]>
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); | |
146 | if (rc->rc_gdata != NULL) | |
147 | abd_free(rc->rc_gdata); | |
148 | if (rc->rc_orig_data != NULL) | |
149 | zio_buf_free(rc->rc_orig_data, rc->rc_size); | |
b2255edc BB |
150 | } |
151 | ||
152 | if (rr->rr_abd_copy != NULL) | |
153 | abd_free(rr->rr_abd_copy); | |
428870ff | 154 | |
b2255edc BB |
155 | if (rr->rr_abd_empty != NULL) |
156 | abd_free(rr->rr_abd_empty); | |
157 | ||
158 | kmem_free(rr, offsetof(raidz_row_t, rr_col[rr->rr_scols])); | |
159 | } | |
428870ff | 160 | |
b2255edc BB |
161 | void |
162 | vdev_raidz_map_free(raidz_map_t *rm) | |
163 | { | |
164 | for (int i = 0; i < rm->rm_nrows; i++) | |
165 | vdev_raidz_row_free(rm->rm_row[i]); | |
428870ff | 166 | |
b2255edc | 167 | kmem_free(rm, offsetof(raidz_map_t, rm_row[rm->rm_nrows])); |
b128c09f BB |
168 | } |
169 | ||
428870ff BB |
170 | static void |
171 | vdev_raidz_map_free_vsd(zio_t *zio) | |
172 | { | |
173 | raidz_map_t *rm = zio->io_vsd; | |
174 | ||
c99c9001 | 175 | ASSERT0(rm->rm_freed); |
b2255edc | 176 | rm->rm_freed = B_TRUE; |
428870ff | 177 | |
b2255edc | 178 | if (rm->rm_reports == 0) { |
428870ff | 179 | vdev_raidz_map_free(rm); |
b2255edc | 180 | } |
428870ff BB |
181 | } |
182 | ||
183 | /*ARGSUSED*/ | |
184 | static void | |
185 | vdev_raidz_cksum_free(void *arg, size_t ignored) | |
186 | { | |
187 | raidz_map_t *rm = arg; | |
188 | ||
189 | ASSERT3U(rm->rm_reports, >, 0); | |
190 | ||
b2255edc | 191 | if (--rm->rm_reports == 0 && rm->rm_freed) |
428870ff BB |
192 | vdev_raidz_map_free(rm); |
193 | } | |
194 | ||
195 | static void | |
84c07ada | 196 | vdev_raidz_cksum_finish(zio_cksum_report_t *zcr, const abd_t *good_data) |
428870ff BB |
197 | { |
198 | raidz_map_t *rm = zcr->zcr_cbdata; | |
84c07ada GN |
199 | const size_t c = zcr->zcr_cbinfo; |
200 | size_t x, offset; | |
428870ff | 201 | |
428870ff BB |
202 | if (good_data == NULL) { |
203 | zfs_ereport_finish_checksum(zcr, NULL, NULL, B_FALSE); | |
204 | return; | |
205 | } | |
206 | ||
b2255edc BB |
207 | ASSERT3U(rm->rm_nrows, ==, 1); |
208 | raidz_row_t *rr = rm->rm_row[0]; | |
209 | ||
210 | const abd_t *good = NULL; | |
211 | const abd_t *bad = rr->rr_col[c].rc_abd; | |
212 | ||
213 | if (c < rr->rr_firstdatacol) { | |
428870ff BB |
214 | /* |
215 | * The first time through, calculate the parity blocks for | |
216 | * the good data (this relies on the fact that the good | |
217 | * data never changes for a given logical ZIO) | |
218 | */ | |
b2255edc | 219 | if (rr->rr_col[0].rc_gdata == NULL) { |
a6255b7f | 220 | abd_t *bad_parity[VDEV_RAIDZ_MAXPARITY]; |
428870ff BB |
221 | |
222 | /* | |
b2255edc | 223 | * Set up the rr_col[]s to generate the parity for |
428870ff BB |
224 | * good_data, first saving the parity bufs and |
225 | * replacing them with buffers to hold the result. | |
226 | */ | |
b2255edc BB |
227 | for (x = 0; x < rr->rr_firstdatacol; x++) { |
228 | bad_parity[x] = rr->rr_col[x].rc_abd; | |
229 | rr->rr_col[x].rc_abd = rr->rr_col[x].rc_gdata = | |
230 | abd_alloc_sametype(rr->rr_col[x].rc_abd, | |
231 | rr->rr_col[x].rc_size); | |
428870ff BB |
232 | } |
233 | ||
234 | /* fill in the data columns from good_data */ | |
84c07ada | 235 | offset = 0; |
b2255edc | 236 | for (; x < rr->rr_cols; x++) { |
e2af2acc | 237 | abd_free(rr->rr_col[x].rc_abd); |
84c07ada | 238 | |
b2255edc | 239 | rr->rr_col[x].rc_abd = |
84c07ada | 240 | abd_get_offset_size((abd_t *)good_data, |
b2255edc BB |
241 | offset, rr->rr_col[x].rc_size); |
242 | offset += rr->rr_col[x].rc_size; | |
428870ff BB |
243 | } |
244 | ||
245 | /* | |
246 | * Construct the parity from the good data. | |
247 | */ | |
b2255edc | 248 | vdev_raidz_generate_parity_row(rm, rr); |
428870ff BB |
249 | |
250 | /* restore everything back to its original state */ | |
b2255edc BB |
251 | for (x = 0; x < rr->rr_firstdatacol; x++) |
252 | rr->rr_col[x].rc_abd = bad_parity[x]; | |
428870ff | 253 | |
a6255b7f | 254 | offset = 0; |
b2255edc | 255 | for (x = rr->rr_firstdatacol; x < rr->rr_cols; x++) { |
e2af2acc | 256 | abd_free(rr->rr_col[x].rc_abd); |
b2255edc BB |
257 | rr->rr_col[x].rc_abd = abd_get_offset_size( |
258 | rr->rr_abd_copy, offset, | |
259 | rr->rr_col[x].rc_size); | |
260 | offset += rr->rr_col[x].rc_size; | |
428870ff BB |
261 | } |
262 | } | |
263 | ||
b2255edc BB |
264 | ASSERT3P(rr->rr_col[c].rc_gdata, !=, NULL); |
265 | good = abd_get_offset_size(rr->rr_col[c].rc_gdata, 0, | |
266 | rr->rr_col[c].rc_size); | |
428870ff BB |
267 | } else { |
268 | /* adjust good_data to point at the start of our column */ | |
84c07ada | 269 | offset = 0; |
b2255edc BB |
270 | for (x = rr->rr_firstdatacol; x < c; x++) |
271 | offset += rr->rr_col[x].rc_size; | |
84c07ada GN |
272 | |
273 | good = abd_get_offset_size((abd_t *)good_data, offset, | |
b2255edc | 274 | rr->rr_col[c].rc_size); |
428870ff BB |
275 | } |
276 | ||
277 | /* we drop the ereport if it ends up that the data was good */ | |
278 | zfs_ereport_finish_checksum(zcr, good, bad, B_TRUE); | |
e2af2acc | 279 | abd_free((abd_t *)good); |
428870ff BB |
280 | } |
281 | ||
282 | /* | |
283 | * Invoked indirectly by zfs_ereport_start_checksum(), called | |
284 | * below when our read operation fails completely. The main point | |
285 | * is to keep a copy of everything we read from disk, so that at | |
286 | * vdev_raidz_cksum_finish() time we can compare it with the good data. | |
287 | */ | |
288 | static void | |
289 | vdev_raidz_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *arg) | |
290 | { | |
291 | size_t c = (size_t)(uintptr_t)arg; | |
428870ff | 292 | raidz_map_t *rm = zio->io_vsd; |
428870ff BB |
293 | |
294 | /* set up the report and bump the refcount */ | |
295 | zcr->zcr_cbdata = rm; | |
296 | zcr->zcr_cbinfo = c; | |
297 | zcr->zcr_finish = vdev_raidz_cksum_finish; | |
298 | zcr->zcr_free = vdev_raidz_cksum_free; | |
299 | ||
300 | rm->rm_reports++; | |
301 | ASSERT3U(rm->rm_reports, >, 0); | |
b2255edc | 302 | ASSERT3U(rm->rm_nrows, ==, 1); |
428870ff | 303 | |
b2255edc | 304 | if (rm->rm_row[0]->rr_abd_copy != NULL) |
428870ff BB |
305 | return; |
306 | ||
307 | /* | |
308 | * It's the first time we're called for this raidz_map_t, so we need | |
309 | * to copy the data aside; there's no guarantee that our zio's buffer | |
310 | * won't be re-used for something else. | |
311 | * | |
312 | * Our parity data is already in separate buffers, so there's no need | |
313 | * to copy them. | |
314 | */ | |
b2255edc BB |
315 | for (int i = 0; i < rm->rm_nrows; i++) { |
316 | raidz_row_t *rr = rm->rm_row[i]; | |
317 | size_t offset = 0; | |
318 | size_t size = 0; | |
428870ff | 319 | |
b2255edc BB |
320 | for (c = rr->rr_firstdatacol; c < rr->rr_cols; c++) |
321 | size += rr->rr_col[c].rc_size; | |
428870ff | 322 | |
b2255edc | 323 | rr->rr_abd_copy = abd_alloc_for_io(size, B_FALSE); |
428870ff | 324 | |
b2255edc BB |
325 | for (c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { |
326 | raidz_col_t *col = &rr->rr_col[c]; | |
327 | abd_t *tmp = abd_get_offset_size(rr->rr_abd_copy, | |
328 | offset, col->rc_size); | |
428870ff | 329 | |
b2255edc | 330 | abd_copy(tmp, col->rc_abd, col->rc_size); |
84c07ada | 331 | |
e2af2acc | 332 | abd_free(col->rc_abd); |
b2255edc | 333 | col->rc_abd = tmp; |
428870ff | 334 | |
b2255edc BB |
335 | offset += col->rc_size; |
336 | } | |
337 | ASSERT3U(offset, ==, size); | |
428870ff | 338 | } |
428870ff BB |
339 | } |
340 | ||
341 | static const zio_vsd_ops_t vdev_raidz_vsd_ops = { | |
56d8d8ac MW |
342 | .vsd_free = vdev_raidz_map_free_vsd, |
343 | .vsd_cksum_report = vdev_raidz_cksum_report | |
428870ff BB |
344 | }; |
345 | ||
e49f1e20 WA |
346 | /* |
347 | * Divides the IO evenly across all child vdevs; usually, dcols is | |
348 | * the number of children in the target vdev. | |
a1687880 BB |
349 | * |
350 | * Avoid inlining the function to keep vdev_raidz_io_start(), which | |
351 | * is this functions only caller, as small as possible on the stack. | |
e49f1e20 | 352 | */ |
ab9f4b0b | 353 | noinline raidz_map_t * |
3d6da72d | 354 | vdev_raidz_map_alloc(zio_t *zio, uint64_t ashift, uint64_t dcols, |
34dc7c2f BB |
355 | uint64_t nparity) |
356 | { | |
b2255edc | 357 | raidz_row_t *rr; |
e49f1e20 | 358 | /* The starting RAIDZ (parent) vdev sector of the block. */ |
3d6da72d | 359 | uint64_t b = zio->io_offset >> ashift; |
e49f1e20 | 360 | /* The zio's size in units of the vdev's minimum sector size. */ |
3d6da72d | 361 | uint64_t s = zio->io_size >> ashift; |
e49f1e20 | 362 | /* The first column for this stripe. */ |
34dc7c2f | 363 | uint64_t f = b % dcols; |
e49f1e20 | 364 | /* The starting byte offset on each child vdev. */ |
3d6da72d | 365 | uint64_t o = (b / dcols) << ashift; |
45d1cae3 | 366 | uint64_t q, r, c, bc, col, acols, scols, coff, devidx, asize, tot; |
34dc7c2f | 367 | |
b2255edc BB |
368 | raidz_map_t *rm = |
369 | kmem_zalloc(offsetof(raidz_map_t, rm_row[1]), KM_SLEEP); | |
370 | rm->rm_nrows = 1; | |
371 | ||
e49f1e20 WA |
372 | /* |
373 | * "Quotient": The number of data sectors for this stripe on all but | |
374 | * the "big column" child vdevs that also contain "remainder" data. | |
375 | */ | |
34dc7c2f | 376 | q = s / (dcols - nparity); |
e49f1e20 WA |
377 | |
378 | /* | |
379 | * "Remainder": The number of partial stripe data sectors in this I/O. | |
380 | * This will add a sector to some, but not all, child vdevs. | |
381 | */ | |
34dc7c2f | 382 | r = s - q * (dcols - nparity); |
e49f1e20 WA |
383 | |
384 | /* The number of "big columns" - those which contain remainder data. */ | |
34dc7c2f | 385 | bc = (r == 0 ? 0 : r + nparity); |
e49f1e20 WA |
386 | |
387 | /* | |
388 | * The total number of data and parity sectors associated with | |
389 | * this I/O. | |
390 | */ | |
45d1cae3 BB |
391 | tot = s + nparity * (q + (r == 0 ? 0 : 1)); |
392 | ||
b2255edc BB |
393 | /* |
394 | * acols: The columns that will be accessed. | |
395 | * scols: The columns that will be accessed or skipped. | |
396 | */ | |
45d1cae3 | 397 | if (q == 0) { |
e49f1e20 | 398 | /* Our I/O request doesn't span all child vdevs. */ |
45d1cae3 BB |
399 | acols = bc; |
400 | scols = MIN(dcols, roundup(bc, nparity + 1)); | |
401 | } else { | |
402 | acols = dcols; | |
403 | scols = dcols; | |
404 | } | |
34dc7c2f | 405 | |
45d1cae3 | 406 | ASSERT3U(acols, <=, scols); |
34dc7c2f | 407 | |
b2255edc BB |
408 | rr = kmem_alloc(offsetof(raidz_row_t, rr_col[scols]), KM_SLEEP); |
409 | rm->rm_row[0] = rr; | |
410 | ||
411 | rr->rr_cols = acols; | |
412 | rr->rr_scols = scols; | |
413 | rr->rr_bigcols = bc; | |
414 | rr->rr_missingdata = 0; | |
415 | rr->rr_missingparity = 0; | |
416 | rr->rr_firstdatacol = nparity; | |
417 | rr->rr_abd_copy = NULL; | |
418 | rr->rr_abd_empty = NULL; | |
419 | rr->rr_nempty = 0; | |
420 | #ifdef ZFS_DEBUG | |
421 | rr->rr_offset = zio->io_offset; | |
422 | rr->rr_size = zio->io_size; | |
423 | #endif | |
34dc7c2f | 424 | |
45d1cae3 BB |
425 | asize = 0; |
426 | ||
427 | for (c = 0; c < scols; c++) { | |
b2255edc | 428 | raidz_col_t *rc = &rr->rr_col[c]; |
34dc7c2f BB |
429 | col = f + c; |
430 | coff = o; | |
431 | if (col >= dcols) { | |
432 | col -= dcols; | |
3d6da72d | 433 | coff += 1ULL << ashift; |
34dc7c2f | 434 | } |
b2255edc BB |
435 | rc->rc_devidx = col; |
436 | rc->rc_offset = coff; | |
437 | rc->rc_abd = NULL; | |
438 | rc->rc_gdata = NULL; | |
439 | rc->rc_orig_data = NULL; | |
440 | rc->rc_error = 0; | |
441 | rc->rc_tried = 0; | |
442 | rc->rc_skipped = 0; | |
443 | rc->rc_repair = 0; | |
444 | rc->rc_need_orig_restore = B_FALSE; | |
45d1cae3 BB |
445 | |
446 | if (c >= acols) | |
b2255edc | 447 | rc->rc_size = 0; |
45d1cae3 | 448 | else if (c < bc) |
b2255edc | 449 | rc->rc_size = (q + 1) << ashift; |
45d1cae3 | 450 | else |
b2255edc | 451 | rc->rc_size = q << ashift; |
45d1cae3 | 452 | |
b2255edc | 453 | asize += rc->rc_size; |
34dc7c2f BB |
454 | } |
455 | ||
3d6da72d | 456 | ASSERT3U(asize, ==, tot << ashift); |
428870ff | 457 | rm->rm_nskip = roundup(tot, nparity + 1) - tot; |
b2255edc | 458 | rm->rm_skipstart = bc; |
34dc7c2f | 459 | |
b2255edc BB |
460 | for (c = 0; c < rr->rr_firstdatacol; c++) |
461 | rr->rr_col[c].rc_abd = | |
462 | abd_alloc_linear(rr->rr_col[c].rc_size, B_FALSE); | |
34dc7c2f | 463 | |
e2af2acc | 464 | for (uint64_t off = 0; c < acols; c++) { |
