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 * Copyright (c) 2018 Intel Corporation. 23 * Copyright (c) 2020 by Lawrence Livermore National Security, LLC. 24 */ 25 26 #include <sys/zfs_context.h> 27 #include <sys/spa.h> 28 #include <sys/spa_impl.h> 29 #include <sys/vdev_impl.h> 30 #include <sys/vdev_draid.h> 31 #include <sys/vdev_raidz.h> 32 #include <sys/vdev_rebuild.h> 33 #include <sys/abd.h> 34 #include <sys/zio.h> 35 #include <sys/nvpair.h> 36 #include <sys/zio_checksum.h> 37 #include <sys/fs/zfs.h> 38 #include <sys/fm/fs/zfs.h> 39 #include <zfs_fletcher.h> 40 41 #ifdef ZFS_DEBUG 42 #include <sys/vdev.h> /* For vdev_xlate() in vdev_draid_io_verify() */ 43 #endif 44 45 /* 46 * dRAID is a distributed spare implementation for ZFS. A dRAID vdev is 47 * comprised of multiple raidz redundancy groups which are spread over the 48 * dRAID children. To ensure an even distribution, and avoid hot spots, a 49 * permutation mapping is applied to the order of the dRAID children. 50 * This mixing effectively distributes the parity columns evenly over all 51 * of the disks in the dRAID. 52 * 53 * This is beneficial because it means when resilvering all of the disks 54 * can participate thereby increasing the available IOPs and bandwidth. 55 * Furthermore, by reserving a small fraction of each child's total capacity 56 * virtual distributed spare disks can be created. These spares similarly 57 * benefit from the performance gains of spanning all of the children. The 58 * consequence of which is that resilvering to a distributed spare can 59 * substantially reduce the time required to restore full parity to pool 60 * with a failed disks. 61 * 62 * === dRAID group layout === 63 * 64 * First, let's define a "row" in the configuration to be a 16M chunk from 65 * each physical drive at the same offset. This is the minimum allowable 66 * size since it must be possible to store a full 16M block when there is 67 * only a single data column. Next, we define a "group" to be a set of 68 * sequential disks containing both the parity and data columns. We allow 69 * groups to span multiple rows in order to align any group size to any 70 * number of physical drives. Finally, a "slice" is comprised of the rows 71 * which contain the target number of groups. The permutation mappings 72 * are applied in a round robin fashion to each slice. 73 * 74 * Given D+P drives in a group (including parity drives) and C-S physical 75 * drives (not including the spare drives), we can distribute the groups 76 * across R rows without remainder by selecting the least common multiple 77 * of D+P and C-S as the number of groups; i.e. ngroups = LCM(D+P, C-S). 78 * 79 * In the example below, there are C=14 physical drives in the configuration 80 * with S=2 drives worth of spare capacity. Each group has a width of 9 81 * which includes D=8 data and P=1 parity drive. There are 4 groups and 82 * 3 rows per slice. Each group has a size of 144M (16M * 9) and a slice 83 * size is 576M (144M * 4). When allocating from a dRAID each group is 84 * filled before moving on to the next as show in slice0 below. 85 * 86 * data disks (8 data + 1 parity) spares (2) 87 * +===+===+===+===+===+===+===+===+===+===+===+===+===+===+ 88 * ^ | 2 | 6 | 1 | 11| 4 | 0 | 7 | 10| 8 | 9 | 13| 5 | 12| 3 | device map 0 89 * | +===+===+===+===+===+===+===+===+===+===+===+===+===+===+ 90 * | | group 0 | group 1..| | 91 * | +-----------------------------------+-----------+-------| 92 * | | 0 1 2 3 4 5 6 7 8 | 36 37 38| | r 93 * | | 9 10 11 12 13 14 15 16 17| 45 46 47| | o 94 * | | 18 19 20 21 22 23 24 25 26| 54 55 56| | w 95 * | 27 28 29 30 31 32 33 34 35| 63 64 65| | 0 96 * s +-----------------------+-----------------------+-------+ 97 * l | ..group 1 | group 2.. | | 98 * i +-----------------------+-----------------------+-------+ 99 * c | 39 40 41 42 43 44| 72 73 74 75 76 77| | r 100 * e | 48 49 50 51 52 53| 81 82 83 84 85 86| | o 101 * 0 | 57 58 59 60 61 62| 90 91 92 93 94 95| | w 102 * | 66 67 68 69 70 71| 99 100 101 102 103 104| | 1 103 * | +-----------+-----------+-----------------------+-------+ 104 * | |..group 2 | group 3 | | 105 * | +-----------+-----------+-----------------------+-------+ 106 * | | 78 79 80|108 109 110 111 112 113 114 115 116| | r 107 * | | 87 88 89|117 118 119 120 121 122 123 124 125| | o 108 * | | 96 97 98|126 127 128 129 130 131 132 133 134| | w 109 * v |105 106 107|135 136 137 138 139 140 141 142 143| | 2 110 * +===+===+===+===+===+===+===+===+===+===+===+===+===+===+ 111 * | 9 | 11| 12| 2 | 4 | 1 | 3 | 0 | 10| 13| 8 | 5 | 6 | 7 | device map 1 112 * s +===+===+===+===+===+===+===+===+===+===+===+===+===+===+ 113 * l | group 4 | group 5..| | row 3 114 * i +-----------------------+-----------+-----------+-------| 115 * c | ..group 5 | group 6.. | | row 4 116 * e +-----------+-----------+-----------------------+-------+ 117 * 1 |..group 6 | group 7 | | row 5 118 * +===+===+===+===+===+===+===+===+===+===+===+===+===+===+ 119 * | 3 | 5 | 10| 8 | 6 | 11| 12| 0 | 2 | 4 | 7 | 1 | 9 | 13| device map 2 120 * s +===+===+===+===+===+===+===+===+===+===+===+===+===+===+ 121 * l | group 8 | group 9..| | row 6 122 * i +-----------------------------------------------+-------| 123 * c | ..group 9 | group 10.. | | row 7 124 * e +-----------------------+-----------------------+-------+ 125 * 2 |..group 10 | group 11 | | row 8 126 * +-----------+-----------------------------------+-------+ 127 * 128 * This layout has several advantages over requiring that each row contain 129 * a whole number of groups. 130 * 131 * 1. The group count is not a relevant parameter when defining a dRAID 132 * layout. Only the group width is needed, and *all* groups will have 133 * the desired size. 134 * 135 * 2. All possible group widths (<= physical disk count) can be supported. 136 * 137 * 3. The logic within vdev_draid.c is simplified when the group width is 138 * the same for all groups (although some of the logic around computing 139 * permutation numbers and drive offsets is more complicated). 140 * 141 * N.B. The following array describes all valid dRAID permutation maps. 142 * Each row is used to generate a permutation map for a different number 143 * of children from a unique seed. The seeds were generated and carefully 144 * evaluated by the 'draid' utility in order to provide balanced mappings. 145 * In addition to the seed a checksum of the in-memory mapping is stored 146 * for verification. 147 * 148 * The imbalance ratio of a given failure (e.g. 5 disks wide, child 3 failed, 149 * with a given permutation map) is the ratio of the amounts of I/O that will 150 * be sent to the least and most busy disks when resilvering. The average 151 * imbalance ratio (of a given number of disks and permutation map) is the 152 * average of the ratios of all possible single and double disk failures. 153 * 154 * In order to achieve a low imbalance ratio the number of permutations in 155 * the mapping must be significantly larger than the number of children. 156 * For dRAID the number of permutations has been limited to 512 to minimize 157 * the map size. This does result in a gradually increasing imbalance ratio 158 * as seen in the table below. Increasing the number of permutations for 159 * larger child counts would reduce the imbalance ratio. However, in practice 160 * when there are a large number of children each child is responsible for 161 * fewer total IOs so it's less of a concern. 162 * 163 * Note these values are hard coded and must never be changed. Existing 164 * pools depend on the same mapping always being generated in order to 165 * read and write from the correct locations. Any change would make 166 * existing pools completely inaccessible. 167 */ 168 static const draid_map_t draid_maps[VDEV_DRAID_MAX_MAPS] = { 169 { 2, 256, 0x89ef3dabbcc7de37, 0x00000000433d433d }, /* 1.000 */ 170 { 3, 256, 0x89a57f3de98121b4, 0x00000000bcd8b7b5 }, /* 1.000 */ 171 { 4, 256, 0xc9ea9ec82340c885, 0x00000001819d7c69 }, /* 1.000 */ 172 { 5, 256, 0xf46733b7f4d47dfd, 0x00000002a1648d74 }, /* 1.010 */ 173 { 6, 256, 0x88c3c62d8585b362, 0x00000003d3b0c2c4 }, /* 1.031 */ 174 { 7, 256, 0x3a65d809b4d1b9d5, 0x000000055c4183ee }, /* 1.043 */ 175 { 8, 256, 0xe98930e3c5d2e90a, 0x00000006edfb0329 }, /* 1.059 */ 176 { 9, 256, 0x5a5430036b982ccb, 0x00000008ceaf6934 }, /* 1.056 */ 177 { 10, 256, 0x92bf389e9eadac74, 0x0000000b26668c09 }, /* 1.072 */ 178 { 11, 256, 0x74ccebf1dcf3ae80, 0x0000000dd691358c }, /* 1.083 */ 179 { 12, 256, 0x8847e41a1a9f5671, 0x00000010a0c63c8e }, /* 1.097 */ 180 { 13, 256, 0x7481b56debf0e637, 0x0000001424121fe4 }, /* 1.100 */ 181 { 14, 256, 0x559b8c44065f8967, 0x00000016ab2ff079 }, /* 1.121 */ 182 { 15, 256, 0x34c49545a2ee7f01, 0x0000001a6028efd6 }, /* 1.103 */ 183 { 16, 256, 0xb85f4fa81a7698f7, 0x0000001e95ff5e66 }, /* 1.111 */ 184 { 17, 256, 0x6353e47b7e47aba0, 0x00000021a81fa0fe }, /* 1.133 */ 185 { 18, 256, 0xaa549746b1cbb81c, 0x00000026f02494c9 }, /* 1.131 */ 186 { 19, 256, 0x892e343f2f31d690, 0x00000029eb392835 }, /* 1.130 */ 187 { 20, 256, 0x76914824db98cc3f, 0x0000003004f31a7c }, /* 1.141 */ 188 { 21, 256, 0x4b3cbabf9cfb1d0f, 0x00000036363a2408 }, /* 1.139 */ 189 { 22, 256, 0xf45c77abb4f035d4, 0x00000038dd0f3e84 }, /* 1.150 */ 190 { 23, 256, 0x5e18bd7f3fd4baf4, 0x0000003f0660391f }, /* 1.174 */ 191 { 24, 256, 0xa7b3a4d285d6503b, 0x000000443dfc9ff6 }, /* 1.168 */ 192 { 25, 256, 0x56ac7dd967521f5a, 0x0000004b03a87eb7 }, /* 1.180 */ 193 { 26, 256, 0x3a42dfda4eb880f7, 0x000000522c719bba }, /* 1.226 */ 194 { 27, 256, 0xd200d2fc6b54bf60, 0x0000005760b4fdf5 }, /* 1.228 */ 195 { 28, 256, 0xc52605bbd486c546, 0x0000005e00d8f74c }, /* 1.217 */ 196 { 29, 256, 0xc761779e63cd762f, 0x00000067be3cd85c }, /* 1.239 */ 197 { 30, 256, 0xca577b1e07f85ca5, 0x0000006f5517f3e4 }, /* 1.238 */ 198 { 31, 256, 0xfd50a593c518b3d4, 0x0000007370e7778f }, /* 1.273 */ 199 { 32, 512, 0xc6c87ba5b042650b, 0x000000f7eb08a156 }, /* 1.191 */ 200 { 33, 512, 0xc3880d0c9d458304, 0x0000010734b5d160 }, /* 1.199 */ 201 { 34, 512, 0xe920927e4d8b2c97, 0x00000118c1edbce0 }, /* 1.195 */ 202 { 35, 512, 0x8da7fcda87bde316, 0x0000012a3e9f9110 }, /* 1.201 */ 203 { 36, 512, 0xcf09937491514a29, 0x0000013bd6a24bef }, /* 1.194 */ 204 { 37, 512, 0x9b5abbf345cbd7cc, 0x0000014b9d90fac3 }, /* 1.237 */ 205 { 38, 512, 0x506312a44668d6a9, 0x0000015e1b5f6148 }, /* 1.242 */ 206 { 39, 512, 0x71659ede62b4755f, 