity; | |
45d1cae3 | 1432 | ntgts = 0; |
b2255edc BB |
1433 | for (i = 0, c = 0; c < rr->rr_cols; c++) { |
1434 | if (c < rr->rr_firstdatacol) | |
45d1cae3 BB |
1435 | parity_valid[c] = B_FALSE; |
1436 | ||
1437 | if (i < nt && c == t[i]) { | |
1438 | tgts[ntgts++] = c; | |
1439 | i++; | |
b2255edc | 1440 | } else if (rr->rr_col[c].rc_error != 0) { |
45d1cae3 | 1441 | tgts[ntgts++] = c; |
b2255edc | 1442 | } else if (c >= rr->rr_firstdatacol) { |
45d1cae3 BB |
1443 | nbaddata--; |
1444 | } else { | |
1445 | parity_valid[c] = B_TRUE; | |
1446 | nbadparity--; | |
1447 | } | |
1448 | } | |
1449 | ||
1450 | ASSERT(ntgts >= nt); | |
1451 | ASSERT(nbaddata >= 0); | |
1452 | ASSERT(nbaddata + nbadparity == ntgts); | |
1453 | ||
1454 | dt = &tgts[nbadparity]; | |
1455 | ||
c9187d86 | 1456 | /* Reconstruct using the new math implementation */ |
b2255edc | 1457 | ret = vdev_raidz_math_reconstruct(rm, rr, parity_valid, dt, nbaddata); |
c9187d86 | 1458 | if (ret != RAIDZ_ORIGINAL_IMPL) |
46df6e98 | 1459 | return; |
ab9f4b0b | 1460 | |
45d1cae3 BB |
1461 | /* |
1462 | * See if we can use any of our optimized reconstruction routines. | |
1463 | */ | |
ab9f4b0b GN |
1464 | switch (nbaddata) { |
1465 | case 1: | |
46df6e98 MA |
1466 | if (parity_valid[VDEV_RAIDZ_P]) { |
1467 | vdev_raidz_reconstruct_p(rr, dt, 1); | |
1468 | return; | |
1469 | } | |
45d1cae3 | 1470 | |
b2255edc | 1471 | ASSERT(rr->rr_firstdatacol > 1); |
45d1cae3 | 1472 | |
46df6e98 MA |
1473 | if (parity_valid[VDEV_RAIDZ_Q]) { |
1474 | vdev_raidz_reconstruct_q(rr, dt, 1); | |
1475 | return; | |
1476 | } | |
45d1cae3 | 1477 | |
b2255edc | 1478 | ASSERT(rr->rr_firstdatacol > 2); |
ab9f4b0b | 1479 | break; |
45d1cae3 | 1480 | |
ab9f4b0b | 1481 | case 2: |
b2255edc | 1482 | ASSERT(rr->rr_firstdatacol > 1); |
45d1cae3 | 1483 | |
ab9f4b0b | 1484 | if (parity_valid[VDEV_RAIDZ_P] && |
46df6e98 MA |
1485 | parity_valid[VDEV_RAIDZ_Q]) { |
1486 | vdev_raidz_reconstruct_pq(rr, dt, 2); | |
1487 | return; | |
1488 | } | |
45d1cae3 | 1489 | |
b2255edc | 1490 | ASSERT(rr->rr_firstdatacol > 2); |
45d1cae3 | 1491 | |
ab9f4b0b | 1492 | break; |
45d1cae3 BB |
1493 | } |
1494 | ||
46df6e98 | 1495 | vdev_raidz_reconstruct_general(rr, tgts, ntgts); |
45d1cae3 | 1496 | } |
34dc7c2f BB |
1497 | |
1498 | static int | |
1bd201e7 | 1499 | vdev_raidz_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize, |
6fe3498c | 1500 | uint64_t *logical_ashift, uint64_t *physical_ashift) |
34dc7c2f | 1501 | { |
b2255edc BB |
1502 | vdev_raidz_t *vdrz = vd->vdev_tsd; |
1503 | uint64_t nparity = vdrz->vd_nparity; | |
45d1cae3 | 1504 | int c; |
34dc7c2f BB |
1505 | int lasterror = 0; |
1506 | int numerrors = 0; | |
1507 | ||
1508 | ASSERT(nparity > 0); | |
1509 | ||
1510 | if (nparity > VDEV_RAIDZ_MAXPARITY || | |
1511 | vd->vdev_children < nparity + 1) { | |
1512 | vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL; | |
2e528b49 | 1513 | return (SET_ERROR(EINVAL)); |
34dc7c2f BB |
1514 | } |
1515 | ||
45d1cae3 BB |
1516 | vdev_open_children(vd); |
1517 | ||
34dc7c2f | 1518 | for (c = 0; c < vd->vdev_children; c++) { |
b2255edc | 1519 | vdev_t *cvd = vd->vdev_child[c]; |
34dc7c2f | 1520 | |
45d1cae3 BB |
1521 | if (cvd->vdev_open_error != 0) { |
1522 | lasterror = cvd->vdev_open_error; | |
34dc7c2f BB |
1523 | numerrors++; |
1524 | continue; | |
1525 | } | |
1526 | ||
1527 | *asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1; | |
1bd201e7 | 1528 | *max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1; |
6fe3498c RM |
1529 | *logical_ashift = MAX(*logical_ashift, cvd->vdev_ashift); |
1530 | *physical_ashift = MAX(*physical_ashift, | |
1531 | cvd->vdev_physical_ashift); | |
34dc7c2f BB |
1532 | } |
1533 | ||
1534 | *asize *= vd->vdev_children; | |
1bd201e7 | 1535 | *max_asize *= vd->vdev_children; |
34dc7c2f BB |
1536 | |
1537 | if (numerrors > nparity) { | |
1538 | vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS; | |
1539 | return (lasterror); | |
1540 | } | |
1541 | ||
1542 | return (0); | |
1543 | } | |
1544 | ||
1545 | static void | |
1546 | vdev_raidz_close(vdev_t *vd) | |
1547 | { | |
b2255edc BB |
1548 | for (int c = 0; c < vd->vdev_children; c++) { |
1549 | if (vd->vdev_child[c] != NULL) | |
1550 | vdev_close(vd->vdev_child[c]); | |
1551 | } | |
34dc7c2f BB |
1552 | } |
1553 | ||
1554 | static uint64_t | |
1555 | vdev_raidz_asize(vdev_t *vd, uint64_t psize) | |
1556 | { | |
b2255edc | 1557 | vdev_raidz_t *vdrz = vd->vdev_tsd; |
34dc7c2f BB |
1558 | uint64_t asize; |
1559 | uint64_t ashift = vd->vdev_top->vdev_ashift; | |
b2255edc BB |
1560 | uint64_t cols = vdrz->vd_logical_width; |
1561 | uint64_t nparity = vdrz->vd_nparity; | |
34dc7c2f BB |
1562 | |
1563 | asize = ((psize - 1) >> ashift) + 1; | |
1564 | asize += nparity * ((asize + cols - nparity - 1) / (cols - nparity)); | |
1565 | asize = roundup(asize, nparity + 1) << ashift; | |
1566 | ||
1567 | return (asize); | |
1568 | } | |
1569 | ||
b2255edc BB |
1570 | /* |
1571 | * The allocatable space for a raidz vdev is N * sizeof(smallest child) | |
1572 | * so each child must provide at least 1/Nth of its asize. | |
1573 | */ | |
1574 | static uint64_t | |
1575 | vdev_raidz_min_asize(vdev_t *vd) | |
1576 | { | |
1577 | return ((vd->vdev_min_asize + vd->vdev_children - 1) / | |
1578 | vd->vdev_children); | |
1579 | } | |
1580 | ||
1581 | void | |
34dc7c2f BB |
1582 | vdev_raidz_child_done(zio_t *zio) |
1583 | { | |
1584 | raidz_col_t *rc = zio->io_private; | |
1585 | ||
345196be | 1586 | ASSERT3P(rc->rc_abd, !=, NULL); |
34dc7c2f BB |
1587 | rc->rc_error = zio->io_error; |
1588 | rc->rc_tried = 1; | |
1589 | rc->rc_skipped = 0; | |
1590 | } | |
1591 | ||
619f0976 | 1592 | static void |
b2255edc | 1593 | vdev_raidz_io_verify(vdev_t *vd, raidz_row_t *rr, int col) |
619f0976 GW |
1594 | { |
1595 | #ifdef ZFS_DEBUG | |
619f0976 GW |
1596 | vdev_t *tvd = vd->vdev_top; |
1597 | ||
b2255edc BB |
1598 | range_seg64_t logical_rs, physical_rs, remain_rs; |
