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 /* 23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 24 * Copyright (c) 2012, 2020 by Delphix. All rights reserved. 25 * Copyright (c) 2016 Gvozden Nešković. All rights reserved. 26 */ 27 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> 33 #include <sys/abd.h> 34 #include <sys/fs/zfs.h> 35 #include <sys/fm/fs/zfs.h> 36 #include <sys/vdev_raidz.h> 37 #include <sys/vdev_raidz_impl.h> 38 #include <sys/vdev_draid.h> 39 40 #ifdef ZFS_DEBUG 41 #include <sys/vdev.h> /* For vdev_xlate() in vdev_raidz_io_verify() */ 42 #endif 43 44 /* 45 * Virtual device vector for RAID-Z. 46 * 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: 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: 72 * 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)). 83 * As an aside, this multiplication is derived from the error correcting 84 * primitive polynomial x^8 + x^4 + x^3 + x^2 + 1. 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 90 * than field addition). The inverse of a field element A (A^-1) is therefore 91 * A ^ (255 - 1) = A^254. 92 * 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: 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 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-1 101 * 102 * We chose 1, 2, and 4 as our generators because 1 corresponds to the trivial 103 * XOR operation, and 2 and 4 can be computed quickly and generate linearly- 104 * independent coefficients. (There are no additional coefficients that have 105 * this property which is why the uncorrected Plank method breaks down.) 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. 109 */ 110 111 #define VDEV_RAIDZ_P 0 112 #define VDEV_RAIDZ_Q 1 113 #define VDEV_RAIDZ_R 2 114 115 #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) ^ \ 129 ((mask) & 0x1d1d1d1d1d1d1d1dULL); \ 130 } 131 132 #define VDEV_RAIDZ_64MUL_4(x, mask) \ 133 { \ 134 VDEV_RAIDZ_64MUL_2((x), mask); \ 135 VDEV_RAIDZ_64MUL_2((x), mask); \ 136 } 137 138 static void 139 vdev_raidz_row_free(raidz_row_t *rr) 140 { 141 for (int c = 0; c < rr->rr_cols; c++) { 142 raidz_col_t *rc = &rr->rr_col[c]; 143 144 if (rc->rc_size != 0) 145 abd_free(rc->rc_abd); 146 if (rc->rc_orig_data != NULL) 147 abd_free(rc->rc_orig_data); 148 } 149 150 if (rr->rr_abd_empty != NULL) 151 abd_free(rr->rr_abd_empty); 152 153 kmem_free(rr, offsetof(raidz_row_t, rr_col[rr->rr_scols])); 154 } 155 156 void 157 vdev_raidz_map_free(raidz_map_t *rm) 158 { 159 for (int i = 0; i < rm->rm_nrows; i++) 160 vdev_raidz_row_free(rm->rm_row[i]); 161 162 kmem_free(rm, offsetof(raidz_map_t, rm_row[rm->rm_nrows])); 163 } 164 165 static void 166 vdev_raidz_map_free_vsd(zio_t *zio) 167 { 168 raidz_map_t *rm = zio->io_vsd; 169 170 vdev_raidz_map_free(rm); 171 } 172 173 const zio_vsd_ops_t vdev_raidz_vsd_ops = { 174 .vsd_free = vdev_raidz_map_free_vsd, 175 }; 176 177 /* 178 * Divides the IO evenly across all child vdevs; usually, dcols is 179 * the number of children in the target vdev. 180 * 181 * Avoid inlining the function to keep vdev_raidz_io_start(), which 182 * is this functions only caller, as small as possible on the stack. 183 */ 184 noinline raidz_map_t * 185 vdev_raidz_map_alloc(zio_t *zio, uint64_t ashift, uint64_t dcols, 186 uint64_t nparity) 187 { 188 raidz_row_t *rr; 189 /* The starting RAIDZ (parent) vdev sector of the block. */ 190 uint64_t b = zio->io_offset >> ashift; 191 /* The zio's size in units of the vdev's minimum sector size. */ 192 uint64_t s = zio->io_size >> ashift; 193 /* The first column for this stripe. */ 194 uint64_t f = b % dcols; 195 /* The starting byte offset on each child vdev. */ 196 uint64_t o = (b / dcols) << ashift; 197 uint64_t q, r, c, bc, col, acols, scols, coff, devidx, asize, tot; 198 199 raidz_map_t *rm = 200 kmem_zalloc(offsetof(raidz_map_t, rm_row[1]), KM_SLEEP); 201 rm->rm_nrows = 1; 202 203 /* 204 * "Quotient": The number of data sectors for this stripe on all but 205 * the "big column" child vdevs that also contain "remainder" data. 206 */ 207 q = s / (dcols - nparity); 208 209 /* 210 * "Remainder": The number of partial stripe data sectors in this I/O. 211 * This will add a sector to some, but not all, child vdevs. 212 */ 213 r = s - q * (dcols - nparity); 214 215 /* The number of "big columns" - those which contain remainder data. */ 216 bc = (r == 0 ? 0 : r + nparity); 217 218 /* 219 * The total number of data and parity sectors associated with 220 * this I/O. 221 */ 222 tot = s + nparity * (q + (r == 0 ? 0 : 1)); 223 224 /* 225 * acols: The columns that will be accessed. 226 * scols: The columns that will be accessed or skipped. 227 */ 228 if (q == 0) { 229 /* Our I/O request doesn't span all child vdevs. */ 230 acols = bc; 231 scols = MIN(dcols, roundup(bc, nparity + 1)); 232 } else { 233 acols = dcols; 234 scols = dcols; 235 } 236 237 ASSERT3U(acols, <=, scols); 238 239 rr = kmem_alloc(offsetof(raidz_row_t, rr_col[scols]), KM_SLEEP); 240 rm->rm_row[0] = rr; 241 242 rr->rr_cols = acols; 243 rr->rr_scols = scols; 244 rr->rr_bigcols = bc; 245 rr->rr_missingdata = 0; 246 rr->rr_missingparity = 0; 247 rr->rr_firstdatacol = nparity; 248 rr->rr_abd_empty = NULL; 249 rr->rr_nempty = 0; 250 #ifdef ZFS_DEBUG 251 rr->rr_offset = zio->io_offset; 252 rr->rr_size = zio->io_size; 253 #endif 254 255 asize = 0; 256 257 for (c = 0; c < scols; c++) { 258 raidz_col_t *rc = &rr->rr_col[c]; 259 col = f + c; 260 coff = o; 261 if (col >= dcols) { 262 col -= dcols; 263 coff += 1ULL << ashift; 264 } 265 rc->rc_devidx = col; 266 rc->rc_offset = coff; 267 rc->rc_abd = NULL; 268 rc->rc_orig_data = NULL; 269 rc->rc_error = 0; 270 rc->rc_tried = 0; 271 rc->rc_skipped = 0; 272 rc->rc_force_repair = 0; 273 rc->rc_allow_repair = 1; 274 rc->rc_need_orig_restore = B_FALSE; 275 276 if (c >= acols) 277 rc->rc_size = 0; 278 else if (c < bc) 279 rc->rc_size = (q + 1) << ashift; 280 else 281 rc->rc_size = q << ashift; 282 283 asize += rc->rc_size; 284 } 285 286 ASSERT3U(asize, ==, tot << ashift); 287 rm->rm_nskip = roundup(tot, nparity + 1) - tot; 288 rm->rm_skipstart = bc; 289 290 for (c = 0; c < rr->rr_firstdatacol; c++) 291 rr->rr_col[c].rc_abd = 292 abd_alloc_linear(rr->rr_col[c].rc_size, B_FALSE); 293 294 for (uint64_t off = 0; c < acols; c++) { 295 raidz_col_t *rc = &rr->rr_col[c]; 296 rc->rc_abd = abd_get_offset_struct(&rc->rc_abdstruct, 297 zio->io_abd, off, rc->rc_size); 298 off += rc->rc_size; 299 } 300 301 /* 302 * If all data stored spans all columns, there's a danger that parity 303 * will always be on the same device and, since parity isn't read 304 * during normal operation, that device's I/O bandwidth won't be 305 * used effectively. We therefore switch the parity every 1MB. 306 * 307 * ... at least that was, ostensibly, the theory. As a practical 308 * matter unless we juggle the parity between all devices evenly, we 309 * won't see any benefit. Further, occasional writes that aren't a 310 * multiple of the LCM of the number of children and the minimum 311 * stripe width are sufficient to avoid pessimal behavior. 312 * Unfortunately, this decision created an implicit on-disk format 313 * requirement that we need to support for all eternity, but only 314 * for single-parity RAID-Z. 315 * 316 * If we intend to skip a sector in the zeroth column for padding 317 * we must make sure to note this swap. We will never intend to 318 * skip the first column since at least one data and one parity 319 * column must appear in each row. 320 */ 321 ASSERT(rr->rr_cols >= 2); 322 ASSERT(rr->rr_col[0].rc_size == rr->rr_col[1].rc_size); 323 324 if (rr->rr_firstdatacol == 1 && (zio->io_offset & (1ULL << 20))) { 325 devidx = rr->rr_col[0].rc_devidx; 326 o = rr->rr_col[0].rc_offset; 327 rr->rr_col[0].rc_devidx = rr->rr_col[1].rc_devidx; 328 rr->rr_col[0].rc_offset = rr->rr_col[1].rc_offset; 329 rr->rr_col[1].rc_devidx = devidx; 330 rr->rr_col[1].rc_offset = o; 331 332 if (rm->rm_skipstart == 0) 333 rm->rm_skipstart = 1; 334 } 335 336 /* init RAIDZ parity ops */ 337 rm->rm_ops = vdev_raidz_math_get_ops(); 338 339 return (rm); 340 } 341 342 struct pqr_struct { 343 uint64_t *p; 344 uint64_t *q; 345 uint64_t *r; 346 }; 347 348 static int 349 vdev_raidz_p_func(void *buf, size_t size, void *private) 350 { 351 struct pqr_struct *pqr = private; 352 const uint64_t *src = buf; 353 int i, cnt = size / sizeof (src[0]); 354 355 ASSERT(pqr->p && !pqr->q && !pqr->r); 356 357 for (i = 0; i < cnt; i++, src++, pqr->p++) 358 *pqr->p ^= *src; 359 360 return (0); 361 } 362 363 static int 364 vdev_raidz_pq_func(void *buf, size_t size, void *private) 365 { 366 struct pqr_struct *pqr = private; 367 const uint64_t *src = buf; 368 uint64_t mask; 369 int i, cnt = size / sizeof (src[0]); 370 371 ASSERT(pqr->p && pqr->q && !pqr->r); 372 373 for (i = 0; i < cnt; i++, src++, pqr->p++, pqr->q++) { 374 *pqr->p ^= *src; 375 VDEV_RAIDZ_64MUL_2(*pqr->q, mask); 376 *pqr->q ^= *src; 377 } 378 379 return (0); 380 } 381 382 static int 383 vdev_raidz_pqr_func(void *buf, size_t size, void *private) 384 { 385 struct pqr_struct *pqr = private; 386 const uint64_t *src = buf; 387 uint64_t mask; 