0x00000173ef029bcd }, /* 1.231 */ 207 { 40, 512, 0xa7fde73fb74cf2d7, 0x000001866fb72748 }, /* 1.233 */ 208 { 41, 512, 0x19e8b461a1dea1d3, 0x000001a046f76b23 }, /* 1.271 */ 209 { 42, 512, 0x031c9b868cc3e976, 0x000001afa64c49d3 }, /* 1.263 */ 210 { 43, 512, 0xbaa5125faa781854, 0x000001c76789e278 }, /* 1.270 */ 211 { 44, 512, 0x4ed55052550d721b, 0x000001d800ccd8eb }, /* 1.281 */ 212 { 45, 512, 0x0fd63ddbdff90677, 0x000001f08ad59ed2 }, /* 1.282 */ 213 { 46, 512, 0x36d66546de7fdd6f, 0x000002016f09574b }, /* 1.286 */ 214 { 47, 512, 0x99f997e7eafb69d7, 0x0000021e42e47cb6 }, /* 1.329 */ 215 { 48, 512, 0xbecd9c2571312c5d, 0x000002320fe2872b }, /* 1.286 */ 216 { 49, 512, 0xd97371329e488a32, 0x0000024cd73f2ca7 }, /* 1.322 */ 217 { 50, 512, 0x30e9b136670749ee, 0x000002681c83b0e0 }, /* 1.335 */ 218 { 51, 512, 0x11ad6bc8f47aaeb4, 0x0000027e9261b5d5 }, /* 1.305 */ 219 { 52, 512, 0x68e445300af432c1, 0x0000029aa0eb7dbf }, /* 1.330 */ 220 { 53, 512, 0x910fb561657ea98c, 0x000002b3dca04853 }, /* 1.365 */ 221 { 54, 512, 0xd619693d8ce5e7a5, 0x000002cc280e9c97 }, /* 1.334 */ 222 { 55, 512, 0x24e281f564dbb60a, 0x000002e9fa842713 }, /* 1.364 */ 223 { 56, 512, 0x947a7d3bdaab44c5, 0x000003046680f72e }, /* 1.374 */ 224 { 57, 512, 0x2d44fec9c093e0de, 0x00000324198ba810 }, /* 1.363 */ 225 { 58, 512, 0x87743c272d29bb4c, 0x0000033ec48c9ac9 }, /* 1.401 */ 226 { 59, 512, 0x96aa3b6f67f5d923, 0x0000034faead902c }, /* 1.392 */ 227 { 60, 512, 0x94a4f1faf520b0d3, 0x0000037d713ab005 }, /* 1.360 */ 228 { 61, 512, 0xb13ed3a272f711a2, 0x00000397368f3cbd }, /* 1.396 */ 229 { 62, 512, 0x3b1b11805fa4a64a, 0x000003b8a5e2840c }, /* 1.453 */ 230 { 63, 512, 0x4c74caad9172ba71, 0x000003d4be280290 }, /* 1.437 */ 231 { 64, 512, 0x035ff643923dd29e, 0x000003fad6c355e1 }, /* 1.402 */ 232 { 65, 512, 0x768e9171b11abd3c, 0x0000040eb07fed20 }, /* 1.459 */ 233 { 66, 512, 0x75880e6f78a13ddd, 0x000004433d6acf14 }, /* 1.423 */ 234 { 67, 512, 0x910b9714f698a877, 0x00000451ea65d5db }, /* 1.447 */ 235 { 68, 512, 0x87f5db6f9fdcf5c7, 0x000004732169e3f7 }, /* 1.450 */ 236 { 69, 512, 0x836d4968fbaa3706, 0x000004954068a380 }, /* 1.455 */ 237 { 70, 512, 0xc567d73a036421ab, 0x000004bd7cb7bd3d }, /* 1.463 */ 238 { 71, 512, 0x619df40f240b8fed, 0x000004e376c2e972 }, /* 1.463 */ 239 { 72, 512, 0x42763a680d5bed8e, 0x000005084275c680 }, /* 1.452 */ 240 { 73, 512, 0x5866f064b3230431, 0x0000052906f2c9ab }, /* 1.498 */ 241 { 74, 512, 0x9fa08548b1621a44, 0x0000054708019247 }, /* 1.526 */ 242 { 75, 512, 0xb6053078ce0fc303, 0x00000572cc5c72b0 }, /* 1.491 */ 243 { 76, 512, 0x4a7aad7bf3890923, 0x0000058e987bc8e9 }, /* 1.470 */ 244 { 77, 512, 0xe165613fd75b5a53, 0x000005c20473a211 }, /* 1.527 */ 245 { 78, 512, 0x3ff154ac878163a6, 0x000005d659194bf3 }, /* 1.509 */ 246 { 79, 512, 0x24b93ade0aa8a532, 0x0000060a201c4f8e }, /* 1.569 */ 247 { 80, 512, 0xc18e2d14cd9bb554, 0x0000062c55cfe48c }, /* 1.555 */ 248 { 81, 512, 0x98cc78302feb58b6, 0x0000066656a07194 }, /* 1.509 */ 249 { 82, 512, 0xc6c5fd5a2abc0543, 0x0000067cff94fbf8 }, /* 1.596 */ 250 { 83, 512, 0xa7962f514acbba21, 0x000006ab7b5afa2e }, /* 1.568 */ 251 { 84, 512, 0xba02545069ddc6dc, 0x000006d19861364f }, /* 1.541 */ 252 { 85, 512, 0x447c73192c35073e, 0x000006fce315ce35 }, /* 1.623 */ 253 { 86, 512, 0x48beef9e2d42b0c2, 0x00000720a8e38b6b }, /* 1.620 */ 254 { 87, 512, 0x4874cf98541a35e0, 0x00000758382a2273 }, /* 1.597 */ 255 { 88, 512, 0xad4cf8333a31127a, 0x00000781e1651b1b }, /* 1.575 */ 256 { 89, 512, 0x47ae4859d57888c1, 0x000007b27edbe5bc }, /* 1.627 */ 257 { 90, 512, 0x06f7723cfe5d1891, 0x000007dc2a96d8eb }, /* 1.596 */ 258 { 91, 512, 0xd4e44218d660576d, 0x0000080ac46f02d5 }, /* 1.622 */ 259 { 92, 512, 0x7066702b0d5be1f2, 0x00000832c96d154e }, /* 1.695 */ 260 { 93, 512, 0x011209b4f9e11fb9, 0x0000085eefda104c }, /* 1.605 */ 261 { 94, 512, 0x47ffba30a0b35708, 0x00000899badc32dc }, /* 1.625 */ 262 { 95, 512, 0x1a95a6ac4538aaa8, 0x000008b6b69a42b2 }, /* 1.687 */ 263 { 96, 512, 0xbda2b239bb2008eb, 0x000008f22d2de38a }, /* 1.621 */ 264 { 97, 512, 0x7ffa0bea90355c6c, 0x0000092e5b23b816 }, /* 1.699 */ 265 { 98, 512, 0x1d56ba34be426795, 0x0000094f482e5d1b }, /* 1.688 */ 266 { 99, 512, 0x0aa89d45c502e93d, 0x00000977d94a98ce }, /* 1.642 */ 267 { 100, 512, 0x54369449f6857774, 0x000009c06c9b34cc }, /* 1.683 */ 268 { 101, 512, 0xf7d4dd8445b46765, 0x000009e5dc542259 }, /* 1.755 */ 269 { 102, 512, 0xfa8866312f169469, 0x00000a16b54eae93 }, /* 1.692 */ 270 { 103, 512, 0xd8a5aea08aef3ff9, 0x00000a381d2cbfe7 }, /* 1.747 */ 271 { 104, 512, 0x66bcd2c3d5f9ef0e, 0x00000a8191817be7 }, /* 1.751 */ 272 { 105, 512, 0x3fb13a47a012ec81, 0x00000ab562b9a254 }, /* 1.751 */ 273 { 106, 512, 0x43100f01c9e5e3ca, 0x00000aeee84c185f }, /* 1.726 */ 274 { 107, 512, 0xca09c50ccee2d054, 0x00000b1c359c047d }, /* 1.788 */ 275 { 108, 512, 0xd7176732ac503f9b, 0x00000b578bc52a73 }, /* 1.740 */ 276 { 109, 512, 0xed206e51f8d9422d, 0x00000b8083e0d960 }, /* 1.780 */ 277 { 110, 512, 0x17ead5dc6ba0dcd6, 0x00000bcfb1a32ca8 }, /* 1.836 */ 278 { 111, 512, 0x5f1dc21e38a969eb, 0x00000c0171becdd6 }, /* 1.778 */ 279 { 112, 512, 0xddaa973de33ec528, 0x00000c3edaba4b95 }, /* 1.831 */ 280 { 113, 512, 0x2a5eccd7735a3630, 0x00000c630664e7df }, /* 1.825 */ 281 { 114, 512, 0xafcccee5c0b71446, 0x00000cb65392f6e4 }, /* 1.826 */ 282 { 115, 512, 0x8fa30c5e7b147e27, 0x00000cd4db391e55 }, /* 1.843 */ 283 { 116, 512, 0x5afe0711fdfafd82, 0x00000d08cb4ec35d }, /* 1.826 */ 284 { 117, 512, 0x533a6090238afd4c, 0x00000d336f115d1b }, /* 1.803 */ 285 { 118, 512, 0x90cf11b595e39a84, 0x00000d8e041c2048 }, /* 1.857 */ 286 { 119, 512, 0x0d61a3b809444009, 0x00000dcb798afe35 }, /* 1.877 */ 287 { 120, 512, 0x7f34da0f54b0d114, 0x00000df3922664e1 }, /* 1.849 */ 288 { 121, 512, 0xa52258d5b72f6551, 0x00000e4d37a9872d }, /* 1.867 */ 289 { 122, 512, 0xc1de54d7672878db, 0x00000e6583a94cf6 }, /* 1.978 */ 290 { 123, 512, 0x1d03354316a414ab, 0x00000ebffc50308d }, /* 1.947 */ 291 { 124, 512, 0xcebdcc377665412c, 0x00000edee1997cea }, /* 1.865 */ 292 { 125, 512, 0x4ddd4c04b1a12344, 0x00000f21d64b373f }, /* 1.881 */ 293 { 126, 512, 0x64fc8f94e3973658, 0x00000f8f87a8896b }, /* 1.882 */ 294 { 127, 512, 0x68765f78034a334e, 0x00000fb8fe62197e }, /* 1.867 */ 295 { 128, 512, 0xaf36b871a303e816, 0x00000fec6f3afb1e }, /* 1.972 */ 296 { 129, 512, 0x2a4cbf73866c3a28, 0x00001027febfe4e5 }, /* 1.896 */ 297 { 130, 512, 0x9cb128aacdcd3b2f, 0x0000106aa8ac569d }, /* 1.965 */ 298 { 131, 512, 0x5511d41c55869124, 0x000010bbd755ddf1 }, /* 1.963 */ 299 { 132, 512, 0x42f92461937f284a, 0x000010fb8bceb3b5 }, /* 1.925 */ 300 { 133, 512, 0xe2d89a1cf6f1f287, 0x0000114cf5331e34 }, /* 1.862 */ 301 { 134, 512, 0xdc631a038956200e, 0x0000116428d2adc5 }, /* 2.042 */ 302 { 135, 512, 0xb2e5ac222cd236be, 0x000011ca88e4d4d2 }, /* 1.935 */ 303 { 136, 512, 0xbc7d8236655d88e7, 0x000011e39cb94e66 }, /* 2.005 */ 304 { 137, 512, 0x073e02d88d2d8e75, 0x0000123136c7933c }, /* 2.041 */ 305 { 138, 512, 0x3ddb9c3873166be0, 0x00001280e4ec6d52 }, /* 1.997 */ 306 { 139, 512, 0x7d3b1a845420e1b5, 0x000012c2e7cd6a44 }, /* 1.996 */ 307 { 140, 512, 0x60102308aa7b2a6c, 0x000012fc490e6c7d }, /* 2.053 */ 308 { 141, 512, 0xdb22bb2f9eb894aa, 0x00001343f5a85a1a }, /* 1.971 */ 309 { 142, 512, 0xd853f879a13b1606, 0x000013bb7d5f9048 }, /* 2.018 */ 310 { 143, 512, 0x001620a03f804b1d, 0x000013e74cc794fd }, /* 1.961 */ 311 { 144, 512, 0xfdb52dda76fbf667, 0x00001442d2f22480 }, /* 2.046 */ 312 { 145, 512, 0xa9160110f66e24ff, 0x0000144b899f9dbb }, /* 1.968 */ 313 { 146, 512, 0x77306a30379ae03b, 0x000014cb98eb1f81 }, /* 2.143 */ 314 { 147, 512, 0x14f5985d2752319d, 0x000014feab821fc9 }, /* 2.064 */ 315 { 148, 512, 0xa4b8ff11de7863f8, 0x0000154a0e60b9c9 }, /* 2.023 */ 316 { 149, 512, 0x44b345426455c1b3, 0x000015999c3c569c }, /* 2.136 */ 317 { 150, 512, 0x272677826049b46c, 0x000015c9697f4b92 }, /* 2.063 */ 318 { 151, 512, 0x2f9216e2cd74fe40, 0x0000162b1f7bbd39 }, /* 1.974 */ 319 { 152, 512, 0x706ae3e763ad8771, 0x00001661371c55e1 }, /* 2.210 */ 320 { 153, 512, 0xf7fd345307c2480e, 0x000016e251f28b6a }, /* 2.006 */ 321 { 154, 512, 0x6e94e3d26b3139eb, 0x000016f2429bb8c6 }, /* 2.193 */ 322 { 155, 512, 0x5458bbfbb781fcba, 0x0000173efdeca1b9 }, /* 2.163 */ 323 { 156, 512, 0xa80e2afeccd93b33, 0x000017bfdcb78adc }, /* 2.046 */ 324 { 157, 512, 0x1e4ccbb22796cf9d, 0x00001826fdcc39c9 }, /* 2.084 */ 325 { 158, 512, 0x8fba4b676aaa3663, 0x00001841a1379480 }, /* 2.264 */ 326 { 159, 512, 0xf82b843814b315fa, 0x000018886e19b8a3 }, /* 2.074 */ 327 { 160, 512, 0x7f21e920ecf753a3, 0x0000191812ca0ea7 }, /* 2.282 */ 328 { 161, 512, 0x48bb8ea2c4caa620, 0x0000192f310faccf }, /* 2.148 */ 329 { 162, 512, 0x5cdb652b4952c91b, 0x0000199e1d7437c7 }, /* 2.355 */ 330 { 163, 512, 0x6ac1ba6f78c06cd4, 0x000019cd11f82c70 }, /* 2.164 */ 331 { 164, 512, 0x9faf5f9ca2669a56, 0x00001a18d5431f6a }, /* 2.393 */ 332 { 165, 512, 0xaa57e9383eb01194, 0x00001a9e7d253d85 }, /* 2.178 */ 333 { 166, 512, 0x896967bf495c34d2, 0x00001afb8319b9fc }, /* 2.334 */ 334 { 167, 512, 0xdfad5f05de225f1b, 0x00001b3a59c3093b }, /* 2.266 */ 335 { 168, 512, 0xfd299a99f9f2abdd, 0x00001bb6f1a10799 }, /* 2.304 */ 336 { 169, 512, 0xdda239e798fe9fd4, 0x00001bfae0c9692d }, /* 2.218 */ 337 { 170, 512, 0x5fca670414a32c3e, 0x00001c22129dbcff }, /* 2.377 */ 338 { 171, 512, 0x1bb8934314b087de, 0x00001c955db36cd0 }, /* 2.155 */ 339 { 172, 512, 0xd96394b4b082200d, 0x00001cfc8619b7e6 }, /* 2.404 */ 340 { 173, 512, 0xb612a7735b1c8cbc, 0x00001d303acdd585 }, /* 2.205 */ 341 { 174, 512, 0x28e7430fe5875fe1, 0x00001d7ed5b3697d }, /* 2.359 */ 342 { 175, 512, 0x5038e89efdd981b9, 0x00001dc40ec35c59 }, /* 2.158 */ 343 { 176, 512, 0x075fd78f1d14db7c, 0x00001e31c83b4a2b }, /* 2.614 */ 344 { 177, 512, 0xc50fafdb5021be15, 0x00001e7cdac82fbc }, /* 2.239 */ 345 { 178, 512, 0xe6dc7572ce7b91c7, 0x00001edd8bb454fc }, /* 2.493 */ 346 { 179, 512, 0x21f7843e7beda537, 0x00001f3a8e019d6c }, /* 2.327 */ 347 { 180, 512, 0xc83385e20b43ec82, 0x00001f70735ec137 }, /* 2.231 */ 348 { 181, 512, 0xca818217dddb21fd, 0x0000201ca44c5a3c }, /* 2.237 */ 349 { 182, 512, 0xe6035defea48f933, 0x00002038e3346658 }, /* 2.691 */ 350 { 183, 512, 0x47262a4f953dac5a, 0x000020c2e554314e }, /* 2.170 */ 351 { 184, 512, 0xe24c7246260873ea, 0x000021197e618d64 }, /* 2.600 */ 352 { 185, 512, 0xeef6b57c9b58e9e1, 0x0000217ea48ecddc }, /* 2.391 */ 353 { 186, 512, 0x2becd3346e386142, 0x000021c496d4a5f9 }, /* 2.677 */ 354 { 187, 512, 0x63c6207bdf3b40a3, 0x0000220e0f2eec0c }, /* 2.410 */ 355 { 188, 512, 0x3056ce8989767d4b, 