1599 | logical_rs.rs_start = rr->rr_offset; | |
619f0976 | 1600 | logical_rs.rs_end = logical_rs.rs_start + |
b2255edc | 1601 | vdev_raidz_asize(vd, rr->rr_size); |
619f0976 | 1602 | |
b2255edc | 1603 | raidz_col_t *rc = &rr->rr_col[col]; |
619f0976 GW |
1604 | vdev_t *cvd = vd->vdev_child[rc->rc_devidx]; |
1605 | ||
b2255edc BB |
1606 | vdev_xlate(cvd, &logical_rs, &physical_rs, &remain_rs); |
1607 | ASSERT(vdev_xlate_is_empty(&remain_rs)); | |
619f0976 GW |
1608 | ASSERT3U(rc->rc_offset, ==, physical_rs.rs_start); |
1609 | ASSERT3U(rc->rc_offset, <, physical_rs.rs_end); | |
1610 | /* | |
1611 | * It would be nice to assert that rs_end is equal | |
1612 | * to rc_offset + rc_size but there might be an | |
1613 | * optional I/O at the end that is not accounted in | |
1614 | * rc_size. | |
1615 | */ | |
1616 | if (physical_rs.rs_end > rc->rc_offset + rc->rc_size) { | |
1617 | ASSERT3U(physical_rs.rs_end, ==, rc->rc_offset + | |
1618 | rc->rc_size + (1 << tvd->vdev_ashift)); | |
1619 | } else { | |
1620 | ASSERT3U(physical_rs.rs_end, ==, rc->rc_offset + rc->rc_size); | |
1621 | } | |
1622 | #endif | |
1623 | } | |
1624 | ||
98b25418 | 1625 | static void |
b2255edc | 1626 | vdev_raidz_io_start_write(zio_t *zio, raidz_row_t *rr, uint64_t ashift) |
34dc7c2f BB |
1627 | { |
1628 | vdev_t *vd = zio->io_vd; | |
b2255edc | 1629 | raidz_map_t *rm = zio->io_vsd; |
34dc7c2f | 1630 | |
b2255edc | 1631 | vdev_raidz_generate_parity_row(rm, rr); |
34dc7c2f | 1632 | |
345196be | 1633 | for (int c = 0; c < rr->rr_scols; c++) { |
b2255edc | 1634 | raidz_col_t *rc = &rr->rr_col[c]; |
345196be | 1635 | vdev_t *cvd = vd->vdev_child[rc->rc_devidx]; |
619f0976 | 1636 | |
b2255edc BB |
1637 | /* Verify physical to logical translation */ |
1638 | vdev_raidz_io_verify(vd, rr, c); | |
34dc7c2f | 1639 | |
345196be BA |
1640 | if (rc->rc_size > 0) { |
1641 | ASSERT3P(rc->rc_abd, !=, NULL); | |
1642 | zio_nowait(zio_vdev_child_io(zio, NULL, cvd, | |
1643 | rc->rc_offset, rc->rc_abd, | |
1644 | abd_get_size(rc->rc_abd), zio->io_type, | |
1645 | zio->io_priority, 0, vdev_raidz_child_done, rc)); | |
1646 | } else { | |
1647 | /* | |
1648 | * Generate optional write for skip sector to improve | |
1649 | * aggregation contiguity. | |
1650 | */ | |
1651 | ASSERT3P(rc->rc_abd, ==, NULL); | |
1652 | zio_nowait(zio_vdev_child_io(zio, NULL, cvd, | |
1653 | rc->rc_offset, NULL, 1ULL << ashift, | |
1654 | zio->io_type, zio->io_priority, | |
1655 | ZIO_FLAG_NODATA | ZIO_FLAG_OPTIONAL, NULL, | |
1656 | NULL)); | |
1657 | } | |
34dc7c2f | 1658 | } |
b2255edc | 1659 | } |
34dc7c2f | 1660 | |
b2255edc BB |
1661 | static void |
1662 | vdev_raidz_io_start_read(zio_t *zio, raidz_row_t *rr) | |
1663 | { | |
1664 | vdev_t *vd = zio->io_vd; | |
34dc7c2f BB |
1665 | |
1666 | /* | |
1667 | * Iterate over the columns in reverse order so that we hit the parity | |
45d1cae3 | 1668 | * last -- any errors along the way will force us to read the parity. |
34dc7c2f | 1669 | */ |
b2255edc BB |
1670 | for (int c = rr->rr_cols - 1; c >= 0; c--) { |
1671 | raidz_col_t *rc = &rr->rr_col[c]; | |
1672 | if (rc->rc_size == 0) | |
1673 | continue; | |
1674 | vdev_t *cvd = vd->vdev_child[rc->rc_devidx]; | |
34dc7c2f | 1675 | if (!vdev_readable(cvd)) { |
b2255edc BB |
1676 | if (c >= rr->rr_firstdatacol) |
1677 | rr->rr_missingdata++; | |
34dc7c2f | 1678 | else |
b2255edc | 1679 | rr->rr_missingparity++; |
2e528b49 | 1680 | rc->rc_error = SET_ERROR(ENXIO); |
34dc7c2f BB |
1681 | rc->rc_tried = 1; /* don't even try */ |
1682 | rc->rc_skipped = 1; | |
1683 | continue; | |
1684 | } | |
428870ff | 1685 | if (vdev_dtl_contains(cvd, DTL_MISSING, zio->io_txg, 1)) { |
b2255edc BB |
1686 | if (c >= rr->rr_firstdatacol) |
1687 | rr->rr_missingdata++; | |
34dc7c2f | 1688 | else |
b2255edc | 1689 | rr->rr_missingparity++; |
2e528b49 | 1690 | rc->rc_error = SET_ERROR(ESTALE); |
34dc7c2f BB |
1691 | rc->rc_skipped = 1; |
1692 | continue; | |
1693 | } | |
b2255edc | 1694 | if (c >= rr->rr_firstdatacol || rr->rr_missingdata > 0 || |
9babb374 | 1695 | (zio->io_flags & (ZIO_FLAG_SCRUB | ZIO_FLAG_RESILVER))) { |
34dc7c2f | 1696 | zio_nowait(zio_vdev_child_io(zio, NULL, cvd, |
a6255b7f | 1697 | rc->rc_offset, rc->rc_abd, rc->rc_size, |
b128c09f | 1698 | zio->io_type, zio->io_priority, 0, |
34dc7c2f BB |
1699 | vdev_raidz_child_done, rc)); |
1700 | } | |
1701 | } | |
b2255edc BB |
1702 | } |
1703 | ||
1704 | /* | |
1705 | * Start an IO operation on a RAIDZ VDev | |
1706 | * | |
1707 | * Outline: | |
1708 | * - For write operations: | |
1709 | * 1. Generate the parity data | |
1710 | * 2. Create child zio write operations to each column's vdev, for both | |
1711 | * data and parity. | |
1712 | * 3. If the column skips any sectors for padding, create optional dummy | |
1713 | * write zio children for those areas to improve aggregation continuity. | |
1714 | * - For read operations: | |
1715 | * 1. Create child zio read operations to each data column's vdev to read | |
1716 | * the range of data required for zio. | |
1717 | * 2. If this is a scrub or resilver operation, or if any of the data | |
1718 | * vdevs have had errors, then create zio read operations to the parity | |
1719 | * columns' VDevs as well. | |
1720 | */ | |
1721 | static void | |
1722 | vdev_raidz_io_start(zio_t *zio) | |
1723 | { | |
1724 | vdev_t *vd = zio->io_vd; | |
1725 | vdev_t *tvd = vd->vdev_top; | |
1726 | vdev_raidz_t *vdrz = vd->vdev_tsd; | |
b2255edc | 1727 | |
330c6c05 | 1728 | raidz_map_t *rm = vdev_raidz_map_alloc(zio, tvd->vdev_ashift, |
b2255edc | 1729 | vdrz->vd_logical_width, vdrz->vd_nparity); |
330c6c05 MA |
1730 | zio->io_vsd = rm; |
1731 | zio->io_vsd_ops = &vdev_raidz_vsd_ops; | |
b2255edc BB |
1732 | |
1733 | /* | |
1734 | * Until raidz expansion is implemented all maps for a raidz vdev | |
1735 | * contain a single row. | |
1736 | */ | |
1737 | ASSERT3U(rm->rm_nrows, ==, 1); | |
1738 | raidz_row_t *rr = rm->rm_row[0]; | |
1739 | ||
b2255edc BB |