b2255edc | 465 | raidz_col_t *rc = &rr->rr_col[c]; |
e2af2acc MA |
466 | rc->rc_abd = abd_get_offset_struct(&rc->rc_abdstruct, |
467 | zio->io_abd, off, rc->rc_size); | |
b2255edc | 468 | off += rc->rc_size; |
a6255b7f | 469 | } |
34dc7c2f BB |
470 | |
471 | /* | |
472 | * If all data stored spans all columns, there's a danger that parity | |
473 | * will always be on the same device and, since parity isn't read | |
e1cfd73f | 474 | * during normal operation, that device's I/O bandwidth won't be |
34dc7c2f BB |
475 | * used effectively. We therefore switch the parity every 1MB. |
476 | * | |
477 | * ... at least that was, ostensibly, the theory. As a practical | |
478 | * matter unless we juggle the parity between all devices evenly, we | |
479 | * won't see any benefit. Further, occasional writes that aren't a | |
480 | * multiple of the LCM of the number of children and the minimum | |
481 | * stripe width are sufficient to avoid pessimal behavior. | |
482 | * Unfortunately, this decision created an implicit on-disk format | |
483 | * requirement that we need to support for all eternity, but only | |
484 | * for single-parity RAID-Z. | |
428870ff BB |
485 | * |
486 | * If we intend to skip a sector in the zeroth column for padding | |
487 | * we must make sure to note this swap. We will never intend to | |
488 | * skip the first column since at least one data and one parity | |
489 | * column must appear in each row. | |
34dc7c2f | 490 | */ |
b2255edc BB |
491 | ASSERT(rr->rr_cols >= 2); |
492 | ASSERT(rr->rr_col[0].rc_size == rr->rr_col[1].rc_size); | |
34dc7c2f | 493 | |
b2255edc BB |
494 | if (rr->rr_firstdatacol == 1 && (zio->io_offset & (1ULL << 20))) { |
495 | devidx = rr->rr_col[0].rc_devidx; | |
496 | o = rr->rr_col[0].rc_offset; | |
497 | rr->rr_col[0].rc_devidx = rr->rr_col[1].rc_devidx; | |
498 | rr->rr_col[0].rc_offset = rr->rr_col[1].rc_offset; | |
499 | rr->rr_col[1].rc_devidx = devidx; | |
500 | rr->rr_col[1].rc_offset = o; | |
428870ff BB |
501 | |
502 | if (rm->rm_skipstart == 0) | |
503 | rm->rm_skipstart = 1; | |
34dc7c2f BB |
504 | } |
505 | ||
c9187d86 GN |
506 | /* init RAIDZ parity ops */ |
507 | rm->rm_ops = vdev_raidz_math_get_ops(); | |
ab9f4b0b | 508 | |
34dc7c2f BB |
509 | return (rm); |
510 | } | |
511 | ||
a6255b7f DQ |
512 | struct pqr_struct { |
513 | uint64_t *p; | |
514 | uint64_t *q; | |
515 | uint64_t *r; | |
516 | }; | |
517 | ||
518 | static int | |
519 | vdev_raidz_p_func(void *buf, size_t size, void *private) | |
520 | { | |
521 | struct pqr_struct *pqr = private; | |
522 | const uint64_t *src = buf; | |
523 | int i, cnt = size / sizeof (src[0]); | |
524 | ||
525 | ASSERT(pqr->p && !pqr->q && !pqr->r); | |
526 | ||
527 | for (i = 0; i < cnt; i++, src++, pqr->p++) | |
528 | *pqr->p ^= *src; | |
529 | ||
530 | return (0); | |
531 | } | |
532 | ||
533 | static int | |
534 | vdev_raidz_pq_func(void *buf, size_t size, void *private) | |
535 | { | |
536 | struct pqr_struct *pqr = private; | |
537 | const uint64_t *src = buf; | |
538 | uint64_t mask; | |
539 | int i, cnt = size / sizeof (src[0]); | |
540 | ||
541 | ASSERT(pqr->p && pqr->q && !pqr->r); | |
542 | ||
543 | for (i = 0; i < cnt; i++, src++, pqr->p++, pqr->q++) { | |
544 | *pqr->p ^= *src; | |
545 | VDEV_RAIDZ_64MUL_2(*pqr->q, mask); | |
546 | *pqr->q ^= *src; | |
547 | } | |
548 | ||
549 | return (0); | |
550 | } | |
551 | ||
552 | static int | |
553 | vdev_raidz_pqr_func(void *buf, size_t size, void *private) | |
554 | { | |
555 | struct pqr_struct *pqr = private; | |
556 | const uint64_t *src = buf; | |
557 | uint64_t mask; | |
558 | int i, cnt = size / sizeof (src[0]); | |
559 | ||
560 | ASSERT(pqr->p && pqr->q && pqr->r); | |
561 | ||
562 | for (i = 0; i < cnt; i++, src++, pqr->p++, pqr->q++, pqr->r++) { | |
563 | *pqr->p ^= *src; | |
564 | VDEV_RAIDZ_64MUL_2(*pqr->q, mask); | |
565 | *pqr->q ^= *src; | |
566 | VDEV_RAIDZ_64MUL_4(*pqr->r, mask); | |
567 | *pqr->r ^= *src; | |
568 | } | |
569 | ||
570 | return (0); | |
571 | } | |
572 | ||
34dc7c2f | 573 | static void |
b2255edc | 574 | vdev_raidz_generate_parity_p(raidz_row_t *rr) |
34dc7c2f | 575 | { |
b2255edc | 576 | uint64_t *p = abd_to_buf(rr->rr_col[VDEV_RAIDZ_P].rc_abd); |
34dc7c2f | 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; | |
34dc7c2f | 580 | |
b2255edc BB |
581 | if (c == rr->rr_firstdatacol) { |
582 | abd_copy_to_buf(p, src, rr->rr_col[c].rc_size); | |
34dc7c2f | 583 | } else { |
a6255b7f | 584 | struct pqr_struct pqr = { p, NULL, NULL }; |
b2255edc | 585 | (void) abd_iterate_func(src, 0, rr->rr_col[c].rc_size, |
a6255b7f | 586 | vdev_raidz_p_func, &pqr); |
34dc7c2f BB |
587 | } |
588 | } | |
589 | } | |
590 | ||
591 | static void | |
b2255edc | 592 | vdev_raidz_generate_parity_pq(raidz_row_t *rr) |
34dc7c2f | 593 | { |
b2255edc BB |
594 | uint64_t *p = abd_to_buf(rr->rr_col[VDEV_RAIDZ_P].rc_abd); |
595 | uint64_t *q = abd_to_buf(rr->rr_col[VDEV_RAIDZ_Q].rc_abd); | |
596 | uint64_t pcnt = rr->rr_col[VDEV_RAIDZ_P].rc_size / sizeof (p[0]); | |
597 | ASSERT(rr->rr_col[VDEV_RAIDZ_P].rc_size == | |
598 | rr->rr_col[VDEV_RAIDZ_Q].rc_size); | |
34dc7c2f | 599 | |
b2255edc BB |
600 | for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { |
601 | abd_t *src = rr->rr_col[c].rc_abd; | |
45d1cae3 | 602 | |
b2255edc | 603 | uint64_t ccnt = rr->rr_col[c].rc_size / sizeof (p[0]); |
34dc7c2f | 604 | |
b2255edc | 605 | if (c == rr->rr_firstdatacol) { |
f7e76821 | 606 | ASSERT(ccnt == pcnt || ccnt == 0); |
b2255edc BB |
607 | abd_copy_to_buf(p, src, rr->rr_col[c].rc_size); |
608 | (void) memcpy(q, p, rr->rr_col[c].rc_size); | |
45d1cae3 | 609 | |
b2255edc | 610 | for (uint64_t i = ccnt; i < pcnt; i++) { |
a6255b7f DQ |
611 | p[i] = 0; |
612 | q[i] = 0; | |
45d1cae3 | 613 | } |
a6255b7f | 614 | } else { |
f7e76821 IH |
615 | struct pqr_struct pqr = { p, q, NULL }; |
616 | ||
617 | ASSERT(ccnt <= pcnt); | |
b2255edc | 618 | (void) abd_iterate_func(src, 0, rr->rr_col[c].rc_size, |
f7e76821 | 619 | vdev_raidz_pq_func, &pqr); |
45d1cae3 BB |
620 | |
621 | /* | |
622 | * Treat short columns as though they are full of 0s. | |
623 | * Note that there's therefore nothing needed for P. | |
624 | */ | |
b2255edc BB |
625 | uint64_t mask; |
626 | for (uint64_t i = ccnt; i < pcnt; i++) { | |
a6255b7f | 627 | VDEV_RAIDZ_64MUL_2(q[i], mask); |
45d1cae3 BB |
628 | } |
629 | } | |
630 | } | |
631 | } | |
632 | ||
633 | static void | |
b2255edc | 634 | vdev_raidz_generate_parity_pqr(raidz_row_t *rr) |
45d1cae3 | 635 | { |
b2255edc BB |
636 | uint64_t *p = abd_to_buf(rr->rr_col[VDEV_RAIDZ_P].rc_abd); |
637 | uint64_t *q = abd_to_buf(rr->rr_col[VDEV_RAIDZ_Q].rc_abd); | |
638 | uint64_t *r = abd_to_buf(rr->rr_col[VDEV_RAIDZ_R].rc_abd); | |
639 | uint64_t pcnt = rr->rr_col[VDEV_RAIDZ_P].rc_size / sizeof (p[0]); | |
640 | ASSERT(rr->rr_col[VDEV_RAIDZ_P].rc_size == | |
641 | rr->rr_col[VDEV_RAIDZ_Q].rc_size); | |
642 | ASSERT(rr->rr_col[VDEV_RAIDZ_P].rc_size == | |
643 | rr->rr_col[VDEV_RAIDZ_R].rc_size); | |
45d1cae3 | 644 | |
b2255edc BB |
645 | for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { |
646 | abd_t *src = rr->rr_col[c].rc_abd; | |
45d1cae3 | 647 | |
b2255edc | 648 | uint64_t ccnt = rr->rr_col[c].rc_size / sizeof (p[0]); |
45d1cae3 | 649 | |
b2255edc | 650 | if (c == rr->rr_firstdatacol) { |
f7e76821 | 651 | ASSERT(ccnt == pcnt || ccnt == 0); |
b2255edc BB |
652 | abd_copy_to_buf(p, src, rr->rr_col[c].rc_size); |
653 | (void) memcpy(q, p, rr->rr_col[c].rc_size); | |
654 | (void) memcpy(r, p, rr->rr_col[c].rc_size); | |
45d1cae3 | 655 | |
b2255edc | 656 | for (uint64_t i = ccnt; i < pcnt; i++) { |
a6255b7f DQ |
657 | p[i] = 0; |
658 | q[i] = 0; | |
659 | r[i] = 0; | |
34dc7c2f | 660 | } |
a6255b7f | 661 | } else { |
f7e76821 IH |
662 | struct pqr_struct pqr = { p, q, r }; |
663 | ||
664 | ASSERT(ccnt <= pcnt); | |
b2255edc | 665 | (void) abd_iterate_func(src, 0, rr->rr_col[c].rc_size, |
f7e76821 IH |
666 | vdev_raidz_pqr_func, &pqr); |
667 | ||
34dc7c2f BB |
668 | /* |
669 | * Treat short columns as though they are full of 0s. | |
45d1cae3 | 670 | * Note that there's therefore nothing needed for P. |
34dc7c2f | 671 | */ |
b2255edc BB |
672 | uint64_t mask; |
673 | for (uint64_t i = ccnt; i < pcnt; i++) { | |
a6255b7f DQ |
674 | VDEV_RAIDZ_64MUL_2(q[i], mask); |
675 | VDEV_RAIDZ_64MUL_4(r[i], mask); | |
34dc7c2f BB |
676 | } |
677 | } | |
678 | } | |
679 | } | |
680 | ||
45d1cae3 BB |
681 | /* |
682 | * Generate RAID parity in the first virtual columns according to the number of | |
683 | * parity columns available. | |
684 | */ | |
ab9f4b0b | 685 | void |
b2255edc | 686 | vdev_raidz_generate_parity_row(raidz_map_t *rm, raidz_row_t *rr) |
45d1cae3 | 687 | { |
b2255edc BB |
688 | ASSERT3U(rr->rr_cols, !=, 0); |
689 | ||
c9187d86 | 690 | /* Generate using the new math implementation */ |
b2255edc | 691 | if (vdev_raidz_math_generate(rm, rr) != RAIDZ_ORIGINAL_IMPL) |
ab9f4b0b | 692 | return; |
ab9f4b0b | 693 | |
b2255edc | 694 | switch (rr->rr_firstdatacol) { |
45d1cae3 | 695 | case 1: |
b2255edc | 696 | vdev_raidz_generate_parity_p(rr); |
45d1cae3 BB |
697 | break; |
698 | case 2: | |
b2255edc | 699 | vdev_raidz_generate_parity_pq(rr); |
45d1cae3 BB |
700 | break; |
701 | case 3: | |
b2255edc | 702 | vdev_raidz_generate_parity_pqr(rr); |
45d1cae3 BB |
703 | break; |
704 | default: | |
705 | cmn_err(CE_PANIC, "invalid RAID-Z configuration"); | |
706 | } | |
707 | } | |
708 | ||
b2255edc BB |
709 | void |
710 | vdev_raidz_generate_parity(raidz_map_t *rm) | |
711 | { | |
712 | for (int i = 0; i < rm->rm_nrows; i++) { | |
713 | raidz_row_t *rr = rm->rm_row[i]; | |
714 | vdev_raidz_generate_parity_row(rm, rr); | |
715 | } | |
716 | } | |
717 | ||
a6255b7f DQ |
718 | /* ARGSUSED */ |
719 | static int | |
720 | vdev_raidz_reconst_p_func(void *dbuf, void *sbuf, size_t size, void *private) | |
721 | { | |
722 | uint64_t *dst = dbuf; | |
723 | uint64_t *src = sbuf; | |
724 | int cnt = size / sizeof (src[0]); | |
a6255b7f | 725 | |
1c27024e | 726 | for (int i = 0; i < cnt; i++) { |
a6255b7f DQ |
727 | dst[i] ^= src[i]; |
728 | } | |
729 | ||
730 | return (0); | |
731 | } | |
732 | ||
733 | /* ARGSUSED */ | |
734 | static int | |
735 | vdev_raidz_reconst_q_pre_func(void *dbuf, void *sbuf, size_t size, | |
736 | void *private) | |
737 | { | |
738 | uint64_t *dst = dbuf; | |
739 | uint64_t *src = sbuf; | |
740 | uint64_t mask; | |
741 | int cnt = size / sizeof (dst[0]); | |
a6255b7f | 742 | |
1c27024e | 743 | for (int i = 0; i < cnt; i++, dst++, src++) { |
a6255b7f DQ |
744 | VDEV_RAIDZ_64MUL_2(*dst, mask); |
745 | *dst ^= *src; | |
746 | } | |
747 | ||
748 | return (0); | |
749 | } | |
750 | ||
751 | /* ARGSUSED */ | |
752 | static int | |
753 | vdev_raidz_reconst_q_pre_tail_func(void *buf, size_t size, void *private) | |
754 | { | |
755 | uint64_t *dst = buf; | |
756 | uint64_t mask; | |
757 | int cnt = size / sizeof (dst[0]); | |
a6255b7f | 758 | |
1c27024e | 759 | for (int i = 0; i < cnt; i++, dst++) { |
a6255b7f DQ |
760 | /* same operation as vdev_raidz_reconst_q_pre_func() on dst */ |
761 | VDEV_RAIDZ_64MUL_2(*dst, mask); | |
762 | } | |
763 | ||
764 | return (0); | |
765 | } | |
766 | ||
767 | struct reconst_q_struct { | |
768 | uint64_t *q; | |
769 | int exp; | |
770 | }; | |
771 | ||
772 | static int | |
773 | vdev_raidz_reconst_q_post_func(void *buf, size_t size, void *private) | |
774 | { | |
775 | struct reconst_q_struct *rq = private; | |
776 | uint64_t *dst = buf; | |
777 | int cnt = size / sizeof (dst[0]); | |
a6255b7f | 778 | |
1c27024e | 779 | for (int i = 0; i < cnt; i++, dst++, rq->q++) { |
a6255b7f DQ |
780 | int j; |
781 | uint8_t *b; | |
782 | ||
783 | *dst ^= *rq->q; | |
784 | for (j = 0, b = (uint8_t *)dst; j < 8; j++, b++) { | |
785 | *b = vdev_raidz_exp2(*b, rq->exp); | |
786 | } | |
787 | } | |
788 | ||
789 | return (0); | |
790 | } | |
791 | ||
792 | struct reconst_pq_struct { | |
793 | uint8_t *p; | |
794 | uint8_t *q; | |
795 | uint8_t *pxy; | |
796 | uint8_t *qxy; | |
797 | int aexp; | |
798 | int bexp; | |
799 | }; | |
800 | ||
801 | static int | |
802 | vdev_raidz_reconst_pq_func(void *xbuf, void *ybuf, size_t size, void *private) | |
803 | { | |
804 | struct reconst_pq_struct *rpq = private; | |
805 | uint8_t *xd = xbuf; | |
806 | uint8_t *yd = ybuf; | |
a6255b7f | 807 | |
1c27024e | 808 | for (int i = 0; i < size; |
a6255b7f DQ |