388 int i, cnt = size / sizeof (src[0]); 389 390 ASSERT(pqr->p && pqr->q && pqr->r); 391 392 for (i = 0; i < cnt; i++, src++, pqr->p++, pqr->q++, pqr->r++) { 393 *pqr->p ^= *src; 394 VDEV_RAIDZ_64MUL_2(*pqr->q, mask); 395 *pqr->q ^= *src; 396 VDEV_RAIDZ_64MUL_4(*pqr->r, mask); 397 *pqr->r ^= *src; 398 } 399 400 return (0); 401 } 402 403 static void 404 vdev_raidz_generate_parity_p(raidz_row_t *rr) 405 { 406 uint64_t *p = abd_to_buf(rr->rr_col[VDEV_RAIDZ_P].rc_abd); 407 408 for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { 409 abd_t *src = rr->rr_col[c].rc_abd; 410 411 if (c == rr->rr_firstdatacol) { 412 abd_copy_to_buf(p, src, rr->rr_col[c].rc_size); 413 } else { 414 struct pqr_struct pqr = { p, NULL, NULL }; 415 (void) abd_iterate_func(src, 0, rr->rr_col[c].rc_size, 416 vdev_raidz_p_func, &pqr); 417 } 418 } 419 } 420 421 static void 422 vdev_raidz_generate_parity_pq(raidz_row_t *rr) 423 { 424 uint64_t *p = abd_to_buf(rr->rr_col[VDEV_RAIDZ_P].rc_abd); 425 uint64_t *q = abd_to_buf(rr->rr_col[VDEV_RAIDZ_Q].rc_abd); 426 uint64_t pcnt = rr->rr_col[VDEV_RAIDZ_P].rc_size / sizeof (p[0]); 427 ASSERT(rr->rr_col[VDEV_RAIDZ_P].rc_size == 428 rr->rr_col[VDEV_RAIDZ_Q].rc_size); 429 430 for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { 431 abd_t *src = rr->rr_col[c].rc_abd; 432 433 uint64_t ccnt = rr->rr_col[c].rc_size / sizeof (p[0]); 434 435 if (c == rr->rr_firstdatacol) { 436 ASSERT(ccnt == pcnt || ccnt == 0); 437 abd_copy_to_buf(p, src, rr->rr_col[c].rc_size); 438 (void) memcpy(q, p, rr->rr_col[c].rc_size); 439 440 for (uint64_t i = ccnt; i < pcnt; i++) { 441 p[i] = 0; 442 q[i] = 0; 443 } 444 } else { 445 struct pqr_struct pqr = { p, q, NULL }; 446 447 ASSERT(ccnt <= pcnt); 448 (void) abd_iterate_func(src, 0, rr->rr_col[c].rc_size, 449 vdev_raidz_pq_func, &pqr); 450 451 /* 452 * Treat short columns as though they are full of 0s. 453 * Note that there's therefore nothing needed for P. 454 */ 455 uint64_t mask; 456 for (uint64_t i = ccnt; i < pcnt; i++) { 457 VDEV_RAIDZ_64MUL_2(q[i], mask); 458 } 459 } 460 } 461 } 462 463 static void 464 vdev_raidz_generate_parity_pqr(raidz_row_t *rr) 465 { 466 uint64_t *p = abd_to_buf(rr->rr_col[VDEV_RAIDZ_P].rc_abd); 467 uint64_t *q = abd_to_buf(rr->rr_col[VDEV_RAIDZ_Q].rc_abd); 468 uint64_t *r = abd_to_buf(rr->rr_col[VDEV_RAIDZ_R].rc_abd); 469 uint64_t pcnt = rr->rr_col[VDEV_RAIDZ_P].rc_size / sizeof (p[0]); 470 ASSERT(rr->rr_col[VDEV_RAIDZ_P].rc_size == 471 rr->rr_col[VDEV_RAIDZ_Q].rc_size); 472 ASSERT(rr->rr_col[VDEV_RAIDZ_P].rc_size == 473 rr->rr_col[VDEV_RAIDZ_R].rc_size); 474 475 for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { 476 abd_t *src = rr->rr_col[c].rc_abd; 477 478 uint64_t ccnt = rr->rr_col[c].rc_size / sizeof (p[0]); 479 480 if (c == rr->rr_firstdatacol) { 481 ASSERT(ccnt == pcnt || ccnt == 0); 482 abd_copy_to_buf(p, src, rr->rr_col[c].rc_size); 483 (void) memcpy(q, p, rr->rr_col[c].rc_size); 484 (void) memcpy(r, p, rr->rr_col[c].rc_size); 485 486 for (uint64_t i = ccnt; i < pcnt; i++) { 487 p[i] = 0; 488 q[i] = 0; 489 r[i] = 0; 490 } 491 } else { 492 struct pqr_struct pqr = { p, q, r }; 493 494 ASSERT(ccnt <= pcnt); 495 (void) abd_iterate_func(src, 0, rr->rr_col[c].rc_size, 496 vdev_raidz_pqr_func, &pqr); 497 498 /* 499 * Treat short columns as though they are full of 0s. 500 * Note that there's therefore nothing needed for P. 501 */ 502 uint64_t mask; 503 for (uint64_t i = ccnt; i < pcnt; i++) { 504 VDEV_RAIDZ_64MUL_2(q[i], mask); 505 VDEV_RAIDZ_64MUL_4(r[i], mask); 506 } 507 } 508 } 509 } 510 511 /* 512 * Generate RAID parity in the first virtual columns according to the number of 513 * parity columns available. 514 */ 515 void 516 vdev_raidz_generate_parity_row(raidz_map_t *rm, raidz_row_t *rr) 517 { 518 ASSERT3U(rr->rr_cols, !=, 0); 519 520 /* Generate using the new math implementation */ 521 if (vdev_raidz_math_generate(rm, rr) != RAIDZ_ORIGINAL_IMPL) 522 return; 523 524 switch (rr->rr_firstdatacol) { 525 case 1: 526 vdev_raidz_generate_parity_p(rr); 527 break; 528 case 2: 529 vdev_raidz_generate_parity_pq(rr); 530 break; 531 case 3: 532 vdev_raidz_generate_parity_pqr(rr); 533 break; 534 default: 535 cmn_err(CE_PANIC, "invalid RAID-Z configuration"); 536 } 537 } 538 539 void 540 vdev_raidz_generate_parity(raidz_map_t *rm) 541 { 542 for (int i = 0; i < rm->rm_nrows; i++) { 543 raidz_row_t *rr = rm->rm_row[i]; 544 vdev_raidz_generate_parity_row(rm, rr); 545 } 546 } 547 548 /* ARGSUSED */ 549 static int 550 vdev_raidz_reconst_p_func(void *dbuf, void *sbuf, size_t size, void *private) 551 { 552 uint64_t *dst = dbuf; 553 uint64_t *src = sbuf; 554 int cnt = size / sizeof (src[0]); 555 556 for (int i = 0; i < cnt; i++) { 557 dst[i] ^= src[i]; 558 } 559 560 return (0); 561 } 562 563 /* ARGSUSED */ 564 static int 565 vdev_raidz_reconst_q_pre_func(void *dbuf, void *sbuf, size_t size, 566 void *private) 567 { 568 uint64_t *dst = dbuf; 569 uint64_t *src = sbuf; 570 uint64_t mask; 571 int cnt = size / sizeof (dst[0]); 572 573 for (int i = 0; i < cnt; i++, dst++, src++) { 574 VDEV_RAIDZ_64MUL_2(*dst, mask); 575 *dst ^= *src; 576 } 577 578 return (0); 579 } 580 581 /* ARGSUSED */ 582 static int 583 vdev_raidz_reconst_q_pre_tail_func(void *buf, size_t size, void *private) 584 { 585 uint64_t *dst = buf; 586 uint64_t mask; 587 int cnt = size / sizeof (dst[0]); 588 589 for (int i = 0; i < cnt; i++, dst++) { 590 /* same operation as vdev_raidz_reconst_q_pre_func() on dst */ 591 VDEV_RAIDZ_64MUL_2(*dst, mask); 592 } 593 594 return (0); 595 } 596 597 struct reconst_q_struct { 598 uint64_t *q; 599 int exp; 600 }; 601 602 static int 603 vdev_raidz_reconst_q_post_func(void *buf, size_t size, void *private) 604 { 605 struct reconst_q_struct *rq = private; 606 uint64_t *dst = buf; 607 int cnt = size / sizeof (dst[0]); 608 609 for (int i = 0; i < cnt; i++, dst++, rq->q++) { 610 int j; 611 uint8_t *b; 612 613 *dst ^= *rq->q; 614 for (j = 0, b = (uint8_t *)dst; j < 8; j++, b++) { 615 *b = vdev_raidz_exp2(*b, rq->exp); 616 } 617 } 618 619 return (0); 620 } 621 622 struct reconst_pq_struct { 623 uint8_t *p; 624 uint8_t *q; 625 uint8_t *pxy; 626 uint8_t *qxy; 627 int aexp; 628 int bexp; 629 }; 630 631 static int 632 vdev_raidz_reconst_pq_func(void *xbuf, void *ybuf, size_t size, void *private) 633 { 634 struct reconst_pq_struct *rpq = private; 635 uint8_t *xd = xbuf; 636 uint8_t *yd = ybuf; 637 638 for (int i = 0; i < size; 639 i++, rpq->p++, rpq->q++, rpq->pxy++, rpq->qxy++, xd++, yd++) { 640 *xd = vdev_raidz_exp2(*rpq->p ^ *rpq->pxy, rpq->aexp) ^ 641 vdev_raidz_exp2(*rpq->q ^ *rpq->qxy, rpq->bexp); 642 *yd = *rpq->p ^ *rpq->pxy ^ *xd; 643 } 644 645 return (0); 646 } 647 648 static int 649 vdev_raidz_reconst_pq_tail_func(void *xbuf, size_t size, void *private) 650 { 651 struct reconst_pq_struct *rpq = private; 652 uint8_t *xd = xbuf; 653 654 for (int i = 0; i < size; 655 i++, rpq->p++, rpq->q++, rpq->pxy++, rpq->qxy++, xd++) { 656 /* same operation as vdev_raidz_reconst_pq_func() on xd */ 657 *xd = vdev_raidz_exp2(*rpq->p ^ *rpq->pxy, rpq->aexp) ^ 658 vdev_raidz_exp2(*rpq->q ^ *rpq->qxy, rpq->bexp); 659 } 660 661 return (0); 662 } 663 664 static void 665 vdev_raidz_reconstruct_p(raidz_row_t *rr, int *tgts, int ntgts) 666 { 667 int x = tgts[0]; 668 abd_t *dst, *src; 669 670 ASSERT3U(ntgts, ==, 1); 671 ASSERT3U(x, >=, rr->rr_firstdatacol); 672 ASSERT3U(x, <, rr->rr_cols); 673 674 ASSERT3U(rr->rr_col[x].rc_size, <=, rr->rr_col[VDEV_RAIDZ_P].rc_size); 675 676 src = rr->rr_col[VDEV_RAIDZ_P].rc_abd; 677 dst = rr->rr_col[x].rc_abd; 678 679 abd_copy_from_buf(dst, abd_to_buf(src), rr->rr_col[x].rc_size); 680 681 for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { 682 uint64_t size = MIN(rr->rr_col[x].rc_size, 683 rr->rr_col[c].rc_size); 684 685 src = rr->rr_col[c].rc_abd; 686 687 if (c == x) 688 continue; 689 690 (void) abd_iterate_func2(dst, src, 0, 0, size, 691 vdev_raidz_reconst_p_func, NULL); 692 } 693 } 694 695 static void 696 vdev_raidz_reconstruct_q(raidz_row_t *rr, int *tgts, int ntgts) 697 { 698 int x = tgts[0]; 699 int c, exp; 700 abd_t *dst, *src; 701 702 ASSERT(ntgts == 1); 703 704 ASSERT(rr->rr_col[x].rc_size <= rr->rr_col[VDEV_RAIDZ_Q].rc_size); 705 706 for (c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { 707 uint64_t size = (c == x) ? 