1740 | if (zio->io_type == ZIO_TYPE_WRITE) { |
1741 | vdev_raidz_io_start_write(zio, rr, tvd->vdev_ashift); | |
1742 | } else { | |
1743 | ASSERT(zio->io_type == ZIO_TYPE_READ); | |
1744 | vdev_raidz_io_start_read(zio, rr); | |
1745 | } | |
34dc7c2f | 1746 | |
98b25418 | 1747 | zio_execute(zio); |
34dc7c2f BB |
1748 | } |
1749 | ||
1750 | /* | |
1751 | * Report a checksum error for a child of a RAID-Z device. | |
1752 | */ | |
3c80e074 BB |
1753 | void |
1754 | vdev_raidz_checksum_error(zio_t *zio, raidz_col_t *rc, abd_t *bad_data) | |
34dc7c2f BB |
1755 | { |
1756 | vdev_t *vd = zio->io_vd->vdev_child[rc->rc_devidx]; | |
34dc7c2f | 1757 | |
b2255edc BB |
1758 | if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE) && |
1759 | zio->io_priority != ZIO_PRIORITY_REBUILD) { | |
428870ff BB |
1760 | zio_bad_cksum_t zbc; |
1761 | raidz_map_t *rm = zio->io_vsd; | |
1762 | ||
428870ff BB |
1763 | zbc.zbc_has_cksum = 0; |
1764 | zbc.zbc_injected = rm->rm_ecksuminjected; | |
1765 | ||
03e02e5b | 1766 | (void) zfs_ereport_post_checksum(zio->io_spa, vd, |
b5256303 TC |
1767 | &zio->io_bookmark, zio, rc->rc_offset, rc->rc_size, |
1768 | rc->rc_abd, bad_data, &zbc); | |
03e02e5b DB |
1769 | mutex_enter(&vd->vdev_stat_lock); |
1770 | vd->vdev_stat.vs_checksum_errors++; | |
1771 | mutex_exit(&vd->vdev_stat_lock); | |
34dc7c2f | 1772 | } |
428870ff BB |
1773 | } |
1774 | ||
1775 | /* | |
1776 | * We keep track of whether or not there were any injected errors, so that | |
1777 | * any ereports we generate can note it. | |
1778 | */ | |
1779 | static int | |
1780 | raidz_checksum_verify(zio_t *zio) | |
1781 | { | |
861166b0 | 1782 | zio_bad_cksum_t zbc = {{{0}}}; |
428870ff BB |
1783 | raidz_map_t *rm = zio->io_vsd; |
1784 | ||
1c27024e | 1785 | int ret = zio_checksum_error(zio, &zbc); |
428870ff BB |
1786 | if (ret != 0 && zbc.zbc_injected != 0) |
1787 | rm->rm_ecksuminjected = 1; | |
34dc7c2f | 1788 | |
428870ff | 1789 | return (ret); |
34dc7c2f BB |
1790 | } |
1791 | ||
1792 | /* | |
1793 | * Generate the parity from the data columns. If we tried and were able to | |
1794 | * read the parity without error, verify that the generated parity matches the | |
1795 | * data we read. If it doesn't, we fire off a checksum error. Return the | |
b2255edc | 1796 | * number of such failures. |
34dc7c2f BB |
1797 | */ |
1798 | static int | |
b2255edc | 1799 | raidz_parity_verify(zio_t *zio, raidz_row_t *rr) |
34dc7c2f | 1800 | { |
84c07ada | 1801 | abd_t *orig[VDEV_RAIDZ_MAXPARITY]; |
34dc7c2f | 1802 | int c, ret = 0; |
b2255edc | 1803 | raidz_map_t *rm = zio->io_vsd; |
34dc7c2f BB |
1804 | raidz_col_t *rc; |
1805 | ||
3c67d83a TH |
1806 | blkptr_t *bp = zio->io_bp; |
1807 | enum zio_checksum checksum = (bp == NULL ? zio->io_prop.zp_checksum : | |
1808 | (BP_IS_GANG(bp) ? ZIO_CHECKSUM_GANG_HEADER : BP_GET_CHECKSUM(bp))); | |
1809 | ||
1810 | if (checksum == ZIO_CHECKSUM_NOPARITY) | |
1811 | return (ret); | |
1812 | ||
b2255edc BB |
1813 | for (c = 0; c < rr->rr_firstdatacol; c++) { |
1814 | rc = &rr->rr_col[c]; | |
34dc7c2f BB |
1815 | if (!rc->rc_tried || rc->rc_error != 0) |
1816 | continue; | |
84c07ada GN |
1817 | |
1818 | orig[c] = abd_alloc_sametype(rc->rc_abd, rc->rc_size); | |
1819 | abd_copy(orig[c], rc->rc_abd, rc->rc_size); | |
34dc7c2f BB |
1820 | } |
1821 | ||
3c80e074 BB |
1822 | /* |
1823 | * Verify any empty sectors are zero filled to ensure the parity | |
1824 | * is calculated correctly even if these non-data sectors are damaged. | |
1825 | */ | |
1826 | if (rr->rr_nempty && rr->rr_abd_empty != NULL) | |
1827 | ret += vdev_draid_map_verify_empty(zio, rr); | |
1828 | ||
b2255edc BB |
1829 | /* |
1830 | * Regenerates parity even for !tried||rc_error!=0 columns. This | |
1831 | * isn't harmful but it does have the side effect of fixing stuff | |
1832 | * we didn't realize was necessary (i.e. even if we return 0). | |
1833 | */ | |
1834 | vdev_raidz_generate_parity_row(rm, rr); | |
1835 | ||
1836 | for (c = 0; c < rr->rr_firstdatacol; c++) { | |
1837 | rc = &rr->rr_col[c]; | |
34dc7c2f | 1838 | |
34dc7c2f BB |
1839 | if (!rc->rc_tried || rc->rc_error != 0) |
1840 | continue; | |
b2255edc | 1841 | |
84c07ada | 1842 | if (abd_cmp(orig[c], rc->rc_abd) != 0) { |
3c80e074 | 1843 | vdev_raidz_checksum_error(zio, rc, orig[c]); |
2e528b49 | 1844 | rc->rc_error = SET_ERROR(ECKSUM); |
34dc7c2f BB |
1845 | ret++; |
1846 | } | |
84c07ada | 1847 | abd_free(orig[c]); |
34dc7c2f BB |
1848 | } |
1849 | ||
1850 | return (ret); | |
1851 | } | |
1852 | ||
34dc7c2f | 1853 | static int |
b2255edc | 1854 | vdev_raidz_worst_error(raidz_row_t *rr) |
b128c09f | 1855 | { |
1c27024e | 1856 | int error = 0; |
b128c09f | 1857 | |
b2255edc BB |
1858 | for (int c = 0; c < rr->rr_cols; c++) |
1859 | error = zio_worst_error(error, rr->rr_col[c].rc_error); | |
b128c09f BB |
1860 | |
1861 | return (error); | |
1862 | } | |
1863 | ||
b2255edc BB |
1864 | static void |
1865 | vdev_raidz_io_done_verified(zio_t *zio, raidz_row_t *rr) | |
45d1cae3 | 1866 | { |
b2255edc BB |
1867 | int unexpected_errors = 0; |
1868 | int parity_errors = 0; | |
1869 | int parity_untried = 0; | |
1870 | int data_errors = 0; | |
45d1cae3 | 1871 | |
b2255edc BB |
1872 | ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ); |
1873 | ||
1874 | for (int c = 0; c < rr->rr_cols; c++) { | |
1875 | raidz_col_t *rc = &rr->rr_col[c]; | |
1876 | ||
1877 | if (rc->rc_error) { | |
1878 | if (c < rr->rr_firstdatacol) | |
1879 | parity_errors++; | |
1880 | else | |
1881 | data_errors++; | |
1882 | ||
1883 | if (!rc->rc_skipped) | |
1884 | unexpected_errors++; | |
1885 | } else if (c < rr->rr_firstdatacol && !rc->rc_tried) { | |
1886 | parity_untried++; | |
1887 | } | |
1888 | } | |
45d1cae3 BB |
1889 | |
1890 | /* | |
b2255edc BB |
1891 | * If we read more parity disks than were used for |
1892 | * reconstruction, confirm that the other parity disks produced | |
1893 | * correct data. | |
1894 | * | |
