]>
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. |
98b25418 | 24 | * Copyright (c) 2012, 2014 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> | |
34dc7c2f BB |
38 | |
39 | /* | |
40 | * Virtual device vector for RAID-Z. | |
41 | * | |
45d1cae3 BB |
42 | * This vdev supports single, double, and triple parity. For single parity, |
43 | * we use a simple XOR of all the data columns. For double or triple parity, | |
44 | * we use a special case of Reed-Solomon coding. This extends the | |
45 | * technique described in "The mathematics of RAID-6" by H. Peter Anvin by | |
46 | * drawing on the system described in "A Tutorial on Reed-Solomon Coding for | |
47 | * Fault-Tolerance in RAID-like Systems" by James S. Plank on which the | |
48 | * former is also based. The latter is designed to provide higher performance | |
49 | * for writes. | |
50 | * | |
51 | * Note that the Plank paper claimed to support arbitrary N+M, but was then | |
52 | * amended six years later identifying a critical flaw that invalidates its | |
53 | * claims. Nevertheless, the technique can be adapted to work for up to | |
54 | * triple parity. For additional parity, the amendment "Note: Correction to | |
55 | * the 1997 Tutorial on Reed-Solomon Coding" by James S. Plank and Ying Ding | |
56 | * is viable, but the additional complexity means that write performance will | |
57 | * suffer. | |
58 | * | |
59 | * All of the methods above operate on a Galois field, defined over the | |
60 | * integers mod 2^N. In our case we choose N=8 for GF(8) so that all elements | |
61 | * can be expressed with a single byte. Briefly, the operations on the | |
62 | * field are defined as follows: | |
34dc7c2f BB |
63 | * |
64 | * o addition (+) is represented by a bitwise XOR | |
65 | * o subtraction (-) is therefore identical to addition: A + B = A - B | |
66 | * o multiplication of A by 2 is defined by the following bitwise expression: | |
d3cc8b15 | 67 | * |
34dc7c2f BB |
68 | * (A * 2)_7 = A_6 |
69 | * (A * 2)_6 = A_5 | |
70 | * (A * 2)_5 = A_4 | |
71 | * (A * 2)_4 = A_3 + A_7 | |
72 | * (A * 2)_3 = A_2 + A_7 | |
73 | * (A * 2)_2 = A_1 + A_7 | |
74 | * (A * 2)_1 = A_0 | |
75 | * (A * 2)_0 = A_7 | |
76 | * | |
77 | * In C, multiplying by 2 is therefore ((a << 1) ^ ((a & 0x80) ? 0x1d : 0)). | |
45d1cae3 BB |
78 | * As an aside, this multiplication is derived from the error correcting |
79 | * primitive polynomial x^8 + x^4 + x^3 + x^2 + 1. | |
34dc7c2f BB |
80 | * |
81 | * Observe that any number in the field (except for 0) can be expressed as a | |
82 | * power of 2 -- a generator for the field. We store a table of the powers of | |
83 | * 2 and logs base 2 for quick look ups, and exploit the fact that A * B can | |
84 | * be rewritten as 2^(log_2(A) + log_2(B)) (where '+' is normal addition rather | |
45d1cae3 BB |
85 | * than field addition). The inverse of a field element A (A^-1) is therefore |
86 | * A ^ (255 - 1) = A^254. | |
34dc7c2f | 87 | * |
45d1cae3 BB |
88 | * The up-to-three parity columns, P, Q, R over several data columns, |
89 | * D_0, ... D_n-1, can be expressed by field operations: | |
34dc7c2f BB |
90 | * |
91 | * P = D_0 + D_1 + ... + D_n-2 + D_n-1 | |
92 | * Q = 2^n-1 * D_0 + 2^n-2 * D_1 + ... + 2^1 * D_n-2 + 2^0 * D_n-1 | |
93 | * = ((...((D_0) * 2 + D_1) * 2 + ...) * 2 + D_n-2) * 2 + D_n-1 | |
45d1cae3 BB |
94 | * R = 4^n-1 * D_0 + 4^n-2 * D_1 + ... + 4^1 * D_n-2 + 4^0 * D_n-1 |
95 | * = ((...((D_0) * 4 + D_1) * 4 + ...) * 4 + D_n-2) * 4 + D_n-1 | |
34dc7c2f | 96 | * |
45d1cae3 BB |
97 | * We chose 1, 2, and 4 as our generators because 1 corresponds to the trival |
98 | * XOR operation, and 2 and 4 can be computed quickly and generate linearly- | |
99 | * independent coefficients. (There are no additional coefficients that have | |
100 | * this property which is why the uncorrected Plank method breaks down.) | |
101 | * | |
102 | * See the reconstruction code below for how P, Q and R can used individually | |
103 | * or in concert to recover missing data columns. | |
34dc7c2f BB |
104 | */ |
105 | ||
34dc7c2f BB |
106 | #define VDEV_RAIDZ_P 0 |
107 | #define VDEV_RAIDZ_Q 1 | |
45d1cae3 | 108 | #define VDEV_RAIDZ_R 2 |
45d1cae3 BB |
109 | |
110 | #define VDEV_RAIDZ_MUL_2(x) (((x) << 1) ^ (((x) & 0x80) ? 0x1d : 0)) | |
111 | #define VDEV_RAIDZ_MUL_4(x) (VDEV_RAIDZ_MUL_2(VDEV_RAIDZ_MUL_2(x))) | |
112 | ||
113 | /* | |
114 | * We provide a mechanism to perform the field multiplication operation on a | |
115 | * 64-bit value all at once rather than a byte at a time. This works by | |
116 | * creating a mask from the top bit in each byte and using that to | |
117 | * conditionally apply the XOR of 0x1d. | |
118 | */ | |
119 | #define VDEV_RAIDZ_64MUL_2(x, mask) \ | |
120 | { \ | |
121 | (mask) = (x) & 0x8080808080808080ULL; \ | |
122 | (mask) = ((mask) << 1) - ((mask) >> 7); \ | |
123 | (x) = (((x) << 1) & 0xfefefefefefefefeULL) ^ \ | |
c5b3a7bb | 124 | ((mask) & 0x1d1d1d1d1d1d1d1dULL); \ |
45d1cae3 | 125 | } |
34dc7c2f | 126 | |
45d1cae3 BB |
127 | #define VDEV_RAIDZ_64MUL_4(x, mask) \ |
128 | { \ | |
129 | VDEV_RAIDZ_64MUL_2((x), mask); \ | |
130 | VDEV_RAIDZ_64MUL_2((x), mask); \ | |
131 | } | |
34dc7c2f | 132 | |
ab9f4b0b | 133 | void |
428870ff | 134 | vdev_raidz_map_free(raidz_map_t *rm) |
b128c09f | 135 | { |
b128c09f BB |
136 | int c; |
137 | ||
428870ff | 138 | for (c = 0; c < rm->rm_firstdatacol; c++) { |
a6255b7f | 139 | abd_free(rm->rm_col[c].rc_abd); |
b128c09f | 140 | |
428870ff | 141 | if (rm->rm_col[c].rc_gdata != NULL) |
84c07ada | 142 | abd_free(rm->rm_col[c].rc_gdata); |
428870ff BB |
143 | } |
144 | ||
84c07ada | 145 | for (c = rm->rm_firstdatacol; c < rm->rm_cols; c++) |
a6255b7f | 146 | abd_put(rm->rm_col[c].rc_abd); |
428870ff | 147 | |
a6255b7f DQ |
148 | if (rm->rm_abd_copy != NULL) |
149 | abd_free(rm->rm_abd_copy); | |
428870ff | 150 | |
45d1cae3 | 151 | kmem_free(rm, offsetof(raidz_map_t, rm_col[rm->rm_scols])); |
b128c09f BB |
152 | } |
153 | ||
428870ff BB |
154 | static void |
155 | vdev_raidz_map_free_vsd(zio_t *zio) | |
156 | { | |
157 | raidz_map_t *rm = zio->io_vsd; | |
158 | ||
c99c9001 | 159 | ASSERT0(rm->rm_freed); |
428870ff BB |
160 | rm->rm_freed = 1; |
161 | ||
162 | if (rm->rm_reports == 0) | |
163 | vdev_raidz_map_free(rm); | |
164 | } | |
165 | ||
166 | /*ARGSUSED*/ | |
167 | static void | |
168 | vdev_raidz_cksum_free(void *arg, size_t ignored) | |
169 | { | |
170 | raidz_map_t *rm = arg; | |
171 | ||
172 | ASSERT3U(rm->rm_reports, >, 0); | |
173 | ||
174 | if (--rm->rm_reports == 0 && rm->rm_freed != 0) | |
175 | vdev_raidz_map_free(rm); | |
176 | } | |
177 | ||
178 | static void | |
84c07ada | 179 | vdev_raidz_cksum_finish(zio_cksum_report_t *zcr, const abd_t *good_data) |
428870ff BB |
180 | { |
181 | raidz_map_t *rm = zcr->zcr_cbdata; | |
84c07ada GN |
182 | const size_t c = zcr->zcr_cbinfo; |
183 | size_t x, offset; | |
428870ff | 184 | |
84c07ada GN |
185 | const abd_t *good = NULL; |
186 | const abd_t *bad = rm->rm_col[c].rc_abd; | |
428870ff BB |
187 | |
188 | if (good_data == NULL) { | |
189 | zfs_ereport_finish_checksum(zcr, NULL, NULL, B_FALSE); | |
190 | return; | |
191 | } | |
192 | ||
193 | if (c < rm->rm_firstdatacol) { | |
194 | /* | |
195 | * The first time through, calculate the parity blocks for | |
196 | * the good data (this relies on the fact that the good | |
197 | * data never changes for a given logical ZIO) | |
198 | */ | |
199 | if (rm->rm_col[0].rc_gdata == NULL) { | |
a6255b7f | 200 | abd_t *bad_parity[VDEV_RAIDZ_MAXPARITY]; |
428870ff BB |
201 | |
202 | /* | |
203 | * Set up the rm_col[]s to generate the parity for | |
204 | * good_data, first saving the parity bufs and | |
205 | * replacing them with buffers to hold the result. | |
206 | */ | |
207 | for (x = 0; x < rm->rm_firstdatacol; x++) { | |
a6255b7f | 208 | bad_parity[x] = rm->rm_col[x].rc_abd; |
a6255b7f | 209 | rm->rm_col[x].rc_abd = |
84c07ada GN |
210 | rm->rm_col[x].rc_gdata = |
211 | abd_alloc_sametype(rm->rm_col[x].rc_abd, | |
a6255b7f | 212 | rm->rm_col[x].rc_size); |
428870ff BB |
213 | } |
214 | ||
215 | /* fill in the data columns from good_data */ | |
84c07ada | 216 | offset = 0; |
428870ff | 217 | for (; x < rm->rm_cols; x++) { |
a6255b7f | 218 | abd_put(rm->rm_col[x].rc_abd); |
84c07ada GN |
219 | |
220 | rm->rm_col[x].rc_abd = | |
221 | abd_get_offset_size((abd_t *)good_data, | |
222 | offset, rm->rm_col[x].rc_size); | |
223 | offset += rm->rm_col[x].rc_size; | |
428870ff BB |
224 | } |
225 | ||
226 | /* | |
227 | * Construct the parity from the good data. | |
228 | */ | |
229 | vdev_raidz_generate_parity(rm); | |
230 | ||
231 | /* restore everything back to its original state */ | |
84c07ada | 232 | for (x = 0; x < rm->rm_firstdatacol; x++) |
a6255b7f | 233 | rm->rm_col[x].rc_abd = bad_parity[x]; |
428870ff | 234 | |
a6255b7f | 235 | offset = 0; |
428870ff | 236 | for (x = rm->rm_firstdatacol; x < rm->rm_cols; x++) { |
a6255b7f | 237 | abd_put(rm->rm_col[x].rc_abd); |
a206522c GN |
238 | rm->rm_col[x].rc_abd = abd_get_offset_size( |
239 | rm->rm_abd_copy, offset, | |
240 | rm->rm_col[x].rc_size); | |
a6255b7f | 241 | offset += rm->rm_col[x].rc_size; |
428870ff BB |
242 | } |
243 | } | |
244 | ||
245 | ASSERT3P(rm->rm_col[c].rc_gdata, !=, NULL); | |
84c07ada GN |
246 | good = abd_get_offset_size(rm->rm_col[c].rc_gdata, 0, |
247 | rm->rm_col[c].rc_size); | |
428870ff BB |
248 | } else { |
249 | /* adjust good_data to point at the start of our column */ | |
84c07ada | 250 | offset = 0; |
428870ff | 251 | for (x = rm->rm_firstdatacol; x < c; x++) |
84c07ada GN |
252 | offset += rm->rm_col[x].rc_size; |
253 | ||
254 | good = abd_get_offset_size((abd_t *)good_data, offset, | |
255 | rm->rm_col[c].rc_size); | |
428870ff BB |
256 | } |
257 | ||
258 | /* we drop the ereport if it ends up that the data was good */ | |
259 | zfs_ereport_finish_checksum(zcr, good, bad, B_TRUE); | |
84c07ada | 260 | abd_put((abd_t *)good); |
428870ff BB |
261 | } |
262 | ||
263 | /* | |
264 | * Invoked indirectly by zfs_ereport_start_checksum(), called | |
265 | * below when our read operation fails completely. The main point | |
266 | * is to keep a copy of everything we read from disk, so that at | |
267 | * vdev_raidz_cksum_finish() time we can compare it with the good data. | |
268 | */ | |
269 | static void | |
270 | vdev_raidz_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *arg) | |
271 | { | |
272 | size_t c = (size_t)(uintptr_t)arg; | |
a6255b7f | 273 | size_t offset; |
428870ff BB |
274 | |
275 | raidz_map_t *rm = zio->io_vsd; | |
276 | size_t size; | |
277 | ||
278 | /* set up the report and bump the refcount */ | |
279 | zcr->zcr_cbdata = rm; | |
280 | zcr->zcr_cbinfo = c; | |
281 | zcr->zcr_finish = vdev_raidz_cksum_finish; | |
282 | zcr->zcr_free = vdev_raidz_cksum_free; | |
283 | ||
284 | rm->rm_reports++; | |
285 | ASSERT3U(rm->rm_reports, >, 0); | |
286 | ||
a6255b7f | 287 | if (rm->rm_abd_copy != NULL) |