0 : MIN(rr->rr_col[x].rc_size, 708 rr->rr_col[c].rc_size); 709 710 src = rr->rr_col[c].rc_abd; 711 dst = rr->rr_col[x].rc_abd; 712 713 if (c == rr->rr_firstdatacol) { 714 abd_copy(dst, src, size); 715 if (rr->rr_col[x].rc_size > size) { 716 abd_zero_off(dst, size, 717 rr->rr_col[x].rc_size - size); 718 } 719 } else { 720 ASSERT3U(size, <=, rr->rr_col[x].rc_size); 721 (void) abd_iterate_func2(dst, src, 0, 0, size, 722 vdev_raidz_reconst_q_pre_func, NULL); 723 (void) abd_iterate_func(dst, 724 size, rr->rr_col[x].rc_size - size, 725 vdev_raidz_reconst_q_pre_tail_func, NULL); 726 } 727 } 728 729 src = rr->rr_col[VDEV_RAIDZ_Q].rc_abd; 730 dst = rr->rr_col[x].rc_abd; 731 exp = 255 - (rr->rr_cols - 1 - x); 732 733 struct reconst_q_struct rq = { abd_to_buf(src), exp }; 734 (void) abd_iterate_func(dst, 0, rr->rr_col[x].rc_size, 735 vdev_raidz_reconst_q_post_func, &rq); 736 } 737 738 static void 739 vdev_raidz_reconstruct_pq(raidz_row_t *rr, int *tgts, int ntgts) 740 { 741 uint8_t *p, *q, *pxy, *qxy, tmp, a, b, aexp, bexp; 742 abd_t *pdata, *qdata; 743 uint64_t xsize, ysize; 744 int x = tgts[0]; 745 int y = tgts[1]; 746 abd_t *xd, *yd; 747 748 ASSERT(ntgts == 2); 749 ASSERT(x < y); 750 ASSERT(x >= rr->rr_firstdatacol); 751 ASSERT(y < rr->rr_cols); 752 753 ASSERT(rr->rr_col[x].rc_size >= rr->rr_col[y].rc_size); 754 755 /* 756 * Move the parity data aside -- we're going to compute parity as 757 * though columns x and y were full of zeros -- Pxy and Qxy. We want to 758 * reuse the parity generation mechanism without trashing the actual 759 * parity so we make those columns appear to be full of zeros by 760 * setting their lengths to zero. 761 */ 762 pdata = rr->rr_col[VDEV_RAIDZ_P].rc_abd; 763 qdata = rr->rr_col[VDEV_RAIDZ_Q].rc_abd; 764 xsize = rr->rr_col[x].rc_size; 765 ysize = rr->rr_col[y].rc_size; 766 767 rr->rr_col[VDEV_RAIDZ_P].rc_abd = 768 abd_alloc_linear(rr->rr_col[VDEV_RAIDZ_P].rc_size, B_TRUE); 769 rr->rr_col[VDEV_RAIDZ_Q].rc_abd = 770 abd_alloc_linear(rr->rr_col[VDEV_RAIDZ_Q].rc_size, B_TRUE); 771 rr->rr_col[x].rc_size = 0; 772 rr->rr_col[y].rc_size = 0; 773 774 vdev_raidz_generate_parity_pq(rr); 775 776 rr->rr_col[x].rc_size = xsize; 777 rr->rr_col[y].rc_size = ysize; 778 779 p = abd_to_buf(pdata); 780 q = abd_to_buf(qdata); 781 pxy = abd_to_buf(rr->rr_col[VDEV_RAIDZ_P].rc_abd); 782 qxy = abd_to_buf(rr->rr_col[VDEV_RAIDZ_Q].rc_abd); 783 xd = rr->rr_col[x].rc_abd; 784 yd = rr->rr_col[y].rc_abd; 785 786 /* 787 * We now have: 788 * Pxy = P + D_x + D_y 789 * Qxy = Q + 2^(ndevs - 1 - x) * D_x + 2^(ndevs - 1 - y) * D_y 790 * 791 * We can then solve for D_x: 792 * D_x = A * (P + Pxy) + B * (Q + Qxy) 793 * where 794 * A = 2^(x - y) * (2^(x - y) + 1)^-1 795 * B = 2^(ndevs - 1 - x) * (2^(x - y) + 1)^-1 796 * 797 * With D_x in hand, we can easily solve for D_y: 798 * D_y = P + Pxy + D_x 799 */ 800 801 a = vdev_raidz_pow2[255 + x - y]; 802 b = vdev_raidz_pow2[255 - (rr->rr_cols - 1 - x)]; 803 tmp = 255 - vdev_raidz_log2[a ^ 1]; 804 805 aexp = vdev_raidz_log2[vdev_raidz_exp2(a, tmp)]; 806 bexp = vdev_raidz_log2[vdev_raidz_exp2(b, tmp)]; 807 808 ASSERT3U(xsize, >=, ysize); 809 struct reconst_pq_struct rpq = { p, q, pxy, qxy, aexp, bexp }; 810 811 (void) abd_iterate_func2(xd, yd, 0, 0, ysize, 812 vdev_raidz_reconst_pq_func, &rpq); 813 (void) abd_iterate_func(xd, ysize, xsize - ysize, 814 vdev_raidz_reconst_pq_tail_func, &rpq); 815 816 abd_free(rr->rr_col[VDEV_RAIDZ_P].rc_abd); 817 abd_free(rr->rr_col[VDEV_RAIDZ_Q].rc_abd); 818 819 /* 820 * Restore the saved parity data. 821 */ 822 rr->rr_col[VDEV_RAIDZ_P].rc_abd = pdata; 823 rr->rr_col[VDEV_RAIDZ_Q].rc_abd = qdata; 824 } 825 826 /* BEGIN CSTYLED */ 827 /* 828 * In the general case of reconstruction, we must solve the system of linear 829 * equations defined by the coefficients used to generate parity as well as 830 * the contents of the data and parity disks. This can be expressed with 831 * vectors for the original data (D) and the actual data (d) and parity (p) 832 * and a matrix composed of the identity matrix (I) and a dispersal matrix (V): 833 * 834 * __ __ __ __ 835 * | | __ __ | p_0 | 836 * | V | | D_0 | | p_m-1 | 837 * | | x | : | = | d_0 | 838 * | I | | D_n-1 | | : | 839 * | | ~~ ~~ | d_n-1 | 840 * ~~ ~~ ~~ ~~ 841 * 842 * I is simply a square identity matrix of size n, and V is a vandermonde 843 * matrix defined by the coefficients we chose for the various parity columns 844 * (1, 2, 4). Note that these values were chosen both for simplicity, speedy 845 * computation as well as linear separability. 846 * 847 * __ __ __ __ 848 * | 1 .. 1 1 1 | | p_0 | 849 * | 2^n-1 .. 4 2 1 | __ __ | : | 850 * | 4^n-1 .. 16 4 1 | | D_0 | | p_m-1 | 851 * | 1 .. 0 0 0 | | D_1 | | d_0 | 852 * | 0 .. 0 0 0 | x | D_2 | = | d_1 | 853 * | : : : : | | : | | d_2 | 854 * | 0 .. 1 0 0 | | D_n-1 | | : | 855 * | 0 .. 0 1 0 | ~~ ~~ | : | 856 * | 0 .. 0 0 1 | | d_n-1 | 857 * ~~ ~~ ~~ ~~ 858 * 859 * Note that I, V, d, and p are known. To compute D, we must invert the 860 * matrix and use the known data and parity values to reconstruct the unknown 861 * data values. We begin by removing the rows in V|I and d|p that correspond 862 * to failed or missing columns; we then make V|I square (n x n) and d|p 863 * sized n by removing rows corresponding to unused parity from the bottom up 864 * to generate (V|I)' and (d|p)'. We can then generate the inverse of (V|I)' 865 * using Gauss-Jordan elimination. In the example below we use m=3 parity 866 * columns, n=8 data columns, with errors in d_1, d_2, and p_1: 867 * __ __ 868 * | 1 1 1 1 1 1 1 1 | 869 * | 128 64 32 16 8 4 2 1 | <-----+-+-- missing disks 870 * | 19 205 116 29 64 16 4 1 | / / 871 * | 1 0 0 0 0 0 0 0 | / / 872 * | 0 1 0 0 0 0 0 0 | <--' / 873 * (V|I) = | 0 0 1 0 0 0 0 0 | <---' 874 * | 0 0 0 1 0 0 0 0 | 875 * | 0 0 0 0 1 0 0 0 | 876 * | 0 0 0 0 0 1 0 0 | 877 * | 0 0 0 0 0 0 1 0 | 878 * | 0 0 0 0 0 0 0 1 | 879 * ~~ ~~ 880 * __ __ 881 * | 1 1 1 1 1 1 1 1 | 882 * | 128 64 32 16 8 4 2 1 | 883 * | 19 205 116 29 64 16 4 1 | 884 * | 1 0 0 0 0 0 0 0 | 885 * | 0 1 0 0 0 0 0 0 | 886 * (V|I)' = | 0 0 1 0 0 0 0 0 | 887 * | 0 0 0 1 0 0 0 0 | 888 * | 0 0 0 0 1 0 0 0 | 889 * | 0 0 0 0 0 1 0 0 | 890 * | 0 0 0 0 0 0 1 0 | 891 * | 0 0 0 0 0 0 0 1 | 892 * ~~ ~~ 893 * 894 * Here we employ Gauss-Jordan elimination to find the inverse of (V|I)'. We 895 * have carefully chosen the seed values 1, 2, and 4 to ensure that this 896 * matrix is not singular. 897 * __ __ 898 * | 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 | 899 * | 19 205 116 29 64 16 4 1 0 1 0 0 0 0 0 0 | 900 * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 | 901 * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 | 902 * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 | 903 * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 | 904 * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 | 905 * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 | 906 * ~~ ~~ 907 * __ __ 908 * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 | 909 * | 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 | 910 * | 19 205 116 29 64 16 4 1 0 1 0 0 0 0 0 0 | 911 * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 | 912 * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 | 913 * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 | 914 * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 | 915 * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 | 916 * ~~ ~~ 917 * __ __ 918 * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 | 919 * | 0 1 1 0 0 0 0 0 1 0 1 1 1 1 1 1 | 920 * | 0 205 116 0 0 0 0 0 0 1 19 29 64 16 4 1 | 921 * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 | 922 * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 | 923 * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 | 924 * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 | 925 * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 | 926 * ~~ ~~ 927 * __ __ 928 * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 | 929 * | 0 1 1 0 0 0 0 0 1 0 1 1 1 1 1 1 | 930 * | 0 0 185 0 0 0 0 0 205 1 222 208 141 221 201 204 | 931 * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 | 932 * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 | 933 * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 | 934 * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 | 935 * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 | 936 * ~~ ~~ 937 * __ __ 938 * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 | 939 * | 0 1 1 0 0 0 0 0 1 0 1 1 1 1 1 1 | 940 * | 0 0 1 0 0 0 0 0 166 100 4 40 158 168 216 209 | 941 * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 | 942 * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 | 943 * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 | 944 * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 | 945 * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 | 946 * ~~ ~~ 947 * __ __ 948 * | 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 | 949 * | 0 1 0 0 0 0 0 0 167 100 5 41 159 169 217 208 | 950 * | 0 0 1 0 0 0 0 0 166 100 4 40 158 168 216 209 | 951 * | 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 | 952 * | 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 | 953 * | 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 | 954 * | 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 | 955 * | 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 | 956 * ~~ ~~ 957 * __ __ 958 * | 0 0 1 0 0 0 0 0 | 959 * | 167 100 5 41 159 169 217 208 | 960 * | 166 100 4 40 158 168 216 209 | 961 * (V|I)'^-1 = | 0 0 0 1 0 0 0 0 | 962 * | 0 0 0 0 1 0 0 0 | 963 * | 0 0 0 0 0 1 0 0 | 964 * | 0 0 0 0 0 0 1 0 | 965 * | 0 0 0 0 0 0 0 1 | 966 * ~~ ~~ 967 * 968 * We can then simply compute D = (V|I)'^-1 x (d|p)' to discover the values 969 * of the missing data. 970 * 971 * As is apparent from the example above, the only non-trivial rows in the 972 * inverse matrix correspond to the data disks that we're trying to 973 * reconstruct. Indeed, those are the only rows we need as the others would 974 * only be useful for reconstructing data known or assumed to be valid. For 975 * that reason, we only build the coefficients in the rows that correspond to 976 * targeted columns. 