1895 | * Note that we also regenerate parity when resilvering so we | |
1896 | * can write it out to failed devices later. | |
45d1cae3 | 1897 | */ |
b2255edc BB |
1898 | if (parity_errors + parity_untried < |
1899 | rr->rr_firstdatacol - data_errors || | |
1900 | (zio->io_flags & ZIO_FLAG_RESILVER)) { | |
1901 | int n = raidz_parity_verify(zio, rr); | |
1902 | unexpected_errors += n; | |
b2255edc | 1903 | } |
45d1cae3 | 1904 | |
b2255edc BB |
1905 | if (zio->io_error == 0 && spa_writeable(zio->io_spa) && |
1906 | (unexpected_errors > 0 || (zio->io_flags & ZIO_FLAG_RESILVER))) { | |
45d1cae3 | 1907 | /* |
b2255edc | 1908 | * Use the good data we have in hand to repair damaged children. |
45d1cae3 | 1909 | */ |
b2255edc BB |
1910 | for (int c = 0; c < rr->rr_cols; c++) { |
1911 | raidz_col_t *rc = &rr->rr_col[c]; | |
1912 | vdev_t *vd = zio->io_vd; | |
1913 | vdev_t *cvd = vd->vdev_child[rc->rc_devidx]; | |
1914 | ||
8fb577ae BB |
1915 | if (!rc->rc_allow_repair) { |
1916 | continue; | |
1917 | } else if (!rc->rc_force_repair && | |
1918 | (rc->rc_error == 0 || rc->rc_size == 0)) { | |
b2255edc | 1919 | continue; |
45d1cae3 BB |
1920 | } |
1921 | ||
b2255edc BB |
1922 | zio_nowait(zio_vdev_child_io(zio, NULL, cvd, |
1923 | rc->rc_offset, rc->rc_abd, rc->rc_size, | |
1924 | ZIO_TYPE_WRITE, | |
1925 | zio->io_priority == ZIO_PRIORITY_REBUILD ? | |
1926 | ZIO_PRIORITY_REBUILD : ZIO_PRIORITY_ASYNC_WRITE, | |
1927 | ZIO_FLAG_IO_REPAIR | (unexpected_errors ? | |
1928 | ZIO_FLAG_SELF_HEAL : 0), NULL, NULL)); | |
1929 | } | |
1930 | } | |
1931 | } | |
1932 | ||
1933 | static void | |
1934 | raidz_restore_orig_data(raidz_map_t *rm) | |
1935 | { | |
1936 | for (int i = 0; i < rm->rm_nrows; i++) { | |
1937 | raidz_row_t *rr = rm->rm_row[i]; | |
1938 | for (int c = 0; c < rr->rr_cols; c++) { | |
1939 | raidz_col_t *rc = &rr->rr_col[c]; | |
1940 | if (rc->rc_need_orig_restore) { | |
330c6c05 | 1941 | abd_copy(rc->rc_abd, |
b2255edc BB |
1942 | rc->rc_orig_data, rc->rc_size); |
1943 | rc->rc_need_orig_restore = B_FALSE; | |
45d1cae3 | 1944 | } |
b2255edc BB |
1945 | } |
1946 | } | |
1947 | } | |
1948 | ||
1949 | /* | |
1950 | * returns EINVAL if reconstruction of the block will not be possible | |
1951 | * returns ECKSUM if this specific reconstruction failed | |
1952 | * returns 0 on successful reconstruction | |
1953 | */ | |
1954 | static int | |
1955 | raidz_reconstruct(zio_t *zio, int *ltgts, int ntgts, int nparity) | |
1956 | { | |
1957 | raidz_map_t *rm = zio->io_vsd; | |
45d1cae3 | 1958 | |
b2255edc BB |
1959 | /* Reconstruct each row */ |
1960 | for (int r = 0; r < rm->rm_nrows; r++) { | |
1961 | raidz_row_t *rr = rm->rm_row[r]; | |
1962 | int my_tgts[VDEV_RAIDZ_MAXPARITY]; /* value is child id */ | |
1963 | int t = 0; | |
1964 | int dead = 0; | |
1965 | int dead_data = 0; | |
1966 | ||
1967 | for (int c = 0; c < rr->rr_cols; c++) { | |
1968 | raidz_col_t *rc = &rr->rr_col[c]; | |
1969 | ASSERT0(rc->rc_need_orig_restore); | |
1970 | if (rc->rc_error != 0) { | |
1971 | dead++; | |
1972 | if (c >= nparity) | |
1973 | dead_data++; | |
1974 | continue; | |
1975 | } | |
1976 | if (rc->rc_size == 0) | |
1977 | continue; | |
1978 | for (int lt = 0; lt < ntgts; lt++) { | |
1979 | if (rc->rc_devidx == ltgts[lt]) { | |
1980 | if (rc->rc_orig_data == NULL) { | |
1981 | rc->rc_orig_data = | |
330c6c05 MA |
1982 | abd_alloc_linear( |
1983 | rc->rc_size, B_TRUE); | |
1984 | abd_copy(rc->rc_orig_data, | |
b2255edc BB |
1985 | rc->rc_abd, rc->rc_size); |
1986 | } | |
1987 | rc->rc_need_orig_restore = B_TRUE; | |
1988 | ||
1989 | dead++; | |
1990 | if (c >= nparity) | |
1991 | dead_data++; | |
1992 | my_tgts[t++] = c; | |
1993 | break; | |
1994 | } | |
1995 | } | |
1996 | } | |
1997 | if (dead > nparity) { | |
1998 | /* reconstruction not possible */ | |
1999 | raidz_restore_orig_data(rm); | |
2000 | return (EINVAL); | |
45d1cae3 | 2001 | } |
b2255edc | 2002 | if (dead_data > 0) |
46df6e98 | 2003 | vdev_raidz_reconstruct_row(rm, rr, my_tgts, t); |
b2255edc | 2004 | } |
45d1cae3 | 2005 | |
b2255edc BB |
2006 | /* Check for success */ |
2007 | if (raidz_checksum_verify(zio) == 0) { | |
2008 | ||
2009 | /* Reconstruction succeeded - report errors */ | |
2010 | for (int i = 0; i < rm->rm_nrows; i++) { | |
2011 | raidz_row_t *rr = rm->rm_row[i]; | |
2012 | ||
2013 | for (int c = 0; c < rr->rr_cols; c++) { | |
2014 | raidz_col_t *rc = &rr->rr_col[c]; | |
2015 | if (rc->rc_need_orig_restore) { | |
2016 | /* | |
2017 | * Note: if this is a parity column, | |
2018 | * we don't really know if it's wrong. | |
2019 | * We need to let | |
2020 | * vdev_raidz_io_done_verified() check | |
2021 | * it, and if we set rc_error, it will | |
2022 | * think that it is a "known" error | |
2023 | * that doesn't need to be checked | |
2024 | * or corrected. | |
2025 | */ | |
2026 | if (rc->rc_error == 0 && | |
2027 | c >= rr->rr_firstdatacol) { | |
3c80e074 | 2028 | vdev_raidz_checksum_error(zio, |
330c6c05 | 2029 | rc, rc->rc_orig_data); |
b2255edc BB |
2030 | rc->rc_error = |
2031 | SET_ERROR(ECKSUM); | |
2032 | } | |
2033 | rc->rc_need_orig_restore = B_FALSE; | |
2034 | } | |
2035 | } | |
45d1cae3 | 2036 | |
b2255edc | 2037 | vdev_raidz_io_done_verified(zio, rr); |
45d1cae3 BB |
2038 | } |
2039 | ||
b2255edc | 2040 | zio_checksum_verified(zio); |
45d1cae3 | 2041 | |
b2255edc BB |
2042 | return (0); |
2043 | } | |
45d1cae3 | 2044 | |
b2255edc BB |
2045 | /* Reconstruction failed - restore original data */ |
2046 | raidz_restore_orig_data(rm); | |
2047 | return (ECKSUM); | |
2048 | } | |
45d1cae3 | 2049 | |
b2255edc BB |
2050 | /* |
2051 | * Iterate over all combinations of N bad vdevs and attempt a reconstruction. | |
2052 | * Note that the algorithm below is non-optimal because it doesn't take into | |
2053 | * account how reconstruction is actually performed. For example, with | |
2054 | * triple-parity RAID-Z the reconstruction procedure is the same if column 4 | |