809 | i++, rpq->p++, rpq->q++, rpq->pxy++, rpq->qxy++, xd++, yd++) { |
810 | *xd = vdev_raidz_exp2(*rpq->p ^ *rpq->pxy, rpq->aexp) ^ | |
811 | vdev_raidz_exp2(*rpq->q ^ *rpq->qxy, rpq->bexp); | |
812 | *yd = *rpq->p ^ *rpq->pxy ^ *xd; | |
813 | } | |
814 | ||
815 | return (0); | |
816 | } | |
817 | ||
818 | static int | |
819 | vdev_raidz_reconst_pq_tail_func(void *xbuf, size_t size, void *private) | |
820 | { | |
821 | struct reconst_pq_struct *rpq = private; | |
822 | uint8_t *xd = xbuf; | |
a6255b7f | 823 | |
1c27024e | 824 | for (int i = 0; i < size; |
a6255b7f DQ |
825 | i++, rpq->p++, rpq->q++, rpq->pxy++, rpq->qxy++, xd++) { |
826 | /* same operation as vdev_raidz_reconst_pq_func() on xd */ | |
827 | *xd = vdev_raidz_exp2(*rpq->p ^ *rpq->pxy, rpq->aexp) ^ | |
828 | vdev_raidz_exp2(*rpq->q ^ *rpq->qxy, rpq->bexp); | |
829 | } | |
830 | ||
831 | return (0); | |
832 | } | |
833 | ||
46df6e98 | 834 | static void |
b2255edc | 835 | vdev_raidz_reconstruct_p(raidz_row_t *rr, int *tgts, int ntgts) |
34dc7c2f | 836 | { |
45d1cae3 | 837 | int x = tgts[0]; |
a6255b7f | 838 | abd_t *dst, *src; |
34dc7c2f | 839 | |
b2255edc BB |
840 | ASSERT3U(ntgts, ==, 1); |
841 | ASSERT3U(x, >=, rr->rr_firstdatacol); | |
842 | ASSERT3U(x, <, rr->rr_cols); | |
45d1cae3 | 843 | |
b2255edc | 844 | ASSERT3U(rr->rr_col[x].rc_size, <=, rr->rr_col[VDEV_RAIDZ_P].rc_size); |
34dc7c2f | 845 | |
b2255edc BB |
846 | src = rr->rr_col[VDEV_RAIDZ_P].rc_abd; |
847 | dst = rr->rr_col[x].rc_abd; | |
a6255b7f | 848 | |
b2255edc | 849 | abd_copy_from_buf(dst, abd_to_buf(src), rr->rr_col[x].rc_size); |
34dc7c2f | 850 | |
b2255edc BB |
851 | for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { |
852 | uint64_t size = MIN(rr->rr_col[x].rc_size, | |
853 | rr->rr_col[c].rc_size); | |
a6255b7f | 854 | |
b2255edc | 855 | src = rr->rr_col[c].rc_abd; |
34dc7c2f BB |
856 | |
857 | if (c == x) | |
858 | continue; | |
859 | ||
a6255b7f DQ |
860 | (void) abd_iterate_func2(dst, src, 0, 0, size, |
861 | vdev_raidz_reconst_p_func, NULL); | |
34dc7c2f BB |
862 | } |
863 | } | |
864 | ||
46df6e98 | 865 | static void |
b2255edc | 866 | vdev_raidz_reconstruct_q(raidz_row_t *rr, int *tgts, int ntgts) |
34dc7c2f | 867 | { |
45d1cae3 | 868 | int x = tgts[0]; |
a6255b7f DQ |
869 | int c, exp; |
870 | abd_t *dst, *src; | |
34dc7c2f | 871 | |
45d1cae3 BB |
872 | ASSERT(ntgts == 1); |
873 | ||
b2255edc | 874 | ASSERT(rr->rr_col[x].rc_size <= rr->rr_col[VDEV_RAIDZ_Q].rc_size); |
34dc7c2f | 875 | |
b2255edc BB |
876 | for (c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { |
877 | uint64_t size = (c == x) ? 0 : MIN(rr->rr_col[x].rc_size, | |
878 | rr->rr_col[c].rc_size); | |
34dc7c2f | 879 | |
b2255edc BB |
880 | src = rr->rr_col[c].rc_abd; |
881 | dst = rr->rr_col[x].rc_abd; | |
34dc7c2f | 882 | |
b2255edc | 883 | if (c == rr->rr_firstdatacol) { |
a6255b7f | 884 | abd_copy(dst, src, size); |
b2255edc | 885 | if (rr->rr_col[x].rc_size > size) { |
a6255b7f | 886 | abd_zero_off(dst, size, |
b2255edc BB |
887 | rr->rr_col[x].rc_size - size); |
888 | } | |
34dc7c2f | 889 | } else { |
b2255edc | 890 | ASSERT3U(size, <=, rr->rr_col[x].rc_size); |
a6255b7f DQ |
891 | (void) abd_iterate_func2(dst, src, 0, 0, size, |
892 | vdev_raidz_reconst_q_pre_func, NULL); | |
893 | (void) abd_iterate_func(dst, | |
b2255edc | 894 | size, rr->rr_col[x].rc_size - size, |
a6255b7f | 895 | vdev_raidz_reconst_q_pre_tail_func, NULL); |
34dc7c2f BB |
896 | } |
897 | } | |
898 | ||
b2255edc BB |
899 | src = rr->rr_col[VDEV_RAIDZ_Q].rc_abd; |
900 | dst = rr->rr_col[x].rc_abd; | |
901 | exp = 255 - (rr->rr_cols - 1 - x); | |
34dc7c2f | 902 | |
1c27024e | 903 | struct reconst_q_struct rq = { abd_to_buf(src), exp }; |
b2255edc | 904 | (void) abd_iterate_func(dst, 0, rr->rr_col[x].rc_size, |
a6255b7f | 905 | vdev_raidz_reconst_q_post_func, &rq); |
34dc7c2f BB |
906 | } |
907 | ||
46df6e98 | 908 | static void |
b2255edc | 909 | vdev_raidz_reconstruct_pq(raidz_row_t *rr, int *tgts, int ntgts) |
34dc7c2f | 910 | { |
a6255b7f DQ |
911 | uint8_t *p, *q, *pxy, *qxy, tmp, a, b, aexp, bexp; |
912 | abd_t *pdata, *qdata; | |
913 | uint64_t xsize, ysize; | |
45d1cae3 BB |
914 | int x = tgts[0]; |
915 | int y = tgts[1]; | |
a6255b7f | 916 | abd_t *xd, *yd; |
34dc7c2f | 917 | |
45d1cae3 | 918 | ASSERT(ntgts == 2); |
34dc7c2f | 919 | ASSERT(x < y); |
b2255edc BB |
920 | ASSERT(x >= rr->rr_firstdatacol); |
921 | ASSERT(y < rr->rr_cols); | |
34dc7c2f | 922 | |
b2255edc | 923 | ASSERT(rr->rr_col[x].rc_size >= rr->rr_col[y].rc_size); |
34dc7c2f BB |
924 | |
925 | /* | |
926 | * Move the parity data aside -- we're going to compute parity as | |
927 | * though columns x and y were full of zeros -- Pxy and Qxy. We want to | |
928 | * reuse the parity generation mechanism without trashing the actual | |
929 | * parity so we make those columns appear to be full of zeros by | |
930 | * setting their lengths to zero. | |
931 | */ | |
b2255edc BB |
932 | pdata = rr->rr_col[VDEV_RAIDZ_P].rc_abd; |
933 | qdata = rr->rr_col[VDEV_RAIDZ_Q].rc_abd; | |
934 | xsize = rr->rr_col[x].rc_size; | |
935 | ysize = rr->rr_col[y].rc_size; | |
34dc7c2f | 936 | |
b2255edc BB |
937 | rr->rr_col[VDEV_RAIDZ_P].rc_abd = |
938 | abd_alloc_linear(rr->rr_col[VDEV_RAIDZ_P].rc_size, B_TRUE); | |
939 | rr->rr_col[VDEV_RAIDZ_Q].rc_abd = | |
940 | abd_alloc_linear(rr->rr_col[VDEV_RAIDZ_Q].rc_size, B_TRUE); | |
941 | rr->rr_col[x].rc_size = 0; | |
942 | rr->rr_col[y].rc_size = 0; | |
34dc7c2f | 943 | |
b2255edc | 944 | vdev_raidz_generate_parity_pq(rr); |
34dc7c2f | 945 | |
b2255edc BB |
946 | rr->rr_col[x].rc_size = xsize; |
947 | rr->rr_col[y].rc_size = ysize; | |
34dc7c2f | 948 | |
a6255b7f DQ |
949 | p = abd_to_buf(pdata); |
950 | q = abd_to_buf(qdata); | |
b2255edc BB |
951 | pxy = abd_to_buf(rr->rr_col[VDEV_RAIDZ_P].rc_abd); |
952 | qxy = abd_to_buf(rr->rr_col[VDEV_RAIDZ_Q].rc_abd); | |
953 | xd = rr->rr_col[x].rc_abd; | |
954 | yd = rr->rr_col[y].rc_abd; | |
34dc7c2f BB |
955 | |
956 | /* | |
957 | * We now have: | |
958 | * Pxy = P + D_x + D_y | |
959 | * Qxy = Q + 2^(ndevs - 1 - x) * D_x + 2^(ndevs - 1 - y) * D_y | |
960 | * | |
961 | * We can then solve for D_x: | |
962 | * D_x = A * (P + Pxy) + B * (Q + Qxy) | |
963 | * where | |
964 | * A = 2^(x - y) * (2^(x - y) + 1)^-1 | |
965 | * B = 2^(ndevs - 1 - x) * (2^(x - y) + 1)^-1 | |
966 | * | |
967 | * With D_x in hand, we can easily solve for D_y: | |
968 | * D_y = P + Pxy + D_x | |
969 | */ | |
970 | ||
971 | a = vdev_raidz_pow2[255 + x - y]; | |
b2255edc | 972 | b = vdev_raidz_pow2[255 - (rr->rr_cols - 1 - x)]; |
34dc7c2f BB |
973 | tmp = 255 - vdev_raidz_log2[a ^ 1]; |
974 | ||
975 | aexp = vdev_raidz_log2[vdev_raidz_exp2(a, tmp)]; | |
976 | bexp = vdev_raidz_log2[vdev_raidz_exp2(b, tmp)]; | |
977 | ||
a6255b7f | 978 | ASSERT3U(xsize, >=, ysize); |
1c27024e | 979 | struct reconst_pq_struct rpq = { p, q, pxy, qxy, aexp, bexp }; |
34dc7c2f | 980 | |
a6255b7f DQ |
981 | (void) abd_iterate_func2(xd, yd, 0, 0, ysize, |
982 | vdev_raidz_reconst_pq_func, &rpq); | |
983 | (void) abd_iterate_func(xd, ysize, xsize - ysize, | |
984 | vdev_raidz_reconst_pq_tail_func, &rpq); | |
34dc7c2f | 985 | |
b2255edc BB |
986 | abd_free(rr->rr_col[VDEV_RAIDZ_P].rc_abd); |
987 | abd_free(rr->rr_col[VDEV_RAIDZ_Q].rc_abd); | |
34dc7c2f BB |
988 | |
989 | /* | |
990 | * Restore the saved parity data. | |
991 | */ | |
b2255edc BB |
992 | rr->rr_col[VDEV_RAIDZ_P].rc_abd = pdata; |
993 | rr->rr_col[VDEV_RAIDZ_Q].rc_abd = qdata; | |
45d1cae3 BB |
994 | } |
995 | ||
996 | /* BEGIN CSTYLED */ | |
997 | /* | |
998 | * In the general case of reconstruction, we must solve the system of linear | |
dd4bc569 | 999 | * equations defined by the coefficients used to generate parity as well as |
45d1cae3 BB |
1000 | * the contents of the data and parity disks. This can be expressed with |
1001 | * vectors for the original data (D) and the actual data (d) and parity (p) | |
1002 | * and a matrix composed of the identity matrix (I) and a dispersal matrix (V): | |
1003 | * | |
1004 | * __ __ __ __ | |
1005 | * | | __ __ | p_0 | | |
1006 | * | V | | D_0 | | p_m-1 | | |
1007 | * | | x | : | = | d_0 | | |
1008 | * | I | | D_n-1 | | : | | |
1009 | * | | ~~ ~~ | d_n-1 | | |
1010 | * ~~ ~~ ~~ ~~ | |
1011 | * | |
1012 | * I is simply a square identity matrix of size n, and V is a vandermonde | |
dd4bc569 | 1013 | * matrix defined by the coefficients we chose for the various parity columns |
45d1cae3 BB |
1014 | * (1, 2, 4). Note that these values were chosen both for simplicity, speedy |
1015 | * computation as well as linear separability. | |
1016 | * | |
1017 | * __ __ __ __ | |
1018 | * | 1 .. 1 1 1 | | p_0 | | |
1019 | * | 2^n-1 .. 4 2 1 | __ __ | : | | |
1020 | * | 4^n-1 .. 16 4 1 | | D_0 | | p_m-1 | | |
1021 | * | 1 .. 0 0 0 | | D_1 | | d_0 | | |
1022 | * | 0 .. 0 0 0 | x | D_2 | = | d_1 | | |
1023 | * | : : : : | | : | | d_2 | | |
1024 | * | 0 .. 1 0 0 | | D_n-1 | | : | | |
1025 | * | 0 .. 0 1 0 | ~~ ~~ | : | | |
1026 | * | 0 .. 0 0 1 | | d_n-1 | | |
1027 | * ~~ ~~ ~~ ~~ | |
1028 | * | |
1029 | * Note that I, V, d, and p are known. To compute D, we must invert the | |
1030 | * matrix and use the known data and parity values to reconstruct the unknown | |
1031 | * data values. We begin by removing the rows in V|I and d|p that correspond | |
1032 | * to failed or missing columns; we then make V|I square (n x n) and d|p | |
1033 | * sized n by removing rows corresponding to unused parity from the bottom up | |
1034 | * to generate (V|I)' and (d|p)'. We can then generate the inverse of (V|I)' | |
1035 | * using Gauss-Jordan elimination. In the example below we use m=3 parity | |
1036 | * columns, n=8 data columns, with errors in d_1, d_2, and p_1: | |
1037 | * __ __ | |
1038 | * | 1 1 1 1 1 1 1 1 | | |
1039 | * | 128 64 32 16 8 4 2 1 | <-----+-+-- missing disks | |
1040 | * | 19 205 116 29 64 16 4 1 | / / | |
1041 | * | 1 0 0 0 0 0 0 0 | / / | |
1042 | * | 0 1 0 0 0 0 0 0 | <--' / | |
1043 | * (V|I) = | 0 0 1 0 0 0 0 0 | <---' | |
1044 | * | 0 0 0 1 0 0 0 0 | | |
1045 | * | 0 0 0 0 1 0 0 0 | | |
1046 | * | 0 0 0 0 0 1 0 0 | | |
1047 | * | 0 0 0 0 0 0 1 0 | | |
1048 | * | 0 0 0 0 0 0 0 1 | | |
1049 | * ~~ ~~ | |
1050 | * __ __ | |
1051 | * | 1 1 1 1 1 1 1 1 | | |
1052 | * | 128 64 32 16 8 4 2 1 | | |
1053 | * | 19 205 116 29 64 16 4 1 | | |
1054 | * | 1 0 0 0 0 0 0 0 | | |
1055 | * | 0 1 0 0 0 0 0 0 | | |
1056 | * (V|I)' = | 0 0 1 0 0 0 0 0 | | |
1057 | * | 0 0 0 1 0 0 0 0 | | |
1058 | * | 0 0 0 0 1 0 0 0 | | |
1059 | * | 0 0 0 0 0 1 0 0 | | |
1060 | * | 0 0 0 0 0 0 1 0 | | |
1061 | * | 0 0 0 0 0 0 0 1 | | |
1062 | * ~~ ~~ | |
1063 | * | |
1064 | * Here we employ Gauss-Jordan elimination to find the inverse of (V|I)'. We | |
1065 | * have carefully chosen the seed values 1, 2, and 4 to ensure that this | |
1066 | * matrix is not singular. | |
1067 | * __ __ | |
1068 | * | 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 | | |
1069 | * | 19 205 116 29 64 16 4 1 0 1 0 0 0 0 0 0 | | |
1070 | * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 | | |
1071 | * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 | | |
1072 | * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 | | |
1073 | * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 | | |
1074 | * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 | | |
1075 | * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 | | |
1076 | * ~~ ~~ | |
1077 | * __ __ | |
1078 | * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 | | |
1079 | * | 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 | | |
1080 | * | 19 205 116 29 64 16 4 1 0 1 0 0 0 0 0 0 | | |
1081 | * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 | | |
1082 | * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 | | |
1083 | * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 | | |
1084 | * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 | | |
1085 | * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 | | |
1086 | * ~~ ~~ | |
1087 | * __ __ | |
1088 | * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 | | |
1089 | * | 0 1 1 0 0 0 0 0 1 0 1 1 1 1 1 1 | | |
1090 | * | 0 205 116 0 0 0 0 0 0 1 19 29 64 16 4 1 | | |
1091 | * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 | | |
1092 | * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 | | |
1093 | * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 | | |