428870ff BB |
288 | return; |
289 | ||
290 | /* | |
291 | * It's the first time we're called for this raidz_map_t, so we need | |
292 | * to copy the data aside; there's no guarantee that our zio's buffer | |
293 | * won't be re-used for something else. | |
294 | * | |
295 | * Our parity data is already in separate buffers, so there's no need | |
296 | * to copy them. | |
297 | */ | |
298 | ||
299 | size = 0; | |
300 | for (c = rm->rm_firstdatacol; c < rm->rm_cols; c++) | |
301 | size += rm->rm_col[c].rc_size; | |
302 | ||
84c07ada | 303 | rm->rm_abd_copy = abd_alloc_for_io(size, B_FALSE); |
428870ff | 304 | |
a6255b7f | 305 | for (offset = 0, c = rm->rm_firstdatacol; c < rm->rm_cols; c++) { |
428870ff | 306 | raidz_col_t *col = &rm->rm_col[c]; |
a206522c GN |
307 | abd_t *tmp = abd_get_offset_size(rm->rm_abd_copy, offset, |
308 | col->rc_size); | |
428870ff | 309 | |
a6255b7f | 310 | abd_copy(tmp, col->rc_abd, col->rc_size); |
84c07ada | 311 | |
a6255b7f DQ |
312 | abd_put(col->rc_abd); |
313 | col->rc_abd = tmp; | |
428870ff | 314 | |
a6255b7f | 315 | offset += col->rc_size; |
428870ff | 316 | } |
a6255b7f | 317 | ASSERT3U(offset, ==, size); |
428870ff BB |
318 | } |
319 | ||
320 | static const zio_vsd_ops_t vdev_raidz_vsd_ops = { | |
321 | vdev_raidz_map_free_vsd, | |
322 | vdev_raidz_cksum_report | |
323 | }; | |
324 | ||
e49f1e20 WA |
325 | /* |
326 | * Divides the IO evenly across all child vdevs; usually, dcols is | |
327 | * the number of children in the target vdev. | |
a1687880 BB |
328 | * |
329 | * Avoid inlining the function to keep vdev_raidz_io_start(), which | |
330 | * is this functions only caller, as small as possible on the stack. | |
e49f1e20 | 331 | */ |
ab9f4b0b | 332 | noinline raidz_map_t * |
3d6da72d | 333 | vdev_raidz_map_alloc(zio_t *zio, uint64_t ashift, uint64_t dcols, |
34dc7c2f BB |
334 | uint64_t nparity) |
335 | { | |
336 | raidz_map_t *rm; | |
e49f1e20 | 337 | /* The starting RAIDZ (parent) vdev sector of the block. */ |
3d6da72d | 338 | uint64_t b = zio->io_offset >> ashift; |
e49f1e20 | 339 | /* The zio's size in units of the vdev's minimum sector size. */ |
3d6da72d | 340 | uint64_t s = zio->io_size >> ashift; |
e49f1e20 | 341 | /* The first column for this stripe. */ |
34dc7c2f | 342 | uint64_t f = b % dcols; |
e49f1e20 | 343 | /* The starting byte offset on each child vdev. */ |
3d6da72d | 344 | uint64_t o = (b / dcols) << ashift; |
45d1cae3 | 345 | uint64_t q, r, c, bc, col, acols, scols, coff, devidx, asize, tot; |
a6255b7f | 346 | uint64_t off = 0; |
34dc7c2f | 347 | |
e49f1e20 WA |
348 | /* |
349 | * "Quotient": The number of data sectors for this stripe on all but | |
350 | * the "big column" child vdevs that also contain "remainder" data. | |
351 | */ | |
34dc7c2f | 352 | q = s / (dcols - nparity); |
e49f1e20 WA |
353 | |
354 | /* | |
355 | * "Remainder": The number of partial stripe data sectors in this I/O. | |
356 | * This will add a sector to some, but not all, child vdevs. | |
357 | */ | |
34dc7c2f | 358 | r = s - q * (dcols - nparity); |
e49f1e20 WA |
359 | |
360 | /* The number of "big columns" - those which contain remainder data. */ | |
34dc7c2f | 361 | bc = (r == 0 ? 0 : r + nparity); |
e49f1e20 WA |
362 | |
363 | /* | |
364 | * The total number of data and parity sectors associated with | |
365 | * this I/O. | |
366 | */ | |
45d1cae3 BB |
367 | tot = s + nparity * (q + (r == 0 ? 0 : 1)); |
368 | ||
e49f1e20 WA |
369 | /* acols: The columns that will be accessed. */ |
370 | /* scols: The columns that will be accessed or skipped. */ | |
45d1cae3 | 371 | if (q == 0) { |
e49f1e20 | 372 | /* Our I/O request doesn't span all child vdevs. */ |
45d1cae3 BB |
373 | acols = bc; |
374 | scols = MIN(dcols, roundup(bc, nparity + 1)); | |
375 | } else { | |
376 | acols = dcols; | |
377 | scols = dcols; | |
378 | } | |
34dc7c2f | 379 | |
45d1cae3 | 380 | ASSERT3U(acols, <=, scols); |
34dc7c2f | 381 | |
79c76d5b | 382 | rm = kmem_alloc(offsetof(raidz_map_t, rm_col[scols]), KM_SLEEP); |
34dc7c2f BB |
383 | |
384 | rm->rm_cols = acols; | |
45d1cae3 | 385 | rm->rm_scols = scols; |
34dc7c2f | 386 | rm->rm_bigcols = bc; |
428870ff | 387 | rm->rm_skipstart = bc; |
34dc7c2f BB |
388 | rm->rm_missingdata = 0; |
389 | rm->rm_missingparity = 0; | |
390 | rm->rm_firstdatacol = nparity; | |
a6255b7f | 391 | rm->rm_abd_copy = NULL; |
428870ff BB |
392 | rm->rm_reports = 0; |
393 | rm->rm_freed = 0; | |
394 | rm->rm_ecksuminjected = 0; | |
34dc7c2f | 395 | |
45d1cae3 BB |
396 | asize = 0; |
397 | ||
398 | for (c = 0; c < scols; c++) { | |
34dc7c2f BB |
399 | col = f + c; |
400 | coff = o; | |
401 | if (col >= dcols) { | |
402 | col -= dcols; | |
3d6da72d | 403 | coff += 1ULL << ashift; |
34dc7c2f BB |
404 | } |
405 | rm->rm_col[c].rc_devidx = col; | |
406 | rm->rm_col[c].rc_offset = coff; | |
a6255b7f | 407 | rm->rm_col[c].rc_abd = NULL; |
428870ff | 408 | rm->rm_col[c].rc_gdata = NULL; |
34dc7c2f BB |
409 | rm->rm_col[c].rc_error = 0; |
410 | rm->rm_col[c].rc_tried = 0; | |
411 | rm->rm_col[c].rc_skipped = 0; | |
45d1cae3 BB |
412 | |
413 | if (c >= acols) | |
414 | rm->rm_col[c].rc_size = 0; | |
415 | else if (c < bc) | |
3d6da72d | 416 | rm->rm_col[c].rc_size = (q + 1) << ashift; |
45d1cae3 | 417 | else |
3d6da72d | 418 | rm->rm_col[c].rc_size = q << ashift; |
45d1cae3 BB |
419 | |
420 | asize += rm->rm_col[c].rc_size; | |
34dc7c2f BB |
421 | } |
422 | ||
3d6da72d IH |
423 | ASSERT3U(asize, ==, tot << ashift); |
424 | rm->rm_asize = roundup(asize, (nparity + 1) << ashift); | |
428870ff | 425 | rm->rm_nskip = roundup(tot, nparity + 1) - tot; |
3d6da72d | 426 | ASSERT3U(rm->rm_asize - asize, ==, rm->rm_nskip << ashift); |
428870ff | 427 | ASSERT3U(rm->rm_nskip, <=, nparity); |
34dc7c2f BB |
428 | |
429 | for (c = 0; c < rm->rm_firstdatacol; c++) | |
a6255b7f | 430 | rm->rm_col[c].rc_abd = |
a206522c | 431 | abd_alloc_linear(rm->rm_col[c].rc_size, B_FALSE); |
34dc7c2f | 432 | |
a206522c GN |
433 | rm->rm_col[c].rc_abd = abd_get_offset_size(zio->io_abd, 0, |
434 | rm->rm_col[c].rc_size); | |
a6255b7f | 435 | off = rm->rm_col[c].rc_size; |
34dc7c2f | 436 | |
a6255b7f | 437 | for (c = c + 1; c < acols; c++) { |
a206522c GN |
438 | rm->rm_col[c].rc_abd = abd_get_offset_size(zio->io_abd, off, |
439 | rm->rm_col[c].rc_size); | |
a6255b7f DQ |
440 | off += rm->rm_col[c].rc_size; |
441 | } | |
34dc7c2f BB |
442 | |
443 | /* | |
444 | * If all data stored spans all columns, there's a danger that parity | |
445 | * will always be on the same device and, since parity isn't read | |
446 | * during normal operation, that that device's I/O bandwidth won't be | |
447 | * used effectively. We therefore switch the parity every 1MB. | |
448 | * | |
449 | * ... at least that was, ostensibly, the theory. As a practical | |
450 | * matter unless we juggle the parity between all devices evenly, we | |
451 | * won't see any benefit. Further, occasional writes that aren't a | |
452 | * multiple of the LCM of the number of children and the minimum | |
453 | * stripe width are sufficient to avoid pessimal behavior. | |
454 | * Unfortunately, this decision created an implicit on-disk format | |
455 | * requirement that we need to support for all eternity, but only | |
456 | * for single-parity RAID-Z. | |
428870ff BB |
457 | * |
458 | * If we intend to skip a sector in the zeroth column for padding | |
459 | * we must make sure to note this swap. We will never intend to | |
460 | * skip the first column since at least one data and one parity | |
461 | * column must appear in each row. | |
34dc7c2f BB |
462 | */ |
463 | ASSERT(rm->rm_cols >= 2); | |
464 | ASSERT(rm->rm_col[0].rc_size == rm->rm_col[1].rc_size); | |
465 | ||
466 | if (rm->rm_firstdatacol == 1 && (zio->io_offset & (1ULL << 20))) { | |
467 | devidx = rm->rm_col[0].rc_devidx; | |
468 | o = rm->rm_col[0].rc_offset; | |
469 | rm->rm_col[0].rc_devidx = rm->rm_col[1].rc_devidx; | |
470 | rm->rm_col[0].rc_offset = rm->rm_col[1].rc_offset; | |
471 | rm->rm_col[1].rc_devidx = devidx; | |
472 | rm->rm_col[1].rc_offset = o; | |
428870ff BB |
473 | |
474 | if (rm->rm_skipstart == 0) | |
475 | rm->rm_skipstart = 1; | |
34dc7c2f BB |
476 | } |
477 | ||
478 | zio->io_vsd = rm; | |
428870ff | 479 | zio->io_vsd_ops = &vdev_raidz_vsd_ops; |
ab9f4b0b | 480 | |
c9187d86 GN |
481 | /* init RAIDZ parity ops */ |
482 | rm->rm_ops = vdev_raidz_math_get_ops(); | |
ab9f4b0b | 483 | |
34dc7c2f BB |
484 | return (rm); |
485 | } | |
486 | ||
a6255b7f DQ |
487 | struct pqr_struct { |
488 | uint64_t *p; | |
489 | uint64_t *q; | |
490 | uint64_t *r; | |
491 | }; | |
492 | ||
493 | static int | |
494 | vdev_raidz_p_func(void *buf, size_t size, void *private) | |
495 | { | |
496 | struct pqr_struct *pqr = private; | |
497 | const uint64_t *src = buf; | |
498 | int i, cnt = size / sizeof (src[0]); | |
499 | ||
500 | ASSERT(pqr->p && !pqr->q && !pqr->r); | |
501 | ||
502 | for (i = 0; i < cnt; i++, src++, pqr->p++) | |
503 | *pqr->p ^= *src; | |
504 | ||
505 | return (0); | |
506 | } | |
507 | ||
508 | static int | |
509 | vdev_raidz_pq_func(void *buf, size_t size, void *private) | |
510 | { | |
511 | struct pqr_struct *pqr = private; | |
512 | const uint64_t *src = buf; | |
513 | uint64_t mask; | |
514 | int i, cnt = size / sizeof (src[0]); | |
515 | ||
516 | ASSERT(pqr->p && pqr->q && !pqr->r); | |
517 | ||
518 | for (i = 0; i < cnt; i++, src++, pqr->p++, pqr->q++) { | |
519 | *pqr->p ^= *src; | |
520 | VDEV_RAIDZ_64MUL_2(*pqr->q, mask); | |
521 | *pqr->q ^= *src; | |
522 | } | |
523 | ||
524 | return (0); | |
525 | } | |
526 | ||
527 | static int | |
528 | vdev_raidz_pqr_func(void *buf, size_t size, void *private) | |
529 | { | |
530 | struct pqr_struct *pqr = private; | |
531 | const uint64_t *src = buf; | |
532 | uint64_t mask; | |
533 | int i, cnt = size / sizeof (src[0]); | |
534 | ||
535 | ASSERT(pqr->p && pqr->q && pqr->r); | |
536 | ||
537 | for (i = 0; i < cnt; i++, src++, pqr->p++, pqr->q++, pqr->r++) { | |
538 | *pqr->p ^= *src; | |
539 | VDEV_RAIDZ_64MUL_2(*pqr->q, mask); | |
540 | *pqr->q ^= *src; | |
541 | VDEV_RAIDZ_64MUL_4(*pqr->r, mask); | |
542 | *pqr->r ^= *src; | |
543 | } | |
544 | ||
545 | return (0); | |
546 | } | |
547 | ||
34dc7c2f BB |
548 | static void |
549 | vdev_raidz_generate_parity_p(raidz_map_t *rm) | |
550 | { | |
a6255b7f | 551 | uint64_t *p; |
34dc7c2f | 552 | int c; |
a6255b7f | 553 | abd_t *src; |
34dc7c2f BB |
554 | |
555 | for (c = rm->rm_firstdatacol; c < rm->rm_cols; c++) { | |
a6255b7f DQ |
556 | src = rm->rm_col[c].rc_abd; |
557 | p = abd_to_buf(rm->rm_col[VDEV_RAIDZ_P].rc_abd); | |
34dc7c2f BB |
558 | |
559 | if (c == rm->rm_firstdatacol) { | |
a6255b7f | 560 | abd_copy_to_buf(p, src, rm->rm_col[c].rc_size); |
34dc7c2f | 561 | } else { |
a6255b7f DQ |
562 | struct pqr_struct pqr = { p, NULL, NULL }; |
563 | (void) abd_iterate_func(src, 0, rm->rm_col[c].rc_size, | |
564 | vdev_raidz_p_func, &pqr); | |
34dc7c2f BB |
565 | } |