0x0000228eb76cd137 }, /* 2.776 */ 356 { 189, 512, 0x91af61c307cee780, 0x000022e17e2ea501 }, /* 2.266 */ 357 { 190, 512, 0xda359da225f6d54f, 0x00002358a2debc19 }, /* 2.717 */ 358 { 191, 512, 0x0a5f7a2a55607ba0, 0x0000238a79dac18c }, /* 2.474 */ 359 { 192, 512, 0x27bb75bf5224638a, 0x00002403a58e2351 }, /* 2.673 */ 360 { 193, 512, 0x1ebfdb94630f5d0f, 0x00002492a10cb339 }, /* 2.420 */ 361 { 194, 512, 0x6eae5e51d9c5f6fb, 0x000024ce4bf98715 }, /* 2.898 */ 362 { 195, 512, 0x08d903b4daedc2e0, 0x0000250d1e15886c }, /* 2.363 */ 363 { 196, 512, 0xc722a2f7fa7cd686, 0x0000258a99ed0c9e }, /* 2.747 */ 364 { 197, 512, 0x8f71faf0e54e361d, 0x000025dee11976f5 }, /* 2.531 */ 365 { 198, 512, 0x87f64695c91a54e7, 0x0000264e00a43da0 }, /* 2.707 */ 366 { 199, 512, 0xc719cbac2c336b92, 0x000026d327277ac1 }, /* 2.315 */ 367 { 200, 512, 0xe7e647afaf771ade, 0x000027523a5c44bf }, /* 3.012 */ 368 { 201, 512, 0x12d4b5c38ce8c946, 0x0000273898432545 }, /* 2.378 */ 369 { 202, 512, 0xf2e0cd4067bdc94a, 0x000027e47bb2c935 }, /* 2.969 */ 370 { 203, 512, 0x21b79f14d6d947d3, 0x0000281e64977f0d }, /* 2.594 */ 371 { 204, 512, 0x515093f952f18cd6, 0x0000289691a473fd }, /* 2.763 */ 372 { 205, 512, 0xd47b160a1b1022c8, 0x00002903e8b52411 }, /* 2.457 */ 373 { 206, 512, 0xc02fc96684715a16, 0x0000297515608601 }, /* 3.057 */ 374 { 207, 512, 0xef51e68efba72ed0, 0x000029ef73604804 }, /* 2.590 */ 375 { 208, 512, 0x9e3be6e5448b4f33, 0x00002a2846ed074b }, /* 3.047 */ 376 { 209, 512, 0x81d446c6d5fec063, 0x00002a92ca693455 }, /* 2.676 */ 377 { 210, 512, 0xff215de8224e57d5, 0x00002b2271fe3729 }, /* 2.993 */ 378 { 211, 512, 0xe2524d9ba8f69796, 0x00002b64b99c3ba2 }, /* 2.457 */ 379 { 212, 512, 0xf6b28e26097b7e4b, 0x00002bd768b6e068 }, /* 3.182 */ 380 { 213, 512, 0x893a487f30ce1644, 0x00002c67f722b4b2 }, /* 2.563 */ 381 { 214, 512, 0x386566c3fc9871df, 0x00002cc1cf8b4037 }, /* 3.025 */ 382 { 215, 512, 0x1e0ed78edf1f558a, 0x00002d3948d36c7f }, /* 2.730 */ 383 { 216, 512, 0xe3bc20c31e61f113, 0x00002d6d6b12e025 }, /* 3.036 */ 384 { 217, 512, 0xd6c3ad2e23021882, 0x00002deff7572241 }, /* 2.722 */ 385 { 218, 512, 0xb4a9f95cf0f69c5a, 0x00002e67d537aa36 }, /* 3.356 */ 386 { 219, 512, 0x6e98ed6f6c38e82f, 0x00002e9720626789 }, /* 2.697 */ 387 { 220, 512, 0x2e01edba33fddac7, 0x00002f407c6b0198 }, /* 2.979 */ 388 { 221, 512, 0x559d02e1f5f57ccc, 0x00002fb6a5ab4f24 }, /* 2.858 */ 389 { 222, 512, 0xac18f5a916adcd8e, 0x0000304ae1c5c57e }, /* 3.258 */ 390 { 223, 512, 0x15789fbaddb86f4b, 0x0000306f6e019c78 }, /* 2.693 */ 391 { 224, 512, 0xf4a9c36d5bc4c408, 0x000030da40434213 }, /* 3.259 */ 392 { 225, 512, 0xf640f90fd2727f44, 0x00003189ed37b90c }, /* 2.733 */ 393 { 226, 512, 0xb5313d390d61884a, 0x000031e152616b37 }, /* 3.235 */ 394 { 227, 512, 0x4bae6b3ce9160939, 0x0000321f40aeac42 }, /* 2.983 */ 395 { 228, 512, 0x838c34480f1a66a1, 0x000032f389c0f78e }, /* 3.308 */ 396 { 229, 512, 0xb1c4a52c8e3d6060, 0x0000330062a40284 }, /* 2.715 */ 397 { 230, 512, 0xe0f1110c6d0ed822, 0x0000338be435644f }, /* 3.540 */ 398 { 231, 512, 0x9f1a8ccdcea68d4b, 0x000034045a4e97e1 }, /* 2.779 */ 399 { 232, 512, 0x3261ed62223f3099, 0x000034702cfc401c }, /* 3.084 */ 400 { 233, 512, 0xf2191e2311022d65, 0x00003509dd19c9fc }, /* 2.987 */ 401 { 234, 512, 0xf102a395c2033abc, 0x000035654dc96fae }, /* 3.341 */ 402 { 235, 512, 0x11fe378f027906b6, 0x000035b5193b0264 }, /* 2.793 */ 403 { 236, 512, 0xf777f2c026b337aa, 0x000036704f5d9297 }, /* 3.518 */ 404 { 237, 512, 0x1b04e9c2ee143f32, 0x000036dfbb7af218 }, /* 2.962 */ 405 { 238, 512, 0x2fcec95266f9352c, 0x00003785c8df24a9 }, /* 3.196 */ 406 { 239, 512, 0xfe2b0e47e427dd85, 0x000037cbdf5da729 }, /* 2.914 */ 407 { 240, 512, 0x72b49bf2225f6c6d, 0x0000382227c15855 }, /* 3.408 */ 408 { 241, 512, 0x50486b43df7df9c7, 0x0000389b88be6453 }, /* 2.903 */ 409 { 242, 512, 0x5192a3e53181c8ab, 0x000038ddf3d67263 }, /* 3.778 */ 410 { 243, 512, 0xe9f5d8365296fd5e, 0x0000399f1c6c9e9c }, /* 3.026 */ 411 { 244, 512, 0xc740263f0301efa8, 0x00003a147146512d }, /* 3.347 */ 412 { 245, 512, 0x23cd0f2b5671e67d, 0x00003ab10bcc0d9d }, /* 3.212 */ 413 { 246, 512, 0x002ccc7e5cd41390, 0x00003ad6cd14a6c0 }, /* 3.482 */ 414 { 247, 512, 0x9aafb3c02544b31b, 0x00003b8cb8779fb0 }, /* 3.146 */ 415 { 248, 512, 0x72ba07a78b121999, 0x00003c24142a5a3f }, /* 3.626 */ 416 { 249, 512, 0x3d784aa58edfc7b4, 0x00003cd084817d99 }, /* 2.952 */ 417 { 250, 512, 0xaab750424d8004af, 0x00003d506a8e098e }, /* 3.463 */ 418 { 251, 512, 0x84403fcf8e6b5ca2, 0x00003d4c54c2aec4 }, /* 3.131 */ 419 { 252, 512, 0x71eb7455ec98e207, 0x00003e655715cf2c }, /* 3.538 */ 420 { 253, 512, 0xd752b4f19301595b, 0x00003ecd7b2ca5ac }, /* 2.974 */ 421 { 254, 512, 0xc4674129750499de, 0x00003e99e86d3e95 }, /* 3.843 */ 422 { 255, 512, 0x9772baff5cd12ef5, 0x00003f895c019841 }, /* 3.088 */ 423 }; 424 425 /* 426 * Verify the map is valid. Each device index must appear exactly 427 * once in every row, and the permutation array checksum must match. 428 */ 429 static int 430 verify_perms(uint8_t *perms, uint64_t children, uint64_t nperms, 431 uint64_t checksum) 432 { 433 int countssz = sizeof (uint16_t) * children; 434 uint16_t *counts = kmem_zalloc(countssz, KM_SLEEP); 435 436 for (int i = 0; i < nperms; i++) { 437 for (int j = 0; j < children; j++) { 438 uint8_t val = perms[(i * children) + j]; 439 440 if (val >= children || counts[val] != i) { 441 kmem_free(counts, countssz); 442 return (EINVAL); 443 } 444 445 counts[val]++; 446 } 447 } 448 449 if (checksum != 0) { 450 int permssz = sizeof (uint8_t) * children * nperms; 451 zio_cksum_t cksum; 452 453 fletcher_4_native_varsize(perms, permssz, &cksum); 454 455 if (checksum != cksum.zc_word[0]) { 456 kmem_free(counts, countssz); 457 return (ECKSUM); 458 } 459 } 460 461 kmem_free(counts, countssz); 462 463 return (0); 464 } 465 466 /* 467 * Generate the permutation array for the draid_map_t. These maps control 468 * the placement of all data in a dRAID. Therefore it's critical that the 469 * seed always generates the same mapping. We provide our own pseudo-random 470 * number generator for this purpose. 471 */ 472 int 473 vdev_draid_generate_perms(const draid_map_t *map, uint8_t **permsp) 474 { 475 VERIFY3U(map->dm_children, >=, VDEV_DRAID_MIN_CHILDREN); 476 VERIFY3U(map->dm_children, <=, VDEV_DRAID_MAX_CHILDREN); 477 VERIFY3U(map->dm_seed, !=, 0); 478 VERIFY3U(map->dm_nperms, !=, 0); 479 VERIFY3P(map->dm_perms, ==, NULL); 480 481 #ifdef _KERNEL 482 /* 483 * The kernel code always provides both a map_seed and checksum. 484 * Only the tests/zfs-tests/cmd/draid/draid.c utility will provide 485 * a zero checksum when generating new candidate maps. 486 */ 487 VERIFY3U(map->dm_checksum, !=, 0); 488 #endif 489 uint64_t children = map->dm_children; 490 uint64_t nperms = map->dm_nperms; 491 int rowsz = sizeof (uint8_t) * children; 492 int permssz = rowsz * nperms; 493 uint8_t *perms; 494 495 /* Allocate the permutation array */ 496 perms = vmem_alloc(permssz, KM_SLEEP); 497 498 /* Setup an initial row with a known pattern */ 499 uint8_t *initial_row = kmem_alloc(rowsz, KM_SLEEP); 500 for (int i = 0; i < children; i++) 501 initial_row[i] = i; 502 503 uint64_t draid_seed[2] = { VDEV_DRAID_SEED, map->dm_seed }; 504 uint8_t *current_row, *previous_row = initial_row; 505 506 /* 507 * Perform a Fisher-Yates shuffle of each row using the previous 508 * row as the starting point. An initial_row with known pattern 509 * is used as the input for the first row. 510 */ 511 for (int i = 0; i < nperms; i++) { 512 current_row = &perms[i * children]; 513 memcpy(current_row, previous_row, rowsz); 514 515 for (int j = children - 1; j > 0; j--) { 516 uint64_t k = vdev_draid_rand(draid_seed) % (j + 1); 517 uint8_t val = current_row[j]; 518 current_row[j] = current_row[k]; 519 current_row[k] = val; 520 } 521 522 previous_row = current_row; 523 } 524 525 kmem_free(initial_row, rowsz); 526 527 int error = verify_perms(perms, children, nperms, map->dm_checksum); 528 if (error) { 529 vmem_free(perms, permssz); 530 return (error); 531 } 532 533 *permsp = perms; 534 535 return (0); 536 } 537 538 /* 539 * Lookup the fixed draid_map_t for the requested number of children. 540 */ 541 int 542 vdev_draid_lookup_map(uint64_t children, const draid_map_t **mapp) 543 { 544 for (int i = 0; i <= VDEV_DRAID_MAX_MAPS; i++) { 545 if (draid_maps[i].dm_children == children) { 546 *mapp = &draid_maps[i]; 547 return (0); 548 } 549 } 550 551 return (ENOENT); 552 } 553 554 /* 555 * Lookup the permutation array and iteration id for the provided offset. 556 */ 557 static void 558 vdev_draid_get_perm(vdev_draid_config_t *vdc, uint64_t pindex, 559 uint8_t **base, uint64_t *iter) 560 { 561 uint64_t ncols = vdc->vdc_children; 562 uint64_t poff = pindex % (vdc->vdc_nperms * ncols); 563 564 *base = vdc->vdc_perms + (poff / ncols) * ncols; 565 *iter = poff % ncols; 566 } 567 568 static inline uint64_t 569 vdev_draid_permute_id(vdev_draid_config_t *vdc, 570 uint8_t *base, uint64_t iter, uint64_t index) 571 { 572 return ((base[index] + iter) % vdc->vdc_children); 573 } 574 575 /* 576 * Return the asize which is the psize rounded up to a full group width. 577 * i.e. vdev_draid_psize_to_asize(). 578 */ 579 static uint64_t 580 vdev_draid_asize(vdev_t *vd, uint64_t psize) 581 { 582 vdev_draid_config_t *vdc = vd->vdev_tsd; 583 uint64_t ashift = vd->vdev_ashift; 584 585 ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops); 586 587 uint64_t rows = ((psize - 1) / (vdc->vdc_ndata << ashift)) + 1; 588 uint64_t asize = (rows * vdc->vdc_groupwidth) << ashift; 589 590 ASSERT3U(asize, !=, 0); 591 ASSERT3U(asize % (vdc->vdc_groupwidth), ==, 0); 592 593 return (asize); 594 } 595 596 /* 597 * Deflate the asize to the psize, this includes stripping parity. 598 */ 599 uint64_t 600 vdev_draid_asize_to_psize(vdev_t *vd, uint64_t asize) 601 { 602 vdev_draid_config_t *vdc = vd->vdev_tsd; 603 604 ASSERT0(asize % vdc->vdc_groupwidth); 605 606 return ((asize / vdc->vdc_groupwidth) * vdc->vdc_ndata); 607 } 608 609 /* 610 * Convert a logical offset to the corresponding group number. 611 */ 612 static uint64_t 613 vdev_draid_offset_to_group(vdev_t *vd, uint64_t offset) 614 { 615 vdev_draid_config_t *vdc = vd->vdev_tsd; 616 617 ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops); 618 619 return (offset / vdc->vdc_groupsz); 620 } 621 622 /* 623 * Convert a group number to the logical starting offset for that group. 