1094 | * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 | | |
1095 | * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 | | |
1096 | * ~~ ~~ | |
1097 | * __ __ | |
1098 | * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 | | |
1099 | * | 0 1 1 0 0 0 0 0 1 0 1 1 1 1 1 1 | | |
1100 | * | 0 0 185 0 0 0 0 0 205 1 222 208 141 221 201 204 | | |
1101 | * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 | | |
1102 | * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 | | |
1103 | * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 | | |
1104 | * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 | | |
1105 | * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 | | |
1106 | * ~~ ~~ | |
1107 | * __ __ | |
1108 | * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 | | |
1109 | * | 0 1 1 0 0 0 0 0 1 0 1 1 1 1 1 1 | | |
1110 | * | 0 0 1 0 0 0 0 0 166 100 4 40 158 168 216 209 | | |
1111 | * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 | | |
1112 | * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 | | |
1113 | * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 | | |
1114 | * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 | | |
1115 | * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 | | |
1116 | * ~~ ~~ | |
1117 | * __ __ | |
1118 | * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 | | |
1119 | * | 0 1 0 0 0 0 0 0 167 100 5 41 159 169 217 208 | | |
1120 | * | 0 0 1 0 0 0 0 0 166 100 4 40 158 168 216 209 | | |
1121 | * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 | | |
1122 | * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 | | |
1123 | * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 | | |
1124 | * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 | | |
1125 | * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 | | |
1126 | * ~~ ~~ | |
1127 | * __ __ | |
1128 | * | 0 0 1 0 0 0 0 0 | | |
1129 | * | 167 100 5 41 159 169 217 208 | | |
1130 | * | 166 100 4 40 158 168 216 209 | | |
1131 | * (V|I)'^-1 = | 0 0 0 1 0 0 0 0 | | |
1132 | * | 0 0 0 0 1 0 0 0 | | |
1133 | * | 0 0 0 0 0 1 0 0 | | |
1134 | * | 0 0 0 0 0 0 1 0 | | |
1135 | * | 0 0 0 0 0 0 0 1 | | |
1136 | * ~~ ~~ | |
1137 | * | |
1138 | * We can then simply compute D = (V|I)'^-1 x (d|p)' to discover the values | |
1139 | * of the missing data. | |
1140 | * | |
1141 | * As is apparent from the example above, the only non-trivial rows in the | |
1142 | * inverse matrix correspond to the data disks that we're trying to | |
1143 | * reconstruct. Indeed, those are the only rows we need as the others would | |
1144 | * only be useful for reconstructing data known or assumed to be valid. For | |
1145 | * that reason, we only build the coefficients in the rows that correspond to | |
1146 | * targeted columns. | |
1147 | */ | |
1148 | /* END CSTYLED */ | |
1149 | ||
1150 | static void | |
b2255edc | 1151 | vdev_raidz_matrix_init(raidz_row_t *rr, int n, int nmap, int *map, |
45d1cae3 BB |
1152 | uint8_t **rows) |
1153 | { | |
1154 | int i, j; | |
1155 | int pow; | |
1156 | ||
b2255edc | 1157 | ASSERT(n == rr->rr_cols - rr->rr_firstdatacol); |
45d1cae3 BB |
1158 | |
1159 | /* | |
1160 | * Fill in the missing rows of interest. | |
1161 | */ | |
1162 | for (i = 0; i < nmap; i++) { | |
1163 | ASSERT3S(0, <=, map[i]); | |
1164 | ASSERT3S(map[i], <=, 2); | |
1165 | ||
1166 | pow = map[i] * n; | |
1167 | if (pow > 255) | |
1168 | pow -= 255; | |
1169 | ASSERT(pow <= 255); | |
1170 | ||
1171 | for (j = 0; j < n; j++) { | |
1172 | pow -= map[i]; | |
1173 | if (pow < 0) | |
1174 | pow += 255; | |
1175 | rows[i][j] = vdev_raidz_pow2[pow]; | |
1176 | } | |
1177 | } | |
1178 | } | |
1179 | ||
1180 | static void | |
b2255edc | 1181 | vdev_raidz_matrix_invert(raidz_row_t *rr, int n, int nmissing, int *missing, |
45d1cae3 BB |
1182 | uint8_t **rows, uint8_t **invrows, const uint8_t *used) |
1183 | { | |
1184 | int i, j, ii, jj; | |
1185 | uint8_t log; | |
1186 | ||
1187 | /* | |
1188 | * Assert that the first nmissing entries from the array of used | |
1189 | * columns correspond to parity columns and that subsequent entries | |
1190 | * correspond to data columns. | |
1191 | */ | |
1192 | for (i = 0; i < nmissing; i++) { | |
b2255edc | 1193 | ASSERT3S(used[i], <, rr->rr_firstdatacol); |
45d1cae3 BB |
1194 | } |
1195 | for (; i < n; i++) { | |
b2255edc | 1196 | ASSERT3S(used[i], >=, rr->rr_firstdatacol); |
45d1cae3 BB |
1197 | } |
1198 | ||
1199 | /* | |
1200 | * First initialize the storage where we'll compute the inverse rows. | |
1201 | */ | |
1202 | for (i = 0; i < nmissing; i++) { | |
1203 | for (j = 0; j < n; j++) { | |
1204 | invrows[i][j] = (i == j) ? 1 : 0; | |
1205 | } | |
1206 | } | |
1207 | ||
1208 | /* | |
1209 | * Subtract all trivial rows from the rows of consequence. | |
1210 | */ | |
1211 | for (i = 0; i < nmissing; i++) { | |
1212 | for (j = nmissing; j < n; j++) { | |
b2255edc BB |
1213 | ASSERT3U(used[j], >=, rr->rr_firstdatacol); |
1214 | jj = used[j] - rr->rr_firstdatacol; | |
45d1cae3 BB |
1215 | ASSERT3S(jj, <, n); |
1216 | invrows[i][j] = rows[i][jj]; | |
1217 | rows[i][jj] = 0; | |
1218 | } | |
1219 | } | |
1220 | ||
1221 | /* | |
1222 | * For each of the rows of interest, we must normalize it and subtract | |
1223 | * a multiple of it from the other rows. | |
1224 | */ | |
1225 | for (i = 0; i < nmissing; i++) { | |
1226 | for (j = 0; j < missing[i]; j++) { | |
c99c9001 | 1227 | ASSERT0(rows[i][j]); |
45d1cae3 BB |
1228 | } |
1229 | ASSERT3U(rows[i][missing[i]], !=, 0); | |
1230 | ||
1231 | /* | |
1232 | * Compute the inverse of the first element and multiply each | |
1233 | * element in the row by that value. | |
1234 | */ | |
1235 | log = 255 - vdev_raidz_log2[rows[i][missing[i]]]; | |
1236 | ||
1237 | for (j = 0; j < n; j++) { | |
1238 | rows[i][j] = vdev_raidz_exp2(rows[i][j], log); | |
1239 | invrows[i][j] = vdev_raidz_exp2(invrows[i][j], log); | |
1240 | } | |
1241 | ||
1242 | for (ii = 0; ii < nmissing; ii++) { | |
1243 | if (i == ii) | |
1244 | continue; | |
1245 | ||
1246 | ASSERT3U(rows[ii][missing[i]], !=, 0); | |
1247 | ||
1248 | log = vdev_raidz_log2[rows[ii][missing[i]]]; | |
1249 | ||
1250 | for (j = 0; j < n; j++) { | |
1251 | rows[ii][j] ^= | |
1252 | vdev_raidz_exp2(rows[i][j], log); | |
1253 | invrows[ii][j] ^= | |
1254 | vdev_raidz_exp2(invrows[i][j], log); | |
1255 | } | |
1256 | } | |
1257 | } | |
1258 | ||
1259 | /* | |
1260 | * Verify that the data that is left in the rows are properly part of | |
1261 | * an identity matrix. | |
1262 | */ | |
1263 | for (i = 0; i < nmissing; i++) { | |
1264 | for (j = 0; j < n; j++) { | |
1265 | if (j == missing[i]) { | |
1266 | ASSERT3U(rows[i][j], ==, 1); | |
1267 | } else { | |
c99c9001 | 1268 | ASSERT0(rows[i][j]); |
45d1cae3 BB |
1269 | } |
1270 | } | |
1271 | } | |
1272 | } | |
1273 | ||
1274 | static void | |
b2255edc | 1275 | vdev_raidz_matrix_reconstruct(raidz_row_t *rr, int n, int nmissing, |
45d1cae3 BB |
1276 | int *missing, uint8_t **invrows, const uint8_t *used) |
1277 | { | |
1278 | int i, j, x, cc, c; | |
1279 | uint8_t *src; | |
1280 | uint64_t ccount; | |
689f093e GN |
1281 | uint8_t *dst[VDEV_RAIDZ_MAXPARITY] = { NULL }; |
1282 | uint64_t dcount[VDEV_RAIDZ_MAXPARITY] = { 0 }; | |
a117a6d6 GW |
1283 | uint8_t log = 0; |
1284 | uint8_t val; | |
45d1cae3 BB |
1285 | int ll; |
1286 | uint8_t *invlog[VDEV_RAIDZ_MAXPARITY]; | |
1287 | uint8_t *p, *pp; | |
1288 | size_t psize; | |
1289 | ||
1290 | psize = sizeof (invlog[0][0]) * n * nmissing; | |
79c76d5b | 1291 | p = kmem_alloc(psize, KM_SLEEP); |
45d1cae3 BB |
1292 | |
1293 | for (pp = p, i = 0; i < nmissing; i++) { | |
1294 | invlog[i] = pp; | |
1295 | pp += n; | |
1296 | } | |
1297 | ||
1298 | for (i = 0; i < nmissing; i++) { | |
1299 | for (j = 0; j < n; j++) { | |
1300 | ASSERT3U(invrows[i][j], !=, 0); | |
1301 | invlog[i][j] = vdev_raidz_log2[invrows[i][j]]; | |
1302 | } | |
1303 | } | |
1304 | ||
1305 | for (i = 0; i < n; i++) { | |
1306 | c = used[i]; | |
b2255edc | 1307 | ASSERT3U(c, <, rr->rr_cols); |
45d1cae3 | 1308 | |
b2255edc BB |
1309 | ccount = rr->rr_col[c].rc_size; |
1310 | ASSERT(ccount >= rr->rr_col[missing[0]].rc_size || i > 0); | |
1311 | if (ccount == 0) | |
1312 | continue; | |
1313 | src = abd_to_buf(rr->rr_col[c].rc_abd); | |
45d1cae3 | 1314 | for (j = 0; j < nmissing; j++) { |
b2255edc BB |
1315 | cc = missing[j] + rr->rr_firstdatacol; |
1316 | ASSERT3U(cc, >=, rr->rr_firstdatacol); | |
1317 | ASSERT3U(cc, <, rr->rr_cols); | |
45d1cae3 BB |
1318 | ASSERT3U(cc, !=, c); |
1319 | ||
b2255edc BB |
1320 | dcount[j] = rr->rr_col[cc].rc_size; |
1321 | if (dcount[j] != 0) | |
1322 | dst[j] = abd_to_buf(rr->rr_col[cc].rc_abd); | |
45d1cae3 BB |
1323 | } |
1324 | ||
45d1cae3 BB |
1325 | for (x = 0; x < ccount; x++, src++) { |
1326 | if (*src != 0) | |
1327 | log = vdev_raidz_log2[*src]; | |
1328 | ||
1329 | for (cc = 0; cc < nmissing; cc++) { | |
1330 | if (x >= dcount[cc]) | |
1331 | continue; | |
1332 | ||
1333 | if (*src == 0) { | |
1334 | val = 0; | |
1335 | } else { | |
1336 | if ((ll = log + invlog[cc][i]) >= 255) | |
1337 | ll -= 255; | |
1338 | val = vdev_raidz_pow2[ll]; | |
1339 | } | |
1340 | ||
1341 | if (i == 0) | |
1342 | dst[cc][x] = val; | |
1343 | else | |
1344 | dst[cc][x] ^= val; | |
1345 | } | |
1346 | } | |
1347 | } | |
1348 | ||
1349 | kmem_free(p, psize); | |
1350 | } | |
1351 | ||
46df6e98 | 1352 | static void |
b2255edc | 1353 | vdev_raidz_reconstruct_general(raidz_row_t *rr, int *tgts, int ntgts) |
45d1cae3 BB |
1354 | { |
1355 | int n, i, c, t, tt; | |
1356 | int nmissing_rows; | |
1357 | int missing_rows[VDEV_RAIDZ_MAXPARITY]; | |
1358 | int parity_map[VDEV_RAIDZ_MAXPARITY]; | |
45d1cae3 BB |
1359 | uint8_t *p, *pp; |
1360 | size_t psize; | |
45d1cae3 BB |
1361 | uint8_t *rows[VDEV_RAIDZ_MAXPARITY]; |
1362 | uint8_t *invrows[VDEV_RAIDZ_MAXPARITY]; | |
1363 | uint8_t *used; | |
1364 | ||
a6255b7f DQ |
1365 | abd_t **bufs = NULL; |
1366 | ||
a6255b7f DQ |
1367 | /* |
1368 | * Matrix reconstruction can't use scatter ABDs yet, so we allocate | |
b2255edc | 1369 | * temporary linear ABDs if any non-linear ABDs are found. |
a6255b7f | 1370 | */ |
b2255edc BB |
1371 | for (i = rr->rr_firstdatacol; i < rr->rr_cols; i++) { |
1372 | if (!abd_is_linear(rr->rr_col[i].rc_abd)) { | |
1373 | bufs = kmem_alloc(rr->rr_cols * sizeof (abd_t *), | |
1374 | KM_PUSHPAGE); | |
1375 | ||
1376 | for (c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { | |
1377 | raidz_col_t *col = &rr->rr_col[c]; | |
1378 | ||
1379 | bufs[c] = col->rc_abd; | |
1380 | if (bufs[c] != NULL) { | |
1381 | col->rc_abd = abd_alloc_linear( | |
1382 | col->rc_size, B_TRUE); | |
1383 | abd_copy(col->rc_abd, bufs[c], | |
1384 | col->rc_size); | |
1385 | } | |
1386 | } | |
a6255b7f | 1387 | |
b2255edc | 1388 | break; |
a6255b7f DQ |
1389 | } |
1390 | } | |
45d1cae3 | 1391 | |
b2255edc | 1392 | n = rr->rr_cols - rr->rr_firstdatacol; |
45d1cae3 BB |
1393 | |
1394 | /* | |
1395 | * Figure out which data columns are missing. | |
1396 | */ | |
1397 | nmissing_rows = 0; | |
1398 | for (t = 0; t < ntgts; t++) { | |
b2255edc | 1399 | if (tgts[t] >= rr->rr_firstdatacol) { |
45d1cae3 | 1400 | missing_rows[nmissing_rows++] = |
b2255edc | 1401 | tgts[t] - rr->rr_firstdatacol; |
45d1cae3 BB |
1402 | } |
1403 | } | |
1404 | ||
1405 | /* | |
1406 | * Figure out which parity columns to use to help generate the missing | |
1407 | * data columns. | |
1408 | */ | |
1409 | for (tt = 0, c = 0, i = 0; i < nmissing_rows; c++) { | |
1410 | ASSERT(tt < ntgts); | |
b2255edc | 1411 | ASSERT(c < rr->rr_firstdatacol); |
45d1cae3 BB |
1412 | |
1413 | /* | |
1414 | * Skip any targeted parity columns. | |
1415 | */ | |
1416 | if (c == tgts[tt]) { | |
1417 | tt++; | |
1418 | continue; | |
1419 | } | |
1420 | ||
45d1cae3 BB |
1421 | parity_map[i] = c; |
1422 | i++; | |
1423 | } | |
1424 | ||
45d1cae3 BB |
1425 | psize = (sizeof (rows[0][0]) + sizeof (invrows[0][0])) * |
1426 | nmissing_rows * n + sizeof (used[0]) * n; | |
79c76d5b | 1427 | p = kmem_alloc(psize, KM_SLEEP); |
45d1cae3 BB |
1428 | |
1429 | for (pp = p, i = 0; i < nmissing_rows; i++) { | |
1430 | rows[i] = pp; | |
1431 | pp += n; | |
1432 | invrows[i] = pp; | |
1433 | pp += n; | |
1434 | } | |
1435 | used = pp; | |
1436 | ||
1437 | for (i = 0; i < nmissing_rows; i++) { | |
1438 | used[i] = parity_map[i]; | |
1439 | } | |
1440 | ||