566 | } | |
567 | } | |
568 | ||
569 | static void | |
570 | vdev_raidz_generate_parity_pq(raidz_map_t *rm) | |
571 | { | |
a6255b7f | 572 | uint64_t *p, *q, pcnt, ccnt, mask, i; |
34dc7c2f | 573 | int c; |
a6255b7f | 574 | abd_t *src; |
34dc7c2f | 575 | |
a6255b7f | 576 | pcnt = rm->rm_col[VDEV_RAIDZ_P].rc_size / sizeof (p[0]); |
34dc7c2f BB |
577 | ASSERT(rm->rm_col[VDEV_RAIDZ_P].rc_size == |
578 | rm->rm_col[VDEV_RAIDZ_Q].rc_size); | |
579 | ||
580 | for (c = rm->rm_firstdatacol; c < rm->rm_cols; c++) { | |
a6255b7f DQ |
581 | src = rm->rm_col[c].rc_abd; |
582 | p = abd_to_buf(rm->rm_col[VDEV_RAIDZ_P].rc_abd); | |
583 | q = abd_to_buf(rm->rm_col[VDEV_RAIDZ_Q].rc_abd); | |
45d1cae3 | 584 | |
a6255b7f | 585 | ccnt = rm->rm_col[c].rc_size / sizeof (p[0]); |
34dc7c2f BB |
586 | |
587 | if (c == rm->rm_firstdatacol) { | |
f7e76821 | 588 | ASSERT(ccnt == pcnt || ccnt == 0); |
a6255b7f DQ |
589 | abd_copy_to_buf(p, src, rm->rm_col[c].rc_size); |
590 | (void) memcpy(q, p, rm->rm_col[c].rc_size); | |
45d1cae3 | 591 | |
a6255b7f DQ |
592 | for (i = ccnt; i < pcnt; i++) { |
593 | p[i] = 0; | |
594 | q[i] = 0; | |
45d1cae3 | 595 | } |
a6255b7f | 596 | } else { |
f7e76821 IH |
597 | struct pqr_struct pqr = { p, q, NULL }; |
598 | ||
599 | ASSERT(ccnt <= pcnt); | |
600 | (void) abd_iterate_func(src, 0, rm->rm_col[c].rc_size, | |
601 | vdev_raidz_pq_func, &pqr); | |
45d1cae3 BB |
602 | |
603 | /* | |
604 | * Treat short columns as though they are full of 0s. | |
605 | * Note that there's therefore nothing needed for P. | |
606 | */ | |
a6255b7f DQ |
607 | for (i = ccnt; i < pcnt; i++) { |
608 | VDEV_RAIDZ_64MUL_2(q[i], mask); | |
45d1cae3 BB |
609 | } |
610 | } | |
611 | } | |
612 | } | |
613 | ||
614 | static void | |
615 | vdev_raidz_generate_parity_pqr(raidz_map_t *rm) | |
616 | { | |
a6255b7f | 617 | uint64_t *p, *q, *r, pcnt, ccnt, mask, i; |
45d1cae3 | 618 | int c; |
a6255b7f | 619 | abd_t *src; |
45d1cae3 | 620 | |
a6255b7f | 621 | pcnt = rm->rm_col[VDEV_RAIDZ_P].rc_size / sizeof (p[0]); |
45d1cae3 BB |
622 | ASSERT(rm->rm_col[VDEV_RAIDZ_P].rc_size == |
623 | rm->rm_col[VDEV_RAIDZ_Q].rc_size); | |
624 | ASSERT(rm->rm_col[VDEV_RAIDZ_P].rc_size == | |
625 | rm->rm_col[VDEV_RAIDZ_R].rc_size); | |
626 | ||
627 | for (c = rm->rm_firstdatacol; c < rm->rm_cols; c++) { | |
a6255b7f DQ |
628 | src = rm->rm_col[c].rc_abd; |
629 | p = abd_to_buf(rm->rm_col[VDEV_RAIDZ_P].rc_abd); | |
630 | q = abd_to_buf(rm->rm_col[VDEV_RAIDZ_Q].rc_abd); | |
631 | r = abd_to_buf(rm->rm_col[VDEV_RAIDZ_R].rc_abd); | |
45d1cae3 | 632 | |
a6255b7f | 633 | ccnt = rm->rm_col[c].rc_size / sizeof (p[0]); |
45d1cae3 BB |
634 | |
635 | if (c == rm->rm_firstdatacol) { | |
f7e76821 | 636 | ASSERT(ccnt == pcnt || ccnt == 0); |
a6255b7f DQ |
637 | abd_copy_to_buf(p, src, rm->rm_col[c].rc_size); |
638 | (void) memcpy(q, p, rm->rm_col[c].rc_size); | |
639 | (void) memcpy(r, p, rm->rm_col[c].rc_size); | |
45d1cae3 | 640 | |
a6255b7f DQ |
641 | for (i = ccnt; i < pcnt; i++) { |
642 | p[i] = 0; | |
643 | q[i] = 0; | |
644 | r[i] = 0; | |
34dc7c2f | 645 | } |
a6255b7f | 646 | } else { |
f7e76821 IH |
647 | struct pqr_struct pqr = { p, q, r }; |
648 | ||
649 | ASSERT(ccnt <= pcnt); | |
650 | (void) abd_iterate_func(src, 0, rm->rm_col[c].rc_size, | |
651 | vdev_raidz_pqr_func, &pqr); | |
652 | ||
34dc7c2f BB |
653 | /* |
654 | * Treat short columns as though they are full of 0s. | |
45d1cae3 | 655 | * Note that there's therefore nothing needed for P. |
34dc7c2f | 656 | */ |
a6255b7f DQ |
657 | for (i = ccnt; i < pcnt; i++) { |
658 | VDEV_RAIDZ_64MUL_2(q[i], mask); | |
659 | VDEV_RAIDZ_64MUL_4(r[i], mask); | |
34dc7c2f BB |
660 | } |
661 | } | |
662 | } | |
663 | } | |
664 | ||
45d1cae3 BB |
665 | /* |
666 | * Generate RAID parity in the first virtual columns according to the number of | |
667 | * parity columns available. | |
668 | */ | |
ab9f4b0b | 669 | void |
45d1cae3 BB |
670 | vdev_raidz_generate_parity(raidz_map_t *rm) |
671 | { | |
c9187d86 GN |
672 | /* Generate using the new math implementation */ |
673 | if (vdev_raidz_math_generate(rm) != RAIDZ_ORIGINAL_IMPL) | |
ab9f4b0b | 674 | return; |
ab9f4b0b | 675 | |
45d1cae3 BB |
676 | switch (rm->rm_firstdatacol) { |
677 | case 1: | |
678 | vdev_raidz_generate_parity_p(rm); | |
679 | break; | |
680 | case 2: | |
681 | vdev_raidz_generate_parity_pq(rm); | |
682 | break; | |
683 | case 3: | |
684 | vdev_raidz_generate_parity_pqr(rm); | |
685 | break; | |
686 | default: | |
687 | cmn_err(CE_PANIC, "invalid RAID-Z configuration"); | |
688 | } | |
689 | } | |
690 | ||
a6255b7f DQ |
691 | /* ARGSUSED */ |
692 | static int | |
693 | vdev_raidz_reconst_p_func(void *dbuf, void *sbuf, size_t size, void *private) | |
694 | { | |
695 | uint64_t *dst = dbuf; | |
696 | uint64_t *src = sbuf; | |
697 | int cnt = size / sizeof (src[0]); | |
a6255b7f | 698 | |
1c27024e | 699 | for (int i = 0; i < cnt; i++) { |
a6255b7f DQ |
700 | dst[i] ^= src[i]; |
701 | } | |
702 | ||
703 | return (0); | |
704 | } | |
705 | ||
706 | /* ARGSUSED */ | |
707 | static int | |
708 | vdev_raidz_reconst_q_pre_func(void *dbuf, void *sbuf, size_t size, | |
709 | void *private) | |
710 | { | |
711 | uint64_t *dst = dbuf; | |
712 | uint64_t *src = sbuf; | |
713 | uint64_t mask; | |
714 | int cnt = size / sizeof (dst[0]); | |
a6255b7f | 715 | |
1c27024e | 716 | for (int i = 0; i < cnt; i++, dst++, src++) { |
a6255b7f DQ |
717 | VDEV_RAIDZ_64MUL_2(*dst, mask); |
718 | *dst ^= *src; | |
719 | } | |
720 | ||
721 | return (0); | |
722 | } | |
723 | ||
724 | /* ARGSUSED */ | |
725 | static int | |
726 | vdev_raidz_reconst_q_pre_tail_func(void *buf, size_t size, void *private) | |
727 | { | |
728 | uint64_t *dst = buf; | |
729 | uint64_t mask; | |
730 | int cnt = size / sizeof (dst[0]); | |
a6255b7f | 731 | |
1c27024e | 732 | for (int i = 0; i < cnt; i++, dst++) { |
a6255b7f DQ |
733 | /* same operation as vdev_raidz_reconst_q_pre_func() on dst */ |
734 | VDEV_RAIDZ_64MUL_2(*dst, mask); | |
735 | } | |
736 | ||
737 | return (0); | |
738 | } | |
739 | ||
740 | struct reconst_q_struct { | |
741 | uint64_t *q; | |
742 | int exp; | |
743 | }; | |
744 | ||
745 | static int | |
746 | vdev_raidz_reconst_q_post_func(void *buf, size_t size, void *private) | |
747 | { | |
748 | struct reconst_q_struct *rq = private; | |
749 | uint64_t *dst = buf; | |
750 | int cnt = size / sizeof (dst[0]); | |
a6255b7f | 751 | |
1c27024e | 752 | for (int i = 0; i < cnt; i++, dst++, rq->q++) { |
a6255b7f DQ |
753 | int j; |
754 | uint8_t *b; | |
755 | ||
756 | *dst ^= *rq->q; | |
757 | for (j = 0, b = (uint8_t *)dst; j < 8; j++, b++) { | |
758 | *b = vdev_raidz_exp2(*b, rq->exp); | |
759 | } | |
760 | } | |
761 | ||
762 | return (0); | |
763 | } | |
764 | ||
765 | struct reconst_pq_struct { | |
766 | uint8_t *p; | |
767 | uint8_t *q; | |
768 | uint8_t *pxy; | |
769 | uint8_t *qxy; | |
770 | int aexp; | |
771 | int bexp; | |
772 | }; | |
773 | ||
774 | static int | |
775 | vdev_raidz_reconst_pq_func(void *xbuf, void *ybuf, size_t size, void *private) | |
776 | { | |
777 | struct reconst_pq_struct *rpq = private; | |
778 | uint8_t *xd = xbuf; | |
779 | uint8_t *yd = ybuf; | |
a6255b7f | 780 | |
1c27024e | 781 | for (int i = 0; i < size; |
a6255b7f DQ |
782 | i++, rpq->p++, rpq->q++, rpq->pxy++, rpq->qxy++, xd++, yd++) { |
783 | *xd = vdev_raidz_exp2(*rpq->p ^ *rpq->pxy, rpq->aexp) ^ | |
784 | vdev_raidz_exp2(*rpq->q ^ *rpq->qxy, rpq->bexp); | |
785 | *yd = *rpq->p ^ *rpq->pxy ^ *xd; | |
786 | } | |
787 | ||
788 | return (0); | |
789 | } | |
790 | ||
791 | static int | |
792 | vdev_raidz_reconst_pq_tail_func(void *xbuf, size_t size, void *private) | |
793 | { | |
794 | struct reconst_pq_struct *rpq = private; | |
795 | uint8_t *xd = xbuf; | |
a6255b7f | 796 | |
1c27024e | 797 | for (int i = 0; i < size; |
a6255b7f DQ |
798 | i++, rpq->p++, rpq->q++, rpq->pxy++, rpq->qxy++, xd++) { |
799 | /* same operation as vdev_raidz_reconst_pq_func() on xd */ | |
800 | *xd = vdev_raidz_exp2(*rpq->p ^ *rpq->pxy, rpq->aexp) ^ | |
801 | vdev_raidz_exp2(*rpq->q ^ *rpq->qxy, rpq->bexp); | |
802 | } | |
803 | ||
804 | return (0); | |
805 | } | |
806 | ||
45d1cae3 BB |
807 | static int |
808 | vdev_raidz_reconstruct_p(raidz_map_t *rm, int *tgts, int ntgts) | |
34dc7c2f | 809 | { |
45d1cae3 | 810 | int x = tgts[0]; |
34dc7c2f | 811 | int c; |
a6255b7f | 812 | abd_t *dst, *src; |
34dc7c2f | 813 | |
45d1cae3 BB |
814 | ASSERT(ntgts == 1); |
815 | ASSERT(x >= rm->rm_firstdatacol); | |
816 | ASSERT(x < rm->rm_cols); | |
817 | ||
a6255b7f DQ |
818 | ASSERT(rm->rm_col[x].rc_size <= rm->rm_col[VDEV_RAIDZ_P].rc_size); |
819 | ASSERT(rm->rm_col[x].rc_size > 0); | |
34dc7c2f | 820 | |
a6255b7f DQ |
821 | src = rm->rm_col[VDEV_RAIDZ_P].rc_abd; |
822 | dst = rm->rm_col[x].rc_abd; | |
823 | ||
824 | abd_copy_from_buf(dst, abd_to_buf(src), rm->rm_col[x].rc_size); | |
34dc7c2f BB |
825 | |
826 | for (c = rm->rm_firstdatacol; c < rm->rm_cols; c++) { | |
a6255b7f DQ |
827 | uint64_t size = MIN(rm->rm_col[x].rc_size, |
828 | rm->rm_col[c].rc_size); | |
829 | ||
830 | src = rm->rm_col[c].rc_abd; | |
831 | dst = rm->rm_col[x].rc_abd; | |
34dc7c2f BB |
832 | |
833 | if (c == x) | |
834 | continue; | |
835 | ||
a6255b7f DQ |
836 | (void) abd_iterate_func2(dst, src, 0, 0, size, |
837 | vdev_raidz_reconst_p_func, NULL); | |
34dc7c2f | 838 | } |
45d1cae3 BB |
839 | |
840 | return (1 << VDEV_RAIDZ_P); | |
34dc7c2f BB |
841 | } |
842 | ||
45d1cae3 BB |
843 | static int |
844 | vdev_raidz_reconstruct_q(raidz_map_t *rm, int *tgts, int ntgts) | |
34dc7c2f | 845 | { |
45d1cae3 | 846 | int x = tgts[0]; |
a6255b7f DQ |
847 | int c, exp; |
848 | abd_t *dst, *src; | |
34dc7c2f | 849 | |
45d1cae3 BB |
850 | ASSERT(ntgts == 1); |
851 | ||
a6255b7f | 852 | ASSERT(rm->rm_col[x].rc_size <= rm->rm_col[VDEV_RAIDZ_Q].rc_size); |
34dc7c2f BB |
853 | |
854 | for (c = rm->rm_firstdatacol; c < rm->rm_cols; c++) { | |
a6255b7f DQ |
855 | uint64_t size = (c == x) ? 0 : MIN(rm->rm_col[x].rc_size, |
856 | rm->rm_col[c].rc_size); | |
34dc7c2f | 857 | |
a6255b7f DQ |
858 | src = rm->rm_col[c].rc_abd; |
859 | dst = rm->rm_col[x].rc_abd; | |
34dc7c2f BB |
860 | |
861 | if (c == rm->rm_firstdatacol) { | |
a6255b7f DQ |
862 | abd_copy(dst, src, size); |
863 | if (rm->rm_col[x].rc_size > size) | |
864 | abd_zero_off(dst, size, | |
865 | rm->rm_col[x].rc_size - size); | |
34dc7c2f BB |
866 | |
867 | } else { | |
a6255b7f DQ |
868 | ASSERT3U(size, <=, rm->rm_col[x].rc_size); |
869 | (void) abd_iterate_func2(dst, src, 0, 0, size, | |
870 | vdev_raidz_reconst_q_pre_func, NULL); | |
871 | (void) abd_iterate_func(dst, | |
872 | size, rm->rm_col[x].rc_size - size, | |
873 | vdev_raidz_reconst_q_pre_tail_func, NULL); | |
34dc7c2f BB |
874 | } |
875 | } | |
876 | ||
a6255b7f DQ |
877 | src = rm->rm_col[VDEV_RAIDZ_Q].rc_abd; |
878 | dst = rm->rm_col[x].rc_abd; | |
34dc7c2f BB |
879 | exp = 255 - (rm->rm_cols - 1 - x); |
880 | ||
1c27024e | 881 | struct reconst_q_struct rq = { abd_to_buf(src), exp }; |