977 */ 978 /* END CSTYLED */ 979 980 static void 981 vdev_raidz_matrix_init(raidz_row_t *rr, int n, int nmap, int *map, 982 uint8_t **rows) 983 { 984 int i, j; 985 int pow; 986 987 ASSERT(n == rr->rr_cols - rr->rr_firstdatacol); 988 989 /* 990 * Fill in the missing rows of interest. 991 */ 992 for (i = 0; i < nmap; i++) { 993 ASSERT3S(0, <=, map[i]); 994 ASSERT3S(map[i], <=, 2); 995 996 pow = map[i] * n; 997 if (pow > 255) 998 pow -= 255; 999 ASSERT(pow <= 255); 1000 1001 for (j = 0; j < n; j++) { 1002 pow -= map[i]; 1003 if (pow < 0) 1004 pow += 255; 1005 rows[i][j] = vdev_raidz_pow2[pow]; 1006 } 1007 } 1008 } 1009 1010 static void 1011 vdev_raidz_matrix_invert(raidz_row_t *rr, int n, int nmissing, int *missing, 1012 uint8_t **rows, uint8_t **invrows, const uint8_t *used) 1013 { 1014 int i, j, ii, jj; 1015 uint8_t log; 1016 1017 /* 1018 * Assert that the first nmissing entries from the array of used 1019 * columns correspond to parity columns and that subsequent entries 1020 * correspond to data columns. 1021 */ 1022 for (i = 0; i < nmissing; i++) { 1023 ASSERT3S(used[i], <, rr->rr_firstdatacol); 1024 } 1025 for (; i < n; i++) { 1026 ASSERT3S(used[i], >=, rr->rr_firstdatacol); 1027 } 1028 1029 /* 1030 * First initialize the storage where we'll compute the inverse rows. 1031 */ 1032 for (i = 0; i < nmissing; i++) { 1033 for (j = 0; j < n; j++) { 1034 invrows[i][j] = (i == j) ? 1 : 0; 1035 } 1036 } 1037 1038 /* 1039 * Subtract all trivial rows from the rows of consequence. 1040 */ 1041 for (i = 0; i < nmissing; i++) { 1042 for (j = nmissing; j < n; j++) { 1043 ASSERT3U(used[j], >=, rr->rr_firstdatacol); 1044 jj = used[j] - rr->rr_firstdatacol; 1045 ASSERT3S(jj, <, n); 1046 invrows[i][j] = rows[i][jj]; 1047 rows[i][jj] = 0; 1048 } 1049 } 1050 1051 /* 1052 * For each of the rows of interest, we must normalize it and subtract 1053 * a multiple of it from the other rows. 1054 */ 1055 for (i = 0; i < nmissing; i++) { 1056 for (j = 0; j < missing[i]; j++) { 1057 ASSERT0(rows[i][j]); 1058 } 1059 ASSERT3U(rows[i][missing[i]], !=, 0); 1060 1061 /* 1062 * Compute the inverse of the first element and multiply each 1063 * element in the row by that value. 1064 */ 1065 log = 255 - vdev_raidz_log2[rows[i][missing[i]]]; 1066 1067 for (j = 0; j < n; j++) { 1068 rows[i][j] = vdev_raidz_exp2(rows[i][j], log); 1069 invrows[i][j] = vdev_raidz_exp2(invrows[i][j], log); 1070 } 1071 1072 for (ii = 0; ii < nmissing; ii++) { 1073 if (i == ii) 1074 continue; 1075 1076 ASSERT3U(rows[ii][missing[i]], !=, 0); 1077 1078 log = vdev_raidz_log2[rows[ii][missing[i]]]; 1079 1080 for (j = 0; j < n; j++) { 1081 rows[ii][j] ^= 1082 vdev_raidz_exp2(rows[i][j], log); 1083 invrows[ii][j] ^= 1084 vdev_raidz_exp2(invrows[i][j], log); 1085 } 1086 } 1087 } 1088 1089 /* 1090 * Verify that the data that is left in the rows are properly part of 1091 * an identity matrix. 1092 */ 1093 for (i = 0; i < nmissing; i++) { 1094 for (j = 0; j < n; j++) { 1095 if (j == missing[i]) { 1096 ASSERT3U(rows[i][j], ==, 1); 1097 } else { 1098 ASSERT0(rows[i][j]); 1099 } 1100 } 1101 } 1102 } 1103 1104 static void 1105 vdev_raidz_matrix_reconstruct(raidz_row_t *rr, int n, int nmissing, 1106 int *missing, uint8_t **invrows, const uint8_t *used) 1107 { 1108 int i, j, x, cc, c; 1109 uint8_t *src; 1110 uint64_t ccount; 1111 uint8_t *dst[VDEV_RAIDZ_MAXPARITY] = { NULL }; 1112 uint64_t dcount[VDEV_RAIDZ_MAXPARITY] = { 0 }; 1113 uint8_t log = 0; 1114 uint8_t val; 1115 int ll; 1116 uint8_t *invlog[VDEV_RAIDZ_MAXPARITY]; 1117 uint8_t *p, *pp; 1118 size_t psize; 1119 1120 psize = sizeof (invlog[0][0]) * n * nmissing; 1121 p = kmem_alloc(psize, KM_SLEEP); 1122 1123 for (pp = p, i = 0; i < nmissing; i++) { 1124 invlog[i] = pp; 1125 pp += n; 1126 } 1127 1128 for (i = 0; i < nmissing; i++) { 1129 for (j = 0; j < n; j++) { 1130 ASSERT3U(invrows[i][j], !=, 0); 1131 invlog[i][j] = vdev_raidz_log2[invrows[i][j]]; 1132 } 1133 } 1134 1135 for (i = 0; i < n; i++) { 1136 c = used[i]; 1137 ASSERT3U(c, <, rr->rr_cols); 1138 1139 ccount = rr->rr_col[c].rc_size; 1140 ASSERT(ccount >= rr->rr_col[missing[0]].rc_size || i > 0); 1141 if (ccount == 0) 1142 continue; 1143 src = abd_to_buf(rr->rr_col[c].rc_abd); 1144 for (j = 0; j < nmissing; j++) { 1145 cc = missing[j] + rr->rr_firstdatacol; 1146 ASSERT3U(cc, >=, rr->rr_firstdatacol); 1147 ASSERT3U(cc, <, rr->rr_cols); 1148 ASSERT3U(cc, !=, c); 1149 1150 dcount[j] = rr->rr_col[cc].rc_size; 1151 if (dcount[j] != 0) 1152 dst[j] = abd_to_buf(rr->rr_col[cc].rc_abd); 1153 } 1154 1155 for (x = 0; x < ccount; x++, src++) { 1156 if (*src != 0) 1157 log = vdev_raidz_log2[*src]; 1158 1159 for (cc = 0; cc < nmissing; cc++) { 1160 if (x >= dcount[cc]) 1161 continue; 1162 1163 if (*src == 0) { 1164 val = 0; 1165 } else { 1166 if ((ll = log + invlog[cc][i]) >= 255) 1167 ll -= 255; 1168 val = vdev_raidz_pow2[ll]; 1169 } 1170 1171 if (i == 0) 1172 dst[cc][x] = val; 1173 else 1174 dst[cc][x] ^= val; 1175 } 1176 } 1177 } 1178 1179 kmem_free(p, psize); 1180 } 1181 1182 static void 1183 vdev_raidz_reconstruct_general(raidz_row_t *rr, int *tgts, int ntgts) 1184 { 1185 int n, i, c, t, tt; 1186 int nmissing_rows; 1187 int missing_rows[VDEV_RAIDZ_MAXPARITY]; 1188 int parity_map[VDEV_RAIDZ_MAXPARITY]; 1189 uint8_t *p, *pp; 1190 size_t psize; 1191 uint8_t *rows[VDEV_RAIDZ_MAXPARITY]; 1192 uint8_t *invrows[VDEV_RAIDZ_MAXPARITY]; 1193 uint8_t *used; 1194 1195 abd_t **bufs = NULL; 1196 1197 /* 1198 * Matrix reconstruction can't use scatter ABDs yet, so we allocate 1199 * temporary linear ABDs if any non-linear ABDs are found. 1200 */ 1201 for (i = rr->rr_firstdatacol; i < rr->rr_cols; i++) { 1202 if (!abd_is_linear(rr->rr_col[i].rc_abd)) { 1203 bufs = kmem_alloc(rr->rr_cols * sizeof (abd_t *), 1204 KM_PUSHPAGE); 1205 1206 for (c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { 1207 raidz_col_t *col = &rr->rr_col[c]; 1208 1209 bufs[c] = col->rc_abd; 1210 if (bufs[c] != NULL) { 1211 col->rc_abd = abd_alloc_linear( 1212 col->rc_size, B_TRUE); 1213 abd_copy(col->rc_abd, bufs[c], 1214 col->rc_size); 1215 } 1216 } 1217 1218 break; 1219 } 1220 } 1221 1222 n = rr->rr_cols - rr->rr_firstdatacol; 1223 1224 /* 1225 * Figure out which data columns are missing. 1226 */ 1227 nmissing_rows = 0; 1228 for (t = 0; t < ntgts; t++) { 1229 if (tgts[t] >= rr->rr_firstdatacol) { 1230 missing_rows[nmissing_rows++] = 1231 tgts[t] - rr->rr_firstdatacol; 1232 } 1233 } 1234 1235 /* 1236 * Figure out which parity columns to use to help generate the missing 1237 * data columns. 1238 */ 1239 for (tt = 0, c = 0, i = 0; i < nmissing_rows; c++) { 1240 ASSERT(tt < ntgts); 1241 ASSERT(c < rr->rr_firstdatacol); 1242 1243 /* 1244 * Skip any targeted parity columns. 1245 */ 1246 if (c == tgts[tt]) { 1247 tt++; 1248 continue; 1249 } 1250 1251 parity_map[i] = c; 1252 i++; 1253 } 1254 1255 psize = (sizeof (rows[0][0]) + sizeof (invrows[0][0])) * 1256 nmissing_rows * n + sizeof (used[0]) * n; 1257 p = kmem_alloc(psize, KM_SLEEP); 1258 1259 for (pp = p, i = 0; i < nmissing_rows; i++) { 1260 rows[i] = pp; 1261 pp += n; 1262 invrows[i] = pp; 1263 pp += n; 1264 } 1265 used = pp; 1266 1267 for (i = 0; i < nmissing_rows; i++) { 1268 used[i] = parity_map[i]; 1269 } 1270 1271 for (tt = 0, c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { 1272 if (tt < nmissing_rows && 1273 c == missing_rows[tt] + rr->rr_firstdatacol) { 1274 tt++; 1275 continue; 1276 } 1277 1278 ASSERT3S(i, <, n); 1279 used[i] = c; 1280 i++; 1281 } 1282 1283 /* 1284 * Initialize the interesting rows of the matrix. 1285 */ 1286 vdev_raidz_matrix_init(rr, n, nmissing_rows, parity_map, rows); 1287 1288 /* 1289 * Invert the matrix. 1290 */ 1291 vdev_raidz_matrix_invert(rr, n, nmissing_rows, missing_rows, rows, 1292 invrows, used); 1293 1294 /* 1295 * Reconstruct the missing data using the generated matrix. 1296 */ 1297 vdev_raidz_matrix_reconstruct(rr, n, nmissing_rows, missing_rows, 1298 invrows, used); 1299 1300 kmem_free(p, psize); 1301 1302 /* 1303 * copy back from temporary linear abds and free them 1304 */ 1305 if (bufs) { 1306 for (c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { 1307 raidz_col_t *col = &rr->rr_col[c]; 1308 1309 if (bufs[c] != NULL) { 1310 abd_copy(bufs[c], col->rc_abd, col->rc_size); 1311 abd_free(col->rc_abd); 1312 } 1313 col->rc_abd = bufs[c]; 1314 } 1315 kmem_free(bufs, rr->rr_cols * sizeof (abd_t *)); 1316 } 1317 } 1318 1319 static void 1320 vdev_raidz_reconstruct_row(raidz_map_t *rm, raidz_row_t *rr, 1321 const int *t, int nt) 1322 { 1323 int tgts[VDEV_RAIDZ_MAXPARITY], *dt; 1324 int ntgts; 1325 int i, c, ret; 1326 int nbadparity, nbaddata; 1327 int parity_valid[VDEV_RAIDZ_MAXPARITY]; 1328 1329 nbadparity = rr->rr_firstdatacol; 1330 nbaddata = rr->rr_cols - nbadparity; 1331 ntgts = 0; 1332 for (i = 0, c = 0; c < rr->rr_cols; c++) { 1333 if (c < rr->rr_firstdatacol) 1334 parity_valid[c] = B_FALSE; 1335 1336 if (i < nt && c == t[i]) { 1337 tgts[ntgts++] = c; 1338 i++; 1339 } else if (rr->rr_col[c].rc_error != 0) { 1340 tgts[ntgts++] = c; 1341 } else if (c >= rr->rr_firstdatacol) { 1342 nbaddata--; 1343 } else { 1344 parity_valid[c] = B_TRUE; 1345 nbadparity--; 1346 } 1347 } 1348 1349 ASSERT(ntgts >= nt); 1350 ASSERT(nbaddata >= 0); 1351 ASSERT(nbaddata + nbadparity == ntgts); 1352 1353 dt = &tgts[nbadparity]; 1354 1355 /* Reconstruct using the new math implementation */ 1356 ret = vdev_raidz_math_reconstruct(rm, rr, parity_valid, dt, nbaddata); 1357 if (ret != RAIDZ_ORIGINAL_IMPL) 1358 return; 1359 1360 /* 1361 * See if we can use any of our optimized reconstruction routines. 