624 */ 625 static uint64_t 626 vdev_draid_group_to_offset(vdev_t *vd, uint64_t group) 627 { 628 vdev_draid_config_t *vdc = vd->vdev_tsd; 629 630 ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops); 631 632 return (group * vdc->vdc_groupsz); 633 } 634 635 636 static void 637 vdev_draid_map_free_vsd(zio_t *zio) 638 { 639 raidz_map_t *rm = zio->io_vsd; 640 641 ASSERT0(rm->rm_freed); 642 rm->rm_freed = B_TRUE; 643 644 if (rm->rm_reports == 0) { 645 vdev_raidz_map_free(rm); 646 } 647 } 648 649 /*ARGSUSED*/ 650 static void 651 vdev_draid_cksum_free(void *arg, size_t ignored) 652 { 653 raidz_map_t *rm = arg; 654 655 ASSERT3U(rm->rm_reports, >, 0); 656 657 if (--rm->rm_reports == 0 && rm->rm_freed) 658 vdev_raidz_map_free(rm); 659 } 660 661 static void 662 vdev_draid_cksum_finish(zio_cksum_report_t *zcr, const abd_t *good_data) 663 { 664 raidz_map_t *rm = zcr->zcr_cbdata; 665 const size_t c = zcr->zcr_cbinfo; 666 uint64_t skip_size = zcr->zcr_sector; 667 uint64_t parity_size; 668 size_t x, offset, size; 669 670 if (good_data == NULL) { 671 zfs_ereport_finish_checksum(zcr, NULL, NULL, B_FALSE); 672 return; 673 } 674 675 /* 676 * Detailed cksum reporting is currently only supported for single 677 * row draid mappings, this covers the vast majority of zios. Only 678 * a dRAID zio which spans groups will have multiple rows. 679 */ 680 if (rm->rm_nrows != 1) { 681 zfs_ereport_finish_checksum(zcr, NULL, NULL, B_FALSE); 682 return; 683 } 684 685 raidz_row_t *rr = rm->rm_row[0]; 686 const abd_t *good = NULL; 687 const abd_t *bad = rr->rr_col[c].rc_abd; 688 689 if (c < rr->rr_firstdatacol) { 690 /* 691 * The first time through, calculate the parity blocks for 692 * the good data (this relies on the fact that the good 693 * data never changes for a given logical zio) 694 */ 695 if (rr->rr_col[0].rc_gdata == NULL) { 696 abd_t *bad_parity[VDEV_DRAID_MAXPARITY]; 697 698 /* 699 * Set up the rr_col[]s to generate the parity for 700 * good_data, first saving the parity bufs and 701 * replacing them with buffers to hold the result. 702 */ 703 for (x = 0; x < rr->rr_firstdatacol; x++) { 704 bad_parity[x] = rr->rr_col[x].rc_abd; 705 rr->rr_col[x].rc_abd = rr->rr_col[x].rc_gdata = 706 abd_alloc_sametype(rr->rr_col[x].rc_abd, 707 rr->rr_col[x].rc_size); 708 } 709 710 /* 711 * Fill in the data columns from good_data being 712 * careful to pad short columns and empty columns 713 * with a skip sector. 714 */ 715 uint64_t good_size = abd_get_size((abd_t *)good_data); 716 717 offset = 0; 718 for (; x < rr->rr_cols; x++) { 719 abd_free(rr->rr_col[x].rc_abd); 720 721 if (offset == good_size) { 722 /* empty data column (small write) */ 723 rr->rr_col[x].rc_abd = 724 abd_get_zeros(skip_size); 725 } else if (x < rr->rr_bigcols) { 726 /* this is a "big column" */ 727 size = rr->rr_col[x].rc_size; 728 rr->rr_col[x].rc_abd = 729 abd_get_offset_size( 730 (abd_t *)good_data, offset, size); 731 offset += size; 732 } else { 733 /* short data column, add skip sector */ 734 size = rr->rr_col[x].rc_size -skip_size; 735 rr->rr_col[x].rc_abd = abd_alloc( 736 rr->rr_col[x].rc_size, B_TRUE); 737 abd_copy_off(rr->rr_col[x].rc_abd, 738 (abd_t *)good_data, 0, offset, 739 size); 740 abd_zero_off(rr->rr_col[x].rc_abd, 741 size, skip_size); 742 offset += size; 743 } 744 } 745 746 /* 747 * Construct the parity from the good data. 748 */ 749 vdev_raidz_generate_parity_row(rm, rr); 750 751 /* restore everything back to its original state */ 752 for (x = 0; x < rr->rr_firstdatacol; x++) 753 rr->rr_col[x].rc_abd = bad_parity[x]; 754 755 offset = 0; 756 for (x = rr->rr_firstdatacol; x < rr->rr_cols; x++) { 757 abd_free(rr->rr_col[x].rc_abd); 758 rr->rr_col[x].rc_abd = abd_get_offset_size( 759 rr->rr_abd_copy, offset, 760 rr->rr_col[x].rc_size); 761 offset += rr->rr_col[x].rc_size; 762 } 763 } 764 765 ASSERT3P(rr->rr_col[c].rc_gdata, !=, NULL); 766 good = abd_get_offset_size(rr->rr_col[c].rc_gdata, 0, 767 rr->rr_col[c].rc_size); 768 } else { 769 /* adjust good_data to point at the start of our column */ 770 parity_size = size = rr->rr_col[0].rc_size; 771 if (c >= rr->rr_bigcols) { 772 size -= skip_size; 773 zcr->zcr_length = size; 774 } 775 776 /* empty column */ 777 if (size == 0) { 778 zfs_ereport_finish_checksum(zcr, NULL, NULL, B_TRUE); 779 return; 780 } 781 782 offset = 0; 783 for (x = rr->rr_firstdatacol; x < c; x++) { 784 if (x < rr->rr_bigcols) { 785 offset += parity_size; 786 } else { 787 offset += parity_size - skip_size; 788 } 789 } 790 791 good = abd_get_offset_size((abd_t *)good_data, offset, size); 792 } 793 794 /* we drop the ereport if it ends up that the data was good */ 795 zfs_ereport_finish_checksum(zcr, good, bad, B_TRUE); 796 abd_free((abd_t *)good); 797 } 798 799 /* 800 * Invoked indirectly by zfs_ereport_start_checksum(), called 801 * below when our read operation fails completely. The main point 802 * is to keep a copy of everything we read from disk, so that at 803 * vdev_draid_cksum_finish() time we can compare it with the good data. 804 */ 805 static void 806 vdev_draid_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *arg) 807 { 808 size_t c = (size_t)(uintptr_t)arg; 809 raidz_map_t *rm = zio->io_vsd; 810 811 /* set up the report and bump the refcount */ 812 zcr->zcr_cbdata = rm; 813 zcr->zcr_cbinfo = c; 814 zcr->zcr_finish = vdev_draid_cksum_finish; 815 zcr->zcr_free = vdev_draid_cksum_free; 816 817 rm->rm_reports++; 818 ASSERT3U(rm->rm_reports, >, 0); 819 820 if (rm->rm_row[0]->rr_abd_copy != NULL) 821 return; 822 823 /* 824 * It's the first time we're called for this raidz_map_t, so we need 825 * to copy the data aside; there's no guarantee that our zio's buffer 826 * won't be re-used for something else. 827 * 828 * Our parity data is already in separate buffers, so there's no need 829 * to copy them. Furthermore, all columns should have been expanded 830 * by vdev_draid_map_alloc_empty() when attempting reconstruction. 831 */ 832 for (int i = 0; i < rm->rm_nrows; i++) { 833 raidz_row_t *rr = rm->rm_row[i]; 834 size_t offset = 0; 835 size_t size = 0; 836 837 for (c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { 838 ASSERT3U(rr->rr_col[c].rc_size, ==, 839 rr->rr_col[0].rc_size); 840 size += rr->rr_col[c].rc_size; 841 } 842 843 rr->rr_abd_copy = abd_alloc_for_io(size, B_FALSE); 844 845 for (c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { 846 raidz_col_t *col = &rr->rr_col[c]; 847 abd_t *tmp = abd_get_offset_size(rr->rr_abd_copy, 848 offset, col->rc_size); 849 850 abd_copy(tmp, col->rc_abd, col->rc_size); 851 abd_free(col->rc_abd); 852 853 col->rc_abd = tmp; 854 offset += col->rc_size; 855 } 856 ASSERT3U(offset, ==, size); 857 } 858 } 859 860 const zio_vsd_ops_t vdev_draid_vsd_ops = { 861 .vsd_free = vdev_draid_map_free_vsd, 862 .vsd_cksum_report = vdev_draid_cksum_report 863 }; 864 865 /* 866 * Full stripe writes. When writing, all columns (D+P) are required. Parity 867 * is calculated over all the columns, including empty zero filled sectors, 868 * and each is written to disk. While only the data columns are needed for 869 * a normal read, all of the columns are required for reconstruction when 870 * performing a sequential resilver. 871 * 872 * For "big columns" it's sufficient to map the correct range of the zio ABD. 873 * Partial columns require allocating a gang ABD in order to zero fill the 874 * empty sectors. When the column is empty a zero filled sector must be 875 * mapped. In all cases the data ABDs must be the same size as the parity 876 * ABDs (e.g. rc->rc_size == parity_size). 877 */ 878 static void 879 vdev_draid_map_alloc_write(zio_t *zio, uint64_t abd_offset, raidz_row_t *rr) 880 { 881 uint64_t skip_size = 1ULL << zio->io_vd->vdev_top->vdev_ashift; 882 uint64_t parity_size = rr->rr_col[0].rc_size; 883 uint64_t abd_off = abd_offset; 884 885 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); 886 ASSERT3U(parity_size, ==, abd_get_size(rr->rr_col[0].rc_abd)); 887 888 for (uint64_t c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { 889 raidz_col_t *rc = &rr->rr_col[c]; 890 891 if (rc->rc_size == 0) { 892 /* empty data column (small write), add a skip sector */ 893 ASSERT3U(skip_size, ==, parity_size); 894 rc->rc_abd = abd_get_zeros(skip_size); 895 } else if (rc->rc_size == parity_size) { 896 /* this is a "big column" */ 897 rc->rc_abd = abd_get_offset_struct(&rc->rc_abdstruct, 898 zio->io_abd, abd_off, rc->rc_size); 899 } else { 900 /* short data column, add a skip sector */ 901 ASSERT3U(rc->rc_size + skip_size, ==, parity_size); 902 rc->rc_abd = abd_alloc_gang(); 903 abd_gang_add(rc->rc_abd, abd_get_offset_size( 904 zio->io_abd, abd_off, rc->rc_size), B_TRUE); 905 abd_gang_add(rc->rc_abd, abd_get_zeros(skip_size), 906 B_TRUE); 907 } 908 909 ASSERT3U(abd_get_size(rc->rc_abd), ==, parity_size); 910 911 abd_off += rc->rc_size; 912 rc->rc_size = parity_size; 913 } 914 915 IMPLY(abd_offset != 0, abd_off == zio->io_size); 916 } 917 918 /* 919 * Scrub/resilver reads. In order to store the contents of the skip sectors 920 * an additional ABD is allocated. The columns are handled in the same way 921 * as a full stripe write except instead of using the zero ABD the newly 922 * allocated skip ABD is used to back the skip sectors. In all cases the 923 * data ABD must be the same size as the parity ABDs. 924 */ 925 static void 926 vdev_draid_map_alloc_scrub(zio_t *zio, uint64_t abd_offset, raidz_row_t *rr) 927 { 928 uint64_t skip_size = 1ULL << zio->io_vd->vdev_top->vdev_ashift; 929 uint64_t parity_size = rr->rr_col[0].rc_size; 930 uint64_t abd_off = abd_offset; 931 uint64_t skip_off = 0; 932 933 ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ); 934 ASSERT3P(rr->rr_abd_empty, ==, NULL); 935 936 if (rr->rr_nempty > 0) { 937 rr->rr_abd_empty = abd_alloc_linear(rr->rr_nempty * skip_size, 938 B_FALSE); 939 } 940 941 for (uint64_t c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { 942 raidz_col_t *rc = &rr->rr_col[c]; 943 944 if (rc->rc_size == 0) { 945 /* empty data column (small read), add a skip sector */ 946 ASSERT3U(skip_size, ==, parity_size); 947 ASSERT3U(rr->rr_nempty, !=, 0); 948 rc->rc_abd = abd_get_offset_size(rr->rr_abd_empty, 949 skip_off, skip_size); 950 skip_off += skip_size; 951 } else if (rc->rc_size == parity_size) { 952 /* this is a "big column" */ 953 rc->rc_abd = abd_get_offset_struct(&rc->rc_abdstruct, 954 zio->io_abd, abd_off, rc->rc_size); 955 } else { 956 /* short data column, add a skip sector */ 957 ASSERT3U(rc->rc_size + skip_size, ==, parity_size); 958 ASSERT3U(rr->rr_nempty, !