b2255edc | 1441 | for (tt = 0, c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { |
45d1cae3 | 1442 | if (tt < nmissing_rows && |
b2255edc | 1443 | c == missing_rows[tt] + rr->rr_firstdatacol) { |
45d1cae3 BB |
1444 | tt++; |
1445 | continue; | |
1446 | } | |
1447 | ||
1448 | ASSERT3S(i, <, n); | |
1449 | used[i] = c; | |
1450 | i++; | |
1451 | } | |
1452 | ||
1453 | /* | |
1454 | * Initialize the interesting rows of the matrix. | |
1455 | */ | |
b2255edc | 1456 | vdev_raidz_matrix_init(rr, n, nmissing_rows, parity_map, rows); |
45d1cae3 BB |
1457 | |
1458 | /* | |
1459 | * Invert the matrix. | |
1460 | */ | |
b2255edc | 1461 | vdev_raidz_matrix_invert(rr, n, nmissing_rows, missing_rows, rows, |
45d1cae3 BB |
1462 | invrows, used); |
1463 | ||
1464 | /* | |
1465 | * Reconstruct the missing data using the generated matrix. | |
1466 | */ | |
b2255edc | 1467 | vdev_raidz_matrix_reconstruct(rr, n, nmissing_rows, missing_rows, |
45d1cae3 BB |
1468 | invrows, used); |
1469 | ||
1470 | kmem_free(p, psize); | |
1471 | ||
a6255b7f DQ |
1472 | /* |
1473 | * copy back from temporary linear abds and free them | |
1474 | */ | |
1475 | if (bufs) { | |
b2255edc BB |
1476 | for (c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { |
1477 | raidz_col_t *col = &rr->rr_col[c]; | |
a6255b7f | 1478 | |
b2255edc BB |
1479 | if (bufs[c] != NULL) { |
1480 | abd_copy(bufs[c], col->rc_abd, col->rc_size); | |
1481 | abd_free(col->rc_abd); | |
1482 | } | |
a6255b7f DQ |
1483 | col->rc_abd = bufs[c]; |
1484 | } | |
b2255edc | 1485 | kmem_free(bufs, rr->rr_cols * sizeof (abd_t *)); |
a6255b7f | 1486 | } |
34dc7c2f BB |
1487 | } |
1488 | ||
46df6e98 | 1489 | static void |
b2255edc BB |
1490 | vdev_raidz_reconstruct_row(raidz_map_t *rm, raidz_row_t *rr, |
1491 | const int *t, int nt) | |
45d1cae3 BB |
1492 | { |
1493 | int tgts[VDEV_RAIDZ_MAXPARITY], *dt; | |
1494 | int ntgts; | |
c9187d86 | 1495 | int i, c, ret; |
45d1cae3 BB |
1496 | int nbadparity, nbaddata; |
1497 | int parity_valid[VDEV_RAIDZ_MAXPARITY]; | |
1498 | ||
b2255edc BB |
1499 | nbadparity = rr->rr_firstdatacol; |
1500 | nbaddata = rr->rr_cols - nbadparity; | |
45d1cae3 | 1501 | ntgts = 0; |
b2255edc BB |
1502 | for (i = 0, c = 0; c < rr->rr_cols; c++) { |
1503 | if (c < rr->rr_firstdatacol) | |
45d1cae3 BB |
1504 | parity_valid[c] = B_FALSE; |
1505 | ||
1506 | if (i < nt && c == t[i]) { | |
1507 | tgts[ntgts++] = c; | |
1508 | i++; | |
b2255edc | 1509 | } else if (rr->rr_col[c].rc_error != 0) { |
45d1cae3 | 1510 | tgts[ntgts++] = c; |
b2255edc | 1511 | } else if (c >= rr->rr_firstdatacol) { |
45d1cae3 BB |
1512 | nbaddata--; |
1513 | } else { | |
1514 | parity_valid[c] = B_TRUE; | |
1515 | nbadparity--; | |
1516 | } | |
1517 | } | |
1518 | ||
1519 | ASSERT(ntgts >= nt); | |
1520 | ASSERT(nbaddata >= 0); | |
1521 | ASSERT(nbaddata + nbadparity == ntgts); | |
1522 | ||
1523 | dt = &tgts[nbadparity]; | |
1524 | ||
c9187d86 | 1525 | /* Reconstruct using the new math implementation */ |
b2255edc | 1526 | ret = vdev_raidz_math_reconstruct(rm, rr, parity_valid, dt, nbaddata); |
c9187d86 | 1527 | if (ret != RAIDZ_ORIGINAL_IMPL) |
46df6e98 | 1528 | return; |
ab9f4b0b | 1529 | |
45d1cae3 BB |
1530 | /* |
1531 | * See if we can use any of our optimized reconstruction routines. | |
1532 | */ | |
ab9f4b0b GN |
1533 | switch (nbaddata) { |
1534 | case 1: | |
46df6e98 MA |
1535 | if (parity_valid[VDEV_RAIDZ_P]) { |
1536 | vdev_raidz_reconstruct_p(rr, dt, 1); | |
1537 | return; | |
1538 | } | |
45d1cae3 | 1539 | |
b2255edc | 1540 | ASSERT(rr->rr_firstdatacol > 1); |
45d1cae3 | 1541 | |
46df6e98 MA |
1542 | if (parity_valid[VDEV_RAIDZ_Q]) { |
1543 | vdev_raidz_reconstruct_q(rr, dt, 1); | |
1544 | return; | |
1545 | } | |
45d1cae3 | 1546 | |
b2255edc | 1547 | ASSERT(rr->rr_firstdatacol > 2); |
ab9f4b0b | 1548 | break; |
45d1cae3 | 1549 | |
ab9f4b0b | 1550 | case 2: |
b2255edc | 1551 | ASSERT(rr->rr_firstdatacol > 1); |
45d1cae3 | 1552 | |
ab9f4b0b | 1553 | if (parity_valid[VDEV_RAIDZ_P] && |
46df6e98 MA |
1554 | parity_valid[VDEV_RAIDZ_Q]) { |
1555 | vdev_raidz_reconstruct_pq(rr, dt, 2); | |
1556 | return; | |
1557 | } | |
45d1cae3 | 1558 | |
b2255edc | 1559 | ASSERT(rr->rr_firstdatacol > 2); |
45d1cae3 | 1560 | |
ab9f4b0b | 1561 | break; |
45d1cae3 BB |
1562 | } |
1563 | ||
46df6e98 | 1564 | vdev_raidz_reconstruct_general(rr, tgts, ntgts); |
45d1cae3 | 1565 | } |
34dc7c2f BB |
1566 | |
1567 | static int | |
1bd201e7 | 1568 | vdev_raidz_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize, |
6fe3498c | 1569 | uint64_t *logical_ashift, uint64_t *physical_ashift) |
34dc7c2f | 1570 | { |
b2255edc BB |
1571 | vdev_raidz_t *vdrz = vd->vdev_tsd; |
1572 | uint64_t nparity = vdrz->vd_nparity; | |
45d1cae3 | 1573 | int c; |
34dc7c2f BB |
1574 | int lasterror = 0; |
1575 | int numerrors = 0; | |
1576 | ||
1577 | ASSERT(nparity > 0); | |
1578 | ||
1579 | if (nparity > VDEV_RAIDZ_MAXPARITY || | |
1580 | vd->vdev_children < nparity + 1) { | |
1581 | vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL; | |
2e528b49 | 1582 | return (SET_ERROR(EINVAL)); |
34dc7c2f BB |
1583 | } |
1584 | ||
45d1cae3 BB |
1585 | vdev_open_children(vd); |
1586 | ||
34dc7c2f | 1587 | for (c = 0; c < vd->vdev_children; c++) { |
b2255edc | 1588 | vdev_t *cvd = vd->vdev_child[c]; |
34dc7c2f | 1589 | |
45d1cae3 BB |
1590 | if (cvd->vdev_open_error != 0) { |
1591 | lasterror = cvd->vdev_open_error; | |
34dc7c2f BB |
1592 | numerrors++; |
1593 | continue; | |
1594 | } | |
1595 | ||
1596 | *asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1; | |
1bd201e7 | 1597 | *max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1; |
6fe3498c RM |
1598 | *logical_ashift = MAX(*logical_ashift, cvd->vdev_ashift); |
1599 | *physical_ashift = MAX(*physical_ashift, | |
1600 | cvd->vdev_physical_ashift); | |
34dc7c2f BB |
1601 | } |
1602 | ||
1603 | *asize *= vd->vdev_children; | |
1bd201e7 | 1604 | *max_asize *= vd->vdev_children; |
34dc7c2f BB |
1605 | |
1606 | if (numerrors > nparity) { | |
1607 | vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS; | |
1608 | return (lasterror); | |
1609 | } | |
1610 | ||
1611 | return (0); | |
1612 | } | |
1613 | ||
1614 | static void | |
1615 | vdev_raidz_close(vdev_t *vd) | |
1616 | { | |
b2255edc BB |
1617 | for (int c = 0; c < vd->vdev_children; c++) { |
1618 | if (vd->vdev_child[c] != NULL) | |
1619 | vdev_close(vd->vdev_child[c]); | |
1620 | } | |
34dc7c2f BB |
1621 | } |
1622 | ||
1623 | static uint64_t | |
1624 | vdev_raidz_asize(vdev_t *vd, uint64_t psize) | |
1625 | { | |
b2255edc | 1626 | vdev_raidz_t *vdrz = vd->vdev_tsd; |
34dc7c2f BB |
1627 | uint64_t asize; |
1628 | uint64_t ashift = vd->vdev_top->vdev_ashift; | |
b2255edc BB |
1629 | uint64_t cols = vdrz->vd_logical_width; |
1630 | uint64_t nparity = vdrz->vd_nparity; | |
34dc7c2f BB |
1631 | |
1632 | asize = ((psize - 1) >> ashift) + 1; | |
1633 | asize += nparity * ((asize + cols - nparity - 1) / (cols - nparity)); | |
1634 | asize = roundup(asize, nparity + 1) << ashift; | |
1635 | ||
1636 | return (asize); | |
1637 | } | |
1638 | ||
b2255edc BB |
1639 | /* |
1640 | * The allocatable space for a raidz vdev is N * sizeof(smallest child) | |
1641 | * so each child must provide at least 1/Nth of its asize. | |
1642 | */ | |
1643 | static uint64_t | |
1644 | vdev_raidz_min_asize(vdev_t *vd) | |
1645 | { | |
1646 | return ((vd->vdev_min_asize + vd->vdev_children - 1) / | |
1647 | vd->vdev_children); | |
1648 | } | |
1649 | ||
1650 | void | |
34dc7c2f BB |
1651 | vdev_raidz_child_done(zio_t *zio) |
1652 | { | |
1653 | raidz_col_t *rc = zio->io_private; | |
1654 | ||
1655 | rc->rc_error = zio->io_error; | |
1656 | rc->rc_tried = 1; | |
1657 | rc->rc_skipped = 0; | |
1658 | } | |
1659 | ||
619f0976 | 1660 | static void |
b2255edc | 1661 | vdev_raidz_io_verify(vdev_t *vd, raidz_row_t *rr, int col) |
619f0976 GW |
1662 | { |
1663 | #ifdef ZFS_DEBUG | |
619f0976 GW |
1664 | vdev_t *tvd = vd->vdev_top; |
1665 | ||
b2255edc BB |
1666 | range_seg64_t logical_rs, physical_rs, remain_rs; |
1667 | logical_rs.rs_start = rr->rr_offset; | |
619f0976 | 1668 | logical_rs.rs_end = logical_rs.rs_start + |
b2255edc | 1669 | vdev_raidz_asize(vd, rr->rr_size); |
619f0976 | 1670 | |
b2255edc | 1671 | raidz_col_t *rc = &rr->rr_col[col]; |
619f0976 GW |
1672 | vdev_t *cvd = vd->vdev_child[rc->rc_devidx]; |
1673 | ||
b2255edc BB |
1674 | vdev_xlate(cvd, &logical_rs, &physical_rs, &remain_rs); |
1675 | ASSERT(vdev_xlate_is_empty(&remain_rs)); | |
619f0976 GW |
1676 | ASSERT3U(rc->rc_offset, ==, physical_rs.rs_start); |
1677 | ASSERT3U(rc->rc_offset, <, physical_rs.rs_end); | |
1678 | /* | |
1679 | * It would be nice to assert that rs_end is equal | |
1680 | * to rc_offset + rc_size but there might be an | |
1681 | * optional I/O at the end that is not accounted in | |
1682 | * rc_size. | |
1683 | */ | |
1684 | if (physical_rs.rs_end > rc->rc_offset + rc->rc_size) { | |
1685 | ASSERT3U(physical_rs.rs_end, ==, rc->rc_offset + | |
1686 | rc->rc_size + (1 << tvd->vdev_ashift)); | |
1687 | } else { | |
1688 | ASSERT3U(physical_rs.rs_end, ==, rc->rc_offset + rc->rc_size); | |
1689 | } | |
1690 | #endif | |
1691 | } | |
1692 | ||
98b25418 | 1693 | static void |
b2255edc | 1694 | vdev_raidz_io_start_write(zio_t *zio, raidz_row_t *rr, uint64_t ashift) |
34dc7c2f BB |
1695 | { |
1696 | vdev_t *vd = zio->io_vd; | |
b2255edc | 1697 | raidz_map_t *rm = zio->io_vsd; |
45d1cae3 | 1698 | int c, i; |
34dc7c2f | 1699 | |
b2255edc | 1700 | vdev_raidz_generate_parity_row(rm, rr); |
34dc7c2f | 1701 | |
b2255edc BB |
1702 | for (int c = 0; c < rr->rr_cols; c++) { |
1703 | raidz_col_t *rc = &rr->rr_col[c]; | |
1704 | if (rc->rc_size == 0) | |
1705 | continue; | |
619f0976 | 1706 | |
b2255edc BB |
1707 | /* Verify physical to logical translation */ |
1708 | vdev_raidz_io_verify(vd, rr, c); | |
34dc7c2f | 1709 | |
b2255edc BB |
1710 | zio_nowait(zio_vdev_child_io(zio, NULL, |
1711 | vd->vdev_child[rc->rc_devidx], rc->rc_offset, | |
1712 | rc->rc_abd, rc->rc_size, zio->io_type, zio->io_priority, | |
1713 | 0, vdev_raidz_child_done, rc)); | |
1714 | } | |
45d1cae3 | 1715 | |
b2255edc BB |
1716 | /* |
1717 | * Generate optional I/Os for skip sectors to improve aggregation | |
1718 | * contiguity. | |
1719 | */ | |
1720 | for (c = rm->rm_skipstart, i = 0; i < rm->rm_nskip; c++, i++) { | |
1721 | ASSERT(c <= rr->rr_scols); | |
1722 | if (c == rr->rr_scols) | |
1723 | c = 0; | |
1724 | ||
1725 | raidz_col_t *rc = &rr->rr_col[c]; | |
1726 | vdev_t *cvd = vd->vdev_child[rc->rc_devidx]; | |
1727 | ||
1728 | zio_nowait(zio_vdev_child_io(zio, NULL, cvd, | |
1729 | rc->rc_offset + rc->rc_size, NULL, 1ULL << ashift, | |
1730 | zio->io_type, zio->io_priority, | |
1731 | ZIO_FLAG_NODATA | ZIO_FLAG_OPTIONAL, NULL, NULL)); | |
34dc7c2f | 1732 | } |
b2255edc | 1733 | } |
34dc7c2f | 1734 | |
b2255edc BB |
1735 | static void |
1736 | vdev_raidz_io_start_read(zio_t *zio, raidz_row_t *rr) | |
1737 | { | |
1738 | vdev_t *vd = zio->io_vd; | |
34dc7c2f BB |
1739 | |
1740 | /* | |
1741 | * Iterate over the columns in reverse order so that we hit the parity | |
45d1cae3 | 1742 | * last -- any errors along the way will force us to read the parity. |
34dc7c2f | 1743 | */ |
b2255edc BB |
1744 | for (int c = rr->rr_cols - 1; c >= 0; c--) { |
1745 | raidz_col_t *rc = &rr->rr_col[c]; | |
1746 | if (rc->rc_size == 0) | |
1747 | continue; | |
1748 | vdev_t *cvd = vd->vdev_child[rc->rc_devidx]; | |
34dc7c2f | 1749 | if (!vdev_readable(cvd)) { |
b2255edc BB |
1750 | if (c >= rr->rr_firstdatacol) |
1751 | rr->rr_missingdata++; | |
34dc7c2f | 1752 | else |
b2255edc | 1753 | rr->rr_missingparity++; |
2e528b49 | 1754 | rc->rc_error = SET_ERROR(ENXIO); |
34dc7c2f BB |
1755 | rc->rc_tried = 1; /* don't even try */ |
1756 | rc->rc_skipped = 1; | |
1757 | continue; | |
1758 | } | |
428870ff | 1759 | if (vdev_dtl_contains(cvd, DTL_MISSING, zio->io_txg, 1)) { |
b2255edc BB |
1760 | if (c >= rr->rr_firstdatacol) |
1761 | rr->rr_missingdata++; | |
34dc7c2f | 1762 | else |
b2255edc | 1763 | rr->rr_missingparity++; |
2e528b49 | 1764 | rc->rc_error = SET_ERROR(ESTALE); |
34dc7c2f BB |
1765 | rc->rc_skipped = 1; |
1766 | continue; | |
1767 | } | |
b2255edc | 1768 | if (c >= rr->rr_firstdatacol || rr->rr_missingdata > 0 || |
9babb374 | 1769 | (zio->io_flags & (ZIO_FLAG_SCRUB | ZIO_FLAG_RESILVER))) { |
34dc7c2f | 1770 | zio_nowait(zio_vdev_child_io(zio, NULL, cvd, |