1362 */ 1363 switch (nbaddata) { 1364 case 1: 1365 if (parity_valid[VDEV_RAIDZ_P]) { 1366 vdev_raidz_reconstruct_p(rr, dt, 1); 1367 return; 1368 } 1369 1370 ASSERT(rr->rr_firstdatacol > 1); 1371 1372 if (parity_valid[VDEV_RAIDZ_Q]) { 1373 vdev_raidz_reconstruct_q(rr, dt, 1); 1374 return; 1375 } 1376 1377 ASSERT(rr->rr_firstdatacol > 2); 1378 break; 1379 1380 case 2: 1381 ASSERT(rr->rr_firstdatacol > 1); 1382 1383 if (parity_valid[VDEV_RAIDZ_P] && 1384 parity_valid[VDEV_RAIDZ_Q]) { 1385 vdev_raidz_reconstruct_pq(rr, dt, 2); 1386 return; 1387 } 1388 1389 ASSERT(rr->rr_firstdatacol > 2); 1390 1391 break; 1392 } 1393 1394 vdev_raidz_reconstruct_general(rr, tgts, ntgts); 1395 } 1396 1397 static int 1398 vdev_raidz_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize, 1399 uint64_t *logical_ashift, uint64_t *physical_ashift) 1400 { 1401 vdev_raidz_t *vdrz = vd->vdev_tsd; 1402 uint64_t nparity = vdrz->vd_nparity; 1403 int c; 1404 int lasterror = 0; 1405 int numerrors = 0; 1406 1407 ASSERT(nparity > 0); 1408 1409 if (nparity > VDEV_RAIDZ_MAXPARITY || 1410 vd->vdev_children < nparity + 1) { 1411 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL; 1412 return (SET_ERROR(EINVAL)); 1413 } 1414 1415 vdev_open_children(vd); 1416 1417 for (c = 0; c < vd->vdev_children; c++) { 1418 vdev_t *cvd = vd->vdev_child[c]; 1419 1420 if (cvd->vdev_open_error != 0) { 1421 lasterror = cvd->vdev_open_error; 1422 numerrors++; 1423 continue; 1424 } 1425 1426 *asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1; 1427 *max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1; 1428 *logical_ashift = MAX(*logical_ashift, cvd->vdev_ashift); 1429 *physical_ashift = MAX(*physical_ashift, 1430 cvd->vdev_physical_ashift); 1431 } 1432 1433 *asize *= vd->vdev_children; 1434 *max_asize *= vd->vdev_children; 1435 1436 if (numerrors > nparity) { 1437 vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS; 1438 return (lasterror); 1439 } 1440 1441 return (0); 1442 } 1443 1444 static void 1445 vdev_raidz_close(vdev_t *vd) 1446 { 1447 for (int c = 0; c < vd->vdev_children; c++) { 1448 if (vd->vdev_child[c] != NULL) 1449 vdev_close(vd->vdev_child[c]); 1450 } 1451 } 1452 1453 static uint64_t 1454 vdev_raidz_asize(vdev_t *vd, uint64_t psize) 1455 { 1456 vdev_raidz_t *vdrz = vd->vdev_tsd; 1457 uint64_t asize; 1458 uint64_t ashift = vd->vdev_top->vdev_ashift; 1459 uint64_t cols = vdrz->vd_logical_width; 1460 uint64_t nparity = vdrz->vd_nparity; 1461 1462 asize = ((psize - 1) >> ashift) + 1; 1463 asize += nparity * ((asize + cols - nparity - 1) / (cols - nparity)); 1464 asize = roundup(asize, nparity + 1) << ashift; 1465 1466 return (asize); 1467 } 1468 1469 /* 1470 * The allocatable space for a raidz vdev is N * sizeof(smallest child) 1471 * so each child must provide at least 1/Nth of its asize. 1472 */ 1473 static uint64_t 1474 vdev_raidz_min_asize(vdev_t *vd) 1475 { 1476 return ((vd->vdev_min_asize + vd->vdev_children - 1) / 1477 vd->vdev_children); 1478 } 1479 1480 void 1481 vdev_raidz_child_done(zio_t *zio) 1482 { 1483 raidz_col_t *rc = zio->io_private; 1484 1485 rc->rc_error = zio->io_error; 1486 rc->rc_tried = 1; 1487 rc->rc_skipped = 0; 1488 } 1489 1490 static void 1491 vdev_raidz_io_verify(vdev_t *vd, raidz_row_t *rr, int col) 1492 { 1493 #ifdef ZFS_DEBUG 1494 vdev_t *tvd = vd->vdev_top; 1495 1496 range_seg64_t logical_rs, physical_rs, remain_rs; 1497 logical_rs.rs_start = rr->rr_offset; 1498 logical_rs.rs_end = logical_rs.rs_start + 1499 vdev_raidz_asize(vd, rr->rr_size); 1500 1501 raidz_col_t *rc = &rr->rr_col[col]; 1502 vdev_t *cvd = vd->vdev_child[rc->rc_devidx]; 1503 1504 vdev_xlate(cvd, &logical_rs, &physical_rs, &remain_rs); 1505 ASSERT(vdev_xlate_is_empty(&remain_rs)); 1506 ASSERT3U(rc->rc_offset, ==, physical_rs.rs_start); 1507 ASSERT3U(rc->rc_offset, <, physical_rs.rs_end); 1508 /* 1509 * It would be nice to assert that rs_end is equal 1510 * to rc_offset + rc_size but there might be an 1511 * optional I/O at the end that is not accounted in 1512 * rc_size. 1513 */ 1514 if (physical_rs.rs_end > rc->rc_offset + rc->rc_size) { 1515 ASSERT3U(physical_rs.rs_end, ==, rc->rc_offset + 1516 rc->rc_size + (1 << tvd->vdev_ashift)); 1517 } else { 1518 ASSERT3U(physical_rs.rs_end, ==, rc->rc_offset + rc->rc_size); 1519 } 1520 #endif 1521 } 1522 1523 static void 1524 vdev_raidz_io_start_write(zio_t *zio, raidz_row_t *rr, uint64_t ashift) 1525 { 1526 vdev_t *vd = zio->io_vd; 1527 raidz_map_t *rm = zio->io_vsd; 1528 int c, i; 1529 1530 vdev_raidz_generate_parity_row(rm, rr); 1531 1532 for (int c = 0; c < rr->rr_cols; c++) { 1533 raidz_col_t *rc = &rr->rr_col[c]; 1534 if (rc->rc_size == 0) 1535 continue; 1536 1537 /* Verify physical to logical translation */ 1538 vdev_raidz_io_verify(vd, rr, c); 1539 1540 zio_nowait(zio_vdev_child_io(zio, NULL, 1541 vd->vdev_child[rc->rc_devidx], rc->rc_offset, 1542 rc->rc_abd, rc->rc_size, zio->io_type, zio->io_priority, 1543 0, vdev_raidz_child_done, rc)); 1544 } 1545 1546 /* 1547 * Generate optional I/Os for skip sectors to improve aggregation 1548 * contiguity. 1549 */ 1550 for (c = rm->rm_skipstart, i = 0; i < rm->rm_nskip; c++, i++) { 1551 ASSERT(c <= rr->rr_scols); 1552 if (c == rr->rr_scols) 1553 c = 0; 1554 1555 raidz_col_t *rc = &rr->rr_col[c]; 1556 vdev_t *cvd = vd->vdev_child[rc->rc_devidx]; 1557 1558 zio_nowait(zio_vdev_child_io(zio, NULL, cvd, 1559 rc->rc_offset + rc->rc_size, NULL, 1ULL << ashift, 1560 zio->io_type, zio->io_priority, 1561 ZIO_FLAG_NODATA | ZIO_FLAG_OPTIONAL, NULL, NULL)); 1562 } 1563 } 1564 1565 static void 1566 vdev_raidz_io_start_read(zio_t *zio, raidz_row_t *rr) 1567 { 1568 vdev_t *vd = zio->io_vd; 1569 1570 /* 1571 * Iterate over the columns in reverse order so that we hit the parity 1572 * last -- any errors along the way will force us to read the parity. 1573 */ 1574 for (int c = rr->rr_cols - 1; c >= 0; c--) { 1575 raidz_col_t *rc = &rr->rr_col[c]; 1576 if (rc->rc_size == 0) 1577 continue; 1578 vdev_t *cvd = vd->vdev_child[rc->rc_devidx]; 1579 if (!vdev_readable(cvd)) { 1580 if (c >= rr->rr_firstdatacol) 1581 rr->rr_missingdata++; 1582 else 1583 rr->rr_missingparity++; 1584 rc->rc_error = SET_ERROR(ENXIO); 1585 rc->rc_tried = 1; /* don't even try */ 1586 rc->rc_skipped = 1; 1587 continue; 1588 } 1589 if (vdev_dtl_contains(cvd, DTL_MISSING, zio->io_txg, 1)) { 1590 if (c >= rr->rr_firstdatacol) 1591 rr->rr_missingdata++; 1592 else 1593 rr->rr_missingparity++; 1594 rc->rc_error = SET_ERROR(ESTALE); 1595 rc->rc_skipped = 1; 1596 continue; 1597 } 1598 if (c >= rr->rr_firstdatacol || rr->rr_missingdata > 0 || 1599 (zio->io_flags & (ZIO_FLAG_SCRUB | ZIO_FLAG_RESILVER))) { 1600 zio_nowait(zio_vdev_child_io(zio, NULL, cvd, 1601 rc->rc_offset, rc->rc_abd, rc->rc_size, 1602 zio->io_type, zio->io_priority, 0, 1603 vdev_raidz_child_done, rc)); 1604 } 1605 } 1606 } 1607 1608 /* 1609 * Start an IO operation on a RAIDZ VDev 1610 * 1611 * Outline: 1612 * - For write operations: 1613 * 1. Generate the parity data 1614 * 2. Create child zio write operations to each column's vdev, for both 1615 * data and parity. 1616 * 3. If the column skips any sectors for padding, create optional dummy 1617 * write zio children for those areas to improve aggregation continuity. 1618 * - For read operations: 1619 * 1. Create child zio read operations to each data column's vdev to read 1620 * the range of data required for zio. 1621 * 2. If this is a scrub or resilver operation, or if any of the data 1622 * vdevs have had errors, then create zio read operations to the parity 1623 * columns' VDevs as well. 1624 */ 1625 static void 1626 vdev_raidz_io_start(zio_t *zio) 1627 { 1628 vdev_t *vd = zio->io_vd; 1629 vdev_t *tvd = vd->vdev_top; 1630 vdev_raidz_t *vdrz = vd->vdev_tsd; 1631 1632 raidz_map_t *rm = vdev_raidz_map_alloc(zio, tvd->vdev_ashift, 1633 vdrz->vd_logical_width, vdrz->vd_nparity); 1634 zio->io_vsd = rm; 1635 zio->io_vsd_ops = &vdev_raidz_vsd_ops; 1636 1637 /* 1638 * Until raidz expansion is implemented all maps for a raidz vdev 1639 * contain a single row. 1640 */ 1641 ASSERT3U(rm->rm_nrows, ==, 1); 1642 raidz_row_t *rr = rm->rm_row[0]; 1643 1644 if (zio->io_type == ZIO_TYPE_WRITE) { 1645 vdev_raidz_io_start_write(zio, rr, tvd->vdev_ashift); 1646 } else { 1647 ASSERT(zio->io_type == ZIO_TYPE_READ); 1648 vdev_raidz_io_start_read(zio, rr); 1649 } 1650 1651 zio_execute(zio); 1652 } 1653 1654 /* 1655 * Report a checksum error for a child of a RAID-Z device. 