a6255b7f DQ |
882 | (void) abd_iterate_func(dst, 0, rm->rm_col[x].rc_size, |
883 | vdev_raidz_reconst_q_post_func, &rq); | |
45d1cae3 BB |
884 | |
885 | return (1 << VDEV_RAIDZ_Q); | |
34dc7c2f BB |
886 | } |
887 | ||
45d1cae3 BB |
888 | static int |
889 | vdev_raidz_reconstruct_pq(raidz_map_t *rm, int *tgts, int ntgts) | |
34dc7c2f | 890 | { |
a6255b7f DQ |
891 | uint8_t *p, *q, *pxy, *qxy, tmp, a, b, aexp, bexp; |
892 | abd_t *pdata, *qdata; | |
893 | uint64_t xsize, ysize; | |
45d1cae3 BB |
894 | int x = tgts[0]; |
895 | int y = tgts[1]; | |
a6255b7f | 896 | abd_t *xd, *yd; |
34dc7c2f | 897 | |
45d1cae3 | 898 | ASSERT(ntgts == 2); |
34dc7c2f BB |
899 | ASSERT(x < y); |
900 | ASSERT(x >= rm->rm_firstdatacol); | |
901 | ASSERT(y < rm->rm_cols); | |
902 | ||
903 | ASSERT(rm->rm_col[x].rc_size >= rm->rm_col[y].rc_size); | |
904 | ||
905 | /* | |
906 | * Move the parity data aside -- we're going to compute parity as | |
907 | * though columns x and y were full of zeros -- Pxy and Qxy. We want to | |
908 | * reuse the parity generation mechanism without trashing the actual | |
909 | * parity so we make those columns appear to be full of zeros by | |
910 | * setting their lengths to zero. | |
911 | */ | |
a6255b7f DQ |
912 | pdata = rm->rm_col[VDEV_RAIDZ_P].rc_abd; |
913 | qdata = rm->rm_col[VDEV_RAIDZ_Q].rc_abd; | |
34dc7c2f BB |
914 | xsize = rm->rm_col[x].rc_size; |
915 | ysize = rm->rm_col[y].rc_size; | |
916 | ||
a6255b7f DQ |
917 | rm->rm_col[VDEV_RAIDZ_P].rc_abd = |
918 | abd_alloc_linear(rm->rm_col[VDEV_RAIDZ_P].rc_size, B_TRUE); | |
919 | rm->rm_col[VDEV_RAIDZ_Q].rc_abd = | |
920 | abd_alloc_linear(rm->rm_col[VDEV_RAIDZ_Q].rc_size, B_TRUE); | |
34dc7c2f BB |
921 | rm->rm_col[x].rc_size = 0; |
922 | rm->rm_col[y].rc_size = 0; | |
923 | ||
924 | vdev_raidz_generate_parity_pq(rm); | |
925 | ||
926 | rm->rm_col[x].rc_size = xsize; | |
927 | rm->rm_col[y].rc_size = ysize; | |
928 | ||
a6255b7f DQ |
929 | p = abd_to_buf(pdata); |
930 | q = abd_to_buf(qdata); | |
931 | pxy = abd_to_buf(rm->rm_col[VDEV_RAIDZ_P].rc_abd); | |
932 | qxy = abd_to_buf(rm->rm_col[VDEV_RAIDZ_Q].rc_abd); | |
933 | xd = rm->rm_col[x].rc_abd; | |
934 | yd = rm->rm_col[y].rc_abd; | |
34dc7c2f BB |
935 | |
936 | /* | |
937 | * We now have: | |
938 | * Pxy = P + D_x + D_y | |
939 | * Qxy = Q + 2^(ndevs - 1 - x) * D_x + 2^(ndevs - 1 - y) * D_y | |
940 | * | |
941 | * We can then solve for D_x: | |
942 | * D_x = A * (P + Pxy) + B * (Q + Qxy) | |
943 | * where | |
944 | * A = 2^(x - y) * (2^(x - y) + 1)^-1 | |
945 | * B = 2^(ndevs - 1 - x) * (2^(x - y) + 1)^-1 | |
946 | * | |
947 | * With D_x in hand, we can easily solve for D_y: | |
948 | * D_y = P + Pxy + D_x | |
949 | */ | |
950 | ||
951 | a = vdev_raidz_pow2[255 + x - y]; | |
952 | b = vdev_raidz_pow2[255 - (rm->rm_cols - 1 - x)]; | |
953 | tmp = 255 - vdev_raidz_log2[a ^ 1]; | |
954 | ||
955 | aexp = vdev_raidz_log2[vdev_raidz_exp2(a, tmp)]; | |
956 | bexp = vdev_raidz_log2[vdev_raidz_exp2(b, tmp)]; | |
957 | ||
a6255b7f | 958 | ASSERT3U(xsize, >=, ysize); |
1c27024e | 959 | struct reconst_pq_struct rpq = { p, q, pxy, qxy, aexp, bexp }; |
34dc7c2f | 960 | |
a6255b7f DQ |
961 | (void) abd_iterate_func2(xd, yd, 0, 0, ysize, |
962 | vdev_raidz_reconst_pq_func, &rpq); | |
963 | (void) abd_iterate_func(xd, ysize, xsize - ysize, | |
964 | vdev_raidz_reconst_pq_tail_func, &rpq); | |
34dc7c2f | 965 | |
a6255b7f DQ |
966 | abd_free(rm->rm_col[VDEV_RAIDZ_P].rc_abd); |
967 | abd_free(rm->rm_col[VDEV_RAIDZ_Q].rc_abd); | |
34dc7c2f BB |
968 | |
969 | /* | |
970 | * Restore the saved parity data. | |
971 | */ | |
a6255b7f DQ |
972 | rm->rm_col[VDEV_RAIDZ_P].rc_abd = pdata; |
973 | rm->rm_col[VDEV_RAIDZ_Q].rc_abd = qdata; | |
45d1cae3 BB |
974 | |
975 | return ((1 << VDEV_RAIDZ_P) | (1 << VDEV_RAIDZ_Q)); | |
976 | } | |
977 | ||
978 | /* BEGIN CSTYLED */ | |
979 | /* | |
980 | * In the general case of reconstruction, we must solve the system of linear | |
981 | * equations defined by the coeffecients used to generate parity as well as | |
982 | * the contents of the data and parity disks. This can be expressed with | |
983 | * vectors for the original data (D) and the actual data (d) and parity (p) | |
984 | * and a matrix composed of the identity matrix (I) and a dispersal matrix (V): | |
985 | * | |
986 | * __ __ __ __ | |
987 | * | | __ __ | p_0 | | |
988 | * | V | | D_0 | | p_m-1 | | |
989 | * | | x | : | = | d_0 | | |
990 | * | I | | D_n-1 | | : | | |
991 | * | | ~~ ~~ | d_n-1 | | |
992 | * ~~ ~~ ~~ ~~ | |
993 | * | |
994 | * I is simply a square identity matrix of size n, and V is a vandermonde | |
995 | * matrix defined by the coeffecients we chose for the various parity columns | |
996 | * (1, 2, 4). Note that these values were chosen both for simplicity, speedy | |
997 | * computation as well as linear separability. | |
998 | * | |
999 | * __ __ __ __ | |
1000 | * | 1 .. 1 1 1 | | p_0 | | |
1001 | * | 2^n-1 .. 4 2 1 | __ __ | : | | |
1002 | * | 4^n-1 .. 16 4 1 | | D_0 | | p_m-1 | | |
1003 | * | 1 .. 0 0 0 | | D_1 | | d_0 | | |
1004 | * | 0 .. 0 0 0 | x | D_2 | = | d_1 | | |
1005 | * | : : : : | | : | | d_2 | | |
1006 | * | 0 .. 1 0 0 | | D_n-1 | | : | | |
1007 | * | 0 .. 0 1 0 | ~~ ~~ | : | | |
1008 | * | 0 .. 0 0 1 | | d_n-1 | | |
1009 | * ~~ ~~ ~~ ~~ | |
1010 | * | |
1011 | * Note that I, V, d, and p are known. To compute D, we must invert the | |
1012 | * matrix and use the known data and parity values to reconstruct the unknown | |
1013 | * data values. We begin by removing the rows in V|I and d|p that correspond | |
1014 | * to failed or missing columns; we then make V|I square (n x n) and d|p | |
1015 | * sized n by removing rows corresponding to unused parity from the bottom up | |
1016 | * to generate (V|I)' and (d|p)'. We can then generate the inverse of (V|I)' | |
1017 | * using Gauss-Jordan elimination. In the example below we use m=3 parity | |
1018 | * columns, n=8 data columns, with errors in d_1, d_2, and p_1: | |
1019 | * __ __ | |
1020 | * | 1 1 1 1 1 1 1 1 | | |
1021 | * | 128 64 32 16 8 4 2 1 | <-----+-+-- missing disks | |
1022 | * | 19 205 116 29 64 16 4 1 | / / | |
1023 | * | 1 0 0 0 0 0 0 0 | / / | |
1024 | * | 0 1 0 0 0 0 0 0 | <--' / | |
1025 | * (V|I) = | 0 0 1 0 0 0 0 0 | <---' | |
1026 | * | 0 0 0 1 0 0 0 0 | | |
1027 | * | 0 0 0 0 1 0 0 0 | | |
1028 | * | 0 0 0 0 0 1 0 0 | | |
1029 | * | 0 0 0 0 0 0 1 0 | | |
1030 | * | 0 0 0 0 0 0 0 1 | | |
1031 | * ~~ ~~ | |
1032 | * __ __ | |
1033 | * | 1 1 1 1 1 1 1 1 | | |
1034 | * | 128 64 32 16 8 4 2 1 | | |
1035 | * | 19 205 116 29 64 16 4 1 | | |
1036 | * | 1 0 0 0 0 0 0 0 | | |
1037 | * | 0 1 0 0 0 0 0 0 | | |
1038 | * (V|I)' = | 0 0 1 0 0 0 0 0 | | |
1039 | * | 0 0 0 1 0 0 0 0 | | |
1040 | * | 0 0 0 0 1 0 0 0 | | |
1041 | * | 0 0 0 0 0 1 0 0 | | |
1042 | * | 0 0 0 0 0 0 1 0 | | |
1043 | * | 0 0 0 0 0 0 0 1 | | |
1044 | * ~~ ~~ | |
1045 | * | |
1046 | * Here we employ Gauss-Jordan elimination to find the inverse of (V|I)'. We | |
1047 | * have carefully chosen the seed values 1, 2, and 4 to ensure that this | |
1048 | * matrix is not singular. | |
1049 | * __ __ | |
1050 | * | 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 | | |
1051 | * | 19 205 116 29 64 16 4 1 0 1 0 0 0 0 0 0 | | |
1052 | * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 | | |
1053 | * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 | | |
1054 | * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 | | |
1055 | * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 | | |
1056 | * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 | | |
1057 | * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 | | |
1058 | * ~~ ~~ | |
1059 | * __ __ | |
1060 | * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 | | |
1061 | * | 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 | | |
1062 | * | 19 205 116 29 64 16 4 1 0 1 0 0 0 0 0 0 | | |
1063 | * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 | | |
1064 | * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 | | |
1065 | * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 | | |
1066 | * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 | | |
1067 | * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 | | |
1068 | * ~~ ~~ | |
1069 | * __ __ | |
1070 | * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 | | |
1071 | * | 0 1 1 0 0 0 0 0 1 0 1 1 1 1 1 1 | | |
1072 | * | 0 205 116 0 0 0 0 0 0 1 19 29 64 16 4 1 | | |
1073 | * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 | | |
1074 | * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 | | |
1075 | * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 | | |
1076 | * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 | | |
1077 | * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 | | |
1078 | * ~~ ~~ | |
1079 | * __ __ | |
1080 | * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 | | |
1081 | * | 0 1 1 0 0 0 0 0 1 0 1 1 1 1 1 1 | | |
1082 | * | 0 0 185 0 0 0 0 0 205 1 222 208 141 221 201 204 | | |
1083 | * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 | | |
1084 | * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 | | |
1085 | * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 | | |
1086 | * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 | | |
1087 | * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 | | |
1088 | * ~~ ~~ | |
1089 | * __ __ | |
1090 | * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 | | |
1091 | * | 0 1 1 0 0 0 0 0 1 0 1 1 1 1 1 1 | | |
1092 | * | 0 0 1 0 0 0 0 0 166 100 4 40 158 168 216 209 | | |
1093 | * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 | | |
1094 | * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 | | |
1095 | * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 | | |
1096 | * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 | | |
1097 | * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 | | |
1098 | * ~~ ~~ | |
1099 | * __ __ | |
1100 | * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 | | |
1101 | * | 0 1 0 0 0 0 0 0 167 100 5 41 159 169 217 208 | | |
1102 | * | 0 0 1 0 0 0 0 0 166 100 4 40 158 168 216 209 | | |
1103 | * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 | | |
1104 | * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 | | |
1105 | * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 | | |
1106 | * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 | | |
1107 | * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 | | |
1108 | * ~~ ~~ | |
1109 | * __ __ | |
1110 | * | 0 0 1 0 0 0 0 0 | | |
1111 | * | 167 100 5 41 159 169 217 208 | | |
1112 | * | 166 100 4 40 158 168 216 209 | | |
1113 | * (V|I)'^-1 = | 0 0 0 1 0 0 0 0 | | |
1114 | * | 0 0 0 0 1 0 0 0 | | |
1115 | * | 0 0 0 0 0 1 0 0 | | |
1116 | * | 0 0 0 0 0 0 1 0 | | |
1117 | * | 0 0 0 0 0 0 0 1 | | |
1118 | * ~~ ~~ | |
1119 | * | |
1120 | * We can then simply compute D = (V|I)'^-1 x (d|p)' to discover the values | |