a6255b7f | 1771 | rc->rc_offset, rc->rc_abd, rc->rc_size, |
b128c09f | 1772 | zio->io_type, zio->io_priority, 0, |
34dc7c2f BB |
1773 | vdev_raidz_child_done, rc)); |
1774 | } | |
1775 | } | |
b2255edc BB |
1776 | } |
1777 | ||
1778 | /* | |
1779 | * Start an IO operation on a RAIDZ VDev | |
1780 | * | |
1781 | * Outline: | |
1782 | * - For write operations: | |
1783 | * 1. Generate the parity data | |
1784 | * 2. Create child zio write operations to each column's vdev, for both | |
1785 | * data and parity. | |
1786 | * 3. If the column skips any sectors for padding, create optional dummy | |
1787 | * write zio children for those areas to improve aggregation continuity. | |
1788 | * - For read operations: | |
1789 | * 1. Create child zio read operations to each data column's vdev to read | |
1790 | * the range of data required for zio. | |
1791 | * 2. If this is a scrub or resilver operation, or if any of the data | |
1792 | * vdevs have had errors, then create zio read operations to the parity | |
1793 | * columns' VDevs as well. | |
1794 | */ | |
1795 | static void | |
1796 | vdev_raidz_io_start(zio_t *zio) | |
1797 | { | |
1798 | vdev_t *vd = zio->io_vd; | |
1799 | vdev_t *tvd = vd->vdev_top; | |
1800 | vdev_raidz_t *vdrz = vd->vdev_tsd; | |
1801 | raidz_map_t *rm; | |
1802 | ||
1803 | rm = vdev_raidz_map_alloc(zio, tvd->vdev_ashift, | |
1804 | vdrz->vd_logical_width, vdrz->vd_nparity); | |
1805 | ||
1806 | /* | |
1807 | * Until raidz expansion is implemented all maps for a raidz vdev | |
1808 | * contain a single row. | |
1809 | */ | |
1810 | ASSERT3U(rm->rm_nrows, ==, 1); | |
1811 | raidz_row_t *rr = rm->rm_row[0]; | |
1812 | ||
1813 | zio->io_vsd = rm; | |
1814 | zio->io_vsd_ops = &vdev_raidz_vsd_ops; | |
1815 | ||
1816 | if (zio->io_type == ZIO_TYPE_WRITE) { | |
1817 | vdev_raidz_io_start_write(zio, rr, tvd->vdev_ashift); | |
1818 | } else { | |
1819 | ASSERT(zio->io_type == ZIO_TYPE_READ); | |
1820 | vdev_raidz_io_start_read(zio, rr); | |
1821 | } | |
34dc7c2f | 1822 | |
98b25418 | 1823 | zio_execute(zio); |
34dc7c2f BB |
1824 | } |
1825 | ||
1826 | /* | |
1827 | * Report a checksum error for a child of a RAID-Z device. | |
1828 | */ | |
1829 | static void | |
84c07ada | 1830 | raidz_checksum_error(zio_t *zio, raidz_col_t *rc, abd_t *bad_data) |
34dc7c2f BB |
1831 | { |
1832 | vdev_t *vd = zio->io_vd->vdev_child[rc->rc_devidx]; | |
34dc7c2f | 1833 | |
b2255edc BB |
1834 | if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE) && |
1835 | zio->io_priority != ZIO_PRIORITY_REBUILD) { | |
428870ff BB |
1836 | zio_bad_cksum_t zbc; |
1837 | raidz_map_t *rm = zio->io_vsd; | |
1838 | ||
428870ff BB |
1839 | zbc.zbc_has_cksum = 0; |
1840 | zbc.zbc_injected = rm->rm_ecksuminjected; | |
1841 | ||
03e02e5b | 1842 | (void) zfs_ereport_post_checksum(zio->io_spa, vd, |
b5256303 TC |
1843 | &zio->io_bookmark, zio, rc->rc_offset, rc->rc_size, |
1844 | rc->rc_abd, bad_data, &zbc); | |
03e02e5b DB |
1845 | mutex_enter(&vd->vdev_stat_lock); |
1846 | vd->vdev_stat.vs_checksum_errors++; | |
1847 | mutex_exit(&vd->vdev_stat_lock); | |
34dc7c2f | 1848 | } |
428870ff BB |
1849 | } |
1850 | ||
1851 | /* | |
1852 | * We keep track of whether or not there were any injected errors, so that | |
1853 | * any ereports we generate can note it. | |
1854 | */ | |
1855 | static int | |
1856 | raidz_checksum_verify(zio_t *zio) | |
1857 | { | |
1858 | zio_bad_cksum_t zbc; | |
1859 | raidz_map_t *rm = zio->io_vsd; | |
1860 | ||
d4ed6673 BB |
1861 | bzero(&zbc, sizeof (zio_bad_cksum_t)); |
1862 | ||
1c27024e | 1863 | int ret = zio_checksum_error(zio, &zbc); |
428870ff BB |
1864 | if (ret != 0 && zbc.zbc_injected != 0) |
1865 | rm->rm_ecksuminjected = 1; | |
34dc7c2f | 1866 | |
428870ff | 1867 | return (ret); |
34dc7c2f BB |
1868 | } |
1869 | ||
1870 | /* | |
1871 | * Generate the parity from the data columns. If we tried and were able to | |
1872 | * read the parity without error, verify that the generated parity matches the | |
1873 | * data we read. If it doesn't, we fire off a checksum error. Return the | |
b2255edc | 1874 | * number of such failures. |
34dc7c2f BB |
1875 | */ |
1876 | static int | |
b2255edc | 1877 | raidz_parity_verify(zio_t *zio, raidz_row_t *rr) |
34dc7c2f | 1878 | { |
84c07ada | 1879 | abd_t *orig[VDEV_RAIDZ_MAXPARITY]; |
34dc7c2f | 1880 | int c, ret = 0; |
b2255edc | 1881 | raidz_map_t *rm = zio->io_vsd; |
34dc7c2f BB |
1882 | raidz_col_t *rc; |
1883 | ||
3c67d83a TH |
1884 | blkptr_t *bp = zio->io_bp; |
1885 | enum zio_checksum checksum = (bp == NULL ? zio->io_prop.zp_checksum : | |
1886 | (BP_IS_GANG(bp) ? ZIO_CHECKSUM_GANG_HEADER : BP_GET_CHECKSUM(bp))); | |
1887 | ||
1888 | if (checksum == ZIO_CHECKSUM_NOPARITY) | |
1889 | return (ret); | |
1890 | ||
b2255edc BB |
1891 | for (c = 0; c < rr->rr_firstdatacol; c++) { |
1892 | rc = &rr->rr_col[c]; | |
34dc7c2f BB |
1893 | if (!rc->rc_tried || rc->rc_error != 0) |
1894 | continue; | |
84c07ada GN |
1895 | |
1896 | orig[c] = abd_alloc_sametype(rc->rc_abd, rc->rc_size); | |
1897 | abd_copy(orig[c], rc->rc_abd, rc->rc_size); | |
34dc7c2f BB |
1898 | } |
1899 | ||
b2255edc BB |
1900 | /* |
1901 | * Regenerates parity even for !tried||rc_error!=0 columns. This | |
1902 | * isn't harmful but it does have the side effect of fixing stuff | |
1903 | * we didn't realize was necessary (i.e. even if we return 0). | |
1904 | */ | |
1905 | vdev_raidz_generate_parity_row(rm, rr); | |
1906 | ||
1907 | for (c = 0; c < rr->rr_firstdatacol; c++) { | |
1908 | rc = &rr->rr_col[c]; | |
34dc7c2f | 1909 | |
34dc7c2f BB |
1910 | if (!rc->rc_tried || rc->rc_error != 0) |
1911 | continue; | |
b2255edc | 1912 | |
84c07ada | 1913 | if (abd_cmp(orig[c], rc->rc_abd) != 0) { |
428870ff | 1914 | raidz_checksum_error(zio, rc, orig[c]); |
2e528b49 | 1915 | rc->rc_error = SET_ERROR(ECKSUM); |
34dc7c2f BB |
1916 | ret++; |
1917 | } | |
84c07ada | 1918 | abd_free(orig[c]); |
34dc7c2f BB |
1919 | } |
1920 | ||
1921 | return (ret); | |
1922 | } | |
1923 | ||
34dc7c2f | 1924 | static int |
b2255edc | 1925 | vdev_raidz_worst_error(raidz_row_t *rr) |
b128c09f | 1926 | { |
1c27024e | 1927 | int error = 0; |
b128c09f | 1928 | |
b2255edc BB |
1929 | for (int c = 0; c < rr->rr_cols; c++) |
1930 | error = zio_worst_error(error, rr->rr_col[c].rc_error); | |
b128c09f BB |
1931 | |
1932 | return (error); | |
1933 | } | |
1934 | ||
b2255edc BB |
1935 | static void |
1936 | vdev_raidz_io_done_verified(zio_t *zio, raidz_row_t *rr) | |
45d1cae3 | 1937 | { |
b2255edc BB |
1938 | int unexpected_errors = 0; |
1939 | int parity_errors = 0; | |
1940 | int parity_untried = 0; | |
1941 | int data_errors = 0; | |
45d1cae3 | 1942 | |
b2255edc BB |
1943 | ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ); |
1944 | ||
1945 | for (int c = 0; c < rr->rr_cols; c++) { | |
1946 | raidz_col_t *rc = &rr->rr_col[c]; | |
1947 | ||
1948 | if (rc->rc_error) { | |
1949 | if (c < rr->rr_firstdatacol) | |
1950 | parity_errors++; | |
1951 | else | |
1952 | data_errors++; | |
1953 | ||
1954 | if (!rc->rc_skipped) | |
1955 | unexpected_errors++; | |
1956 | } else if (c < rr->rr_firstdatacol && !rc->rc_tried) { | |
1957 | parity_untried++; | |
1958 | } | |
1959 | } | |
45d1cae3 BB |
1960 | |
1961 | /* | |
b2255edc BB |
1962 | * If we read more parity disks than were used for |
1963 | * reconstruction, confirm that the other parity disks produced | |
1964 | * correct data. | |
1965 | * | |
1966 | * Note that we also regenerate parity when resilvering so we | |
1967 | * can write it out to failed devices later. | |
45d1cae3 | 1968 | */ |
b2255edc BB |
1969 | if (parity_errors + parity_untried < |
1970 | rr->rr_firstdatacol - data_errors || | |
1971 | (zio->io_flags & ZIO_FLAG_RESILVER)) { | |
1972 | int n = raidz_parity_verify(zio, rr); | |
1973 | unexpected_errors += n; | |
1974 | ASSERT3U(parity_errors + n, <=, rr->rr_firstdatacol); | |
1975 | } | |
45d1cae3 | 1976 | |
b2255edc BB |
1977 | if (zio->io_error == 0 && spa_writeable(zio->io_spa) && |
1978 | (unexpected_errors > 0 || (zio->io_flags & ZIO_FLAG_RESILVER))) { | |
45d1cae3 | 1979 | /* |
b2255edc | 1980 | * Use the good data we have in hand to repair damaged children. |
45d1cae3 | 1981 | */ |
b2255edc BB |
1982 | for (int c = 0; c < rr->rr_cols; c++) { |
1983 | raidz_col_t *rc = &rr->rr_col[c]; | |
1984 | vdev_t *vd = zio->io_vd; | |
1985 | vdev_t *cvd = vd->vdev_child[rc->rc_devidx]; | |
1986 | ||
1987 | if ((rc->rc_error == 0 || rc->rc_size == 0) && | |
1988 | (rc->rc_repair == 0)) { | |
1989 | continue; | |
45d1cae3 BB |
1990 | } |
1991 | ||
b2255edc BB |
1992 | zio_nowait(zio_vdev_child_io(zio, NULL, cvd, |
1993 | rc->rc_offset, rc->rc_abd, rc->rc_size, | |
1994 | ZIO_TYPE_WRITE, | |
1995 | zio->io_priority == ZIO_PRIORITY_REBUILD ? | |
1996 | ZIO_PRIORITY_REBUILD : ZIO_PRIORITY_ASYNC_WRITE, | |
1997 | ZIO_FLAG_IO_REPAIR | (unexpected_errors ? | |
1998 | ZIO_FLAG_SELF_HEAL : 0), NULL, NULL)); | |
1999 | } | |
2000 | } | |
2001 | } | |
2002 | ||
2003 | static void | |
2004 | raidz_restore_orig_data(raidz_map_t *rm) | |
2005 | { | |
2006 | for (int i = 0; i < rm->rm_nrows; i++) { | |
2007 | raidz_row_t *rr = rm->rm_row[i]; | |
2008 | for (int c = 0; c < rr->rr_cols; c++) { | |
2009 | raidz_col_t *rc = &rr->rr_col[c]; | |
2010 | if (rc->rc_need_orig_restore) { | |
2011 | abd_copy_from_buf(rc->rc_abd, | |
2012 | rc->rc_orig_data, rc->rc_size); | |
2013 | rc->rc_need_orig_restore = B_FALSE; | |
45d1cae3 | 2014 | } |
b2255edc BB |
2015 | } |
2016 | } | |
2017 | } | |
2018 | ||
2019 | /* | |
2020 | * returns EINVAL if reconstruction of the block will not be possible | |
2021 | * returns ECKSUM if this specific reconstruction failed | |
2022 | * returns 0 on successful reconstruction | |
2023 | */ | |
2024 | static int | |
2025 | raidz_reconstruct(zio_t *zio, int *ltgts, int ntgts, int nparity) | |
2026 | { | |
2027 | raidz_map_t *rm = zio->io_vsd; | |
45d1cae3 | 2028 | |
b2255edc BB |
2029 | /* Reconstruct each row */ |
2030 | for (int r = 0; r < rm->rm_nrows; r++) { | |
2031 | raidz_row_t *rr = rm->rm_row[r]; | |
2032 | int my_tgts[VDEV_RAIDZ_MAXPARITY]; /* value is child id */ | |
2033 | int t = 0; | |
2034 | int dead = 0; | |
2035 | int dead_data = 0; | |
2036 | ||
2037 | for (int c = 0; c < rr->rr_cols; c++) { | |
2038 | raidz_col_t *rc = &rr->rr_col[c]; | |
2039 | ASSERT0(rc->rc_need_orig_restore); | |
2040 | if (rc->rc_error != 0) { | |
2041 | dead++; | |
2042 | if (c >= nparity) | |
2043 | dead_data++; | |
2044 | continue; | |
2045 | } | |
2046 | if (rc->rc_size == 0) | |
2047 | continue; | |
2048 | for (int lt = 0; lt < ntgts; lt++) { | |
2049 | if (rc->rc_devidx == ltgts[lt]) { | |
2050 | if (rc->rc_orig_data == NULL) { | |
2051 | rc->rc_orig_data = | |
2052 | zio_buf_alloc(rc->rc_size); | |
2053 | abd_copy_to_buf( | |
2054 | rc->rc_orig_data, | |
2055 | rc->rc_abd, rc->rc_size); | |
2056 | } | |
2057 | rc->rc_need_orig_restore = B_TRUE; | |
2058 | ||
2059 | dead++; | |
2060 | if (c >= nparity) | |
2061 | dead_data++; | |
2062 | my_tgts[t++] = c; | |
2063 | break; | |
2064 | } | |
2065 | } | |
2066 | } | |
2067 | if (dead > nparity) { | |
2068 | /* reconstruction not possible */ | |
2069 | raidz_restore_orig_data(rm); | |
2070 | return (EINVAL); | |
45d1cae3 | 2071 | } |
b2255edc | 2072 | if (dead_data > 0) |
46df6e98 | 2073 | vdev_raidz_reconstruct_row(rm, rr, my_tgts, t); |
b2255edc | 2074 | } |
45d1cae3 | 2075 | |
b2255edc BB |
2076 | /* Check for success */ |
2077 | if (raidz_checksum_verify(zio) == 0) { | |
2078 | ||
2079 | /* Reconstruction succeeded - report errors */ | |
2080 | for (int i = 0; i < rm->rm_nrows; i++) { | |
2081 | raidz_row_t *rr = rm->rm_row[i]; | |
2082 | ||
2083 | for (int c = 0; c < rr->rr_cols; c++) { | |
2084 | raidz_col_t *rc = &rr->rr_col[c]; | |
2085 | if (rc->rc_need_orig_restore) { | |
2086 | /* | |
2087 | * Note: if this is a parity column, | |
2088 | * we don't really know if it's wrong. | |
2089 | * We need to let | |
2090 | * vdev_raidz_io_done_verified() check | |
2091 | * it, and if we set rc_error, it will | |
2092 | * think that it is a "known" error | |
2093 | * that doesn't need to be checked | |
2094 | * or corrected. | |
2095 | */ | |
2096 | if (rc->rc_error == 0 && | |
2097 | c >= rr->rr_firstdatacol) { | |
2098 | raidz_checksum_error(zio, | |
2099 | rc, rc->rc_gdata); | |
2100 | rc->rc_error = | |