1656 */ 1657 static void 1658 raidz_checksum_error(zio_t *zio, raidz_col_t *rc, abd_t *bad_data) 1659 { 1660 vdev_t *vd = zio->io_vd->vdev_child[rc->rc_devidx]; 1661 1662 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE) && 1663 zio->io_priority != ZIO_PRIORITY_REBUILD) { 1664 zio_bad_cksum_t zbc; 1665 raidz_map_t *rm = zio->io_vsd; 1666 1667 zbc.zbc_has_cksum = 0; 1668 zbc.zbc_injected = rm->rm_ecksuminjected; 1669 1670 (void) zfs_ereport_post_checksum(zio->io_spa, vd, 1671 &zio->io_bookmark, zio, rc->rc_offset, rc->rc_size, 1672 rc->rc_abd, bad_data, &zbc); 1673 mutex_enter(&vd->vdev_stat_lock); 1674 vd->vdev_stat.vs_checksum_errors++; 1675 mutex_exit(&vd->vdev_stat_lock); 1676 } 1677 } 1678 1679 /* 1680 * We keep track of whether or not there were any injected errors, so that 1681 * any ereports we generate can note it. 1682 */ 1683 static int 1684 raidz_checksum_verify(zio_t *zio) 1685 { 1686 zio_bad_cksum_t zbc; 1687 raidz_map_t *rm = zio->io_vsd; 1688 1689 bzero(&zbc, sizeof (zio_bad_cksum_t)); 1690 1691 int ret = zio_checksum_error(zio, &zbc); 1692 if (ret != 0 && zbc.zbc_injected != 0) 1693 rm->rm_ecksuminjected = 1; 1694 1695 return (ret); 1696 } 1697 1698 /* 1699 * Generate the parity from the data columns. If we tried and were able to 1700 * read the parity without error, verify that the generated parity matches the 1701 * data we read. If it doesn't, we fire off a checksum error. Return the 1702 * number of such failures. 1703 */ 1704 static int 1705 raidz_parity_verify(zio_t *zio, raidz_row_t *rr) 1706 { 1707 abd_t *orig[VDEV_RAIDZ_MAXPARITY]; 1708 int c, ret = 0; 1709 raidz_map_t *rm = zio->io_vsd; 1710 raidz_col_t *rc; 1711 1712 blkptr_t *bp = zio->io_bp; 1713 enum zio_checksum checksum = (bp == NULL ? zio->io_prop.zp_checksum : 1714 (BP_IS_GANG(bp) ? ZIO_CHECKSUM_GANG_HEADER : BP_GET_CHECKSUM(bp))); 1715 1716 if (checksum == ZIO_CHECKSUM_NOPARITY) 1717 return (ret); 1718 1719 for (c = 0; c < rr->rr_firstdatacol; c++) { 1720 rc = &rr->rr_col[c]; 1721 if (!rc->rc_tried || rc->rc_error != 0) 1722 continue; 1723 1724 orig[c] = abd_alloc_sametype(rc->rc_abd, rc->rc_size); 1725 abd_copy(orig[c], rc->rc_abd, rc->rc_size); 1726 } 1727 1728 /* 1729 * Regenerates parity even for !tried||rc_error!=0 columns. This 1730 * isn't harmful but it does have the side effect of fixing stuff 1731 * we didn't realize was necessary (i.e. even if we return 0). 1732 */ 1733 vdev_raidz_generate_parity_row(rm, rr); 1734 1735 for (c = 0; c < rr->rr_firstdatacol; c++) { 1736 rc = &rr->rr_col[c]; 1737 1738 if (!rc->rc_tried || rc->rc_error != 0) 1739 continue; 1740 1741 if (abd_cmp(orig[c], rc->rc_abd) != 0) { 1742 raidz_checksum_error(zio, rc, orig[c]); 1743 rc->rc_error = SET_ERROR(ECKSUM); 1744 ret++; 1745 } 1746 abd_free(orig[c]); 1747 } 1748 1749 return (ret); 1750 } 1751 1752 static int 1753 vdev_raidz_worst_error(raidz_row_t *rr) 1754 { 1755 int error = 0; 1756 1757 for (int c = 0; c < rr->rr_cols; c++) 1758 error = zio_worst_error(error, rr->rr_col[c].rc_error); 1759 1760 return (error); 1761 } 1762 1763 static void 1764 vdev_raidz_io_done_verified(zio_t *zio, raidz_row_t *rr) 1765 { 1766 int unexpected_errors = 0; 1767 int parity_errors = 0; 1768 int parity_untried = 0; 1769 int data_errors = 0; 1770 1771 ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ); 1772 1773 for (int c = 0; c < rr->rr_cols; c++) { 1774 raidz_col_t *rc = &rr->rr_col[c]; 1775 1776 if (rc->rc_error) { 1777 if (c < rr->rr_firstdatacol) 1778 parity_errors++; 1779 else 1780 data_errors++; 1781 1782 if (!rc->rc_skipped) 1783 unexpected_errors++; 1784 } else if (c < rr->rr_firstdatacol && !rc->rc_tried) { 1785 parity_untried++; 1786 } 1787 } 1788 1789 /* 1790 * If we read more parity disks than were used for 1791 * reconstruction, confirm that the other parity disks produced 1792 * correct data. 1793 * 1794 * Note that we also regenerate parity when resilvering so we 1795 * can write it out to failed devices later. 1796 */ 1797 if (parity_errors + parity_untried < 1798 rr->rr_firstdatacol - data_errors || 1799 (zio->io_flags & ZIO_FLAG_RESILVER)) { 1800 int n = raidz_parity_verify(zio, rr); 1801 unexpected_errors += n; 1802 ASSERT3U(parity_errors + n, <=, rr->rr_firstdatacol); 1803 } 1804 1805 if (zio->io_error == 0 && spa_writeable(zio->io_spa) && 1806 (unexpected_errors > 0 || (zio->io_flags & ZIO_FLAG_RESILVER))) { 1807 /* 1808 * Use the good data we have in hand to repair damaged children. 1809 */ 1810 for (int c = 0; c < rr->rr_cols; c++) { 1811 raidz_col_t *rc = &rr->rr_col[c]; 1812 vdev_t *vd = zio->io_vd; 1813 vdev_t *cvd = vd->vdev_child[rc->rc_devidx]; 1814 1815 if (!rc->rc_allow_repair) { 1816 continue; 1817 } else if (!rc->rc_force_repair && 1818 (rc->rc_error == 0 || rc->rc_size == 0)) { 1819 continue; 1820 } 1821 1822 zio_nowait(zio_vdev_child_io(zio, NULL, cvd, 1823 rc->rc_offset, rc->rc_abd, rc->rc_size, 1824 ZIO_TYPE_WRITE, 1825 zio->io_priority == ZIO_PRIORITY_REBUILD ? 1826 ZIO_PRIORITY_REBUILD : ZIO_PRIORITY_ASYNC_WRITE, 1827 ZIO_FLAG_IO_REPAIR | (unexpected_errors ? 1828 ZIO_FLAG_SELF_HEAL : 0), NULL, NULL)); 1829 } 1830 } 1831 } 1832 1833 static void 1834 raidz_restore_orig_data(raidz_map_t *rm) 1835 { 1836 for (int i = 0; i < rm->rm_nrows; i++) { 1837 raidz_row_t *rr = rm->rm_row[i]; 1838 for (int c = 0; c < rr->rr_cols; c++) { 1839 raidz_col_t *rc = &rr->rr_col[c]; 1840 if (rc->rc_need_orig_restore) { 1841 abd_copy(rc->rc_abd, 1842 rc->rc_orig_data, rc->rc_size); 1843 rc->rc_need_orig_restore = B_FALSE; 1844 } 1845 } 1846 } 1847 } 1848 1849 /* 1850 * returns EINVAL if reconstruction of the block will not be possible 1851 * returns ECKSUM if this specific reconstruction failed 1852 * returns 0 on successful reconstruction 1853 */ 1854 static int 1855 raidz_reconstruct(zio_t *zio, int *ltgts, int ntgts, int nparity) 1856 { 1857 raidz_map_t *rm = zio->io_vsd; 1858 1859 /* Reconstruct each row */ 1860 for (int r = 0; r < rm->rm_nrows; r++) { 1861 raidz_row_t *rr = rm->rm_row[r]; 1862 int my_tgts[VDEV_RAIDZ_MAXPARITY]; /* value is child id */ 1863 int t = 0; 1864 int dead = 0; 1865 int dead_data = 0; 1866 1867 for (int c = 0; c < rr->rr_cols; c++) { 1868 raidz_col_t *rc = &rr->rr_col[c]; 1869 ASSERT0(rc->rc_need_orig_restore); 1870 if (rc->rc_error != 0) { 1871 dead++; 1872 if (c >= nparity) 1873 dead_data++; 1874 continue; 1875 } 1876 if (rc->rc_size == 0) 1877 continue; 1878 for (int lt = 0; lt < ntgts; lt++) { 1879 if (rc->rc_devidx == ltgts[lt]) { 1880 if (rc->rc_orig_data == NULL) { 1881 rc->rc_orig_data = 1882 abd_alloc_linear( 1883 rc->rc_size, B_TRUE); 1884 abd_copy(rc->rc_orig_data, 1885 rc->rc_abd, rc->rc_size); 1886 } 1887 rc->rc_need_orig_restore = B_TRUE; 1888 1889 dead++; 1890 if (c >= nparity) 1891 dead_data++; 1892 my_tgts[t++] = c; 1893 break; 1894 } 1895 } 1896 } 1897 if (dead > nparity) { 1898 /* reconstruction not possible */ 1899 raidz_restore_orig_data(rm); 1900 return (EINVAL); 1901 } 1902 if (dead_data > 0) 1903 vdev_raidz_reconstruct_row(rm, rr, my_tgts, t); 1904 } 1905 1906 /* Check for success */ 1907 if (raidz_checksum_verify(zio) == 0) { 1908 1909 /* Reconstruction succeeded - report errors */ 1910 for (int i = 0; i < rm->rm_nrows; i++) { 1911 raidz_row_t *rr = rm->rm_row[i]; 1912 1913 for (int c = 0; c < rr->rr_cols; c++) { 1914 raidz_col_t *rc = &rr->rr_col[c]; 1915 if (rc->rc_need_orig_restore) { 1916 /* 1917 * Note: if this is a parity column, 1918 * we don't really know if it's wrong. 1919 * We need to let 1920 * vdev_raidz_io_done_verified() check 1921 * it, and if we set rc_error, it will 1922 * think that it is a "known" error 1923 * that doesn't need to be checked 1924 * or corrected. 1925 */ 1926 if (rc->rc_error == 0 && 1927 c >= rr->rr_firstdatacol) { 1928 raidz_checksum_error(zio, 1929 rc, rc->rc_orig_data); 1930 rc->rc_error = 1931 SET_ERROR(ECKSUM); 1932 } 1933 rc->rc_need_orig_restore = B_FALSE; 1934 } 1935 } 1936 1937 vdev_raidz_io_done_verified(zio, rr); 1938 } 1939 1940 zio_checksum_verified(zio); 1941 1942 return (0); 1943 } 1944 1945 /* Reconstruction failed - restore original data */ 1946 raidz_restore_orig_data(rm); 1947 return (ECKSUM); 1948 } 1949 1950 /* 1951 * Iterate over all combinations of N bad vdevs and attempt a reconstruction. 1952 * Note that the algorithm below is non-optimal because it doesn't take into 1953 * account how reconstruction is actually performed. For example, with 1954 * triple-parity RAID-Z the reconstruction procedure is the same if column 4 1955 * is targeted as invalid as if columns 1 and 4 are targeted since in both 1956 * cases we'd only use parity information in column 0. 