2055 | * is targeted as invalid as if columns 1 and 4 are targeted since in both | |
2056 | * cases we'd only use parity information in column 0. | |
2057 | * | |
2058 | * The order that we find the various possible combinations of failed | |
2059 | * disks is dictated by these rules: | |
2060 | * - Examine each "slot" (the "i" in tgts[i]) | |
2061 | * - Try to increment this slot (tgts[i] = tgts[i] + 1) | |
2062 | * - if we can't increment because it runs into the next slot, | |
2063 | * reset our slot to the minimum, and examine the next slot | |
2064 | * | |
2065 | * For example, with a 6-wide RAIDZ3, and no known errors (so we have to choose | |
2066 | * 3 columns to reconstruct), we will generate the following sequence: | |
2067 | * | |
2068 | * STATE ACTION | |
2069 | * 0 1 2 special case: skip since these are all parity | |
2070 | * 0 1 3 first slot: reset to 0; middle slot: increment to 2 | |
2071 | * 0 2 3 first slot: increment to 1 | |
2072 | * 1 2 3 first: reset to 0; middle: reset to 1; last: increment to 4 | |
2073 | * 0 1 4 first: reset to 0; middle: increment to 2 | |
2074 | * 0 2 4 first: increment to 1 | |
2075 | * 1 2 4 first: reset to 0; middle: increment to 3 | |
2076 | * 0 3 4 first: increment to 1 | |
2077 | * 1 3 4 first: increment to 2 | |
2078 | * 2 3 4 first: reset to 0; middle: reset to 1; last: increment to 5 | |
2079 | * 0 1 5 first: reset to 0; middle: increment to 2 | |
2080 | * 0 2 5 first: increment to 1 | |
2081 | * 1 2 5 first: reset to 0; middle: increment to 3 | |
2082 | * 0 3 5 first: increment to 1 | |
2083 | * 1 3 5 first: increment to 2 | |
2084 | * 2 3 5 first: reset to 0; middle: increment to 4 | |
2085 | * 0 4 5 first: increment to 1 | |
2086 | * 1 4 5 first: increment to 2 | |
2087 | * 2 4 5 first: increment to 3 | |
2088 | * 3 4 5 done | |
2089 | * | |
bf169e9f | 2090 | * This strategy works for dRAID but is less efficient when there are a large |
b2255edc BB |
2091 | * number of child vdevs and therefore permutations to check. Furthermore, |
2092 | * since the raidz_map_t rows likely do not overlap reconstruction would be | |
2093 | * possible as long as there are no more than nparity data errors per row. | |
2094 | * These additional permutations are not currently checked but could be as | |
2095 | * a future improvement. | |
2096 | */ | |
2097 | static int | |
2098 | vdev_raidz_combrec(zio_t *zio) | |
2099 | { | |
2100 | int nparity = vdev_get_nparity(zio->io_vd); | |
2101 | raidz_map_t *rm = zio->io_vsd; | |
45d1cae3 | 2102 | |
b2255edc BB |
2103 | /* Check if there's enough data to attempt reconstrution. */ |
2104 | for (int i = 0; i < rm->rm_nrows; i++) { | |
2105 | raidz_row_t *rr = rm->rm_row[i]; | |
2106 | int total_errors = 0; | |
45d1cae3 | 2107 | |
b2255edc BB |
2108 | for (int c = 0; c < rr->rr_cols; c++) { |
2109 | if (rr->rr_col[c].rc_error) | |
2110 | total_errors++; | |
2111 | } | |
45d1cae3 | 2112 | |
b2255edc BB |
2113 | if (total_errors > nparity) |
2114 | return (vdev_raidz_worst_error(rr)); | |
2115 | } | |
45d1cae3 | 2116 | |
b2255edc BB |
2117 | for (int num_failures = 1; num_failures <= nparity; num_failures++) { |
2118 | int tstore[VDEV_RAIDZ_MAXPARITY + 2]; | |
2119 | int *ltgts = &tstore[1]; /* value is logical child ID */ | |
2120 | ||
2121 | /* Determine number of logical children, n */ | |
2122 | int n = zio->io_vd->vdev_children; | |
2123 | ||
2124 | ASSERT3U(num_failures, <=, nparity); | |
2125 | ASSERT3U(num_failures, <=, VDEV_RAIDZ_MAXPARITY); | |
2126 | ||
2127 | /* Handle corner cases in combrec logic */ | |
2128 | ltgts[-1] = -1; | |
2129 | for (int i = 0; i < num_failures; i++) { | |
2130 | ltgts[i] = i; | |
2131 | } | |
2132 | ltgts[num_failures] = n; | |
2133 | ||
2134 | for (;;) { | |
2135 | int err = raidz_reconstruct(zio, ltgts, num_failures, | |
2136 | nparity); | |
2137 | if (err == EINVAL) { | |
45d1cae3 | 2138 | /* |
b2255edc BB |
2139 | * Reconstruction not possible with this # |
2140 | * failures; try more failures. | |
45d1cae3 | 2141 | */ |
b2255edc BB |
2142 | break; |
2143 | } else if (err == 0) | |
2144 | return (0); | |
2145 | ||
2146 | /* Compute next targets to try */ | |
2147 | for (int t = 0; ; t++) { | |
2148 | ASSERT3U(t, <, num_failures); | |
2149 | ltgts[t]++; | |
2150 | if (ltgts[t] == n) { | |
2151 | /* try more failures */ | |
2152 | ASSERT3U(t, ==, num_failures - 1); | |
2153 | break; | |
2154 | } | |
45d1cae3 | 2155 | |
b2255edc BB |
2156 | ASSERT3U(ltgts[t], <, n); |
2157 | ASSERT3U(ltgts[t], <=, ltgts[t + 1]); | |
45d1cae3 BB |
2158 | |
2159 | /* | |
2160 | * If that spot is available, we're done here. | |
b2255edc | 2161 | * Try the next combination. |
45d1cae3 | 2162 | */ |
b2255edc | 2163 | if (ltgts[t] != ltgts[t + 1]) |
45d1cae3 BB |
2164 | break; |
2165 | ||
2166 | /* | |
b2255edc BB |
2167 | * Otherwise, reset this tgt to the minimum, |
2168 | * and move on to the next tgt. | |
45d1cae3 | 2169 | */ |
b2255edc BB |
2170 | ltgts[t] = ltgts[t - 1] + 1; |
2171 | ASSERT3U(ltgts[t], ==, t); | |
2172 | } | |
45d1cae3 | 2173 | |
b2255edc BB |
2174 | /* Increase the number of failures and keep trying. */ |
2175 | if (ltgts[num_failures - 1] == n) | |
2176 | break; | |
45d1cae3 BB |
2177 | } |
2178 | } | |
45d1cae3 | 2179 | |
b2255edc BB |
2180 | return (ECKSUM); |
2181 | } | |
2182 | ||
2183 | void | |
2184 | vdev_raidz_reconstruct(raidz_map_t *rm, const int *t, int nt) | |
2185 | { | |
2186 | for (uint64_t row = 0; row < rm->rm_nrows; row++) { | |
2187 | raidz_row_t *rr = rm->rm_row[row]; | |
2188 | vdev_raidz_reconstruct_row(rm, rr, t, nt); | |
2189 | } | |
45d1cae3 BB |
2190 | } |
2191 | ||
e49f1e20 | 2192 | /* |
b2255edc | 2193 | * Complete a write IO operation on a RAIDZ VDev |
e49f1e20 WA |
2194 | * |
2195 | * Outline: | |
e49f1e20 WA |
2196 | * 1. Check for errors on the child IOs. |
2197 | * 2. Return, setting an error code if too few child VDevs were written | |
2198 | * to reconstruct the data later. Note that partial writes are | |