=, 0); 959 rc->rc_abd = abd_alloc_gang(); 960 abd_gang_add(rc->rc_abd, abd_get_offset_size( 961 zio->io_abd, abd_off, rc->rc_size), B_TRUE); 962 abd_gang_add(rc->rc_abd, abd_get_offset_size( 963 rr->rr_abd_empty, skip_off, skip_size), B_TRUE); 964 skip_off += skip_size; 965 } 966 967 uint64_t abd_size = abd_get_size(rc->rc_abd); 968 ASSERT3U(abd_size, ==, abd_get_size(rr->rr_col[0].rc_abd)); 969 970 /* 971 * Increase rc_size so the skip ABD is included in subsequent 972 * parity calculations. 973 */ 974 abd_off += rc->rc_size; 975 rc->rc_size = abd_size; 976 } 977 978 IMPLY(abd_offset != 0, abd_off == zio->io_size); 979 ASSERT3U(skip_off, ==, rr->rr_nempty * skip_size); 980 } 981 982 /* 983 * Normal reads. In this common case only the columns containing data 984 * are read in to the zio ABDs. Neither the parity columns or empty skip 985 * sectors are read unless the checksum fails verification. In which case 986 * vdev_raidz_read_all() will call vdev_draid_map_alloc_empty() to expand 987 * the raid map in order to allow reconstruction using the parity data and 988 * skip sectors. 989 */ 990 static void 991 vdev_draid_map_alloc_read(zio_t *zio, uint64_t abd_offset, raidz_row_t *rr) 992 { 993 uint64_t abd_off = abd_offset; 994 995 ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ); 996 997 for (uint64_t c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { 998 raidz_col_t *rc = &rr->rr_col[c]; 999 1000 if (rc->rc_size > 0) { 1001 rc->rc_abd = abd_get_offset_struct(&rc->rc_abdstruct, 1002 zio->io_abd, abd_off, rc->rc_size); 1003 abd_off += rc->rc_size; 1004 } 1005 } 1006 1007 IMPLY(abd_offset != 0, abd_off == zio->io_size); 1008 } 1009 1010 /* 1011 * Converts a normal "read" raidz_row_t to a "scrub" raidz_row_t. The key 1012 * difference is that an ABD is allocated to back skip sectors so they may 1013 * be read in to memory, verified, and repaired if needed. 1014 */ 1015 void 1016 vdev_draid_map_alloc_empty(zio_t *zio, raidz_row_t *rr) 1017 { 1018 uint64_t skip_size = 1ULL << zio->io_vd->vdev_top->vdev_ashift; 1019 uint64_t parity_size = rr->rr_col[0].rc_size; 1020 uint64_t skip_off = 0; 1021 1022 ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ); 1023 ASSERT3P(rr->rr_abd_empty, ==, NULL); 1024 1025 if (rr->rr_nempty > 0) { 1026 rr->rr_abd_empty = abd_alloc_linear(rr->rr_nempty * skip_size, 1027 B_FALSE); 1028 } 1029 1030 for (uint64_t c = rr->rr_firstdatacol; c < rr->rr_cols; c++) { 1031 raidz_col_t *rc = &rr->rr_col[c]; 1032 1033 if (rc->rc_size == 0) { 1034 /* empty data column (small read), add a skip sector */ 1035 ASSERT3U(skip_size, ==, parity_size); 1036 ASSERT3U(rr->rr_nempty, !=, 0); 1037 ASSERT3P(rc->rc_abd, ==, NULL); 1038 rc->rc_abd = abd_get_offset_size(rr->rr_abd_empty, 1039 skip_off, skip_size); 1040 skip_off += skip_size; 1041 } else if (rc->rc_size == parity_size) { 1042 /* this is a "big column", nothing to add */ 1043 ASSERT3P(rc->rc_abd, !=, NULL); 1044 } else { 1045 /* short data column, add a skip sector */ 1046 ASSERT3U(rc->rc_size + skip_size, ==, parity_size); 1047 ASSERT3U(rr->rr_nempty, !=, 0); 1048 ASSERT3P(rc->rc_abd, !=, NULL); 1049 ASSERT(!abd_is_gang(rc->rc_abd)); 1050 abd_t *read_abd = rc->rc_abd; 1051 rc->rc_abd = abd_alloc_gang(); 1052 abd_gang_add(rc->rc_abd, read_abd, B_TRUE); 1053 abd_gang_add(rc->rc_abd, abd_get_offset_size( 1054 rr->rr_abd_empty, skip_off, skip_size), B_TRUE); 1055 skip_off += skip_size; 1056 } 1057 1058 /* 1059 * Increase rc_size so the empty ABD is included in subsequent 1060 * parity calculations. 1061 */ 1062 rc->rc_size = parity_size; 1063 } 1064 1065 ASSERT3U(skip_off, ==, rr->rr_nempty * skip_size); 1066 } 1067 1068 /* 1069 * Given a logical address within a dRAID configuration, return the physical 1070 * address on the first drive in the group that this address maps to 1071 * (at position 'start' in permutation number 'perm'). 1072 */ 1073 static uint64_t 1074 vdev_draid_logical_to_physical(vdev_t *vd, uint64_t logical_offset, 1075 uint64_t *perm, uint64_t *start) 1076 { 1077 vdev_draid_config_t *vdc = vd->vdev_tsd; 1078 1079 /* b is the dRAID (parent) sector offset. */ 1080 uint64_t ashift = vd->vdev_top->vdev_ashift; 1081 uint64_t b_offset = logical_offset >> ashift; 1082 1083 /* 1084 * The height of a row in units of the vdev's minimum sector size. 1085 * This is the amount of data written to each disk of each group 1086 * in a given permutation. 1087 */ 1088 uint64_t rowheight_sectors = VDEV_DRAID_ROWHEIGHT >> ashift; 1089 1090 /* 1091 * We cycle through a disk permutation every groupsz * ngroups chunk 1092 * of address space. Note that ngroups * groupsz must be a multiple 1093 * of the number of data drives (ndisks) in order to guarantee 1094 * alignment. So, for example, if our row height is 16MB, our group 1095 * size is 10, and there are 13 data drives in the draid, then ngroups 1096 * will be 13, we will change permutation every 2.08GB and each 1097 * disk will have 160MB of data per chunk. 1098 */ 1099 uint64_t groupwidth = vdc->vdc_groupwidth; 1100 uint64_t ngroups = vdc->vdc_ngroups; 1101 uint64_t ndisks = vdc->vdc_ndisks; 1102 1103 /* 1104 * groupstart is where the group this IO will land in "starts" in 1105 * the permutation array. 1106 */ 1107 uint64_t group = logical_offset / vdc->vdc_groupsz; 1108 uint64_t groupstart = (group * groupwidth) % ndisks; 1109 ASSERT3U(groupstart + groupwidth, <=, ndisks + groupstart); 1110 *start = groupstart; 1111 1112 /* b_offset is the sector offset within a group chunk */ 1113 b_offset = b_offset % (rowheight_sectors * groupwidth); 1114 ASSERT0(b_offset % groupwidth); 1115 1116 /* 1117 * Find the starting byte offset on each child vdev: 1118 * - within a permutation there are ngroups groups spread over the 1119 * rows, where each row covers a slice portion of the disk 1120 * - each permutation has (groupwidth * ngroups) / ndisks rows 1121 * - so each permutation covers rows * slice portion of the disk 1122 * - so we need to find the row where this IO group target begins 1123 */ 1124 *perm = group / ngroups; 1125 uint64_t row = (*perm * ((groupwidth * ngroups) / ndisks)) + 1126 (((group % ngroups) * groupwidth) / ndisks); 1127 1128 return (((rowheight_sectors * row) + 1129 (b_offset / groupwidth)) << ashift); 1130 } 1131 1132 static uint64_t 1133 vdev_draid_map_alloc_row(zio_t *zio, raidz_row_t **rrp, uint64_t io_offset, 1134 uint64_t abd_offset, uint64_t abd_size) 1135 { 1136 vdev_t *vd = zio->io_vd; 1137 vdev_draid_config_t *vdc = vd->vdev_tsd; 1138 uint64_t ashift = vd->vdev_top->vdev_ashift; 1139 uint64_t io_size = abd_size; 1140 uint64_t io_asize = vdev_draid_asize(vd, io_size); 1141 uint64_t group = vdev_draid_offset_to_group(vd, io_offset); 1142 uint64_t start_offset = vdev_draid_group_to_offset(vd, group + 1); 1143 1144 /* 1145 * Limit the io_size to the space remaining in the group. A second 1146 * row in the raidz_map_t is created for the remainder. 1147 */ 1148 if (io_offset + io_asize > start_offset) { 1149 io_size = vdev_draid_asize_to_psize(vd, 1150 start_offset - io_offset); 1151 } 1152 1153 /* 1154 * At most a block may span the logical end of one group and the start 1155 * of the next group. Therefore, at the end of a group the io_size must 1156 * span the group width evenly and the remainder must be aligned to the 1157 * start of the next group. 1158 */ 1159 IMPLY(abd_offset == 0 && io_size < zio->io_size, 1160 (io_asize >> ashift) % vdc->vdc_groupwidth == 0); 1161 IMPLY(abd_offset != 0, 1162 vdev_draid_group_to_offset(vd, group) == io_offset); 1163 1164 /* Lookup starting byte offset on each child vdev */ 1165 uint64_t groupstart, perm; 1166 uint64_t physical_offset = vdev_draid_logical_to_physical(vd, 1167 io_offset, &perm, &groupstart); 1168 1169 /* 1170 * If there is less than groupwidth drives available after the group 1171 * start, the group is going to wrap onto the next row. 'wrap' is the 1172 * group disk number that starts on the next row. 1173 */ 1174 uint64_t ndisks = vdc->vdc_ndisks; 1175 uint64_t groupwidth = vdc->vdc_groupwidth; 1176 uint64_t wrap = groupwidth; 1177 1178 if (groupstart + groupwidth > ndisks) 1179 wrap = ndisks - groupstart; 1180 1181 /* The io size in units of the vdev's minimum sector size. */ 1182 const uint64_t psize = io_size >> ashift; 1183 1184 /* 1185 * "Quotient": The number of data sectors for this stripe on all but 1186 * the "big column" child vdevs that also contain "remainder" data. 1187 */ 1188 uint64_t q = psize / vdc->vdc_ndata; 1189 1190 /* 1191 * "Remainder": The number of partial stripe data sectors in this I/O. 1192 * This will add a sector to some, but not all, child vdevs. 1193 */ 1194 uint64_t r = psize - q * vdc->vdc_ndata; 1195 1196 /* The number of "big columns" - those which contain remainder data. */ 1197 uint64_t bc = (r == 0 ? 0 : r + vdc->vdc_nparity); 1198 ASSERT3U(bc, <, groupwidth); 1199 1200 /* The total number of data and parity sectors for this I/O. */ 1201 uint64_t tot = psize + (vdc->vdc_nparity * (q + (r == 0 ? 0 : 1))); 1202 1203 raidz_row_t *rr; 1204 rr = kmem_alloc(offsetof(raidz_row_t, rr_col[groupwidth]), KM_SLEEP); 1205 rr->rr_cols = groupwidth; 1206 rr->rr_scols = groupwidth; 1207 rr->rr_bigcols = bc; 1208 rr->rr_missingdata = 0; 1209 rr->rr_missingparity = 0; 1210 rr->rr_firstdatacol = vdc->vdc_nparity; 1211 rr->rr_abd_copy = NULL; 1212 rr->rr_abd_empty = NULL; 1213 #ifdef ZFS_DEBUG 1214 rr->rr_offset = io_offset; 1215 rr->rr_size = io_size; 1216 #endif 1217 *rrp = rr; 1218 1219 uint8_t *base; 1220 uint64_t iter, asize = 0; 1221 vdev_draid_get_perm(vdc, perm, &base, &iter); 1222 for (uint64_t i = 0; i < groupwidth; i++) { 1223 raidz_col_t *rc = &rr->rr_col[i]; 1224 uint64_t c = (groupstart + i) % ndisks; 1225 1226 /* increment the offset if we wrap to the next row */ 1227 if (i == wrap) 1228 physical_offset += VDEV_DRAID_ROWHEIGHT; 1229 1230 rc->rc_devidx = vdev_draid_permute_id(vdc, base, iter, c); 1231 rc->rc_offset = physical_offset; 1232 rc->rc_abd = NULL; 1233 rc->rc_gdata = NULL; 1234 rc->rc_orig_data = NULL; 1235 rc->rc_error = 0; 1236 rc->rc_tried = 0; 1237 rc->rc_skipped = 0; 1238 rc->rc_repair = 0; 1239 rc->rc_need_orig_restore = B_FALSE; 1240 1241 if (q == 0 && i >= bc) 1242 rc->rc_size = 0; 1243 else if (i < bc) 1244 rc->rc_size = (q + 1) << ashift; 1245 else 1246 rc->rc_size = q << ashift; 1247 1248 asize += rc->rc_size; 1249 } 1250 1251 ASSERT3U(asize, ==, tot << ashift); 1252 rr->rr_nempty = roundup(tot, groupwidth) - tot; 1253 IMPLY(bc > 0, rr->rr_nempty == groupwidth - bc); 1254 1255 /* Allocate buffers for the parity columns */ 1256 for (uint64_t c = 0; c < rr->rr_firstdatacol; c++) { 1257 raidz_col_t *rc = &rr->rr_col[c]; 1258 rc->rc_abd = abd_alloc_linear(rc->rc_size, B_FALSE); 1259 } 1260 1261 /* 1262 * Map buffers for data columns and allocate/map buffers for skip 1263 * sectors. There are three distinct cases for dRAID which are 1264 * required to support sequential rebuild. 