1957 * 1958 * The order that we find the various possible combinations of failed 1959 * disks is dictated by these rules: 1960 * - Examine each "slot" (the "i" in tgts[i]) 1961 * - Try to increment this slot (tgts[i] = tgts[i] + 1) 1962 * - if we can't increment because it runs into the next slot, 1963 * reset our slot to the minimum, and examine the next slot 1964 * 1965 * For example, with a 6-wide RAIDZ3, and no known errors (so we have to choose 1966 * 3 columns to reconstruct), we will generate the following sequence: 1967 * 1968 * STATE ACTION 1969 * 0 1 2 special case: skip since these are all parity 1970 * 0 1 3 first slot: reset to 0; middle slot: increment to 2 1971 * 0 2 3 first slot: increment to 1 1972 * 1 2 3 first: reset to 0; middle: reset to 1; last: increment to 4 1973 * 0 1 4 first: reset to 0; middle: increment to 2 1974 * 0 2 4 first: increment to 1 1975 * 1 2 4 first: reset to 0; middle: increment to 3 1976 * 0 3 4 first: increment to 1 1977 * 1 3 4 first: increment to 2 1978 * 2 3 4 first: reset to 0; middle: reset to 1; last: increment to 5 1979 * 0 1 5 first: reset to 0; middle: increment to 2 1980 * 0 2 5 first: increment to 1 1981 * 1 2 5 first: reset to 0; middle: increment to 3 1982 * 0 3 5 first: increment to 1 1983 * 1 3 5 first: increment to 2 1984 * 2 3 5 first: reset to 0; middle: increment to 4 1985 * 0 4 5 first: increment to 1 1986 * 1 4 5 first: increment to 2 1987 * 2 4 5 first: increment to 3 1988 * 3 4 5 done 1989 * 1990 * This strategy works for dRAID but is less efficient when there are a large 1991 * number of child vdevs and therefore permutations to check. Furthermore, 1992 * since the raidz_map_t rows likely do not overlap reconstruction would be 1993 * possible as long as there are no more than nparity data errors per row. 1994 * These additional permutations are not currently checked but could be as 1995 * a future improvement. 1996 */ 1997 static int 1998 vdev_raidz_combrec(zio_t *zio) 1999 { 2000 int nparity = vdev_get_nparity(zio->io_vd); 2001 raidz_map_t *rm = zio->io_vsd; 2002 2003 /* Check if there's enough data to attempt reconstrution. */ 2004 for (int i = 0; i < rm->rm_nrows; i++) { 2005 raidz_row_t *rr = rm->rm_row[i]; 2006 int total_errors = 0; 2007 2008 for (int c = 0; c < rr->rr_cols; c++) { 2009 if (rr->rr_col[c].rc_error) 2010 total_errors++; 2011 } 2012 2013 if (total_errors > nparity) 2014 return (vdev_raidz_worst_error(rr)); 2015 } 2016 2017 for (int num_failures = 1; num_failures <= nparity; num_failures++) { 2018 int tstore[VDEV_RAIDZ_MAXPARITY + 2]; 2019 int *ltgts = &tstore[1]; /* value is logical child ID */ 2020 2021 /* Determine number of logical children, n */ 2022 int n = zio->io_vd->vdev_children; 2023 2024 ASSERT3U(num_failures, <=, nparity); 2025 ASSERT3U(num_failures, <=, VDEV_RAIDZ_MAXPARITY); 2026 2027 /* Handle corner cases in combrec logic */ 2028 ltgts[-1] = -1; 2029 for (int i = 0; i < num_failures; i++) { 2030 ltgts[i] = i; 2031 } 2032 ltgts[num_failures] = n; 2033 2034 for (;;) { 2035 int err = raidz_reconstruct(zio, ltgts, num_failures, 2036 nparity); 2037 if (err == EINVAL) { 2038 /* 2039 * Reconstruction not possible with this # 2040 * failures; try more failures. 2041 */ 2042 break; 2043 } else if (err == 0) 2044 return (0); 2045 2046 /* Compute next targets to try */ 2047 for (int t = 0; ; t++) { 2048 ASSERT3U(t, <, num_failures); 2049 ltgts[t]++; 2050 if (ltgts[t] == n) { 2051 /* try more failures */ 2052 ASSERT3U(t, ==, num_failures - 1); 2053 break; 2054 } 2055 2056 ASSERT3U(ltgts[t], <, n); 2057 ASSERT3U(ltgts[t], <=, ltgts[t + 1]); 2058 2059 /* 2060 * If that spot is available, we're done here. 2061 * Try the next combination. 2062 */ 2063 if (ltgts[t] != ltgts[t + 1]) 2064 break; 2065 2066 /* 2067 * Otherwise, reset this tgt to the minimum, 2068 * and move on to the next tgt. 2069 */ 2070 ltgts[t] = ltgts[t - 1] + 1; 2071 ASSERT3U(ltgts[t], ==, t); 2072 } 2073 2074 /* Increase the number of failures and keep trying. */ 2075 if (ltgts[num_failures - 1] == n) 2076 break; 2077 } 2078 } 2079 2080 return (ECKSUM); 2081 } 2082 2083 void 2084 vdev_raidz_reconstruct(raidz_map_t *rm, const int *t, int nt) 2085 { 2086 for (uint64_t row = 0; row < rm->rm_nrows; row++) { 2087 raidz_row_t *rr = rm->rm_row[row]; 2088 vdev_raidz_reconstruct_row(rm, rr, t, nt); 2089 } 2090 } 2091 2092 /* 2093 * Complete a write IO operation on a RAIDZ VDev 2094 * 2095 * Outline: 2096 * 1. Check for errors on the child IOs. 2097 * 2. Return, setting an error code if too few child VDevs were written 2098 * to reconstruct the data later. Note that partial writes are 2099 * considered successful if they can be reconstructed at all. 2100 */ 2101 static void 2102 vdev_raidz_io_done_write_impl(zio_t *zio, raidz_row_t *rr) 2103 { 2104 int total_errors = 0; 2105 2106 ASSERT3U(rr->rr_missingparity, <=, rr->rr_firstdatacol); 2107 ASSERT3U(rr->rr_missingdata, <=, rr->rr_cols - rr->rr_firstdatacol); 2108 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); 2109 2110 for (int c = 0; c < rr->rr_cols; c++) { 2111 raidz_col_t *rc = &rr->rr_col[c]; 2112 2113 if (rc->rc_error) { 2114 ASSERT(rc->rc_error != ECKSUM); /* child has no bp */ 2115 2116 total_errors++; 2117 } 2118 } 2119 2120 /* 2121 * Treat partial writes as a success. If we couldn't write enough 2122 * columns to reconstruct the data, the I/O failed. Otherwise, 2123 * good enough. 2124 * 2125 * Now that we support write reallocation, it would be better 2126 * to treat partial failure as real failure unless there are 2127 * no non-degraded top-level vdevs left, and not update DTLs 2128 * if we intend to reallocate. 2129 */ 2130 if (total_errors > rr->rr_firstdatacol) { 2131 zio->io_error = zio_worst_error(zio->io_error, 2132 vdev_raidz_worst_error(rr)); 2133 } 2134 } 2135 2136 static void 2137 vdev_raidz_io_done_reconstruct_known_missing(zio_t *zio, raidz_map_t *rm, 2138 raidz_row_t *rr) 2139 { 2140 int parity_errors = 0; 2141 int parity_untried = 0; 2142 int data_errors = 0; 2143 int total_errors = 0; 2144 2145 ASSERT3U(rr->rr_missingparity, <=, rr->rr_firstdatacol); 2146 ASSERT3U(rr->rr_missingdata, <=, rr->rr_cols - rr->rr_firstdatacol); 2147 ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ); 2148 2149 for (int c = 0; c < rr->rr_cols; c++) { 2150 raidz_col_t *rc = &rr->rr_col[c]; 2151 2152 if (rc->rc_error) { 2153 ASSERT(rc->rc_error != ECKSUM); /* child has no bp */ 2154 2155 if (c < rr->rr_firstdatacol) 2156 parity_errors++; 2157 else 2158 data_errors++; 2159 2160 total_errors++; 2161 } else if (c < rr->rr_firstdatacol && !rc->rc_tried) { 2162 parity_untried++; 2163 } 2164 } 2165 2166 /* 2167 * If there were data errors and the number of errors we saw was 2168 * correctable -- less than or equal to the number of parity disks read 2169 * -- reconstruct based on the missing data. 2170 */ 2171 if (data_errors != 0 && 2172 total_errors <= rr->rr_firstdatacol - parity_untried) { 2173 /* 2174 * We either attempt to read all the parity columns or 2175 * none of them. If we didn't try to read parity, we 2176 * wouldn't be here in the correctable case. There must 2177 * also have been fewer parity errors than parity 2178 * columns or, again, we wouldn't be in this code path. 2179 */ 2180 ASSERT(parity_untried == 0); 2181 ASSERT(parity_errors < rr->rr_firstdatacol); 2182 2183 /* 2184 * Identify the data columns that reported an error. 2185 */ 2186 int n = 0; 2187 int tgts[VDEV_RAIDZ_MAXPARITY]; 2188 for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { 2189 raidz_col_t *rc = &rr->rr_col[c]; 2190 if (rc->rc_error != 0) { 2191 ASSERT(n < VDEV_RAIDZ_MAXPARITY); 2192 tgts[n++] = c; 2193 } 2194 } 2195 2196 ASSERT(rr->rr_firstdatacol >= n); 2197 2198 vdev_raidz_reconstruct_row(rm, rr, tgts, n); 2199 } 2200 } 2201 2202 /* 2203 * Return the number of reads issued. 2204 */ 2205 static int 2206 vdev_raidz_read_all(zio_t *zio, raidz_row_t *rr) 2207 { 2208 vdev_t *vd = zio->io_vd; 2209 int nread = 0; 2210 2211 rr->rr_missingdata = 0; 2212 rr->rr_missingparity = 0; 2213 2214 /* 2215 * If this rows contains empty sectors which are not required 2216 * for a normal read then allocate an ABD for them now so they 2217 * may be read, verified, and any needed repairs performed. 2218 */ 2219 if (rr->rr_nempty && rr->rr_abd_empty == NULL) 2220 vdev_draid_map_alloc_empty(zio, rr); 2221 2222 for (int c = 0; c < rr->rr_cols; c++) { 2223 raidz_col_t *rc = &rr->rr_col[c]; 2224 if (rc->rc_tried || rc->rc_size == 0) 2225 continue; 2226 2227 zio_nowait(zio_vdev_child_io(zio, NULL, 2228 vd->vdev_child[rc->rc_devidx], 2229 rc->rc_offset, rc->rc_abd, rc->rc_size, 2230 zio->io_type, zio->io_priority, 0, 2231 vdev_raidz_child_done, rc)); 2232 nread++; 2233 } 2234 return (nread); 2235 } 2236 2237 /* 2238 * We're here because either there were too many errors to even attempt 2239 * reconstruction (total_errors == rm_first_datacol), or vdev_*_combrec() 2240 * failed. In either case, there is enough bad data to prevent reconstruction. 2241 * Start checksum ereports for all children which haven't failed. 