2199 | * considered successful if they can be reconstructed at all. | |
e49f1e20 | 2200 | */ |
b128c09f | 2201 | static void |
b2255edc BB |
2202 | vdev_raidz_io_done_write_impl(zio_t *zio, raidz_row_t *rr) |
2203 | { | |
2204 | int total_errors = 0; | |
2205 | ||
2206 | ASSERT3U(rr->rr_missingparity, <=, rr->rr_firstdatacol); | |
2207 | ASSERT3U(rr->rr_missingdata, <=, rr->rr_cols - rr->rr_firstdatacol); | |
2208 | ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); | |
2209 | ||
2210 | for (int c = 0; c < rr->rr_cols; c++) { | |
2211 | raidz_col_t *rc = &rr->rr_col[c]; | |
2212 | ||
2213 | if (rc->rc_error) { | |
2214 | ASSERT(rc->rc_error != ECKSUM); /* child has no bp */ | |
2215 | ||
2216 | total_errors++; | |
2217 | } | |
2218 | } | |
2219 | ||
2220 | /* | |
2221 | * Treat partial writes as a success. If we couldn't write enough | |
2222 | * columns to reconstruct the data, the I/O failed. Otherwise, | |
2223 | * good enough. | |
2224 | * | |
2225 | * Now that we support write reallocation, it would be better | |
2226 | * to treat partial failure as real failure unless there are | |
2227 | * no non-degraded top-level vdevs left, and not update DTLs | |
2228 | * if we intend to reallocate. | |
2229 | */ | |
2230 | if (total_errors > rr->rr_firstdatacol) { | |
2231 | zio->io_error = zio_worst_error(zio->io_error, | |
2232 | vdev_raidz_worst_error(rr)); | |
2233 | } | |
2234 | } | |
2235 | ||
46df6e98 | 2236 | static void |
b2255edc BB |
2237 | vdev_raidz_io_done_reconstruct_known_missing(zio_t *zio, raidz_map_t *rm, |
2238 | raidz_row_t *rr) | |
34dc7c2f | 2239 | { |
34dc7c2f BB |
2240 | int parity_errors = 0; |
2241 | int parity_untried = 0; | |
2242 | int data_errors = 0; | |
b128c09f | 2243 | int total_errors = 0; |
34dc7c2f | 2244 | |
b2255edc BB |
2245 | ASSERT3U(rr->rr_missingparity, <=, rr->rr_firstdatacol); |
2246 | ASSERT3U(rr->rr_missingdata, <=, rr->rr_cols - rr->rr_firstdatacol); | |
2247 | ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ); | |
34dc7c2f | 2248 | |
b2255edc BB |
2249 | for (int c = 0; c < rr->rr_cols; c++) { |
2250 | raidz_col_t *rc = &rr->rr_col[c]; | |
34dc7c2f | 2251 | |
34dc7c2f | 2252 | if (rc->rc_error) { |
b128c09f | 2253 | ASSERT(rc->rc_error != ECKSUM); /* child has no bp */ |
34dc7c2f | 2254 | |
b2255edc | 2255 | if (c < rr->rr_firstdatacol) |
34dc7c2f BB |
2256 | parity_errors++; |
2257 | else | |
2258 | data_errors++; | |
2259 | ||
b128c09f | 2260 | total_errors++; |
b2255edc | 2261 | } else if (c < rr->rr_firstdatacol && !rc->rc_tried) { |
34dc7c2f BB |
2262 | parity_untried++; |
2263 | } | |
2264 | } | |
2265 | ||
34dc7c2f | 2266 | /* |
b2255edc BB |
2267 | * If there were data errors and the number of errors we saw was |
2268 | * correctable -- less than or equal to the number of parity disks read | |
2269 | * -- reconstruct based on the missing data. | |
34dc7c2f | 2270 | */ |
b2255edc BB |
2271 | if (data_errors != 0 && |
2272 | total_errors <= rr->rr_firstdatacol - parity_untried) { | |
2273 | /* | |
2274 | * We either attempt to read all the parity columns or | |
2275 | * none of them. If we didn't try to read parity, we | |
2276 | * wouldn't be here in the correctable case. There must | |
2277 | * also have been fewer parity errors than parity | |
2278 | * columns or, again, we wouldn't be in this code path. | |
2279 | */ | |
2280 | ASSERT(parity_untried == 0); | |
2281 | ASSERT(parity_errors < rr->rr_firstdatacol); | |
34dc7c2f | 2282 | |
b2255edc BB |
2283 | /* |
2284 | * Identify the data columns that reported an error. | |
2285 | */ | |
2286 | int n = 0; | |
2287 | int tgts[VDEV_RAIDZ_MAXPARITY]; | |
2288 | for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { | |
2289 | raidz_col_t *rc = &rr->rr_col[c]; | |
2290 | if (rc->rc_error != 0) { | |
2291 | ASSERT(n < VDEV_RAIDZ_MAXPARITY); | |
2292 | tgts[n++] = c; | |
34dc7c2f | 2293 | } |
b2255edc | 2294 | } |
34dc7c2f | 2295 | |
b2255edc | 2296 | ASSERT(rr->rr_firstdatacol >= n); |
34dc7c2f | 2297 | |
46df6e98 | 2298 | vdev_raidz_reconstruct_row(rm, rr, tgts, n); |
b2255edc | 2299 | } |
b2255edc | 2300 | } |
34dc7c2f | 2301 | |
b2255edc BB |
2302 | /* |
2303 | * Return the number of reads issued. | |
2304 | */ | |
2305 | static int | |
2306 | vdev_raidz_read_all(zio_t *zio, raidz_row_t *rr) | |
2307 | { | |
2308 | vdev_t *vd = zio->io_vd; | |
2309 | int nread = 0; | |
34dc7c2f | 2310 | |
b2255edc BB |
2311 | rr->rr_missingdata = 0; |
2312 | rr->rr_missingparity = 0; | |
34dc7c2f BB |
2313 | |
2314 | /* | |
b2255edc BB |
2315 | * If this rows contains empty sectors which are not required |
2316 | * for a normal read then allocate an ABD for them now so they | |
2317 | * may be read, verified, and any needed repairs performed. | |
34dc7c2f | 2318 | */ |
b2255edc BB |
2319 | if (rr->rr_nempty && rr->rr_abd_empty == NULL) |
2320 | vdev_draid_map_alloc_empty(zio, rr); | |
34dc7c2f | 2321 | |
b2255edc BB |
2322 | for (int c = 0; c < rr->rr_cols; c++) { |
2323 | raidz_col_t *rc = &rr->rr_col[c]; | |
2324 | if (rc->rc_tried || rc->rc_size == 0) | |
34dc7c2f BB |
2325 | continue; |
2326 | ||
b2255edc BB |
2327 | zio_nowait(zio_vdev_child_io(zio, NULL, |
2328 | vd->vdev_child[rc->rc_devidx], | |
2329 | rc->rc_offset, rc->rc_abd, rc->rc_size, | |
2330 | zio->io_type, zio->io_priority, 0, | |
2331 | vdev_raidz_child_done, rc)); | |
2332 | nread++; | |
34dc7c2f | 2333 | } |
b2255edc BB |
2334 | return (nread); |
2335 | } | |
34dc7c2f | 2336 | |
b2255edc BB |
2337 | /* |
2338 | * We're here because either there were too many errors to even attempt | |
2339 | * reconstruction (total_errors == rm_first_datacol), or vdev_*_combrec() | |
2340 | * failed. In either case, there is enough bad data to prevent reconstruction. | |
2341 | * Start checksum ereports for all children which haven't failed. | |
2342 | */ | |
2343 | static void | |
2344 | vdev_raidz_io_done_unrecoverable(zio_t *zio) | |
2345 | { | |
2346 | raidz_map_t *rm = zio->io_vsd; | |
34dc7c2f | 2347 | |