1265 */ 1266 if (zio->io_type == ZIO_TYPE_WRITE) { 1267 vdev_draid_map_alloc_write(zio, abd_offset, rr); 1268 } else if ((rr->rr_nempty > 0) && 1269 (zio->io_flags & (ZIO_FLAG_SCRUB | ZIO_FLAG_RESILVER))) { 1270 vdev_draid_map_alloc_scrub(zio, abd_offset, rr); 1271 } else { 1272 ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ); 1273 vdev_draid_map_alloc_read(zio, abd_offset, rr); 1274 } 1275 1276 return (io_size); 1277 } 1278 1279 /* 1280 * Allocate the raidz mapping to be applied to the dRAID I/O. The parity 1281 * calculations for dRAID are identical to raidz however there are a few 1282 * differences in the layout. 1283 * 1284 * - dRAID always allocates a full stripe width. Any extra sectors due 1285 * this padding are zero filled and written to disk. They will be read 1286 * back during a scrub or repair operation since they are included in 1287 * the parity calculation. This property enables sequential resilvering. 1288 * 1289 * - When the block at the logical offset spans redundancy groups then two 1290 * rows are allocated in the raidz_map_t. One row resides at the end of 1291 * the first group and the other at the start of the following group. 1292 */ 1293 static raidz_map_t * 1294 vdev_draid_map_alloc(zio_t *zio) 1295 { 1296 raidz_row_t *rr[2]; 1297 uint64_t abd_offset = 0; 1298 uint64_t abd_size = zio->io_size; 1299 uint64_t io_offset = zio->io_offset; 1300 uint64_t size; 1301 int nrows = 1; 1302 1303 size = vdev_draid_map_alloc_row(zio, &rr[0], io_offset, 1304 abd_offset, abd_size); 1305 if (size < abd_size) { 1306 vdev_t *vd = zio->io_vd; 1307 1308 io_offset += vdev_draid_asize(vd, size); 1309 abd_offset += size; 1310 abd_size -= size; 1311 nrows++; 1312 1313 ASSERT3U(io_offset, ==, vdev_draid_group_to_offset( 1314 vd, vdev_draid_offset_to_group(vd, io_offset))); 1315 ASSERT3U(abd_offset, <, zio->io_size); 1316 ASSERT3U(abd_size, !=, 0); 1317 1318 size = vdev_draid_map_alloc_row(zio, &rr[1], 1319 io_offset, abd_offset, abd_size); 1320 VERIFY3U(size, ==, abd_size); 1321 } 1322 1323 raidz_map_t *rm; 1324 rm = kmem_zalloc(offsetof(raidz_map_t, rm_row[nrows]), KM_SLEEP); 1325 rm->rm_ops = vdev_raidz_math_get_ops(); 1326 rm->rm_nrows = nrows; 1327 rm->rm_row[0] = rr[0]; 1328 if (nrows == 2) 1329 rm->rm_row[1] = rr[1]; 1330 1331 zio->io_vsd = rm; 1332 zio->io_vsd_ops = &vdev_draid_vsd_ops; 1333 1334 return (rm); 1335 } 1336 1337 /* 1338 * Given an offset into a dRAID return the next group width aligned offset 1339 * which can be used to start an allocation. 1340 */ 1341 static uint64_t 1342 vdev_draid_get_astart(vdev_t *vd, const uint64_t start) 1343 { 1344 vdev_draid_config_t *vdc = vd->vdev_tsd; 1345 1346 ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops); 1347 1348 return (roundup(start, vdc->vdc_groupwidth << vd->vdev_ashift)); 1349 } 1350 1351 /* 1352 * Allocatable space for dRAID is (children - nspares) * sizeof(smallest child) 1353 * rounded down to the last full slice. So each child must provide at least 1354 * 1 / (children - nspares) of its asize. 1355 */ 1356 static uint64_t 1357 vdev_draid_min_asize(vdev_t *vd) 1358 { 1359 vdev_draid_config_t *vdc = vd->vdev_tsd; 1360 1361 ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops); 1362 1363 return ((vd->vdev_min_asize + vdc->vdc_ndisks - 1) / (vdc->vdc_ndisks)); 1364 } 1365 1366 /* 1367 * When using dRAID the minimum allocation size is determined by the number 1368 * of data disks in the redundancy group. Full stripes are always used. 1369 */ 1370 static uint64_t 1371 vdev_draid_min_alloc(vdev_t *vd) 1372 { 1373 vdev_draid_config_t *vdc = vd->vdev_tsd; 1374 1375 ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops); 1376 1377 return (vdc->vdc_ndata << vd->vdev_ashift); 1378 } 1379 1380 /* 1381 * Returns true if the txg range does not exist on any leaf vdev. 1382 * 1383 * A dRAID spare does not fit into the DTL model. While it has child vdevs 1384 * there is no redundancy among them, and the effective child vdev is 1385 * determined by offset. Essentially we do a vdev_dtl_reassess() on the 1386 * fly by replacing a dRAID spare with the child vdev under the offset. 1387 * Note that it is a recursive process because the child vdev can be 1388 * another dRAID spare and so on. 1389 */ 1390 boolean_t 1391 vdev_draid_missing(vdev_t *vd, uint64_t physical_offset, uint64_t txg, 1392 uint64_t size) 1393 { 1394 if (vd->vdev_ops == &vdev_spare_ops || 1395 vd->vdev_ops == &vdev_replacing_ops) { 1396 /* 1397 * Check all of the readable children, if any child 1398 * contains the txg range the data it is not missing. 1399 */ 1400 for (int c = 0; c < vd->vdev_children; c++) { 1401 vdev_t *cvd = vd->vdev_child[c]; 1402 1403 if (!vdev_readable(cvd)) 1404 continue; 1405 1406 if (!vdev_draid_missing(cvd, physical_offset, 1407 txg, size)) 1408 return (B_FALSE); 1409 } 1410 1411 return (B_TRUE); 1412 } 1413 1414 if (vd->vdev_ops == &vdev_draid_spare_ops) { 1415 /* 1416 * When sequentially resilvering we don't have a proper 1417 * txg range so instead we must presume all txgs are 1418 * missing on this vdev until the resilver completes. 1419 */ 1420 if (vd->vdev_rebuild_txg != 0) 1421 return (B_TRUE); 1422 1423 /* 1424 * DTL_MISSING is set for all prior txgs when a resilver 1425 * is started in spa_vdev_attach(). 1426 */ 1427 if (vdev_dtl_contains(vd, DTL_MISSING, txg, size)) 1428 return (B_TRUE); 1429 1430 /* 1431 * Consult the DTL on the relevant vdev. Either a vdev 1432 * leaf or spare/replace mirror child may be returned so 1433 * we must recursively call vdev_draid_missing_impl(). 1434 */ 1435 vd = vdev_draid_spare_get_child(vd, physical_offset); 1436 if (vd == NULL) 1437 return (B_TRUE); 1438 1439 return (vdev_draid_missing(vd, physical_offset, 1440 txg, size)); 1441 } 1442 1443 return (vdev_dtl_contains(vd, DTL_MISSING, txg, size)); 1444 } 1445 1446 /* 1447 * Returns true if the txg is only partially replicated on the leaf vdevs. 1448 */ 1449 static boolean_t 1450 vdev_draid_partial(vdev_t *vd, uint64_t physical_offset, uint64_t txg, 1451 uint64_t size) 1452 { 1453 if (vd->vdev_ops == &vdev_spare_ops || 1454 vd->vdev_ops == &vdev_replacing_ops) { 1455 /* 1456 * Check all of the readable children, if any child is 1457 * missing the txg range then it is partially replicated. 1458 */ 1459 for (int c = 0; c < vd->vdev_children; c++) { 1460 vdev_t *cvd = vd->vdev_child[c]; 1461 1462 if (!vdev_readable(cvd)) 1463 continue; 1464 1465 if (vdev_draid_partial(cvd, physical_offset, txg, size)) 1466 return (B_TRUE); 1467 } 1468 1469 return (B_FALSE); 1470 } 1471 1472 if (vd->vdev_ops == &vdev_draid_spare_ops) { 1473 /* 1474 * When sequentially resilvering we don't have a proper 1475 * txg range so instead we must presume all txgs are 1476 * missing on this vdev until the resilver completes. 1477 */ 1478 if (vd->vdev_rebuild_txg != 0) 1479 return (B_TRUE); 1480 1481 /* 1482 * DTL_MISSING is set for all prior txgs when a resilver 1483 * is started in spa_vdev_attach(). 1484 */ 1485 if (vdev_dtl_contains(vd, DTL_MISSING, txg, size)) 1486 return (B_TRUE); 1487 1488 /* 1489 * Consult the DTL on the relevant vdev. Either a vdev 1490 * leaf or spare/replace mirror child may be returned so 1491 * we must recursively call vdev_draid_missing_impl(). 1492 */ 1493 vd = vdev_draid_spare_get_child(vd, physical_offset); 1494 if (vd == NULL) 1495 return (B_TRUE); 1496 1497 return (vdev_draid_partial(vd, physical_offset, txg, size)); 1498 } 1499 1500 return (vdev_dtl_contains(vd, DTL_MISSING, txg, size)); 1501 } 1502 1503 /* 1504 * Determine if the vdev is readable at the given offset. 1505 */ 1506 boolean_t 1507 vdev_draid_readable(vdev_t *vd, uint64_t physical_offset) 1508 { 1509 if (vd->vdev_ops == &vdev_draid_spare_ops) { 1510 vd = vdev_draid_spare_get_child(vd, physical_offset); 1511 if (vd == NULL) 1512 return (B_FALSE); 1513 } 1514 1515 if (vd->vdev_ops == &vdev_spare_ops || 1516 vd->vdev_ops == &vdev_replacing_ops) { 1517 1518 for (int c = 0; c < vd->vdev_children; c++) { 1519 vdev_t *cvd = vd->vdev_child[c]; 1520 1521 if (!vdev_readable(cvd)) 1522 continue; 1523 1524 if (vdev_draid_readable(cvd, physical_offset)) 1525 return (B_TRUE); 1526 } 1527 1528 return (B_FALSE); 1529 } 1530 1531 return (vdev_readable(vd)); 1532 } 1533 1534 /* 1535 * Returns the first distributed spare found under the provided vdev tree. 1536 */ 1537 static vdev_t * 1538 vdev_draid_find_spare(vdev_t *vd) 1539 { 1540 if (vd->vdev_ops == &vdev_draid_spare_ops) 1541 return (vd); 1542 1543 for (int c = 0; c < vd->vdev_children; c++) { 1544 vdev_t *svd = vdev_draid_find_spare(vd->vdev_child[c]); 1545 if (svd != NULL) 1546 return (svd); 1547 } 1548 1549 return (NULL); 1550 } 1551 1552 /* 1553 * Returns B_TRUE if the passed in vdev is currently "faulted". 1554 * Faulted, in this context, means that the vdev represents a 1555 * replacing or sparing vdev tree. 1556 */ 1557 static boolean_t 1558 vdev_draid_faulted(vdev_t *vd, uint64_t physical_offset) 1559 { 1560 if (vd->vdev_ops == &vdev_draid_spare_ops) { 1561 vd = vdev_draid_spare_get_child(vd, physical_offset); 1562 if (vd == NULL) 1563 return (B_FALSE); 1564 1565 /* 1566 * After resolving the distributed spare to a leaf vdev 1567 * check the parent to determine if it's "faulted". 1568 */ 1569 vd = vd->vdev_parent; 1570 } 1571 1572 return (vd->vdev_ops == &vdev_replacing_ops || 1573 vd->vdev_ops == &vdev_spare_ops); 1574 } 1575 1576 /* 1577 * Determine if the dRAID block at the logical offset is degraded. 1578 * Used by sequential resilver. 