2242 */ 2243 static void 2244 vdev_raidz_io_done_unrecoverable(zio_t *zio) 2245 { 2246 raidz_map_t *rm = zio->io_vsd; 2247 2248 for (int i = 0; i < rm->rm_nrows; i++) { 2249 raidz_row_t *rr = rm->rm_row[i]; 2250 2251 for (int c = 0; c < rr->rr_cols; c++) { 2252 raidz_col_t *rc = &rr->rr_col[c]; 2253 vdev_t *cvd = zio->io_vd->vdev_child[rc->rc_devidx]; 2254 2255 if (rc->rc_error != 0) 2256 continue; 2257 2258 zio_bad_cksum_t zbc; 2259 zbc.zbc_has_cksum = 0; 2260 zbc.zbc_injected = rm->rm_ecksuminjected; 2261 2262 (void) zfs_ereport_start_checksum(zio->io_spa, 2263 cvd, &zio->io_bookmark, zio, rc->rc_offset, 2264 rc->rc_size, &zbc); 2265 mutex_enter(&cvd->vdev_stat_lock); 2266 cvd->vdev_stat.vs_checksum_errors++; 2267 mutex_exit(&cvd->vdev_stat_lock); 2268 } 2269 } 2270 } 2271 2272 void 2273 vdev_raidz_io_done(zio_t *zio) 2274 { 2275 raidz_map_t *rm = zio->io_vsd; 2276 2277 if (zio->io_type == ZIO_TYPE_WRITE) { 2278 for (int i = 0; i < rm->rm_nrows; i++) { 2279 vdev_raidz_io_done_write_impl(zio, rm->rm_row[i]); 2280 } 2281 } else { 2282 for (int i = 0; i < rm->rm_nrows; i++) { 2283 raidz_row_t *rr = rm->rm_row[i]; 2284 vdev_raidz_io_done_reconstruct_known_missing(zio, 2285 rm, rr); 2286 } 2287 2288 if (raidz_checksum_verify(zio) == 0) { 2289 for (int i = 0; i < rm->rm_nrows; i++) { 2290 raidz_row_t *rr = rm->rm_row[i]; 2291 vdev_raidz_io_done_verified(zio, rr); 2292 } 2293 zio_checksum_verified(zio); 2294 } else { 2295 /* 2296 * A sequential resilver has no checksum which makes 2297 * combinatoral reconstruction impossible. This code 2298 * path is unreachable since raidz_checksum_verify() 2299 * has no checksum to verify and must succeed. 2300 */ 2301 ASSERT3U(zio->io_priority, !=, ZIO_PRIORITY_REBUILD); 2302 2303 /* 2304 * This isn't a typical situation -- either we got a 2305 * read error or a child silently returned bad data. 2306 * Read every block so we can try again with as much 2307 * data and parity as we can track down. If we've 2308 * already been through once before, all children will 2309 * be marked as tried so we'll proceed to combinatorial 2310 * reconstruction. 2311 */ 2312 int nread = 0; 2313 for (int i = 0; i < rm->rm_nrows; i++) { 2314 nread += vdev_raidz_read_all(zio, 2315 rm->rm_row[i]); 2316 } 2317 if (nread != 0) { 2318 /* 2319 * Normally our stage is VDEV_IO_DONE, but if 2320 * we've already called redone(), it will have 2321 * changed to VDEV_IO_START, in which case we 2322 * don't want to call redone() again. 2323 */ 2324 if (zio->io_stage != ZIO_STAGE_VDEV_IO_START) 2325 zio_vdev_io_redone(zio); 2326 return; 2327 } 2328 2329 zio->io_error = vdev_raidz_combrec(zio); 2330 if (zio->io_error == ECKSUM && 2331 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { 2332 vdev_raidz_io_done_unrecoverable(zio); 2333 } 2334 } 2335 } 2336 } 2337 2338 static void 2339 vdev_raidz_state_change(vdev_t *vd, int faulted, int degraded) 2340 { 2341 vdev_raidz_t *vdrz = vd->vdev_tsd; 2342 if (faulted > vdrz->vd_nparity) 2343 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN, 2344 VDEV_AUX_NO_REPLICAS); 2345 else if (degraded + faulted != 0) 2346 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE); 2347 else 2348 vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE); 2349 } 2350 2351 /* 2352 * Determine if any portion of the provided block resides on a child vdev 2353 * with a dirty DTL and therefore needs to be resilvered. The function 2354 * assumes that at least one DTL is dirty which implies that full stripe 2355 * width blocks must be resilvered. 2356 */ 2357 static boolean_t 2358 vdev_raidz_need_resilver(vdev_t *vd, const dva_t *dva, size_t psize, 2359 uint64_t phys_birth) 2360 { 2361 vdev_raidz_t *vdrz = vd->vdev_tsd; 2362 uint64_t dcols = vd->vdev_children; 2363 uint64_t nparity = vdrz->vd_nparity; 2364 uint64_t ashift = vd->vdev_top->vdev_ashift; 2365 /* The starting RAIDZ (parent) vdev sector of the block. */ 2366 uint64_t b = DVA_GET_OFFSET(dva) >> ashift; 2367 /* The zio's size in units of the vdev's minimum sector size. */ 2368 uint64_t s = ((psize - 1) >> ashift) + 1; 2369 /* The first column for this stripe. */ 2370 uint64_t f = b % dcols; 2371 2372 /* Unreachable by sequential resilver. */ 2373 ASSERT3U(phys_birth, !=, TXG_UNKNOWN); 2374 2375 if (!vdev_dtl_contains(vd, DTL_PARTIAL, phys_birth, 1)) 2376 return (B_FALSE); 2377 2378 if (s + nparity >= dcols) 2379 return (B_TRUE); 2380 2381 for (uint64_t c = 0; c < s + nparity; c++) { 2382 uint64_t devidx = (f + c) % dcols; 2383 vdev_t *cvd = vd->vdev_child[devidx]; 2384 2385 /* 2386 * dsl_scan_need_resilver() already checked vd with 2387 * vdev_dtl_contains(). So here just check cvd with 2388 * vdev_dtl_empty(), cheaper and a good approximation. 2389 */ 2390 if (!vdev_dtl_empty(cvd, DTL_PARTIAL)) 2391 return (B_TRUE); 2392 } 2393 2394 return (B_FALSE); 2395 } 2396 2397 static void 2398 vdev_raidz_xlate(vdev_t *cvd, const range_seg64_t *logical_rs, 2399 range_seg64_t *physical_rs, range_seg64_t *remain_rs) 2400 { 2401 vdev_t *raidvd = cvd->vdev_parent; 2402 ASSERT(raidvd->vdev_ops == &vdev_raidz_ops); 2403 2404 uint64_t width = raidvd->vdev_children; 2405 uint64_t tgt_col = cvd->vdev_id; 2406 uint64_t ashift = raidvd->vdev_top->vdev_ashift; 2407 2408 /* make sure the offsets are block-aligned */ 2409 ASSERT0(logical_rs->rs_start % (1 << ashift)); 2410 ASSERT0(logical_rs->rs_end % (1 << ashift)); 2411 uint64_t b_start = logical_rs->rs_start >> ashift; 2412 uint64_t b_end = logical_rs->rs_end >> ashift; 2413 2414 uint64_t start_row = 0; 2415 if (b_start > tgt_col) /* avoid underflow */ 2416 start_row = ((b_start - tgt_col - 1) / width) + 1; 2417 2418 uint64_t end_row = 0; 2419 if (b_end > tgt_col) 2420 end_row = ((b_end - tgt_col - 1) / width) + 1; 2421 2422 physical_rs->rs_start = start_row << ashift; 2423 physical_rs->rs_end = end_row << ashift; 2424 2425 ASSERT3U(physical_rs->rs_start, <=, logical_rs->rs_start); 2426 ASSERT3U(physical_rs->rs_end - physical_rs->rs_start, <=, 2427 logical_rs->rs_end - logical_rs->rs_start); 2428 } 2429 2430 /* 2431 * Initialize private RAIDZ specific fields from the nvlist. 2432 */ 2433 static int 2434 vdev_raidz_init(spa_t *spa, nvlist_t *nv, void **tsd) 2435 { 2436 vdev_raidz_t *vdrz; 2437 uint64_t nparity; 2438 2439 uint_t children; 2440 nvlist_t **child; 2441 int error = nvlist_lookup_nvlist_array(nv, 2442 ZPOOL_CONFIG_CHILDREN, &child, &children); 2443 if (error != 0) 2444 return (SET_ERROR(EINVAL)); 2445 2446 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY, &nparity) == 0) { 2447 if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY) 2448 return (SET_ERROR(EINVAL)); 2449 2450 /* 2451 * Previous versions could only support 1 or 2 parity 2452 * device. 2453 */ 2454 if (nparity > 1 && spa_version(spa) < SPA_VERSION_RAIDZ2) 2455 return (SET_ERROR(EINVAL)); 2456 else if (nparity > 2 && spa_version(spa) < SPA_VERSION_RAIDZ3) 2457 return (SET_ERROR(EINVAL)); 2458 } else { 2459 /* 2460 * We require the parity to be specified for SPAs that 2461 * support multiple parity levels. 2462 */ 2463 if (spa_version(spa) >= SPA_VERSION_RAIDZ2) 2464 return (SET_ERROR(EINVAL)); 2465 2466 /* 2467 * Otherwise, we default to 1 parity device for RAID-Z. 2468 */ 2469 nparity = 1; 2470 } 2471 2472 vdrz = kmem_zalloc(sizeof (*vdrz), KM_SLEEP); 2473 vdrz->vd_logical_width = children; 2474 vdrz->vd_nparity = nparity; 2475 2476 *tsd = vdrz; 2477 2478 return (0); 2479 } 2480 2481 static void 2482 vdev_raidz_fini(vdev_t *vd) 2483 { 2484 kmem_free(vd->vdev_tsd, sizeof (vdev_raidz_t)); 2485 } 2486 2487 /* 2488 * Add RAIDZ specific fields to the config nvlist. 2489 */ 2490 static void 2491 vdev_raidz_config_generate(vdev_t *vd, nvlist_t *nv) 2492 { 2493 ASSERT3P(vd->vdev_ops, ==, &vdev_raidz_ops); 2494 vdev_raidz_t *vdrz = vd->vdev_tsd; 2495 2496 /* 2497 * Make sure someone hasn't managed to sneak a fancy new vdev 2498 * into a crufty old storage pool. 2499 */ 2500 ASSERT(vdrz->vd_nparity == 1 || 2501 (vdrz->vd_nparity <= 2 && 2502 spa_version(vd->vdev_spa) >= SPA_VERSION_RAIDZ2) || 2503 (vdrz->vd_nparity <= 3 && 2504 spa_version(vd->vdev_spa) >= SPA_VERSION_RAIDZ3)); 2505 2506 /* 2507 * Note that we'll add these even on storage pools where they 2508 * aren't strictly required -- older software will just ignore 2509 * it. 2510 */ 2511 fnvlist_add_uint64(nv, ZPOOL_CONFIG_NPARITY, vdrz->vd_nparity); 2512 } 2513 2514 static uint64_t 2515 vdev_raidz_nparity(vdev_t *vd) 2516 { 2517 vdev_raidz_t *vdrz = vd->vdev_tsd; 2518 return (vdrz->vd_nparity); 2519 } 2520 2521 static uint64_t 2522 vdev_raidz_ndisks(vdev_t *vd) 2523 { 2524 return (vd->vdev_children); 2525 } 2526 2527 vdev_ops_t vdev_raidz_ops = { 2528 .vdev_op_init = vdev_raidz_init, 2529 .vdev_op_fini = vdev_raidz_fini, 2530 .vdev_op_open = vdev_raidz_open, 2531 .vdev_op_close = vdev_raidz_close, 2532 .vdev_op_asize = vdev_raidz_asize, 2533 .vdev_op_min_asize = vdev_raidz_min_asize, 2534 .vdev_op_min_alloc = NULL, 2535 .vdev_op_io_start = vdev_raidz_io_start, 2536 .vdev_op_io_done = vdev_raidz_io_done, 2537 .vdev_op_state_change = vdev_raidz_state_change, 2538 .vdev_op_need_resilver = vdev_raidz_need_resilver, 2539 .vdev_op_hold = NULL, 2540 .vdev_op_rele = NULL, 2541 .vdev_op_remap = NULL, 2542 .vdev_op_xlate = vdev_raidz_xlate, 2543 .vdev_op_rebuild_asize = NULL, 2544 .vdev_op_metaslab_init = NULL, 2545 .vdev_op_config_generate = vdev_raidz_config_generate, 2546 .vdev_op_nparity = vdev_raidz_nparity, 2547 .vdev_op_ndisks = vdev_raidz_ndisks, 2548 .vdev_op_type = VDEV_TYPE_RAIDZ, /* name of this vdev type */ 2549 .vdev_op_leaf = B_FALSE /* not a leaf vdev */ 2550 }; 2551