1121 | * of the missing data. | |
1122 | * | |
1123 | * As is apparent from the example above, the only non-trivial rows in the | |
1124 | * inverse matrix correspond to the data disks that we're trying to | |
1125 | * reconstruct. Indeed, those are the only rows we need as the others would | |
1126 | * only be useful for reconstructing data known or assumed to be valid. For | |
1127 | * that reason, we only build the coefficients in the rows that correspond to | |
1128 | * targeted columns. | |
1129 | */ | |
1130 | /* END CSTYLED */ | |
1131 | ||
1132 | static void | |
1133 | vdev_raidz_matrix_init(raidz_map_t *rm, int n, int nmap, int *map, | |
1134 | uint8_t **rows) | |
1135 | { | |
1136 | int i, j; | |
1137 | int pow; | |
1138 | ||
1139 | ASSERT(n == rm->rm_cols - rm->rm_firstdatacol); | |
1140 | ||
1141 | /* | |
1142 | * Fill in the missing rows of interest. | |
1143 | */ | |
1144 | for (i = 0; i < nmap; i++) { | |
1145 | ASSERT3S(0, <=, map[i]); | |
1146 | ASSERT3S(map[i], <=, 2); | |
1147 | ||
1148 | pow = map[i] * n; | |
1149 | if (pow > 255) | |
1150 | pow -= 255; | |
1151 | ASSERT(pow <= 255); | |
1152 | ||
1153 | for (j = 0; j < n; j++) { | |
1154 | pow -= map[i]; | |
1155 | if (pow < 0) | |
1156 | pow += 255; | |
1157 | rows[i][j] = vdev_raidz_pow2[pow]; | |
1158 | } | |
1159 | } | |
1160 | } | |
1161 | ||
1162 | static void | |
1163 | vdev_raidz_matrix_invert(raidz_map_t *rm, int n, int nmissing, int *missing, | |
1164 | uint8_t **rows, uint8_t **invrows, const uint8_t *used) | |
1165 | { | |
1166 | int i, j, ii, jj; | |
1167 | uint8_t log; | |
1168 | ||
1169 | /* | |
1170 | * Assert that the first nmissing entries from the array of used | |
1171 | * columns correspond to parity columns and that subsequent entries | |
1172 | * correspond to data columns. | |
1173 | */ | |
1174 | for (i = 0; i < nmissing; i++) { | |
1175 | ASSERT3S(used[i], <, rm->rm_firstdatacol); | |
1176 | } | |
1177 | for (; i < n; i++) { | |
1178 | ASSERT3S(used[i], >=, rm->rm_firstdatacol); | |
1179 | } | |
1180 | ||
1181 | /* | |
1182 | * First initialize the storage where we'll compute the inverse rows. | |
1183 | */ | |
1184 | for (i = 0; i < nmissing; i++) { | |
1185 | for (j = 0; j < n; j++) { | |
1186 | invrows[i][j] = (i == j) ? 1 : 0; | |
1187 | } | |
1188 | } | |
1189 | ||
1190 | /* | |
1191 | * Subtract all trivial rows from the rows of consequence. | |
1192 | */ | |
1193 | for (i = 0; i < nmissing; i++) { | |
1194 | for (j = nmissing; j < n; j++) { | |
1195 | ASSERT3U(used[j], >=, rm->rm_firstdatacol); | |
1196 | jj = used[j] - rm->rm_firstdatacol; | |
1197 | ASSERT3S(jj, <, n); | |
1198 | invrows[i][j] = rows[i][jj]; | |
1199 | rows[i][jj] = 0; | |
1200 | } | |
1201 | } | |
1202 | ||
1203 | /* | |
1204 | * For each of the rows of interest, we must normalize it and subtract | |
1205 | * a multiple of it from the other rows. | |
1206 | */ | |
1207 | for (i = 0; i < nmissing; i++) { | |
1208 | for (j = 0; j < missing[i]; j++) { | |
c99c9001 | 1209 | ASSERT0(rows[i][j]); |
45d1cae3 BB |
1210 | } |
1211 | ASSERT3U(rows[i][missing[i]], !=, 0); | |
1212 | ||
1213 | /* | |
1214 | * Compute the inverse of the first element and multiply each | |
1215 | * element in the row by that value. | |
1216 | */ | |
1217 | log = 255 - vdev_raidz_log2[rows[i][missing[i]]]; | |
1218 | ||
1219 | for (j = 0; j < n; j++) { | |
1220 | rows[i][j] = vdev_raidz_exp2(rows[i][j], log); | |
1221 | invrows[i][j] = vdev_raidz_exp2(invrows[i][j], log); | |
1222 | } | |
1223 | ||
1224 | for (ii = 0; ii < nmissing; ii++) { | |
1225 | if (i == ii) | |
1226 | continue; | |
1227 | ||
1228 | ASSERT3U(rows[ii][missing[i]], !=, 0); | |
1229 | ||
1230 | log = vdev_raidz_log2[rows[ii][missing[i]]]; | |
1231 | ||
1232 | for (j = 0; j < n; j++) { | |
1233 | rows[ii][j] ^= | |
1234 | vdev_raidz_exp2(rows[i][j], log); | |
1235 | invrows[ii][j] ^= | |
1236 | vdev_raidz_exp2(invrows[i][j], log); | |
1237 | } | |
1238 | } | |
1239 | } | |
1240 | ||
1241 | /* | |
1242 | * Verify that the data that is left in the rows are properly part of | |
1243 | * an identity matrix. | |
1244 | */ | |
1245 | for (i = 0; i < nmissing; i++) { | |
1246 | for (j = 0; j < n; j++) { | |
1247 | if (j == missing[i]) { | |
1248 | ASSERT3U(rows[i][j], ==, 1); | |
1249 | } else { | |
c99c9001 | 1250 | ASSERT0(rows[i][j]); |
45d1cae3 BB |
1251 | } |
1252 | } | |
1253 | } | |
1254 | } | |
1255 | ||
1256 | static void | |
1257 | vdev_raidz_matrix_reconstruct(raidz_map_t *rm, int n, int nmissing, | |
1258 | int *missing, uint8_t **invrows, const uint8_t *used) | |
1259 | { | |
1260 | int i, j, x, cc, c; | |
1261 | uint8_t *src; | |
1262 | uint64_t ccount; | |
689f093e GN |
1263 | uint8_t *dst[VDEV_RAIDZ_MAXPARITY] = { NULL }; |
1264 | uint64_t dcount[VDEV_RAIDZ_MAXPARITY] = { 0 }; | |
a117a6d6 GW |
1265 | uint8_t log = 0; |
1266 | uint8_t val; | |
45d1cae3 BB |
1267 | int ll; |
1268 | uint8_t *invlog[VDEV_RAIDZ_MAXPARITY]; | |
1269 | uint8_t *p, *pp; | |
1270 | size_t psize; | |
1271 | ||
1272 | psize = sizeof (invlog[0][0]) * n * nmissing; | |
79c76d5b | 1273 | p = kmem_alloc(psize, KM_SLEEP); |
45d1cae3 BB |
1274 | |
1275 | for (pp = p, i = 0; i < nmissing; i++) { | |
1276 | invlog[i] = pp; | |
1277 | pp += n; | |
1278 | } | |
1279 | ||
1280 | for (i = 0; i < nmissing; i++) { | |
1281 | for (j = 0; j < n; j++) { | |
1282 | ASSERT3U(invrows[i][j], !=, 0); | |
1283 | invlog[i][j] = vdev_raidz_log2[invrows[i][j]]; | |
1284 | } | |
1285 | } | |
1286 | ||
1287 | for (i = 0; i < n; i++) { | |
1288 | c = used[i]; | |
1289 | ASSERT3U(c, <, rm->rm_cols); | |
1290 | ||
a6255b7f | 1291 | src = abd_to_buf(rm->rm_col[c].rc_abd); |
45d1cae3 BB |
1292 | ccount = rm->rm_col[c].rc_size; |
1293 | for (j = 0; j < nmissing; j++) { | |
1294 | cc = missing[j] + rm->rm_firstdatacol; | |
1295 | ASSERT3U(cc, >=, rm->rm_firstdatacol); | |
1296 | ASSERT3U(cc, <, rm->rm_cols); | |
1297 | ASSERT3U(cc, !=, c); | |
1298 | ||
a6255b7f | 1299 | dst[j] = abd_to_buf(rm->rm_col[cc].rc_abd); |
45d1cae3 BB |
1300 | dcount[j] = rm->rm_col[cc].rc_size; |
1301 | } | |
1302 | ||
1303 | ASSERT(ccount >= rm->rm_col[missing[0]].rc_size || i > 0); | |
1304 | ||
1305 | for (x = 0; x < ccount; x++, src++) { | |
1306 | if (*src != 0) | |
1307 | log = vdev_raidz_log2[*src]; | |
1308 | ||
1309 | for (cc = 0; cc < nmissing; cc++) { | |
1310 | if (x >= dcount[cc]) | |
1311 | continue; | |
1312 | ||
1313 | if (*src == 0) { | |
1314 | val = 0; | |
1315 | } else { | |
1316 | if ((ll = log + invlog[cc][i]) >= 255) | |
1317 | ll -= 255; | |
1318 | val = vdev_raidz_pow2[ll]; | |
1319 | } | |
1320 | ||
1321 | if (i == 0) | |
1322 | dst[cc][x] = val; | |
1323 | else | |
1324 | dst[cc][x] ^= val; | |
1325 | } | |
1326 | } | |
1327 | } | |
1328 | ||
1329 | kmem_free(p, psize); | |
1330 | } | |
1331 | ||
1332 | static int | |
1333 | vdev_raidz_reconstruct_general(raidz_map_t *rm, int *tgts, int ntgts) | |
1334 | { | |
1335 | int n, i, c, t, tt; | |
1336 | int nmissing_rows; | |
1337 | int missing_rows[VDEV_RAIDZ_MAXPARITY]; | |
1338 | int parity_map[VDEV_RAIDZ_MAXPARITY]; | |
1339 | ||
1340 | uint8_t *p, *pp; | |
1341 | size_t psize; | |
1342 | ||
1343 | uint8_t *rows[VDEV_RAIDZ_MAXPARITY]; | |
1344 | uint8_t *invrows[VDEV_RAIDZ_MAXPARITY]; | |
1345 | uint8_t *used; | |
1346 | ||
a6255b7f DQ |
1347 | abd_t **bufs = NULL; |
1348 | ||
45d1cae3 BB |
1349 | int code = 0; |
1350 | ||
a6255b7f DQ |
1351 | /* |
1352 | * Matrix reconstruction can't use scatter ABDs yet, so we allocate | |
1353 | * temporary linear ABDs. | |
1354 | */ | |
1355 | if (!abd_is_linear(rm->rm_col[rm->rm_firstdatacol].rc_abd)) { | |
1356 | bufs = kmem_alloc(rm->rm_cols * sizeof (abd_t *), KM_PUSHPAGE); | |
1357 | ||
1358 | for (c = rm->rm_firstdatacol; c < rm->rm_cols; c++) { | |
1359 | raidz_col_t *col = &rm->rm_col[c]; | |
1360 | ||
1361 | bufs[c] = col->rc_abd; | |
1362 | col->rc_abd = abd_alloc_linear(col->rc_size, B_TRUE); | |
1363 | abd_copy(col->rc_abd, bufs[c], col->rc_size); | |
1364 | } | |
1365 | } | |
45d1cae3 BB |
1366 | |
1367 | n = rm->rm_cols - rm->rm_firstdatacol; | |
1368 | ||
1369 | /* | |
1370 | * Figure out which data columns are missing. | |
1371 | */ | |
1372 | nmissing_rows = 0; | |
1373 | for (t = 0; t < ntgts; t++) { | |
1374 | if (tgts[t] >= rm->rm_firstdatacol) { | |
1375 | missing_rows[nmissing_rows++] = | |
1376 | tgts[t] - rm->rm_firstdatacol; | |
1377 | } | |
1378 | } | |
1379 | ||
1380 | /* | |
1381 | * Figure out which parity columns to use to help generate the missing | |
1382 | * data columns. | |
1383 | */ | |
1384 | for (tt = 0, c = 0, i = 0; i < nmissing_rows; c++) { | |
1385 | ASSERT(tt < ntgts); | |
1386 | ASSERT(c < rm->rm_firstdatacol); | |
1387 | ||
1388 | /* | |
1389 | * Skip any targeted parity columns. | |
1390 | */ | |
1391 | if (c == tgts[tt]) { | |
1392 | tt++; | |
1393 | continue; | |
1394 | } | |
1395 | ||
1396 | code |= 1 << c; | |
1397 | ||
1398 | parity_map[i] = c; | |
1399 | i++; | |
1400 | } | |
1401 | ||
1402 | ASSERT(code != 0); | |
1403 | ASSERT3U(code, <, 1 << VDEV_RAIDZ_MAXPARITY); | |
1404 | ||
1405 | psize = (sizeof (rows[0][0]) + sizeof (invrows[0][0])) * | |
1406 | nmissing_rows * n + sizeof (used[0]) * n; | |
79c76d5b | 1407 | p = kmem_alloc(psize, KM_SLEEP); |
45d1cae3 BB |
1408 | |
1409 | for (pp = p, i = 0; i < nmissing_rows; i++) { | |
1410 | rows[i] = pp; | |
1411 | pp += n; | |
1412 | invrows[i] = pp; | |
1413 | pp += n; | |
1414 | } | |
1415 | used = pp; | |
1416 | ||
1417 | for (i = 0; i < nmissing_rows; i++) { | |
1418 | used[i] = parity_map[i]; | |
1419 | } | |
1420 | ||
1421 | for (tt = 0, c = rm->rm_firstdatacol; c < rm->rm_cols; c++) { | |
1422 | if (tt < nmissing_rows && | |
1423 | c == missing_rows[tt] + rm->rm_firstdatacol) { | |
1424 | tt++; | |
1425 | continue; | |
1426 | } | |
1427 | ||
1428 | ASSERT3S(i, <, n); | |
1429 | used[i] = c; | |
1430 | i++; | |
1431 | } | |
1432 | ||
1433 | /* | |
1434 | * Initialize the interesting rows of the matrix. | |
1435 | */ | |
1436 | vdev_raidz_matrix_init(rm, n, nmissing_rows, parity_map, rows); | |
1437 | ||
1438 | /* | |
1439 | * Invert the matrix. | |
1440 | */ | |
1441 | vdev_raidz_matrix_invert(rm, n, nmissing_rows, missing_rows, rows, | |
1442 | invrows, used); | |
1443 | ||
1444 | /* | |
1445 | * Reconstruct the missing data using the generated matrix. | |
1446 | */ | |
1447 | vdev_raidz_matrix_reconstruct(rm, n, nmissing_rows, missing_rows, | |
1448 | invrows, used); | |
1449 | ||
1450 | kmem_free(p, psize); | |
1451 | ||
a6255b7f DQ |
1452 | /* |
1453 | * copy back from temporary linear abds and free them | |
1454 | */ | |
1455 | if (bufs) { | |
1456 | for (c = rm->rm_firstdatacol; c < rm->rm_cols; c++) { | |
1457 | raidz_col_t *col = &rm->rm_col[c]; | |
1458 | ||
1459 | abd_copy(bufs[c], col->rc_abd, col->rc_size); | |
1460 | abd_free(col->rc_abd); | |
1461 | col->rc_abd = bufs[c]; | |
1462 | } | |
1463 | kmem_free(bufs, rm->rm_cols * sizeof (abd_t *)); | |
1464 | } | |