2101 | SET_ERROR(ECKSUM); | |
2102 | } | |
2103 | rc->rc_need_orig_restore = B_FALSE; | |
2104 | } | |
2105 | } | |
45d1cae3 | 2106 | |
b2255edc | 2107 | vdev_raidz_io_done_verified(zio, rr); |
45d1cae3 BB |
2108 | } |
2109 | ||
b2255edc | 2110 | zio_checksum_verified(zio); |
45d1cae3 | 2111 | |
b2255edc BB |
2112 | return (0); |
2113 | } | |
45d1cae3 | 2114 | |
b2255edc BB |
2115 | /* Reconstruction failed - restore original data */ |
2116 | raidz_restore_orig_data(rm); | |
2117 | return (ECKSUM); | |
2118 | } | |
45d1cae3 | 2119 | |
b2255edc BB |
2120 | /* |
2121 | * Iterate over all combinations of N bad vdevs and attempt a reconstruction. | |
2122 | * Note that the algorithm below is non-optimal because it doesn't take into | |
2123 | * account how reconstruction is actually performed. For example, with | |
2124 | * triple-parity RAID-Z the reconstruction procedure is the same if column 4 | |
2125 | * is targeted as invalid as if columns 1 and 4 are targeted since in both | |
2126 | * cases we'd only use parity information in column 0. | |
2127 | * | |
2128 | * The order that we find the various possible combinations of failed | |
2129 | * disks is dictated by these rules: | |
2130 | * - Examine each "slot" (the "i" in tgts[i]) | |
2131 | * - Try to increment this slot (tgts[i] = tgts[i] + 1) | |
2132 | * - if we can't increment because it runs into the next slot, | |
2133 | * reset our slot to the minimum, and examine the next slot | |
2134 | * | |
2135 | * For example, with a 6-wide RAIDZ3, and no known errors (so we have to choose | |
2136 | * 3 columns to reconstruct), we will generate the following sequence: | |
2137 | * | |
2138 | * STATE ACTION | |
2139 | * 0 1 2 special case: skip since these are all parity | |
2140 | * 0 1 3 first slot: reset to 0; middle slot: increment to 2 | |
2141 | * 0 2 3 first slot: increment to 1 | |
2142 | * 1 2 3 first: reset to 0; middle: reset to 1; last: increment to 4 | |
2143 | * 0 1 4 first: reset to 0; middle: increment to 2 | |
2144 | * 0 2 4 first: increment to 1 | |
2145 | * 1 2 4 first: reset to 0; middle: increment to 3 | |
2146 | * 0 3 4 first: increment to 1 | |
2147 | * 1 3 4 first: increment to 2 | |
2148 | * 2 3 4 first: reset to 0; middle: reset to 1; last: increment to 5 | |
2149 | * 0 1 5 first: reset to 0; middle: increment to 2 | |
2150 | * 0 2 5 first: increment to 1 | |
2151 | * 1 2 5 first: reset to 0; middle: increment to 3 | |
2152 | * 0 3 5 first: increment to 1 | |
2153 | * 1 3 5 first: increment to 2 | |
2154 | * 2 3 5 first: reset to 0; middle: increment to 4 | |
2155 | * 0 4 5 first: increment to 1 | |
2156 | * 1 4 5 first: increment to 2 | |
2157 | * 2 4 5 first: increment to 3 | |
2158 | * 3 4 5 done | |
2159 | * | |
2160 | * This strategy works for dRAID but is less effecient when there are a large | |
2161 | * number of child vdevs and therefore permutations to check. Furthermore, | |
2162 | * since the raidz_map_t rows likely do not overlap reconstruction would be | |
2163 | * possible as long as there are no more than nparity data errors per row. | |
2164 | * These additional permutations are not currently checked but could be as | |
2165 | * a future improvement. | |
2166 | */ | |
2167 | static int | |
2168 | vdev_raidz_combrec(zio_t *zio) | |
2169 | { | |
2170 | int nparity = vdev_get_nparity(zio->io_vd); | |
2171 | raidz_map_t *rm = zio->io_vsd; | |
45d1cae3 | 2172 | |
b2255edc BB |
2173 | /* Check if there's enough data to attempt reconstrution. */ |
2174 | for (int i = 0; i < rm->rm_nrows; i++) { | |
2175 | raidz_row_t *rr = rm->rm_row[i]; | |
2176 | int total_errors = 0; | |
45d1cae3 | 2177 | |
b2255edc BB |
2178 | for (int c = 0; c < rr->rr_cols; c++) { |
2179 | if (rr->rr_col[c].rc_error) | |
2180 | total_errors++; | |
2181 | } | |
45d1cae3 | 2182 | |
b2255edc BB |
2183 | if (total_errors > nparity) |
2184 | return (vdev_raidz_worst_error(rr)); | |
2185 | } | |
45d1cae3 | 2186 | |
b2255edc BB |
2187 | for (int num_failures = 1; num_failures <= nparity; num_failures++) { |
2188 | int tstore[VDEV_RAIDZ_MAXPARITY + 2]; | |
2189 | int *ltgts = &tstore[1]; /* value is logical child ID */ | |
2190 | ||
2191 | /* Determine number of logical children, n */ | |
2192 | int n = zio->io_vd->vdev_children; | |
2193 | ||
2194 | ASSERT3U(num_failures, <=, nparity); | |
2195 | ASSERT3U(num_failures, <=, VDEV_RAIDZ_MAXPARITY); | |
2196 | ||
2197 | /* Handle corner cases in combrec logic */ | |
2198 | ltgts[-1] = -1; | |
2199 | for (int i = 0; i < num_failures; i++) { | |
2200 | ltgts[i] = i; | |
2201 | } | |
2202 | ltgts[num_failures] = n; | |
2203 | ||
2204 | for (;;) { | |
2205 | int err = raidz_reconstruct(zio, ltgts, num_failures, | |
2206 | nparity); | |
2207 | if (err == EINVAL) { | |
45d1cae3 | 2208 | /* |
b2255edc BB |
2209 | * Reconstruction not possible with this # |
2210 | * failures; try more failures. | |
45d1cae3 | 2211 | */ |
b2255edc BB |
2212 | break; |
2213 | } else if (err == 0) | |
2214 | return (0); | |
2215 | ||
2216 | /* Compute next targets to try */ | |
2217 | for (int t = 0; ; t++) { | |
2218 | ASSERT3U(t, <, num_failures); | |
2219 | ltgts[t]++; | |
2220 | if (ltgts[t] == n) { | |
2221 | /* try more failures */ | |
2222 | ASSERT3U(t, ==, num_failures - 1); | |
2223 | break; | |
2224 | } | |
45d1cae3 | 2225 | |
b2255edc BB |
2226 | ASSERT3U(ltgts[t], <, n); |
2227 | ASSERT3U(ltgts[t], <=, ltgts[t + 1]); | |
45d1cae3 BB |
2228 | |
2229 | /* | |
2230 | * If that spot is available, we're done here. | |
b2255edc | 2231 | * Try the next combination. |
45d1cae3 | 2232 | */ |
b2255edc | 2233 | if (ltgts[t] != ltgts[t + 1]) |
45d1cae3 BB |
2234 | break; |
2235 | ||
2236 | /* | |
b2255edc BB |
2237 | * Otherwise, reset this tgt to the minimum, |
2238 | * and move on to the next tgt. | |
45d1cae3 | 2239 | */ |
b2255edc BB |
2240 | ltgts[t] = ltgts[t - 1] + 1; |
2241 | ASSERT3U(ltgts[t], ==, t); | |
2242 | } | |
45d1cae3 | 2243 | |
b2255edc BB |
2244 | /* Increase the number of failures and keep trying. */ |
2245 | if (ltgts[num_failures - 1] == n) | |
2246 | break; | |
45d1cae3 BB |
2247 | } |
2248 | } | |
45d1cae3 | 2249 | |
b2255edc BB |
2250 | return (ECKSUM); |
2251 | } | |
2252 | ||
2253 | void | |
2254 | vdev_raidz_reconstruct(raidz_map_t *rm, const int *t, int nt) | |
2255 | { | |
2256 | for (uint64_t row = 0; row < rm->rm_nrows; row++) { | |
2257 | raidz_row_t *rr = rm->rm_row[row]; | |
2258 | vdev_raidz_reconstruct_row(rm, rr, t, nt); | |
2259 | } | |
45d1cae3 BB |
2260 | } |
2261 | ||
e49f1e20 | 2262 | /* |
b2255edc | 2263 | * Complete a write IO operation on a RAIDZ VDev |
e49f1e20 WA |
2264 | * |
2265 | * Outline: | |
e49f1e20 WA |
2266 | * 1. Check for errors on the child IOs. |
2267 | * 2. Return, setting an error code if too few child VDevs were written | |
2268 | * to reconstruct the data later. Note that partial writes are | |
2269 | * considered successful if they can be reconstructed at all. | |
e49f1e20 | 2270 | */ |
b128c09f | 2271 | static void |
b2255edc BB |
2272 | vdev_raidz_io_done_write_impl(zio_t *zio, raidz_row_t *rr) |
2273 | { | |
2274 | int total_errors = 0; | |
2275 | ||
2276 | ASSERT3U(rr->rr_missingparity, <=, rr->rr_firstdatacol); | |
2277 | ASSERT3U(rr->rr_missingdata, <=, rr->rr_cols - rr->rr_firstdatacol); | |
2278 | ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); | |
2279 | ||
2280 | for (int c = 0; c < rr->rr_cols; c++) { | |
2281 | raidz_col_t *rc = &rr->rr_col[c]; | |
2282 | ||
2283 | if (rc->rc_error) { | |
2284 | ASSERT(rc->rc_error != ECKSUM); /* child has no bp */ | |
2285 | ||
2286 | total_errors++; | |
2287 | } | |
2288 | } | |
2289 | ||
2290 | /* | |
2291 | * Treat partial writes as a success. If we couldn't write enough | |
2292 | * columns to reconstruct the data, the I/O failed. Otherwise, | |
2293 | * good enough. | |
2294 | * | |
2295 | * Now that we support write reallocation, it would be better | |
2296 | * to treat partial failure as real failure unless there are | |
2297 | * no non-degraded top-level vdevs left, and not update DTLs | |
2298 | * if we intend to reallocate. | |
2299 | */ | |
2300 | if (total_errors > rr->rr_firstdatacol) { | |
2301 | zio->io_error = zio_worst_error(zio->io_error, | |
2302 | vdev_raidz_worst_error(rr)); | |
2303 | } | |
2304 | } | |
2305 | ||
46df6e98 | 2306 | static void |
b2255edc BB |
2307 | vdev_raidz_io_done_reconstruct_known_missing(zio_t *zio, raidz_map_t *rm, |
2308 | raidz_row_t *rr) | |
34dc7c2f | 2309 | { |
34dc7c2f BB |
2310 | int parity_errors = 0; |
2311 | int parity_untried = 0; | |
2312 | int data_errors = 0; | |
b128c09f | 2313 | int total_errors = 0; |
34dc7c2f | 2314 | |
b2255edc BB |
2315 | ASSERT3U(rr->rr_missingparity, <=, rr->rr_firstdatacol); |
2316 | ASSERT3U(rr->rr_missingdata, <=, rr->rr_cols - rr->rr_firstdatacol); | |
2317 | ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ); | |
34dc7c2f | 2318 | |
b2255edc BB |
2319 | for (int c = 0; c < rr->rr_cols; c++) { |
2320 | raidz_col_t *rc = &rr->rr_col[c]; | |
34dc7c2f | 2321 | |
34dc7c2f | 2322 | if (rc->rc_error) { |
b128c09f | 2323 | ASSERT(rc->rc_error != ECKSUM); /* child has no bp */ |
34dc7c2f | 2324 | |
b2255edc | 2325 | if (c < rr->rr_firstdatacol) |
34dc7c2f BB |
2326 | parity_errors++; |
2327 | else | |
2328 | data_errors++; | |
2329 | ||
b128c09f | 2330 | total_errors++; |
b2255edc | 2331 | } else if (c < rr->rr_firstdatacol && !rc->rc_tried) { |
34dc7c2f BB |
2332 | parity_untried++; |
2333 | } | |
2334 | } | |
2335 | ||
34dc7c2f | 2336 | /* |
b2255edc BB |
2337 | * If there were data errors and the number of errors we saw was |
2338 | * correctable -- less than or equal to the number of parity disks read | |
2339 | * -- reconstruct based on the missing data. | |
34dc7c2f | 2340 | */ |
b2255edc BB |
2341 | if (data_errors != 0 && |
2342 | total_errors <= rr->rr_firstdatacol - parity_untried) { | |
2343 | /* | |
2344 | * We either attempt to read all the parity columns or | |
2345 | * none of them. If we didn't try to read parity, we | |
2346 | * wouldn't be here in the correctable case. There must | |
2347 | * also have been fewer parity errors than parity | |
2348 | * columns or, again, we wouldn't be in this code path. | |
2349 | */ | |
2350 | ASSERT(parity_untried == 0); | |
2351 | ASSERT(parity_errors < rr->rr_firstdatacol); | |
34dc7c2f | 2352 | |
b2255edc BB |
2353 | /* |
2354 | * Identify the data columns that reported an error. | |
2355 | */ | |
2356 | int n = 0; | |
2357 | int tgts[VDEV_RAIDZ_MAXPARITY]; | |
2358 | for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { | |
2359 | raidz_col_t *rc = &rr->rr_col[c]; | |
2360 | if (rc->rc_error != 0) { | |
2361 | ASSERT(n < VDEV_RAIDZ_MAXPARITY); | |
2362 | tgts[n++] = c; | |
34dc7c2f | 2363 | } |
b2255edc | 2364 | } |
34dc7c2f | 2365 | |
b2255edc | 2366 | ASSERT(rr->rr_firstdatacol >= n); |
34dc7c2f | 2367 | |
46df6e98 | 2368 | vdev_raidz_reconstruct_row(rm, rr, tgts, n); |
b2255edc | 2369 | } |
b2255edc | 2370 | } |
34dc7c2f | 2371 | |
b2255edc BB |
2372 | /* |
2373 | * Return the number of reads issued. | |
2374 | */ | |
2375 | static int | |
2376 | vdev_raidz_read_all(zio_t *zio, raidz_row_t *rr) | |
2377 | { | |
2378 | vdev_t *vd = zio->io_vd; | |
2379 | int nread = 0; | |
34dc7c2f | 2380 | |
b2255edc BB |
2381 | rr->rr_missingdata = 0; |
2382 | rr->rr_missingparity = 0; | |
34dc7c2f BB |
2383 | |
2384 | /* | |
b2255edc BB |
2385 | * If this rows contains empty sectors which are not required |
2386 | * for a normal read then allocate an ABD for them now so they | |
2387 | * may be read, verified, and any needed repairs performed. | |
34dc7c2f | 2388 | */ |
b2255edc BB |
2389 | if (rr->rr_nempty && rr->rr_abd_empty == NULL) |
2390 | vdev_draid_map_alloc_empty(zio, rr); | |
34dc7c2f | 2391 | |
b2255edc BB |
2392 | for (int c = 0; c < rr->rr_cols; c++) { |
2393 | raidz_col_t *rc = &rr->rr_col[c]; | |
2394 | if (rc->rc_tried || rc->rc_size == 0) | |
34dc7c2f BB |
2395 | continue; |
2396 | ||
b2255edc BB |
2397 | zio_nowait(zio_vdev_child_io(zio, NULL, |
2398 | vd->vdev_child[rc->rc_devidx], | |
2399 | rc->rc_offset, rc->rc_abd, rc->rc_size, | |
2400 | zio->io_type, zio->io_priority, 0, | |
2401 | vdev_raidz_child_done, rc)); | |
2402 | nread++; | |
34dc7c2f | 2403 | } |
b2255edc BB |
2404 | return (nread); |
2405 | } | |
34dc7c2f | 2406 | |
b2255edc BB |
2407 | /* |