b2255edc BB |
2348 | for (int i = 0; i < rm->rm_nrows; i++) { |
2349 | raidz_row_t *rr = rm->rm_row[i]; | |
428870ff | 2350 | |
b2255edc BB |
2351 | for (int c = 0; c < rr->rr_cols; c++) { |
2352 | raidz_col_t *rc = &rr->rr_col[c]; | |
2353 | vdev_t *cvd = zio->io_vd->vdev_child[rc->rc_devidx]; | |
2354 | ||
2355 | if (rc->rc_error != 0) | |
2356 | continue; | |
2357 | ||
2358 | zio_bad_cksum_t zbc; | |
2359 | zbc.zbc_has_cksum = 0; | |
2360 | zbc.zbc_injected = rm->rm_ecksuminjected; | |
2361 | ||
03e02e5b | 2362 | (void) zfs_ereport_start_checksum(zio->io_spa, |
b2255edc | 2363 | cvd, &zio->io_bookmark, zio, rc->rc_offset, |
330c6c05 | 2364 | rc->rc_size, &zbc); |
03e02e5b DB |
2365 | mutex_enter(&cvd->vdev_stat_lock); |
2366 | cvd->vdev_stat.vs_checksum_errors++; | |
2367 | mutex_exit(&cvd->vdev_stat_lock); | |
34dc7c2f BB |
2368 | } |
2369 | } | |
b2255edc | 2370 | } |
34dc7c2f | 2371 | |
b2255edc BB |
2372 | void |
2373 | vdev_raidz_io_done(zio_t *zio) | |
2374 | { | |
2375 | raidz_map_t *rm = zio->io_vsd; | |
34dc7c2f | 2376 | |
b2255edc BB |
2377 | if (zio->io_type == ZIO_TYPE_WRITE) { |
2378 | for (int i = 0; i < rm->rm_nrows; i++) { | |
2379 | vdev_raidz_io_done_write_impl(zio, rm->rm_row[i]); | |
2380 | } | |
2381 | } else { | |
2382 | for (int i = 0; i < rm->rm_nrows; i++) { | |
2383 | raidz_row_t *rr = rm->rm_row[i]; | |
46df6e98 | 2384 | vdev_raidz_io_done_reconstruct_known_missing(zio, |
b2255edc BB |
2385 | rm, rr); |
2386 | } | |
34dc7c2f | 2387 | |
b2255edc BB |
2388 | if (raidz_checksum_verify(zio) == 0) { |
2389 | for (int i = 0; i < rm->rm_nrows; i++) { | |
2390 | raidz_row_t *rr = rm->rm_row[i]; | |
2391 | vdev_raidz_io_done_verified(zio, rr); | |
2392 | } | |
2393 | zio_checksum_verified(zio); | |
2394 | } else { | |
2395 | /* | |
2396 | * A sequential resilver has no checksum which makes | |
2397 | * combinatoral reconstruction impossible. This code | |
2398 | * path is unreachable since raidz_checksum_verify() | |
2399 | * has no checksum to verify and must succeed. | |
2400 | */ | |
2401 | ASSERT3U(zio->io_priority, !=, ZIO_PRIORITY_REBUILD); | |
34dc7c2f | 2402 | |
b2255edc BB |
2403 | /* |
2404 | * This isn't a typical situation -- either we got a | |
2405 | * read error or a child silently returned bad data. | |
2406 | * Read every block so we can try again with as much | |
2407 | * data and parity as we can track down. If we've | |
2408 | * already been through once before, all children will | |
2409 | * be marked as tried so we'll proceed to combinatorial | |
2410 | * reconstruction. | |
2411 | */ | |
2412 | int nread = 0; | |
2413 | for (int i = 0; i < rm->rm_nrows; i++) { | |
2414 | nread += vdev_raidz_read_all(zio, | |
2415 | rm->rm_row[i]); | |
2416 | } | |
2417 | if (nread != 0) { | |
2418 | /* | |
2419 | * Normally our stage is VDEV_IO_DONE, but if | |
2420 | * we've already called redone(), it will have | |
2421 | * changed to VDEV_IO_START, in which case we | |
2422 | * don't want to call redone() again. | |
2423 | */ | |
2424 | if (zio->io_stage != ZIO_STAGE_VDEV_IO_START) | |
2425 | zio_vdev_io_redone(zio); | |
2426 | return; | |
2427 | } | |
2428 | ||
2429 | zio->io_error = vdev_raidz_combrec(zio); | |
2430 | if (zio->io_error == ECKSUM && | |
2431 | !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { | |
2432 | vdev_raidz_io_done_unrecoverable(zio); | |
2433 | } | |
34dc7c2f | 2434 | } |
34dc7c2f | 2435 | } |
34dc7c2f BB |
2436 | } |
2437 | ||
2438 | static void | |
2439 | vdev_raidz_state_change(vdev_t *vd, int faulted, int degraded) | |
2440 | { | |
b2255edc BB |
2441 | vdev_raidz_t *vdrz = vd->vdev_tsd; |
2442 | if (faulted > vdrz->vd_nparity) | |
34dc7c2f BB |
2443 | vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN, |
2444 | VDEV_AUX_NO_REPLICAS); | |
2445 | else if (degraded + faulted != 0) | |
2446 | vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE); | |
2447 | else | |
2448 | vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE); | |
2449 | } | |
2450 | ||
3d6da72d IH |
2451 | /* |
2452 | * Determine if any portion of the provided block resides on a child vdev | |
2453 | * with a dirty DTL and therefore needs to be resilvered. The function | |
e1cfd73f | 2454 | * assumes that at least one DTL is dirty which implies that full stripe |
3d6da72d IH |
2455 | * width blocks must be resilvered. |
2456 | */ | |
2457 | static boolean_t | |
b2255edc BB |
2458 | vdev_raidz_need_resilver(vdev_t *vd, const dva_t *dva, size_t psize, |
2459 | uint64_t phys_birth) | |
3d6da72d | 2460 | { |
b2255edc | 2461 | vdev_raidz_t *vdrz = vd->vdev_tsd; |
3d6da72d | 2462 | uint64_t dcols = vd->vdev_children; |
b2255edc | 2463 | uint64_t nparity = vdrz->vd_nparity; |
3d6da72d IH |
2464 | uint64_t ashift = vd->vdev_top->vdev_ashift; |
2465 | /* The starting RAIDZ (parent) vdev sector of the block. */ | |
b2255edc | 2466 | uint64_t b = DVA_GET_OFFSET(dva) >> ashift; |
3d6da72d IH |
2467 | /* The zio's size in units of the vdev's minimum sector size. */ |
2468 | uint64_t s = ((psize - 1) >> ashift) + 1; | |
2469 | /* The first column for this stripe. */ | |
2470 | uint64_t f = b % dcols; | |
2471 | ||
b2255edc BB |
2472 | /* Unreachable by sequential resilver. */ |
2473 | ASSERT3U(phys_birth, !=, TXG_UNKNOWN); | |
2474 | ||
2475 | if (!vdev_dtl_contains(vd, DTL_PARTIAL, phys_birth, 1)) | |
2476 | return (B_FALSE); | |
2477 | ||
3d6da72d IH |
2478 | if (s + nparity >= dcols) |
2479 | return (B_TRUE); | |
2480 | ||
2481 | for (uint64_t c = 0; c < s + nparity; c++) { | |
2482 | uint64_t devidx = (f + c) % dcols; | |
2483 | vdev_t *cvd = vd->vdev_child[devidx]; | |
2484 | ||
2485 | /* | |
2486 | * dsl_scan_need_resilver() already checked vd with | |
2487 | * vdev_dtl_contains(). So here just check cvd with | |
2488 | * vdev_dtl_empty(), cheaper and a good approximation. | |
2489 | */ | |
2490 | if (!vdev_dtl_empty(cvd, DTL_PARTIAL)) | |
2491 | return (B_TRUE); | |
2492 | } | |
2493 | ||