1465 | ||
45d1cae3 | 1466 | return (code); |
34dc7c2f BB |
1467 | } |
1468 | ||
ab9f4b0b GN |
1469 | int |
1470 | vdev_raidz_reconstruct(raidz_map_t *rm, const int *t, int nt) | |
45d1cae3 BB |
1471 | { |
1472 | int tgts[VDEV_RAIDZ_MAXPARITY], *dt; | |
1473 | int ntgts; | |
c9187d86 | 1474 | int i, c, ret; |
45d1cae3 BB |
1475 | int code; |
1476 | int nbadparity, nbaddata; | |
1477 | int parity_valid[VDEV_RAIDZ_MAXPARITY]; | |
1478 | ||
1479 | /* | |
1480 | * The tgts list must already be sorted. | |
1481 | */ | |
1482 | for (i = 1; i < nt; i++) { | |
1483 | ASSERT(t[i] > t[i - 1]); | |
1484 | } | |
1485 | ||
1486 | nbadparity = rm->rm_firstdatacol; | |
1487 | nbaddata = rm->rm_cols - nbadparity; | |
1488 | ntgts = 0; | |
1489 | for (i = 0, c = 0; c < rm->rm_cols; c++) { | |
1490 | if (c < rm->rm_firstdatacol) | |
1491 | parity_valid[c] = B_FALSE; | |
1492 | ||
1493 | if (i < nt && c == t[i]) { | |
1494 | tgts[ntgts++] = c; | |
1495 | i++; | |
1496 | } else if (rm->rm_col[c].rc_error != 0) { | |
1497 | tgts[ntgts++] = c; | |
1498 | } else if (c >= rm->rm_firstdatacol) { | |
1499 | nbaddata--; | |
1500 | } else { | |
1501 | parity_valid[c] = B_TRUE; | |
1502 | nbadparity--; | |
1503 | } | |
1504 | } | |
1505 | ||
1506 | ASSERT(ntgts >= nt); | |
1507 | ASSERT(nbaddata >= 0); | |
1508 | ASSERT(nbaddata + nbadparity == ntgts); | |
1509 | ||
1510 | dt = &tgts[nbadparity]; | |
1511 | ||
c9187d86 GN |
1512 | /* Reconstruct using the new math implementation */ |
1513 | ret = vdev_raidz_math_reconstruct(rm, parity_valid, dt, nbaddata); | |
1514 | if (ret != RAIDZ_ORIGINAL_IMPL) | |
1515 | return (ret); | |
ab9f4b0b | 1516 | |
45d1cae3 BB |
1517 | /* |
1518 | * See if we can use any of our optimized reconstruction routines. | |
1519 | */ | |
ab9f4b0b GN |
1520 | switch (nbaddata) { |
1521 | case 1: | |
1522 | if (parity_valid[VDEV_RAIDZ_P]) | |
1523 | return (vdev_raidz_reconstruct_p(rm, dt, 1)); | |
45d1cae3 | 1524 | |
ab9f4b0b | 1525 | ASSERT(rm->rm_firstdatacol > 1); |
45d1cae3 | 1526 | |
ab9f4b0b GN |
1527 | if (parity_valid[VDEV_RAIDZ_Q]) |
1528 | return (vdev_raidz_reconstruct_q(rm, dt, 1)); | |
45d1cae3 | 1529 | |
ab9f4b0b GN |
1530 | ASSERT(rm->rm_firstdatacol > 2); |
1531 | break; | |
45d1cae3 | 1532 | |
ab9f4b0b GN |
1533 | case 2: |
1534 | ASSERT(rm->rm_firstdatacol > 1); | |
45d1cae3 | 1535 | |
ab9f4b0b GN |
1536 | if (parity_valid[VDEV_RAIDZ_P] && |
1537 | parity_valid[VDEV_RAIDZ_Q]) | |
1538 | return (vdev_raidz_reconstruct_pq(rm, dt, 2)); | |
45d1cae3 | 1539 | |
ab9f4b0b | 1540 | ASSERT(rm->rm_firstdatacol > 2); |
45d1cae3 | 1541 | |
ab9f4b0b | 1542 | break; |
45d1cae3 BB |
1543 | } |
1544 | ||
1545 | code = vdev_raidz_reconstruct_general(rm, tgts, ntgts); | |
1546 | ASSERT(code < (1 << VDEV_RAIDZ_MAXPARITY)); | |
1547 | ASSERT(code > 0); | |
1548 | return (code); | |
1549 | } | |
34dc7c2f BB |
1550 | |
1551 | static int | |
1bd201e7 CS |
1552 | vdev_raidz_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize, |
1553 | uint64_t *ashift) | |
34dc7c2f BB |
1554 | { |
1555 | vdev_t *cvd; | |
1556 | uint64_t nparity = vd->vdev_nparity; | |
45d1cae3 | 1557 | int c; |
34dc7c2f BB |
1558 | int lasterror = 0; |
1559 | int numerrors = 0; | |
1560 | ||
1561 | ASSERT(nparity > 0); | |
1562 | ||
1563 | if (nparity > VDEV_RAIDZ_MAXPARITY || | |
1564 | vd->vdev_children < nparity + 1) { | |
1565 | vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL; | |
2e528b49 | 1566 | return (SET_ERROR(EINVAL)); |
34dc7c2f BB |
1567 | } |
1568 | ||
45d1cae3 BB |
1569 | vdev_open_children(vd); |
1570 | ||
34dc7c2f BB |
1571 | for (c = 0; c < vd->vdev_children; c++) { |
1572 | cvd = vd->vdev_child[c]; | |
1573 | ||
45d1cae3 BB |
1574 | if (cvd->vdev_open_error != 0) { |
1575 | lasterror = cvd->vdev_open_error; | |
34dc7c2f BB |
1576 | numerrors++; |
1577 | continue; | |
1578 | } | |
1579 | ||
1580 | *asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1; | |
1bd201e7 | 1581 | *max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1; |
34dc7c2f BB |
1582 | *ashift = MAX(*ashift, cvd->vdev_ashift); |
1583 | } | |
1584 | ||
1585 | *asize *= vd->vdev_children; | |
1bd201e7 | 1586 | *max_asize *= vd->vdev_children; |
34dc7c2f BB |
1587 | |
1588 | if (numerrors > nparity) { | |
1589 | vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS; | |
1590 | return (lasterror); | |
1591 | } | |
1592 | ||
1593 | return (0); | |
1594 | } | |
1595 | ||
1596 | static void | |
1597 | vdev_raidz_close(vdev_t *vd) | |
1598 | { | |
1599 | int c; | |
1600 | ||
1601 | for (c = 0; c < vd->vdev_children; c++) | |
1602 | vdev_close(vd->vdev_child[c]); | |
1603 | } | |
1604 | ||
1605 | static uint64_t | |
1606 | vdev_raidz_asize(vdev_t *vd, uint64_t psize) | |
1607 | { | |
1608 | uint64_t asize; | |
1609 | uint64_t ashift = vd->vdev_top->vdev_ashift; | |
1610 | uint64_t cols = vd->vdev_children; | |
1611 | uint64_t nparity = vd->vdev_nparity; | |
1612 | ||
1613 | asize = ((psize - 1) >> ashift) + 1; | |
1614 | asize += nparity * ((asize + cols - nparity - 1) / (cols - nparity)); | |
1615 | asize = roundup(asize, nparity + 1) << ashift; | |
1616 | ||
1617 | return (asize); | |
1618 | } | |
1619 | ||
1620 | static void | |
1621 | vdev_raidz_child_done(zio_t *zio) | |
1622 | { | |
1623 | raidz_col_t *rc = zio->io_private; | |
1624 | ||
1625 | rc->rc_error = zio->io_error; | |
1626 | rc->rc_tried = 1; | |
1627 | rc->rc_skipped = 0; | |
1628 | } | |
1629 | ||
e49f1e20 WA |
1630 | /* |
1631 | * Start an IO operation on a RAIDZ VDev | |
1632 | * | |
1633 | * Outline: | |
1634 | * - For write operations: | |
1635 | * 1. Generate the parity data | |
1636 | * 2. Create child zio write operations to each column's vdev, for both | |
1637 | * data and parity. | |
1638 | * 3. If the column skips any sectors for padding, create optional dummy | |
1639 | * write zio children for those areas to improve aggregation continuity. | |
1640 | * - For read operations: | |
1641 | * 1. Create child zio read operations to each data column's vdev to read | |
1642 | * the range of data required for zio. | |
1643 | * 2. If this is a scrub or resilver operation, or if any of the data | |
1644 | * vdevs have had errors, then create zio read operations to the parity | |
1645 | * columns' VDevs as well. | |
1646 | */ | |
98b25418 | 1647 | static void |
34dc7c2f BB |
1648 | vdev_raidz_io_start(zio_t *zio) |
1649 | { | |
1650 | vdev_t *vd = zio->io_vd; | |
1651 | vdev_t *tvd = vd->vdev_top; | |
1652 | vdev_t *cvd; | |
34dc7c2f BB |
1653 | raidz_map_t *rm; |
1654 | raidz_col_t *rc; | |
45d1cae3 | 1655 | int c, i; |
34dc7c2f BB |
1656 | |
1657 | rm = vdev_raidz_map_alloc(zio, tvd->vdev_ashift, vd->vdev_children, | |
1658 | vd->vdev_nparity); | |
1659 | ||
1660 | ASSERT3U(rm->rm_asize, ==, vdev_psize_to_asize(vd, zio->io_size)); | |
1661 | ||
1662 | if (zio->io_type == ZIO_TYPE_WRITE) { | |
45d1cae3 | 1663 | vdev_raidz_generate_parity(rm); |
34dc7c2f BB |
1664 | |
1665 | for (c = 0; c < rm->rm_cols; c++) { | |
1666 | rc = &rm->rm_col[c]; | |
1667 | cvd = vd->vdev_child[rc->rc_devidx]; | |
1668 | zio_nowait(zio_vdev_child_io(zio, NULL, cvd, | |
a6255b7f | 1669 | rc->rc_offset, rc->rc_abd, rc->rc_size, |
b128c09f | 1670 | zio->io_type, zio->io_priority, 0, |
34dc7c2f BB |
1671 | vdev_raidz_child_done, rc)); |
1672 | } | |
1673 | ||
45d1cae3 BB |
1674 | /* |
1675 | * Generate optional I/Os for any skipped sectors to improve | |
1676 | * aggregation contiguity. | |
1677 | */ | |
428870ff | 1678 | for (c = rm->rm_skipstart, i = 0; i < rm->rm_nskip; c++, i++) { |
45d1cae3 BB |
1679 | ASSERT(c <= rm->rm_scols); |
1680 | if (c == rm->rm_scols) | |
1681 | c = 0; | |
1682 | rc = &rm->rm_col[c]; | |
1683 | cvd = vd->vdev_child[rc->rc_devidx]; | |
1684 | zio_nowait(zio_vdev_child_io(zio, NULL, cvd, | |
1685 | rc->rc_offset + rc->rc_size, NULL, | |
1686 | 1 << tvd->vdev_ashift, | |
1687 | zio->io_type, zio->io_priority, | |
1688 | ZIO_FLAG_NODATA | ZIO_FLAG_OPTIONAL, NULL, NULL)); | |
1689 | } | |
1690 | ||
98b25418 GW |
1691 | zio_execute(zio); |
1692 | return; | |
34dc7c2f BB |
1693 | } |
1694 | ||
1695 | ASSERT(zio->io_type == ZIO_TYPE_READ); | |
1696 | ||
1697 | /* | |
1698 | * Iterate over the columns in reverse order so that we hit the parity | |
45d1cae3 | 1699 | * last -- any errors along the way will force us to read the parity. |
34dc7c2f BB |
1700 | */ |
1701 | for (c = rm->rm_cols - 1; c >= 0; c--) { | |
1702 | rc = &rm->rm_col[c]; | |
1703 | cvd = vd->vdev_child[rc->rc_devidx]; | |
1704 | if (!vdev_readable(cvd)) { | |
1705 | if (c >= rm->rm_firstdatacol) | |
1706 | rm->rm_missingdata++; | |
1707 | else | |
1708 | rm->rm_missingparity++; | |
2e528b49 | 1709 | rc->rc_error = SET_ERROR(ENXIO); |
34dc7c2f BB |
1710 | rc->rc_tried = 1; /* don't even try */ |
1711 | rc->rc_skipped = 1; | |
1712 | continue; | |
1713 | } | |
428870ff | 1714 | if (vdev_dtl_contains(cvd, DTL_MISSING, zio->io_txg, 1)) { |
34dc7c2f BB |
1715 | if (c >= rm->rm_firstdatacol) |
1716 | rm->rm_missingdata++; | |
1717 | else | |
1718 | rm->rm_missingparity++; | |
2e528b49 | 1719 | rc->rc_error = SET_ERROR(ESTALE); |
34dc7c2f BB |
1720 | rc->rc_skipped = 1; |
1721 | continue; | |
1722 | } | |
1723 | if (c >= rm->rm_firstdatacol || rm->rm_missingdata > 0 || | |
9babb374 | 1724 | (zio->io_flags & (ZIO_FLAG_SCRUB | ZIO_FLAG_RESILVER))) { |
34dc7c2f | 1725 | zio_nowait(zio_vdev_child_io(zio, NULL, cvd, |
a6255b7f | 1726 | rc->rc_offset, rc->rc_abd, rc->rc_size, |
b128c09f | 1727 | zio->io_type, zio->io_priority, 0, |
34dc7c2f BB |
1728 | vdev_raidz_child_done, rc)); |
1729 | } | |
1730 | } | |
1731 | ||
98b25418 | 1732 | zio_execute(zio); |
34dc7c2f BB |
1733 | } |
1734 | ||
428870ff | 1735 | |
34dc7c2f BB |
1736 | /* |
1737 | * Report a checksum error for a child of a RAID-Z device. | |
1738 | */ | |
1739 | static void | |
84c07ada | 1740 | raidz_checksum_error(zio_t *zio, raidz_col_t *rc, abd_t *bad_data) |
34dc7c2f BB |
1741 | { |
1742 | vdev_t *vd = zio->io_vd->vdev_child[rc->rc_devidx]; | |
34dc7c2f BB |
1743 | |
1744 | if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { | |
428870ff BB |
1745 | zio_bad_cksum_t zbc; |
1746 | raidz_map_t *rm = zio->io_vsd; | |
1747 | ||
34dc7c2f BB |
1748 | mutex_enter(&vd->vdev_stat_lock); |
1749 | vd->vdev_stat.vs_checksum_errors++; | |
1750 | mutex_exit(&vd->vdev_stat_lock); | |
428870ff BB |
1751 | |
1752 | zbc.zbc_has_cksum = 0; | |
1753 | zbc.zbc_injected = rm->rm_ecksuminjected; | |
1754 | ||
b5256303 TC |
1755 | zfs_ereport_post_checksum(zio->io_spa, vd, |
1756 | &zio->io_bookmark, zio, rc->rc_offset, rc->rc_size, | |
1757 | rc->rc_abd, bad_data, &zbc); | |
34dc7c2f | 1758 | } |
428870ff BB |
1759 | } |
1760 | ||
1761 | /* | |
1762 | * We keep track of whether or not there were any injected errors, so that | |
1763 | * any ereports we generate can note it. | |
1764 | */ | |
1765 | static int | |
1766 | raidz_checksum_verify(zio_t *zio) | |
1767 | { | |
1768 | zio_bad_cksum_t zbc; | |