1579 */ 1580 static boolean_t 1581 vdev_draid_group_degraded(vdev_t *vd, uint64_t offset) 1582 { 1583 vdev_draid_config_t *vdc = vd->vdev_tsd; 1584 1585 ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops); 1586 ASSERT3U(vdev_draid_get_astart(vd, offset), ==, offset); 1587 1588 uint64_t groupstart, perm; 1589 uint64_t physical_offset = vdev_draid_logical_to_physical(vd, 1590 offset, &perm, &groupstart); 1591 1592 uint8_t *base; 1593 uint64_t iter; 1594 vdev_draid_get_perm(vdc, perm, &base, &iter); 1595 1596 for (uint64_t i = 0; i < vdc->vdc_groupwidth; i++) { 1597 uint64_t c = (groupstart + i) % vdc->vdc_ndisks; 1598 uint64_t cid = vdev_draid_permute_id(vdc, base, iter, c); 1599 vdev_t *cvd = vd->vdev_child[cid]; 1600 1601 /* Group contains a faulted vdev. */ 1602 if (vdev_draid_faulted(cvd, physical_offset)) 1603 return (B_TRUE); 1604 1605 /* 1606 * Always check groups with active distributed spares 1607 * because any vdev failure in the pool will affect them. 1608 */ 1609 if (vdev_draid_find_spare(cvd) != NULL) 1610 return (B_TRUE); 1611 } 1612 1613 return (B_FALSE); 1614 } 1615 1616 /* 1617 * Determine if the txg is missing. Used by healing resilver. 1618 */ 1619 static boolean_t 1620 vdev_draid_group_missing(vdev_t *vd, uint64_t offset, uint64_t txg, 1621 uint64_t size) 1622 { 1623 vdev_draid_config_t *vdc = vd->vdev_tsd; 1624 1625 ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops); 1626 ASSERT3U(vdev_draid_get_astart(vd, offset), ==, offset); 1627 1628 uint64_t groupstart, perm; 1629 uint64_t physical_offset = vdev_draid_logical_to_physical(vd, 1630 offset, &perm, &groupstart); 1631 1632 uint8_t *base; 1633 uint64_t iter; 1634 vdev_draid_get_perm(vdc, perm, &base, &iter); 1635 1636 for (uint64_t i = 0; i < vdc->vdc_groupwidth; i++) { 1637 uint64_t c = (groupstart + i) % vdc->vdc_ndisks; 1638 uint64_t cid = vdev_draid_permute_id(vdc, base, iter, c); 1639 vdev_t *cvd = vd->vdev_child[cid]; 1640 1641 /* Transaction group is known to be partially replicated. */ 1642 if (vdev_draid_partial(cvd, physical_offset, txg, size)) 1643 return (B_TRUE); 1644 1645 /* 1646 * Always check groups with active distributed spares 1647 * because any vdev failure in the pool will affect them. 1648 */ 1649 if (vdev_draid_find_spare(cvd) != NULL) 1650 return (B_TRUE); 1651 } 1652 1653 return (B_FALSE); 1654 } 1655 1656 /* 1657 * Find the smallest child asize and largest sector size to calculate the 1658 * available capacity. Distributed spares are ignored since their capacity 1659 * is also based of the minimum child size in the top-level dRAID. 1660 */ 1661 static void 1662 vdev_draid_calculate_asize(vdev_t *vd, uint64_t *asizep, uint64_t *max_asizep, 1663 uint64_t *logical_ashiftp, uint64_t *physical_ashiftp) 1664 { 1665 uint64_t logical_ashift = 0, physical_ashift = 0; 1666 uint64_t asize = 0, max_asize = 0; 1667 1668 ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops); 1669 1670 for (int c = 0; c < vd->vdev_children; c++) { 1671 vdev_t *cvd = vd->vdev_child[c]; 1672 1673 if (cvd->vdev_ops == &vdev_draid_spare_ops) 1674 continue; 1675 1676 asize = MIN(asize - 1, cvd->vdev_asize - 1) + 1; 1677 max_asize = MIN(max_asize - 1, cvd->vdev_max_asize - 1) + 1; 1678 logical_ashift = MAX(logical_ashift, cvd->vdev_ashift); 1679 physical_ashift = MAX(physical_ashift, 1680 cvd->vdev_physical_ashift); 1681 } 1682 1683 *asizep = asize; 1684 *max_asizep = max_asize; 1685 *logical_ashiftp = logical_ashift; 1686 *physical_ashiftp = physical_ashift; 1687 } 1688 1689 /* 1690 * Open spare vdevs. 1691 */ 1692 static boolean_t 1693 vdev_draid_open_spares(vdev_t *vd) 1694 { 1695 return (vd->vdev_ops == &vdev_draid_spare_ops || 1696 vd->vdev_ops == &vdev_replacing_ops || 1697 vd->vdev_ops == &vdev_spare_ops); 1698 } 1699 1700 /* 1701 * Open all children, excluding spares. 1702 */ 1703 static boolean_t 1704 vdev_draid_open_children(vdev_t *vd) 1705 { 1706 return (!vdev_draid_open_spares(vd)); 1707 } 1708 1709 /* 1710 * Open a top-level dRAID vdev. 1711 */ 1712 static int 1713 vdev_draid_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize, 1714 uint64_t *logical_ashift, uint64_t *physical_ashift) 1715 { 1716 vdev_draid_config_t *vdc = vd->vdev_tsd; 1717 uint64_t nparity = vdc->vdc_nparity; 1718 int open_errors = 0; 1719 1720 if (nparity > VDEV_DRAID_MAXPARITY || 1721 vd->vdev_children < nparity + 1) { 1722 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL; 1723 return (SET_ERROR(EINVAL)); 1724 } 1725 1726 /* 1727 * First open the normal children then the distributed spares. This 1728 * ordering is important to ensure the distributed spares calculate 1729 * the correct psize in the event that the dRAID vdevs were expanded. 1730 */ 1731 vdev_open_children_subset(vd, vdev_draid_open_children); 1732 vdev_open_children_subset(vd, vdev_draid_open_spares); 1733 1734 /* Verify enough of the children are available to continue. */ 1735 for (int c = 0; c < vd->vdev_children; c++) { 1736 if (vd->vdev_child[c]->vdev_open_error != 0) { 1737 if ((++open_errors) > nparity) { 1738 vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS; 1739 return (SET_ERROR(ENXIO)); 1740 } 1741 } 1742 } 1743 1744 /* 1745 * Allocatable capacity is the sum of the space on all children less 1746 * the number of distributed spares rounded down to last full row 1747 * and then to the last full group. An additional 32MB of scratch 1748 * space is reserved at the end of each child for use by the dRAID 1749 * expansion feature. 1750 */ 1751 uint64_t child_asize, child_max_asize; 1752 vdev_draid_calculate_asize(vd, &child_asize, &child_max_asize, 1753 logical_ashift, physical_ashift); 1754 1755 /* 1756 * Should be unreachable since the minimum child size is 64MB, but 1757 * we want to make sure an underflow absolutely cannot occur here. 1758 */ 1759 if (child_asize < VDEV_DRAID_REFLOW_RESERVE || 1760 child_max_asize < VDEV_DRAID_REFLOW_RESERVE) { 1761 return (SET_ERROR(ENXIO)); 1762 } 1763 1764 child_asize = ((child_asize - VDEV_DRAID_REFLOW_RESERVE) / 1765 VDEV_DRAID_ROWHEIGHT) * VDEV_DRAID_ROWHEIGHT; 1766 child_max_asize = ((child_max_asize - VDEV_DRAID_REFLOW_RESERVE) / 1767 VDEV_DRAID_ROWHEIGHT) * VDEV_DRAID_ROWHEIGHT; 1768 1769 *asize = (((child_asize * vdc->vdc_ndisks) / vdc->vdc_groupsz) * 1770 vdc->vdc_groupsz); 1771 *max_asize = (((child_max_asize * vdc->vdc_ndisks) / vdc->vdc_groupsz) * 1772 vdc->vdc_groupsz); 1773 1774 return (0); 1775 } 1776 1777 /* 1778 * Close a top-level dRAID vdev. 1779 */ 1780 static void 1781 vdev_draid_close(vdev_t *vd) 1782 { 1783 for (int c = 0; c < vd->vdev_children; c++) { 1784 if (vd->vdev_child[c] != NULL) 1785 vdev_close(vd->vdev_child[c]); 1786 } 1787 } 1788 1789 /* 1790 * Return the maximum asize for a rebuild zio in the provided range 1791 * given the following constraints. A dRAID chunks may not: 1792 * 1793 * - Exceed the maximum allowed block size (SPA_MAXBLOCKSIZE), or 1794 * - Span dRAID redundancy groups. 1795 */ 1796 static uint64_t 1797 vdev_draid_rebuild_asize(vdev_t *vd, uint64_t start, uint64_t asize, 1798 uint64_t max_segment) 1799 { 1800 vdev_draid_config_t *vdc = vd->vdev_tsd; 1801 1802 ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops); 1803 1804 uint64_t ashift = vd->vdev_ashift; 1805 uint64_t ndata = vdc->vdc_ndata; 1806 uint64_t psize = MIN(P2ROUNDUP(max_segment * ndata, 1 << ashift), 1807 SPA_MAXBLOCKSIZE); 1808 1809 ASSERT3U(vdev_draid_get_astart(vd, start), ==, start); 1810 ASSERT3U(asize % (vdc->vdc_groupwidth << ashift), ==, 0); 1811 1812 /* Chunks must evenly span all data columns in the group. */ 1813 psize = (((psize >> ashift) / ndata) * ndata) << ashift; 1814 uint64_t chunk_size = MIN(asize, vdev_psize_to_asize(vd, psize)); 1815 1816 /* Reduce the chunk size to the group space remaining. */ 1817 uint64_t group = vdev_draid_offset_to_group(vd, start); 1818 uint64_t left = vdev_draid_group_to_offset(vd, group + 1) - start; 1819 chunk_size = MIN(chunk_size, left); 1820 1821 ASSERT3U(chunk_size % (vdc->vdc_groupwidth << ashift), ==, 0); 1822 ASSERT3U(vdev_draid_offset_to_group(vd, start), ==, 1823 vdev_draid_offset_to_group(vd, start + chunk_size - 1)); 1824 1825 return (chunk_size); 1826 } 1827 1828 /* 1829 * Align the start of the metaslab to the group width and slightly reduce 1830 * its size to a multiple of the group width. Since full stripe writes are 1831 * required by dRAID this space is unallocable. Furthermore, aligning the 1832 * metaslab start is important for vdev initialize and TRIM which both operate 1833 * on metaslab boundaries which vdev_xlate() expects to be aligned. 1834 */ 1835 static void 1836 vdev_draid_metaslab_init(vdev_t *vd, uint64_t *ms_start, uint64_t *ms_size) 1837 { 1838 vdev_draid_config_t *vdc = vd->vdev_tsd; 1839 1840 ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops); 1841 1842 uint64_t sz = vdc->vdc_groupwidth << vd->vdev_ashift; 1843 uint64_t astart = vdev_draid_get_astart(vd, *ms_start); 1844 uint64_t asize = ((*ms_size - (astart - *ms_start)) / sz) * sz; 1845 1846 *ms_start = astart; 1847 *ms_size = asize; 1848 1849 ASSERT0(*ms_start % sz); 1850 ASSERT0(*ms_size % sz); 1851 } 1852 1853 /* 1854 * Add virtual dRAID spares to the list of valid spares. In order to accomplish 1855 * this the existing array must be freed and reallocated with the additional 1856 * entries. 