2408 | * We're here because either there were too many errors to even attempt | |
2409 | * reconstruction (total_errors == rm_first_datacol), or vdev_*_combrec() | |
2410 | * failed. In either case, there is enough bad data to prevent reconstruction. | |
2411 | * Start checksum ereports for all children which haven't failed. | |
2412 | */ | |
2413 | static void | |
2414 | vdev_raidz_io_done_unrecoverable(zio_t *zio) | |
2415 | { | |
2416 | raidz_map_t *rm = zio->io_vsd; | |
34dc7c2f | 2417 | |
b2255edc BB |
2418 | for (int i = 0; i < rm->rm_nrows; i++) { |
2419 | raidz_row_t *rr = rm->rm_row[i]; | |
428870ff | 2420 | |
b2255edc BB |
2421 | for (int c = 0; c < rr->rr_cols; c++) { |
2422 | raidz_col_t *rc = &rr->rr_col[c]; | |
2423 | vdev_t *cvd = zio->io_vd->vdev_child[rc->rc_devidx]; | |
2424 | ||
2425 | if (rc->rc_error != 0) | |
2426 | continue; | |
2427 | ||
2428 | zio_bad_cksum_t zbc; | |
2429 | zbc.zbc_has_cksum = 0; | |
2430 | zbc.zbc_injected = rm->rm_ecksuminjected; | |
2431 | ||
03e02e5b | 2432 | (void) zfs_ereport_start_checksum(zio->io_spa, |
b2255edc BB |
2433 | cvd, &zio->io_bookmark, zio, rc->rc_offset, |
2434 | rc->rc_size, (void *)(uintptr_t)c, &zbc); | |
03e02e5b DB |
2435 | mutex_enter(&cvd->vdev_stat_lock); |
2436 | cvd->vdev_stat.vs_checksum_errors++; | |
2437 | mutex_exit(&cvd->vdev_stat_lock); | |
34dc7c2f BB |
2438 | } |
2439 | } | |
b2255edc | 2440 | } |
34dc7c2f | 2441 | |
b2255edc BB |
2442 | void |
2443 | vdev_raidz_io_done(zio_t *zio) | |
2444 | { | |
2445 | raidz_map_t *rm = zio->io_vsd; | |
34dc7c2f | 2446 | |
b2255edc BB |
2447 | if (zio->io_type == ZIO_TYPE_WRITE) { |
2448 | for (int i = 0; i < rm->rm_nrows; i++) { | |
2449 | vdev_raidz_io_done_write_impl(zio, rm->rm_row[i]); | |
2450 | } | |
2451 | } else { | |
2452 | for (int i = 0; i < rm->rm_nrows; i++) { | |
2453 | raidz_row_t *rr = rm->rm_row[i]; | |
46df6e98 | 2454 | vdev_raidz_io_done_reconstruct_known_missing(zio, |
b2255edc BB |
2455 | rm, rr); |
2456 | } | |
34dc7c2f | 2457 | |
b2255edc BB |
2458 | if (raidz_checksum_verify(zio) == 0) { |
2459 | for (int i = 0; i < rm->rm_nrows; i++) { | |
2460 | raidz_row_t *rr = rm->rm_row[i]; | |
2461 | vdev_raidz_io_done_verified(zio, rr); | |
2462 | } | |
2463 | zio_checksum_verified(zio); | |
2464 | } else { | |
2465 | /* | |
2466 | * A sequential resilver has no checksum which makes | |
2467 | * combinatoral reconstruction impossible. This code | |
2468 | * path is unreachable since raidz_checksum_verify() | |
2469 | * has no checksum to verify and must succeed. | |
2470 | */ | |
2471 | ASSERT3U(zio->io_priority, !=, ZIO_PRIORITY_REBUILD); | |
34dc7c2f | 2472 | |
b2255edc BB |
2473 | /* |
2474 | * This isn't a typical situation -- either we got a | |
2475 | * read error or a child silently returned bad data. | |
2476 | * Read every block so we can try again with as much | |
2477 | * data and parity as we can track down. If we've | |
2478 | * already been through once before, all children will | |
2479 | * be marked as tried so we'll proceed to combinatorial | |
2480 | * reconstruction. | |
2481 | */ | |
2482 | int nread = 0; | |
2483 | for (int i = 0; i < rm->rm_nrows; i++) { | |
2484 | nread += vdev_raidz_read_all(zio, | |
2485 | rm->rm_row[i]); | |
2486 | } | |
2487 | if (nread != 0) { | |
2488 | /* | |
2489 | * Normally our stage is VDEV_IO_DONE, but if | |
2490 | * we've already called redone(), it will have | |
2491 | * changed to VDEV_IO_START, in which case we | |
2492 | * don't want to call redone() again. | |
2493 | */ | |
2494 | if (zio->io_stage != ZIO_STAGE_VDEV_IO_START) | |
2495 | zio_vdev_io_redone(zio); | |
2496 | return; | |
2497 | } | |
2498 | ||
2499 | zio->io_error = vdev_raidz_combrec(zio); | |
2500 | if (zio->io_error == ECKSUM && | |
2501 | !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { | |
2502 | vdev_raidz_io_done_unrecoverable(zio); | |
2503 | } | |
34dc7c2f | 2504 | } |
34dc7c2f | 2505 | } |
34dc7c2f BB |
2506 | } |
2507 | ||
2508 | static void | |
2509 | vdev_raidz_state_change(vdev_t *vd, int faulted, int degraded) | |
2510 | { | |
b2255edc BB |
2511 | vdev_raidz_t *vdrz = vd->vdev_tsd; |
2512 | if (faulted > vdrz->vd_nparity) | |
34dc7c2f BB |
2513 | vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN, |
2514 | VDEV_AUX_NO_REPLICAS); | |
2515 | else if (degraded + faulted != 0) | |
2516 | vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE); | |
2517 | else | |
2518 | vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE); | |
2519 | } | |
2520 | ||
3d6da72d IH |
2521 | /* |
2522 | * Determine if any portion of the provided block resides on a child vdev | |
2523 | * with a dirty DTL and therefore needs to be resilvered. The function | |
e1cfd73f | 2524 | * assumes that at least one DTL is dirty which implies that full stripe |
3d6da72d IH |
2525 | * width blocks must be resilvered. |
2526 | */ | |
2527 | static boolean_t | |
b2255edc BB |
2528 | vdev_raidz_need_resilver(vdev_t *vd, const dva_t *dva, size_t psize, |
2529 | uint64_t phys_birth) | |
3d6da72d | 2530 | { |
b2255edc | 2531 | vdev_raidz_t *vdrz = vd->vdev_tsd; |
3d6da72d | 2532 | uint64_t dcols = vd->vdev_children; |
b2255edc | 2533 | uint64_t nparity = vdrz->vd_nparity; |
3d6da72d IH |
2534 | uint64_t ashift = vd->vdev_top->vdev_ashift; |
2535 | /* The starting RAIDZ (parent) vdev sector of the block. */ | |
b2255edc | 2536 | uint64_t b = DVA_GET_OFFSET(dva) >> ashift; |
3d6da72d IH |
2537 | /* The zio's size in units of the vdev's minimum sector size. */ |
2538 | uint64_t s = ((psize - 1) >> ashift) + 1; | |
2539 | /* The first column for this stripe. */ | |
2540 | uint64_t f = b % dcols; | |
2541 | ||
b2255edc BB |
2542 | /* Unreachable by sequential resilver. */ |
2543 | ASSERT3U(phys_birth, !=, TXG_UNKNOWN); | |
2544 | ||
2545 | if (!vdev_dtl_contains(vd, DTL_PARTIAL, phys_birth, 1)) | |
2546 | return (B_FALSE); | |
2547 | ||
3d6da72d IH |
2548 | if (s + nparity >= dcols) |
2549 | return (B_TRUE); | |
2550 | ||
2551 | for (uint64_t c = 0; c < s + nparity; c++) { | |
2552 | uint64_t devidx = (f + c) % dcols; | |
2553 | vdev_t *cvd = vd->vdev_child[devidx]; | |
2554 | ||
2555 | /* | |
2556 | * dsl_scan_need_resilver() already checked vd with | |
2557 | * vdev_dtl_contains(). So here just check cvd with | |
2558 | * vdev_dtl_empty(), cheaper and a good approximation. | |
2559 | */ | |
2560 | if (!vdev_dtl_empty(cvd, DTL_PARTIAL)) | |
2561 | return (B_TRUE); | |
2562 | } | |
2563 | ||
2564 | return (B_FALSE); | |
2565 | } | |
2566 | ||
619f0976 | 2567 | static void |
b2255edc BB |
2568 | vdev_raidz_xlate(vdev_t *cvd, const range_seg64_t *logical_rs, |
2569 | range_seg64_t *physical_rs, range_seg64_t *remain_rs) | |
619f0976 GW |
2570 | { |
2571 | vdev_t *raidvd = cvd->vdev_parent; | |
2572 | ASSERT(raidvd->vdev_ops == &vdev_raidz_ops); | |
2573 | ||
2574 | uint64_t width = raidvd->vdev_children; | |
2575 | uint64_t tgt_col = cvd->vdev_id; | |
2576 | uint64_t ashift = raidvd->vdev_top->vdev_ashift; | |
2577 | ||
2578 | /* make sure the offsets are block-aligned */ | |
b2255edc BB |
2579 | ASSERT0(logical_rs->rs_start % (1 << ashift)); |
2580 | ASSERT0(logical_rs->rs_end % (1 << ashift)); | |
2581 | uint64_t b_start = logical_rs->rs_start >> ashift; | |
2582 | uint64_t b_end = logical_rs->rs_end >> ashift; | |
619f0976 GW |
2583 | |
2584 | uint64_t start_row = 0; | |
2585 | if (b_start > tgt_col) /* avoid underflow */ | |
2586 | start_row = ((b_start - tgt_col - 1) / width) + 1; | |
2587 | ||
2588 | uint64_t end_row = 0; | |
2589 | if (b_end > tgt_col) | |
2590 | end_row = ((b_end - tgt_col - 1) / width) + 1; | |
2591 | ||
b2255edc BB |
2592 | physical_rs->rs_start = start_row << ashift; |
2593 | physical_rs->rs_end = end_row << ashift; | |
619f0976 | 2594 | |
b2255edc BB |
2595 | ASSERT3U(physical_rs->rs_start, <=, logical_rs->rs_start); |
2596 | ASSERT3U(physical_rs->rs_end - physical_rs->rs_start, <=, | |
2597 | logical_rs->rs_end - logical_rs->rs_start); | |
2598 | } | |
2599 | ||
2600 | /* | |
2601 | * Initialize private RAIDZ specific fields from the nvlist. | |
2602 | */ | |
2603 | static int | |
2604 | vdev_raidz_init(spa_t *spa, nvlist_t *nv, void **tsd) | |
2605 | { | |
2606 | vdev_raidz_t *vdrz; | |
2607 | uint64_t nparity; | |
2608 | ||
2609 | uint_t children; | |
2610 | nvlist_t **child; | |
2611 | int error = nvlist_lookup_nvlist_array(nv, | |
2612 | ZPOOL_CONFIG_CHILDREN, &child, &children); | |
2613 | if (error != 0) | |
2614 | return (SET_ERROR(EINVAL)); | |
2615 | ||
2616 | if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY, &nparity) == 0) { | |
2617 | if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY) | |
2618 | return (SET_ERROR(EINVAL)); | |
2619 | ||
2620 | /* | |
2621 | * Previous versions could only support 1 or 2 parity | |
2622 | * device. | |
2623 | */ | |
2624 | if (nparity > 1 && spa_version(spa) < SPA_VERSION_RAIDZ2) | |
2625 | return (SET_ERROR(EINVAL)); | |
2626 | else if (nparity > 2 && spa_version(spa) < SPA_VERSION_RAIDZ3) | |
2627 | return (SET_ERROR(EINVAL)); | |
2628 | } else { | |
2629 | /* | |
2630 | * We require the parity to be specified for SPAs that | |
2631 | * support multiple parity levels. | |
2632 | */ | |
2633 | if (spa_version(spa) >= SPA_VERSION_RAIDZ2) | |
2634 | return (SET_ERROR(EINVAL)); | |
2635 | ||
2636 | /* | |
2637 | * Otherwise, we default to 1 parity device for RAID-Z. | |
2638 | */ | |
2639 | nparity = 1; | |
2640 | } | |
2641 | ||
2642 | vdrz = kmem_zalloc(sizeof (*vdrz), KM_SLEEP); | |
2643 | vdrz->vd_logical_width = children; | |
2644 | vdrz->vd_nparity = nparity; | |
2645 | ||
2646 | *tsd = vdrz; | |
2647 | ||
2648 | return (0); | |
2649 | } | |
2650 | ||
2651 | static void | |
2652 | vdev_raidz_fini(vdev_t *vd) | |
2653 | { | |
2654 | kmem_free(vd->vdev_tsd, sizeof (vdev_raidz_t)); | |
2655 | } | |
2656 | ||
2657 | /* | |
2658 | * Add RAIDZ specific fields to the config nvlist. | |
2659 | */ | |
2660 | static void | |
2661 | vdev_raidz_config_generate(vdev_t *vd, nvlist_t *nv) | |
2662 | { | |
2663 | ASSERT3P(vd->vdev_ops, ==, &vdev_raidz_ops); | |
2664 | vdev_raidz_t *vdrz = vd->vdev_tsd; | |
2665 | ||
2666 | /* | |
2667 | * Make sure someone hasn't managed to sneak a fancy new vdev | |
2668 | * into a crufty old storage pool. | |
2669 | */ | |
2670 | ASSERT(vdrz->vd_nparity == 1 || | |
2671 | (vdrz->vd_nparity <= 2 && | |
2672 | spa_version(vd->vdev_spa) >= SPA_VERSION_RAIDZ2) || | |
2673 | (vdrz->vd_nparity <= 3 && | |
2674 | spa_version(vd->vdev_spa) >= SPA_VERSION_RAIDZ3)); | |
2675 | ||
2676 | /* | |
2677 | * Note that we'll add these even on storage pools where they | |
2678 | * aren't strictly required -- older software will just ignore | |
2679 | * it. | |
2680 | */ | |
2681 | fnvlist_add_uint64(nv, ZPOOL_CONFIG_NPARITY, vdrz->vd_nparity); | |
2682 | } | |
2683 | ||
2684 | static uint64_t | |
2685 | vdev_raidz_nparity(vdev_t *vd) | |
2686 | { | |
2687 | vdev_raidz_t *vdrz = vd->vdev_tsd; | |
2688 | return (vdrz->vd_nparity); | |
2689 | } | |
2690 | ||
2691 | static uint64_t | |
2692 | vdev_raidz_ndisks(vdev_t *vd) | |
2693 | { | |
2694 | return (vd->vdev_children); | |
619f0976 GW |
2695 | } |
2696 | ||
34dc7c2f | 2697 | vdev_ops_t vdev_raidz_ops = { |
b2255edc BB |
2698 | .vdev_op_init = vdev_raidz_init, |
2699 | .vdev_op_fini = vdev_raidz_fini, | |
a64f8276 I |
2700 | .vdev_op_open = vdev_raidz_open, |
2701 | .vdev_op_close = vdev_raidz_close, | |
2702 | .vdev_op_asize = vdev_raidz_asize, | |
b2255edc BB |
2703 | .vdev_op_min_asize = vdev_raidz_min_asize, |
2704 | .vdev_op_min_alloc = NULL, | |
a64f8276 I |
2705 | .vdev_op_io_start = vdev_raidz_io_start, |
2706 | .vdev_op_io_done = vdev_raidz_io_done, | |
2707 | .vdev_op_state_change = vdev_raidz_state_change, | |
2708 | .vdev_op_need_resilver = vdev_raidz_need_resilver, | |
2709 | .vdev_op_hold = NULL, | |
2710 | .vdev_op_rele = NULL, | |
2711 | .vdev_op_remap = NULL, | |
2712 | .vdev_op_xlate = vdev_raidz_xlate, | |
b2255edc BB |
2713 | .vdev_op_rebuild_asize = NULL, |
2714 | .vdev_op_metaslab_init = NULL, | |
2715 | .vdev_op_config_generate = vdev_raidz_config_generate, | |
2716 | .vdev_op_nparity = vdev_raidz_nparity, | |
2717 | .vdev_op_ndisks = vdev_raidz_ndisks, | |
a64f8276 I |
2718 | .vdev_op_type = VDEV_TYPE_RAIDZ, /* name of this vdev type */ |
2719 | .vdev_op_leaf = B_FALSE /* not a leaf vdev */ | |
34dc7c2f | 2720 | }; |