1769 | raidz_map_t *rm = zio->io_vsd; | |
1770 | ||
d4ed6673 BB |
1771 | bzero(&zbc, sizeof (zio_bad_cksum_t)); |
1772 | ||
1c27024e | 1773 | int ret = zio_checksum_error(zio, &zbc); |
428870ff BB |
1774 | if (ret != 0 && zbc.zbc_injected != 0) |
1775 | rm->rm_ecksuminjected = 1; | |
34dc7c2f | 1776 | |
428870ff | 1777 | return (ret); |
34dc7c2f BB |
1778 | } |
1779 | ||
1780 | /* | |
1781 | * Generate the parity from the data columns. If we tried and were able to | |
1782 | * read the parity without error, verify that the generated parity matches the | |
1783 | * data we read. If it doesn't, we fire off a checksum error. Return the | |
1784 | * number such failures. | |
1785 | */ | |
1786 | static int | |
1787 | raidz_parity_verify(zio_t *zio, raidz_map_t *rm) | |
1788 | { | |
84c07ada | 1789 | abd_t *orig[VDEV_RAIDZ_MAXPARITY]; |
34dc7c2f BB |
1790 | int c, ret = 0; |
1791 | raidz_col_t *rc; | |
1792 | ||
3c67d83a TH |
1793 | blkptr_t *bp = zio->io_bp; |
1794 | enum zio_checksum checksum = (bp == NULL ? zio->io_prop.zp_checksum : | |
1795 | (BP_IS_GANG(bp) ? ZIO_CHECKSUM_GANG_HEADER : BP_GET_CHECKSUM(bp))); | |
1796 | ||
1797 | if (checksum == ZIO_CHECKSUM_NOPARITY) | |
1798 | return (ret); | |
1799 | ||
34dc7c2f BB |
1800 | for (c = 0; c < rm->rm_firstdatacol; c++) { |
1801 | rc = &rm->rm_col[c]; | |
1802 | if (!rc->rc_tried || rc->rc_error != 0) | |
1803 | continue; | |
84c07ada GN |
1804 | |
1805 | orig[c] = abd_alloc_sametype(rc->rc_abd, rc->rc_size); | |
1806 | abd_copy(orig[c], rc->rc_abd, rc->rc_size); | |
34dc7c2f BB |
1807 | } |
1808 | ||
45d1cae3 | 1809 | vdev_raidz_generate_parity(rm); |
34dc7c2f BB |
1810 | |
1811 | for (c = 0; c < rm->rm_firstdatacol; c++) { | |
1812 | rc = &rm->rm_col[c]; | |
1813 | if (!rc->rc_tried || rc->rc_error != 0) | |
1814 | continue; | |
84c07ada | 1815 | if (abd_cmp(orig[c], rc->rc_abd) != 0) { |
428870ff | 1816 | raidz_checksum_error(zio, rc, orig[c]); |
2e528b49 | 1817 | rc->rc_error = SET_ERROR(ECKSUM); |
34dc7c2f BB |
1818 | ret++; |
1819 | } | |
84c07ada | 1820 | abd_free(orig[c]); |
34dc7c2f BB |
1821 | } |
1822 | ||
1823 | return (ret); | |
1824 | } | |
1825 | ||
34dc7c2f | 1826 | static int |
b128c09f BB |
1827 | vdev_raidz_worst_error(raidz_map_t *rm) |
1828 | { | |
1c27024e | 1829 | int error = 0; |
b128c09f | 1830 | |
1c27024e | 1831 | for (int c = 0; c < rm->rm_cols; c++) |
b128c09f BB |
1832 | error = zio_worst_error(error, rm->rm_col[c].rc_error); |
1833 | ||
1834 | return (error); | |
1835 | } | |
1836 | ||
45d1cae3 BB |
1837 | /* |
1838 | * Iterate over all combinations of bad data and attempt a reconstruction. | |
1839 | * Note that the algorithm below is non-optimal because it doesn't take into | |
1840 | * account how reconstruction is actually performed. For example, with | |
1841 | * triple-parity RAID-Z the reconstruction procedure is the same if column 4 | |
1842 | * is targeted as invalid as if columns 1 and 4 are targeted since in both | |
1843 | * cases we'd only use parity information in column 0. | |
1844 | */ | |
1845 | static int | |
1846 | vdev_raidz_combrec(zio_t *zio, int total_errors, int data_errors) | |
1847 | { | |
1848 | raidz_map_t *rm = zio->io_vsd; | |
1849 | raidz_col_t *rc; | |
84c07ada | 1850 | abd_t *orig[VDEV_RAIDZ_MAXPARITY]; |
45d1cae3 BB |
1851 | int tstore[VDEV_RAIDZ_MAXPARITY + 2]; |
1852 | int *tgts = &tstore[1]; | |
5631c038 | 1853 | int curr, next, i, c, n; |
45d1cae3 BB |
1854 | int code, ret = 0; |
1855 | ||
1856 | ASSERT(total_errors < rm->rm_firstdatacol); | |
1857 | ||
1858 | /* | |
1859 | * This simplifies one edge condition. | |
1860 | */ | |
1861 | tgts[-1] = -1; | |
1862 | ||
1863 | for (n = 1; n <= rm->rm_firstdatacol - total_errors; n++) { | |
1864 | /* | |
1865 | * Initialize the targets array by finding the first n columns | |
1866 | * that contain no error. | |
1867 | * | |
1868 | * If there were no data errors, we need to ensure that we're | |
1869 | * always explicitly attempting to reconstruct at least one | |
1870 | * data column. To do this, we simply push the highest target | |
1871 | * up into the data columns. | |
1872 | */ | |
1873 | for (c = 0, i = 0; i < n; i++) { | |
1874 | if (i == n - 1 && data_errors == 0 && | |
1875 | c < rm->rm_firstdatacol) { | |
1876 | c = rm->rm_firstdatacol; | |
1877 | } | |
1878 | ||
1879 | while (rm->rm_col[c].rc_error != 0) { | |
1880 | c++; | |
1881 | ASSERT3S(c, <, rm->rm_cols); | |
1882 | } | |
1883 | ||
1884 | tgts[i] = c++; | |
1885 | } | |
1886 | ||
1887 | /* | |
1888 | * Setting tgts[n] simplifies the other edge condition. | |
1889 | */ | |
1890 | tgts[n] = rm->rm_cols; | |
1891 | ||
1892 | /* | |
1893 | * These buffers were allocated in previous iterations. | |
1894 | */ | |
1895 | for (i = 0; i < n - 1; i++) { | |
1896 | ASSERT(orig[i] != NULL); | |
1897 | } | |
1898 | ||
84c07ada GN |
1899 | orig[n - 1] = abd_alloc_sametype(rm->rm_col[0].rc_abd, |
1900 | rm->rm_col[0].rc_size); | |
45d1cae3 | 1901 | |
5631c038 BB |
1902 | curr = 0; |
1903 | next = tgts[curr]; | |
45d1cae3 | 1904 | |
5631c038 BB |
1905 | while (curr != n) { |
1906 | tgts[curr] = next; | |
1907 | curr = 0; | |
45d1cae3 BB |
1908 | |
1909 | /* | |
1910 | * Save off the original data that we're going to | |
1911 | * attempt to reconstruct. | |
1912 | */ | |
1913 | for (i = 0; i < n; i++) { | |
1914 | ASSERT(orig[i] != NULL); | |
1915 | c = tgts[i]; | |
1916 | ASSERT3S(c, >=, 0); | |
1917 | ASSERT3S(c, <, rm->rm_cols); | |
1918 | rc = &rm->rm_col[c]; | |
84c07ada | 1919 | abd_copy(orig[i], rc->rc_abd, rc->rc_size); |
45d1cae3 BB |
1920 | } |
1921 | ||
1922 | /* | |
1923 | * Attempt a reconstruction and exit the outer loop on | |
1924 | * success. | |
1925 | */ | |
1926 | code = vdev_raidz_reconstruct(rm, tgts, n); | |
428870ff | 1927 | if (raidz_checksum_verify(zio) == 0) { |
45d1cae3 BB |
1928 | |
1929 | for (i = 0; i < n; i++) { | |
1930 | c = tgts[i]; | |
1931 | rc = &rm->rm_col[c]; | |
1932 | ASSERT(rc->rc_error == 0); | |
1933 | if (rc->rc_tried) | |
428870ff BB |
1934 | raidz_checksum_error(zio, rc, |
1935 | orig[i]); | |
2e528b49 | 1936 | rc->rc_error = SET_ERROR(ECKSUM); |
45d1cae3 BB |
1937 | } |
1938 | ||
1939 | ret = code; | |
1940 | goto done; | |
1941 | } | |
1942 | ||
1943 | /* | |
1944 | * Restore the original data. | |
1945 | */ | |
1946 | for (i = 0; i < n; i++) { | |
1947 | c = tgts[i]; | |
1948 | rc = &rm->rm_col[c]; | |
84c07ada | 1949 | abd_copy(rc->rc_abd, orig[i], rc->rc_size); |
45d1cae3 BB |
1950 | } |
1951 | ||
1952 | do { | |
1953 | /* | |
5631c038 | 1954 | * Find the next valid column after the curr |
45d1cae3 BB |
1955 | * position.. |
1956 | */ | |
5631c038 | 1957 | for (next = tgts[curr] + 1; |
45d1cae3 BB |
1958 | next < rm->rm_cols && |
1959 | rm->rm_col[next].rc_error != 0; next++) | |
1960 | continue; | |
1961 | ||
5631c038 | 1962 | ASSERT(next <= tgts[curr + 1]); |
45d1cae3 BB |
1963 | |
1964 | /* | |
1965 | * If that spot is available, we're done here. | |
1966 | */ | |
5631c038 | 1967 | if (next != tgts[curr + 1]) |
45d1cae3 BB |
1968 | break; |
1969 | ||
1970 | /* | |
1971 | * Otherwise, find the next valid column after | |
1972 | * the previous position. | |
1973 | */ | |
5631c038 | 1974 | for (c = tgts[curr - 1] + 1; |
45d1cae3 BB |
1975 | rm->rm_col[c].rc_error != 0; c++) |
1976 | continue; | |
1977 | ||
5631c038 BB |
1978 | tgts[curr] = c; |
1979 | curr++; | |
45d1cae3 | 1980 | |
5631c038 | 1981 | } while (curr != n); |
45d1cae3 BB |
1982 | } |
1983 | } | |
1984 | n--; | |
1985 | done: | |
84c07ada GN |
1986 | for (i = 0; i < n; i++) |
1987 | abd_free(orig[i]); | |
45d1cae3 BB |
1988 | |
1989 | return (ret); | |
1990 | } | |
1991 | ||
e49f1e20 WA |
1992 | /* |
1993 | * Complete an IO operation on a RAIDZ VDev | |
1994 | * | |
1995 | * Outline: | |
1996 | * - For write operations: | |
1997 | * 1. Check for errors on the child IOs. | |
1998 | * 2. Return, setting an error code if too few child VDevs were written | |
1999 | * to reconstruct the data later. Note that partial writes are | |
2000 | * considered successful if they can be reconstructed at all. | |
2001 | * - For read operations: | |
2002 | * 1. Check for errors on the child IOs. | |
2003 | * 2. If data errors occurred: | |
2004 | * a. Try to reassemble the data from the parity available. | |
2005 | * b. If we haven't yet read the parity drives, read them now. | |
2006 | * c. If all parity drives have been read but the data still doesn't | |
2007 | * reassemble with a correct checksum, then try combinatorial | |
2008 | * reconstruction. | |
2009 | * d. If that doesn't work, return an error. | |
2010 | * 3. If there were unexpected errors or this is a resilver operation, | |
2011 | * rewrite the vdevs that had errors. | |
2012 | */ | |
b128c09f | 2013 | static void |
34dc7c2f BB |
2014 | vdev_raidz_io_done(zio_t *zio) |
2015 | { | |
2016 | vdev_t *vd = zio->io_vd; | |
2017 | vdev_t *cvd; | |
2018 | raidz_map_t *rm = zio->io_vsd; | |
d4ed6673 | 2019 | raidz_col_t *rc = NULL; |
34dc7c2f BB |
2020 | int unexpected_errors = 0; |
2021 | int parity_errors = 0; | |
2022 | int parity_untried = 0; | |
2023 | int data_errors = 0; | |
b128c09f | 2024 | int total_errors = 0; |
45d1cae3 BB |
2025 | int n, c; |
2026 | int tgts[VDEV_RAIDZ_MAXPARITY]; | |
2027 | int code; | |
34dc7c2f BB |
2028 | |
2029 | ASSERT(zio->io_bp != NULL); /* XXX need to add code to enforce this */ | |
2030 | ||
34dc7c2f BB |
2031 | ASSERT(rm->rm_missingparity <= rm->rm_firstdatacol); |
2032 | ASSERT(rm->rm_missingdata <= rm->rm_cols - rm->rm_firstdatacol); | |
2033 | ||
2034 | for (c = 0; c < rm->rm_cols; c++) { | |
2035 | rc = &rm->rm_col[c]; | |
2036 | ||
34dc7c2f | 2037 | if (rc->rc_error) { |
b128c09f | 2038 | ASSERT(rc->rc_error != ECKSUM); /* child has no bp */ |
34dc7c2f BB |
2039 | |
2040 | if (c < rm->rm_firstdatacol) | |
2041 | parity_errors++; | |
2042 | else | |
2043 | data_errors++; | |
2044 | ||
2045 | if (!rc->rc_skipped) | |
2046 | unexpected_errors++; | |
2047 | ||
b128c09f | 2048 | total_errors++; |
34dc7c2f BB |
2049 | } else if (c < rm->rm_firstdatacol && !rc->rc_tried) { |
2050 | parity_untried++; | |
2051 | } | |
2052 | } | |
2053 | ||
2054 | if (zio->io_type == ZIO_TYPE_WRITE) { | |
2055 | /* | |
b128c09f BB |
2056 | * XXX -- for now, treat partial writes as a success. |
2057 | * (If we couldn't write enough columns to reconstruct | |
2058 | * the data, the I/O failed. Otherwise, good enough.) | |
2059 | * | |
2060 | * Now that we support write reallocation, it would be better | |
2061 | * to treat partial failure as real failure unless there are | |
2062 | * no non-degraded top-level vdevs left, and not update DTLs | |
2063 | * if we intend to reallocate. | |
34dc7c2f BB |
2064 | */ |
2065 | /* XXPOLICY */ | |
b128c09f BB |
2066 | if (total_errors > rm->rm_firstdatacol) |
2067 | zio->io_error = vdev_raidz_worst_error(rm); | |
34dc7c2f | 2068 | |
b128c09f | 2069 | return; |
34dc7c2f BB |
2070 | } |