2494 | return (B_FALSE); | |
2495 | } | |
2496 | ||
619f0976 | 2497 | static void |
b2255edc BB |
2498 | vdev_raidz_xlate(vdev_t *cvd, const range_seg64_t *logical_rs, |
2499 | range_seg64_t *physical_rs, range_seg64_t *remain_rs) | |
619f0976 | 2500 | { |
14e4e3cb AZ |
2501 | (void) remain_rs; |
2502 | ||
619f0976 GW |
2503 | vdev_t *raidvd = cvd->vdev_parent; |
2504 | ASSERT(raidvd->vdev_ops == &vdev_raidz_ops); | |
2505 | ||
2506 | uint64_t width = raidvd->vdev_children; | |
2507 | uint64_t tgt_col = cvd->vdev_id; | |
2508 | uint64_t ashift = raidvd->vdev_top->vdev_ashift; | |
2509 | ||
2510 | /* make sure the offsets are block-aligned */ | |
b2255edc BB |
2511 | ASSERT0(logical_rs->rs_start % (1 << ashift)); |
2512 | ASSERT0(logical_rs->rs_end % (1 << ashift)); | |
2513 | uint64_t b_start = logical_rs->rs_start >> ashift; | |
2514 | uint64_t b_end = logical_rs->rs_end >> ashift; | |
619f0976 GW |
2515 | |
2516 | uint64_t start_row = 0; | |
2517 | if (b_start > tgt_col) /* avoid underflow */ | |
2518 | start_row = ((b_start - tgt_col - 1) / width) + 1; | |
2519 | ||
2520 | uint64_t end_row = 0; | |
2521 | if (b_end > tgt_col) | |
2522 | end_row = ((b_end - tgt_col - 1) / width) + 1; | |
2523 | ||
b2255edc BB |
2524 | physical_rs->rs_start = start_row << ashift; |
2525 | physical_rs->rs_end = end_row << ashift; | |
619f0976 | 2526 | |
b2255edc BB |
2527 | ASSERT3U(physical_rs->rs_start, <=, logical_rs->rs_start); |
2528 | ASSERT3U(physical_rs->rs_end - physical_rs->rs_start, <=, | |
2529 | logical_rs->rs_end - logical_rs->rs_start); | |
2530 | } | |
2531 | ||
2532 | /* | |
2533 | * Initialize private RAIDZ specific fields from the nvlist. | |
2534 | */ | |
2535 | static int | |
2536 | vdev_raidz_init(spa_t *spa, nvlist_t *nv, void **tsd) | |
2537 | { | |
2538 | vdev_raidz_t *vdrz; | |
2539 | uint64_t nparity; | |
2540 | ||
2541 | uint_t children; | |
2542 | nvlist_t **child; | |
2543 | int error = nvlist_lookup_nvlist_array(nv, | |
2544 | ZPOOL_CONFIG_CHILDREN, &child, &children); | |
2545 | if (error != 0) | |
2546 | return (SET_ERROR(EINVAL)); | |
2547 | ||
2548 | if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY, &nparity) == 0) { | |
2549 | if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY) | |
2550 | return (SET_ERROR(EINVAL)); | |
2551 | ||
2552 | /* | |
2553 | * Previous versions could only support 1 or 2 parity | |
2554 | * device. | |
2555 | */ | |
2556 | if (nparity > 1 && spa_version(spa) < SPA_VERSION_RAIDZ2) | |
2557 | return (SET_ERROR(EINVAL)); | |
2558 | else if (nparity > 2 && spa_version(spa) < SPA_VERSION_RAIDZ3) | |
2559 | return (SET_ERROR(EINVAL)); | |
2560 | } else { | |
2561 | /* | |
2562 | * We require the parity to be specified for SPAs that | |
2563 | * support multiple parity levels. | |
2564 | */ | |
2565 | if (spa_version(spa) >= SPA_VERSION_RAIDZ2) | |
2566 | return (SET_ERROR(EINVAL)); | |
2567 | ||
2568 | /* | |
2569 | * Otherwise, we default to 1 parity device for RAID-Z. | |
2570 | */ | |
2571 | nparity = 1; | |
2572 | } | |
2573 | ||
2574 | vdrz = kmem_zalloc(sizeof (*vdrz), KM_SLEEP); | |
2575 | vdrz->vd_logical_width = children; | |
2576 | vdrz->vd_nparity = nparity; | |
2577 | ||
2578 | *tsd = vdrz; | |
2579 | ||
2580 | return (0); | |
2581 | } | |
2582 | ||
2583 | static void | |
2584 | vdev_raidz_fini(vdev_t *vd) | |
2585 | { | |
2586 | kmem_free(vd->vdev_tsd, sizeof (vdev_raidz_t)); | |
2587 | } | |
2588 | ||
2589 | /* | |
2590 | * Add RAIDZ specific fields to the config nvlist. | |
2591 | */ | |
2592 | static void | |
2593 | vdev_raidz_config_generate(vdev_t *vd, nvlist_t *nv) | |
2594 | { | |
2595 | ASSERT3P(vd->vdev_ops, ==, &vdev_raidz_ops); | |
2596 | vdev_raidz_t *vdrz = vd->vdev_tsd; | |
2597 | ||
2598 | /* | |
2599 | * Make sure someone hasn't managed to sneak a fancy new vdev | |
2600 | * into a crufty old storage pool. | |
2601 | */ | |
2602 | ASSERT(vdrz->vd_nparity == 1 || | |
2603 | (vdrz->vd_nparity <= 2 && | |
2604 | spa_version(vd->vdev_spa) >= SPA_VERSION_RAIDZ2) || | |
2605 | (vdrz->vd_nparity <= 3 && | |
2606 | spa_version(vd->vdev_spa) >= SPA_VERSION_RAIDZ3)); | |
2607 | ||
2608 | /* | |
2609 | * Note that we'll add these even on storage pools where they | |
2610 | * aren't strictly required -- older software will just ignore | |
2611 | * it. | |
2612 | */ | |
2613 | fnvlist_add_uint64(nv, ZPOOL_CONFIG_NPARITY, vdrz->vd_nparity); | |
2614 | } | |
2615 | ||
2616 | static uint64_t | |
2617 | vdev_raidz_nparity(vdev_t *vd) | |
2618 | { | |
2619 | vdev_raidz_t *vdrz = vd->vdev_tsd; | |
2620 | return (vdrz->vd_nparity); | |
2621 | } | |
2622 | ||
2623 | static uint64_t | |
2624 | vdev_raidz_ndisks(vdev_t *vd) | |
2625 | { | |
2626 | return (vd->vdev_children); | |
619f0976 GW |
2627 | } |
2628 | ||
34dc7c2f | 2629 | vdev_ops_t vdev_raidz_ops = { |
b2255edc BB |
2630 | .vdev_op_init = vdev_raidz_init, |
2631 | .vdev_op_fini = vdev_raidz_fini, | |
a64f8276 I |
2632 | .vdev_op_open = vdev_raidz_open, |
2633 | .vdev_op_close = vdev_raidz_close, | |
2634 | .vdev_op_asize = vdev_raidz_asize, | |
b2255edc BB |
2635 | .vdev_op_min_asize = vdev_raidz_min_asize, |
2636 | .vdev_op_min_alloc = NULL, | |
a64f8276 I |
2637 | .vdev_op_io_start = vdev_raidz_io_start, |
2638 | .vdev_op_io_done = vdev_raidz_io_done, | |
2639 | .vdev_op_state_change = vdev_raidz_state_change, | |
2640 | .vdev_op_need_resilver = vdev_raidz_need_resilver, | |
2641 | .vdev_op_hold = NULL, | |
2642 | .vdev_op_rele = NULL, | |
2643 | .vdev_op_remap = NULL, | |
2644 | .vdev_op_xlate = vdev_raidz_xlate, | |
b2255edc BB |
2645 | .vdev_op_rebuild_asize = NULL, |
2646 | .vdev_op_metaslab_init = NULL, | |
2647 | .vdev_op_config_generate = vdev_raidz_config_generate, | |
2648 | .vdev_op_nparity = vdev_raidz_nparity, | |
2649 | .vdev_op_ndisks = vdev_raidz_ndisks, | |
a64f8276 I |
2650 | .vdev_op_type = VDEV_TYPE_RAIDZ, /* name of this vdev type */ |
2651 | .vdev_op_leaf = B_FALSE /* not a leaf vdev */ | |
34dc7c2f | 2652 | }; |