2071 | ||
2072 | ASSERT(zio->io_type == ZIO_TYPE_READ); | |
2073 | /* | |
2074 | * There are three potential phases for a read: | |
2075 | * 1. produce valid data from the columns read | |
2076 | * 2. read all disks and try again | |
2077 | * 3. perform combinatorial reconstruction | |
2078 | * | |
2079 | * Each phase is progressively both more expensive and less likely to | |
2080 | * occur. If we encounter more errors than we can repair or all phases | |
2081 | * fail, we have no choice but to return an error. | |
2082 | */ | |
2083 | ||
2084 | /* | |
2085 | * If the number of errors we saw was correctable -- less than or equal | |
2086 | * to the number of parity disks read -- attempt to produce data that | |
2087 | * has a valid checksum. Naturally, this case applies in the absence of | |
2088 | * any errors. | |
2089 | */ | |
b128c09f | 2090 | if (total_errors <= rm->rm_firstdatacol - parity_untried) { |
45d1cae3 | 2091 | if (data_errors == 0) { |
428870ff | 2092 | if (raidz_checksum_verify(zio) == 0) { |
34dc7c2f BB |
2093 | /* |
2094 | * If we read parity information (unnecessarily | |
2095 | * as it happens since no reconstruction was | |
2096 | * needed) regenerate and verify the parity. | |
2097 | * We also regenerate parity when resilvering | |
2098 | * so we can write it out to the failed device | |
2099 | * later. | |
2100 | */ | |
2101 | if (parity_errors + parity_untried < | |
2102 | rm->rm_firstdatacol || | |
2103 | (zio->io_flags & ZIO_FLAG_RESILVER)) { | |
2104 | n = raidz_parity_verify(zio, rm); | |
2105 | unexpected_errors += n; | |
2106 | ASSERT(parity_errors + n <= | |
2107 | rm->rm_firstdatacol); | |
2108 | } | |
2109 | goto done; | |
2110 | } | |
45d1cae3 | 2111 | } else { |
34dc7c2f BB |
2112 | /* |
2113 | * We either attempt to read all the parity columns or | |
2114 | * none of them. If we didn't try to read parity, we | |
2115 | * wouldn't be here in the correctable case. There must | |
2116 | * also have been fewer parity errors than parity | |
2117 | * columns or, again, we wouldn't be in this code path. | |
2118 | */ | |
2119 | ASSERT(parity_untried == 0); | |
2120 | ASSERT(parity_errors < rm->rm_firstdatacol); | |
2121 | ||
2122 | /* | |
45d1cae3 | 2123 | * Identify the data columns that reported an error. |
34dc7c2f | 2124 | */ |
45d1cae3 | 2125 | n = 0; |
34dc7c2f BB |
2126 | for (c = rm->rm_firstdatacol; c < rm->rm_cols; c++) { |
2127 | rc = &rm->rm_col[c]; | |
45d1cae3 BB |
2128 | if (rc->rc_error != 0) { |
2129 | ASSERT(n < VDEV_RAIDZ_MAXPARITY); | |
2130 | tgts[n++] = c; | |
2131 | } | |
34dc7c2f | 2132 | } |
34dc7c2f | 2133 | |
45d1cae3 BB |
2134 | ASSERT(rm->rm_firstdatacol >= n); |
2135 | ||
2136 | code = vdev_raidz_reconstruct(rm, tgts, n); | |
34dc7c2f | 2137 | |
428870ff | 2138 | if (raidz_checksum_verify(zio) == 0) { |
34dc7c2f | 2139 | /* |
45d1cae3 BB |
2140 | * If we read more parity disks than were used |
2141 | * for reconstruction, confirm that the other | |
2142 | * parity disks produced correct data. This | |
2143 | * routine is suboptimal in that it regenerates | |
2144 | * the parity that we already used in addition | |
2145 | * to the parity that we're attempting to | |
2146 | * verify, but this should be a relatively | |
2147 | * uncommon case, and can be optimized if it | |
2148 | * becomes a problem. Note that we regenerate | |
2149 | * parity when resilvering so we can write it | |
2150 | * out to failed devices later. | |
34dc7c2f | 2151 | */ |
45d1cae3 | 2152 | if (parity_errors < rm->rm_firstdatacol - n || |
34dc7c2f BB |
2153 | (zio->io_flags & ZIO_FLAG_RESILVER)) { |
2154 | n = raidz_parity_verify(zio, rm); | |
2155 | unexpected_errors += n; | |
2156 | ASSERT(parity_errors + n <= | |
2157 | rm->rm_firstdatacol); | |
2158 | } | |
2159 | ||
2160 | goto done; | |
2161 | } | |
34dc7c2f BB |
2162 | } |
2163 | } | |
2164 | ||
2165 | /* | |
2166 | * This isn't a typical situation -- either we got a read error or | |
2167 | * a child silently returned bad data. Read every block so we can | |
2168 | * try again with as much data and parity as we can track down. If | |
2169 | * we've already been through once before, all children will be marked | |
2170 | * as tried so we'll proceed to combinatorial reconstruction. | |
2171 | */ | |
2172 | unexpected_errors = 1; | |
2173 | rm->rm_missingdata = 0; | |
2174 | rm->rm_missingparity = 0; | |
2175 | ||
2176 | for (c = 0; c < rm->rm_cols; c++) { | |
2177 | if (rm->rm_col[c].rc_tried) | |
2178 | continue; | |
2179 | ||
34dc7c2f BB |
2180 | zio_vdev_io_redone(zio); |
2181 | do { | |
2182 | rc = &rm->rm_col[c]; | |
2183 | if (rc->rc_tried) | |
2184 | continue; | |
2185 | zio_nowait(zio_vdev_child_io(zio, NULL, | |
2186 | vd->vdev_child[rc->rc_devidx], | |
a6255b7f | 2187 | rc->rc_offset, rc->rc_abd, rc->rc_size, |
b128c09f | 2188 | zio->io_type, zio->io_priority, 0, |
34dc7c2f BB |
2189 | vdev_raidz_child_done, rc)); |
2190 | } while (++c < rm->rm_cols); | |
34dc7c2f | 2191 | |
b128c09f | 2192 | return; |
34dc7c2f BB |
2193 | } |
2194 | ||
2195 | /* | |
2196 | * At this point we've attempted to reconstruct the data given the | |
2197 | * errors we detected, and we've attempted to read all columns. There | |
2198 | * must, therefore, be one or more additional problems -- silent errors | |
2199 | * resulting in invalid data rather than explicit I/O errors resulting | |
45d1cae3 BB |
2200 | * in absent data. We check if there is enough additional data to |
2201 | * possibly reconstruct the data and then perform combinatorial | |
2202 | * reconstruction over all possible combinations. If that fails, | |
2203 | * we're cooked. | |
34dc7c2f | 2204 | */ |
428870ff | 2205 | if (total_errors > rm->rm_firstdatacol) { |
b128c09f | 2206 | zio->io_error = vdev_raidz_worst_error(rm); |
34dc7c2f | 2207 | |
428870ff BB |
2208 | } else if (total_errors < rm->rm_firstdatacol && |
2209 | (code = vdev_raidz_combrec(zio, total_errors, data_errors)) != 0) { | |
34dc7c2f | 2210 | /* |
45d1cae3 BB |
2211 | * If we didn't use all the available parity for the |
2212 | * combinatorial reconstruction, verify that the remaining | |
2213 | * parity is correct. | |
34dc7c2f | 2214 | */ |
45d1cae3 BB |
2215 | if (code != (1 << rm->rm_firstdatacol) - 1) |
2216 | (void) raidz_parity_verify(zio, rm); | |
2217 | } else { | |
34dc7c2f | 2218 | /* |
428870ff BB |
2219 | * We're here because either: |
2220 | * | |
2221 | * total_errors == rm_first_datacol, or | |
2222 | * vdev_raidz_combrec() failed | |
2223 | * | |
2224 | * In either case, there is enough bad data to prevent | |
2225 | * reconstruction. | |
2226 | * | |
2227 | * Start checksum ereports for all children which haven't | |
2228 | * failed, and the IO wasn't speculative. | |
34dc7c2f | 2229 | */ |
2e528b49 | 2230 | zio->io_error = SET_ERROR(ECKSUM); |
34dc7c2f | 2231 | |
45d1cae3 BB |
2232 | if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { |
2233 | for (c = 0; c < rm->rm_cols; c++) { | |
2234 | rc = &rm->rm_col[c]; | |
428870ff BB |
2235 | if (rc->rc_error == 0) { |
2236 | zio_bad_cksum_t zbc; | |
2237 | zbc.zbc_has_cksum = 0; | |
2238 | zbc.zbc_injected = | |
2239 | rm->rm_ecksuminjected; | |
2240 | ||
2241 | zfs_ereport_start_checksum( | |
2242 | zio->io_spa, | |
2243 | vd->vdev_child[rc->rc_devidx], | |
b5256303 TC |
2244 | &zio->io_bookmark, zio, |
2245 | rc->rc_offset, rc->rc_size, | |
428870ff BB |
2246 | (void *)(uintptr_t)c, &zbc); |
2247 | } | |
34dc7c2f | 2248 | } |
34dc7c2f BB |
2249 | } |
2250 | } | |
2251 | ||
2252 | done: | |
2253 | zio_checksum_verified(zio); | |
2254 | ||
fb5f0bc8 | 2255 | if (zio->io_error == 0 && spa_writeable(zio->io_spa) && |
34dc7c2f | 2256 | (unexpected_errors || (zio->io_flags & ZIO_FLAG_RESILVER))) { |
34dc7c2f BB |
2257 | /* |
2258 | * Use the good data we have in hand to repair damaged children. | |
34dc7c2f | 2259 | */ |
34dc7c2f BB |
2260 | for (c = 0; c < rm->rm_cols; c++) { |
2261 | rc = &rm->rm_col[c]; | |
2262 | cvd = vd->vdev_child[rc->rc_devidx]; | |
2263 | ||
2264 | if (rc->rc_error == 0) | |
2265 | continue; | |
2266 | ||
b128c09f | 2267 | zio_nowait(zio_vdev_child_io(zio, NULL, cvd, |
a6255b7f | 2268 | rc->rc_offset, rc->rc_abd, rc->rc_size, |
e8b96c60 | 2269 | ZIO_TYPE_WRITE, ZIO_PRIORITY_ASYNC_WRITE, |
fb5f0bc8 BB |
2270 | ZIO_FLAG_IO_REPAIR | (unexpected_errors ? |
2271 | ZIO_FLAG_SELF_HEAL : 0), NULL, NULL)); | |
34dc7c2f | 2272 | } |
34dc7c2f | 2273 | } |
34dc7c2f BB |
2274 | } |
2275 | ||
2276 | static void | |
2277 | vdev_raidz_state_change(vdev_t *vd, int faulted, int degraded) | |
2278 | { | |
2279 | if (faulted > vd->vdev_nparity) | |
2280 | vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN, | |
2281 | VDEV_AUX_NO_REPLICAS); | |
2282 | else if (degraded + faulted != 0) | |
2283 | vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE); | |
2284 | else | |
2285 | vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE); | |
2286 | } | |
2287 | ||
3d6da72d IH |
2288 | /* |
2289 | * Determine if any portion of the provided block resides on a child vdev | |
2290 | * with a dirty DTL and therefore needs to be resilvered. The function | |
2291 | * assumes that at least one DTL is dirty which imples that full stripe | |
2292 | * width blocks must be resilvered. | |
2293 | */ | |
2294 | static boolean_t | |
2295 | vdev_raidz_need_resilver(vdev_t *vd, uint64_t offset, size_t psize) | |
2296 | { | |
2297 | uint64_t dcols = vd->vdev_children; | |
2298 | uint64_t nparity = vd->vdev_nparity; | |
2299 | uint64_t ashift = vd->vdev_top->vdev_ashift; | |
2300 | /* The starting RAIDZ (parent) vdev sector of the block. */ | |
2301 | uint64_t b = offset >> ashift; | |
2302 | /* The zio's size in units of the vdev's minimum sector size. */ | |
2303 | uint64_t s = ((psize - 1) >> ashift) + 1; | |
2304 | /* The first column for this stripe. */ | |
2305 | uint64_t f = b % dcols; | |
2306 | ||
2307 | if (s + nparity >= dcols) | |
2308 | return (B_TRUE); | |
2309 | ||
2310 | for (uint64_t c = 0; c < s + nparity; c++) { | |
2311 | uint64_t devidx = (f + c) % dcols; | |
2312 | vdev_t *cvd = vd->vdev_child[devidx]; | |
2313 | ||
2314 | /* | |
2315 | * dsl_scan_need_resilver() already checked vd with | |
2316 | * vdev_dtl_contains(). So here just check cvd with | |
2317 | * vdev_dtl_empty(), cheaper and a good approximation. | |
2318 | */ | |
2319 | if (!vdev_dtl_empty(cvd, DTL_PARTIAL)) | |
2320 | return (B_TRUE); | |
2321 | } | |
2322 | ||
2323 | return (B_FALSE); | |
2324 | } | |
2325 | ||
34dc7c2f BB |
2326 | vdev_ops_t vdev_raidz_ops = { |
2327 | vdev_raidz_open, | |
2328 | vdev_raidz_close, | |
34dc7c2f BB |
2329 | vdev_raidz_asize, |
2330 | vdev_raidz_io_start, | |
2331 | vdev_raidz_io_done, | |
2332 | vdev_raidz_state_change, | |
3d6da72d | 2333 | vdev_raidz_need_resilver, |
428870ff BB |
2334 | NULL, |
2335 | NULL, | |
34dc7c2f BB |
2336 | VDEV_TYPE_RAIDZ, /* name of this vdev type */ |
2337 | B_FALSE /* not a leaf vdev */ | |
2338 | }; |