1857 */ 1858 int 1859 vdev_draid_spare_create(nvlist_t *nvroot, vdev_t *vd, uint64_t *ndraidp, 1860 uint64_t next_vdev_id) 1861 { 1862 uint64_t draid_nspares = 0; 1863 uint64_t ndraid = 0; 1864 int error; 1865 1866 for (uint64_t i = 0; i < vd->vdev_children; i++) { 1867 vdev_t *cvd = vd->vdev_child[i]; 1868 1869 if (cvd->vdev_ops == &vdev_draid_ops) { 1870 vdev_draid_config_t *vdc = cvd->vdev_tsd; 1871 draid_nspares += vdc->vdc_nspares; 1872 ndraid++; 1873 } 1874 } 1875 1876 if (draid_nspares == 0) { 1877 *ndraidp = ndraid; 1878 return (0); 1879 } 1880 1881 nvlist_t **old_spares, **new_spares; 1882 uint_t old_nspares; 1883 error = nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, 1884 &old_spares, &old_nspares); 1885 if (error) 1886 old_nspares = 0; 1887 1888 /* Allocate memory and copy of the existing spares. */ 1889 new_spares = kmem_alloc(sizeof (nvlist_t *) * 1890 (draid_nspares + old_nspares), KM_SLEEP); 1891 for (uint_t i = 0; i < old_nspares; i++) 1892 new_spares[i] = fnvlist_dup(old_spares[i]); 1893 1894 /* Add new distributed spares to ZPOOL_CONFIG_SPARES. */ 1895 uint64_t n = old_nspares; 1896 for (uint64_t vdev_id = 0; vdev_id < vd->vdev_children; vdev_id++) { 1897 vdev_t *cvd = vd->vdev_child[vdev_id]; 1898 char path[64]; 1899 1900 if (cvd->vdev_ops != &vdev_draid_ops) 1901 continue; 1902 1903 vdev_draid_config_t *vdc = cvd->vdev_tsd; 1904 uint64_t nspares = vdc->vdc_nspares; 1905 uint64_t nparity = vdc->vdc_nparity; 1906 1907 for (uint64_t spare_id = 0; spare_id < nspares; spare_id++) { 1908 bzero(path, sizeof (path)); 1909 (void) snprintf(path, sizeof (path) - 1, 1910 "%s%llu-%llu-%llu", VDEV_TYPE_DRAID, 1911 (u_longlong_t)nparity, 1912 (u_longlong_t)next_vdev_id + vdev_id, 1913 (u_longlong_t)spare_id); 1914 1915 nvlist_t *spare = fnvlist_alloc(); 1916 fnvlist_add_string(spare, ZPOOL_CONFIG_PATH, path); 1917 fnvlist_add_string(spare, ZPOOL_CONFIG_TYPE, 1918 VDEV_TYPE_DRAID_SPARE); 1919 fnvlist_add_uint64(spare, ZPOOL_CONFIG_TOP_GUID, 1920 cvd->vdev_guid); 1921 fnvlist_add_uint64(spare, ZPOOL_CONFIG_SPARE_ID, 1922 spare_id); 1923 fnvlist_add_uint64(spare, ZPOOL_CONFIG_IS_LOG, 0); 1924 fnvlist_add_uint64(spare, ZPOOL_CONFIG_IS_SPARE, 1); 1925 fnvlist_add_uint64(spare, ZPOOL_CONFIG_WHOLE_DISK, 1); 1926 fnvlist_add_uint64(spare, ZPOOL_CONFIG_ASHIFT, 1927 cvd->vdev_ashift); 1928 1929 new_spares[n] = spare; 1930 n++; 1931 } 1932 } 1933 1934 if (n > 0) { 1935 (void) nvlist_remove_all(nvroot, ZPOOL_CONFIG_SPARES); 1936 fnvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, 1937 new_spares, n); 1938 } 1939 1940 for (int i = 0; i < n; i++) 1941 nvlist_free(new_spares[i]); 1942 1943 kmem_free(new_spares, sizeof (*new_spares) * n); 1944 *ndraidp = ndraid; 1945 1946 return (0); 1947 } 1948 1949 /* 1950 * Determine if any portion of the provided block resides on a child vdev 1951 * with a dirty DTL and therefore needs to be resilvered. 1952 */ 1953 static boolean_t 1954 vdev_draid_need_resilver(vdev_t *vd, const dva_t *dva, size_t psize, 1955 uint64_t phys_birth) 1956 { 1957 uint64_t offset = DVA_GET_OFFSET(dva); 1958 uint64_t asize = vdev_draid_asize(vd, psize); 1959 1960 if (phys_birth == TXG_UNKNOWN) { 1961 /* 1962 * Sequential resilver. There is no meaningful phys_birth 1963 * for this block, we can only determine if block resides 1964 * in a degraded group in which case it must be resilvered. 1965 */ 1966 ASSERT3U(vdev_draid_offset_to_group(vd, offset), ==, 1967 vdev_draid_offset_to_group(vd, offset + asize - 1)); 1968 1969 return (vdev_draid_group_degraded(vd, offset)); 1970 } else { 1971 /* 1972 * Healing resilver. TXGs not in DTL_PARTIAL are intact, 1973 * as are blocks in non-degraded groups. 1974 */ 1975 if (!vdev_dtl_contains(vd, DTL_PARTIAL, phys_birth, 1)) 1976 return (B_FALSE); 1977 1978 if (vdev_draid_group_missing(vd, offset, phys_birth, 1)) 1979 return (B_TRUE); 1980 1981 /* The block may span groups in which case check both. */ 1982 if (vdev_draid_offset_to_group(vd, offset) != 1983 vdev_draid_offset_to_group(vd, offset + asize - 1)) { 1984 if (vdev_draid_group_missing(vd, 1985 offset + asize, phys_birth, 1)) 1986 return (B_TRUE); 1987 } 1988 1989 return (B_FALSE); 1990 } 1991 } 1992 1993 static boolean_t 1994 vdev_draid_rebuilding(vdev_t *vd) 1995 { 1996 if (vd->vdev_ops->vdev_op_leaf && vd->vdev_rebuild_txg) 1997 return (B_TRUE); 1998 1999 for (int i = 0; i < vd->vdev_children; i++) { 2000 if (vdev_draid_rebuilding(vd->vdev_child[i])) { 2001 return (B_TRUE); 2002 } 2003 } 2004 2005 return (B_FALSE); 2006 } 2007 2008 static void 2009 vdev_draid_io_verify(vdev_t *vd, raidz_row_t *rr, int col) 2010 { 2011 #ifdef ZFS_DEBUG 2012 range_seg64_t logical_rs, physical_rs, remain_rs; 2013 logical_rs.rs_start = rr->rr_offset; 2014 logical_rs.rs_end = logical_rs.rs_start + 2015 vdev_draid_asize(vd, rr->rr_size); 2016 2017 raidz_col_t *rc = &rr->rr_col[col]; 2018 vdev_t *cvd = vd->vdev_child[rc->rc_devidx]; 2019 2020 vdev_xlate(cvd, &logical_rs, &physical_rs, &remain_rs); 2021 ASSERT(vdev_xlate_is_empty(&remain_rs)); 2022 ASSERT3U(rc->rc_offset, ==, physical_rs.rs_start); 2023 ASSERT3U(rc->rc_offset, <, physical_rs.rs_end); 2024 ASSERT3U(rc->rc_offset + rc->rc_size, ==, physical_rs.rs_end); 2025 #endif 2026 } 2027 2028 /* 2029 * For write operations: 2030 * 1. Generate the parity data 2031 * 2. Create child zio write operations to each column's vdev, for both 2032 * data and parity. A gang ABD is allocated by vdev_draid_map_alloc() 2033 * if a skip sector needs to be added to a column. 2034 */ 2035 static void 2036 vdev_draid_io_start_write(zio_t *zio, raidz_row_t *rr) 2037 { 2038 vdev_t *vd = zio->io_vd; 2039 raidz_map_t *rm = zio->io_vsd; 2040 2041 vdev_raidz_generate_parity_row(rm, rr); 2042 2043 for (int c = 0; c < rr->rr_cols; c++) { 2044 raidz_col_t *rc = &rr->rr_col[c]; 2045 2046 /* 2047 * Empty columns are zero filled and included in the parity 2048 * calculation and therefore must be written. 2049 */ 2050 ASSERT3U(rc->rc_size, !=, 0); 2051 2052 /* Verify physical to logical translation */ 2053 vdev_draid_io_verify(vd, rr, c); 2054 2055 zio_nowait(zio_vdev_child_io(zio, NULL, 2056 vd->vdev_child[rc->rc_devidx], rc->rc_offset, 2057 rc->rc_abd, rc->rc_size, zio->io_type, zio->io_priority, 2058 0, vdev_raidz_child_done, rc)); 2059 } 2060 } 2061 2062 /* 2063 * For read operations: 2064 * 1. The vdev_draid_map_alloc() function will create a minimal raidz 2065 * mapping for the read based on the zio->io_flags. There are two 2066 * possible mappings either 1) a normal read, or 2) a scrub/resilver. 2067 * 2. Create the zio read operations. This will include all parity 2068 * columns and skip sectors for a scrub/resilver. 2069 */ 2070 static void 2071 vdev_draid_io_start_read(zio_t *zio, raidz_row_t *rr) 2072 { 2073 vdev_t *vd = zio->io_vd; 2074 2075 /* Sequential rebuild must do IO at redundancy group boundary. */ 2076 IMPLY(zio->io_priority == ZIO_PRIORITY_REBUILD, rr->rr_nempty == 0); 2077 2078 /* 2079 * Iterate over the columns in reverse order so that we hit the parity 2080 * last. Any errors along the way will force us to read the parity. 2081 * For scrub/resilver IOs which verify skip sectors, a gang ABD will 2082 * have been allocated to store them and rc->rc_size is increased. 2083 */ 2084 for (int c = rr->rr_cols - 1; c >= 0; c--) { 2085 raidz_col_t *rc = &rr->rr_col[c]; 2086 vdev_t *cvd = vd->vdev_child[rc->rc_devidx]; 2087 2088 if (!vdev_draid_readable(cvd, rc->rc_offset)) { 2089 if (c >= rr->rr_firstdatacol) 2090 rr->rr_missingdata++; 2091 else 2092 rr->rr_missingparity++; 2093 rc->rc_error = SET_ERROR(ENXIO); 2094 rc->rc_tried = 1; 2095 rc->rc_skipped = 1; 2096 continue; 2097 } 2098 2099 if (vdev_draid_missing(cvd, rc->rc_offset, zio->io_txg, 1)) { 2100 if (c >= rr->rr_firstdatacol) 2101 rr->rr_missingdata++; 2102 else 2103 rr->rr_missingparity++; 2104 rc->rc_error = SET_ERROR(ESTALE); 2105 rc->rc_skipped = 1; 2106 continue; 2107 } 2108 2109 /* 2110 * Empty columns may be read during vdev_draid_io_done(). 2111 * Only skip them after the readable and missing checks 2112 * verify they are available. 2113 */ 2114 if (rc->rc_size == 0) { 2115 rc->rc_skipped = 1; 2116 continue; 2117 } 2118 2119 if (zio->io_flags & ZIO_FLAG_RESILVER) { 2120 vdev_t *svd; 2121 2122 /* 2123 * If this child is a distributed spare then the 2124 * offset might reside on the vdev being replaced. 2125 * In which case this data must be written to the 2126 * new device. Failure to do so would result in 2127 * checksum errors when the old device is detached 2128 * and the pool is scrubbed. 2129 */ 2130 if ((svd = vdev_draid_find_spare(cvd)) != NULL) { 2131 svd = vdev_draid_spare_get_child(svd, 2132 rc->rc_offset); 2133 if (svd && (svd->vdev_ops == &vdev_spare_ops || 2134 svd->vdev_ops == &vdev_replacing_ops)) { 2135 rc->rc_repair = 1; 2136 } 2137 } 2138 2139 /* 2140 * Always issue a repair IO to this child when its 2141 * a spare or replacing vdev with an active rebuild. 2142 */ 2143 if ((cvd->vdev_ops == &vdev_spare_ops || 2144 cvd->vdev_ops == &vdev_replacing_ops) && 2145 vdev_draid_rebuilding(cvd)) { 2146 rc->rc_repair = 1; 2147 } 2148 } 2149 } 2150 2151 /* 2152 * Either a parity or data column is missing this means a repair 2153 * may be attempted by vdev_draid_io_done(). Expand the raid map 2154 * to read in empty columns which are needed along with the parity 2155 * during reconstruction. 2156 */ 2157 if ((rr->rr_missingdata > 0 || rr->rr_missingparity > 0) && 2158 rr->rr_nempty > 0 && rr->rr_abd_empty == NULL) { 2159 vdev_draid_map_alloc_empty(zio, rr); 2160 } 2161 2162 for (int c = rr->rr_cols - 1; c >= 0; c--) { 2163 raidz_col_t *rc = &rr->rr_col[c]; 2164 vdev_t *cvd = vd->vdev_child[rc->rc_devidx]; 2165 2166 if (rc->rc_error || rc->rc_size == 0) 2167 continue; 2168 2169 if (c >= rr->rr_firstdatacol || rr->rr_missingdata > 0 || 2170 (zio->io_flags & (ZIO_FLAG_SCRUB | ZIO_FLAG_RESILVER))) { 2171 zio_nowait(zio_vdev_child_io(zio, NULL, cvd, 2172 rc->rc_offset, rc->rc_abd, rc->rc_size, 2173 zio->io_type, zio->io_priority, 0, 2174 vdev_raidz_child_done, rc)); 2175 } 2176 } 2177 } 2178 2179 /* 2180 * Start an IO operation to a dRAID vdev. 2181 */ 2182 static void 2183 vdev_draid_io_start(zio_t *zio) 2184 { 2185 vdev_t *vd __maybe_unused = zio->io_vd; 2186 raidz_map_t *rm; 2187 2188 ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops); 2189 ASSERT3U(zio->io_offset, ==, vdev_draid_get_astart(vd, zio->io_offset)); 2190 2191 rm = vdev_draid_map_alloc(zio); 2192 2193 if (zio->io_type == ZIO_TYPE_WRITE) { 2194 for (int i = 0; i < rm->rm_nrows; i++) { 2195 vdev_draid_io_start_write(zio, rm->rm_row[i]); 2196 } 2197 } else { 2198 ASSERT(zio->io_type == ZIO_TYPE_READ); 2199 2200 for (int i = 0; i < rm->rm_nrows; i++) { 2201 vdev_draid_io_start_read(zio, rm->rm_row[i]); 2202 } 2203 } 2204 2205 zio_execute(zio); 2206 } 2207 2208 /* 2209 * Complete an IO operation on a dRAID vdev. The raidz logic can be applied 2210 * to dRAID since the layout is fully described by the raidz_map_t. 2211 */ 2212 static void 2213 vdev_draid_io_done(zio_t *zio) 2214 { 2215 vdev_raidz_io_done(zio); 2216 } 2217 2218 static void 2219 vdev_draid_state_change(vdev_t *vd, int faulted, int degraded) 2220 { 2221 vdev_draid_config_t *vdc = vd->vdev_tsd; 2222 ASSERT(vd->vdev_ops == &vdev_draid_ops); 2223 2224 if (faulted > vdc->vdc_nparity) 2225 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN, 2226 VDEV_AUX_NO_REPLICAS); 2227 else if (degraded + faulted != 0) 2228 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE); 2229 else 2230 vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE); 2231 } 2232 2233 static void 2234 vdev_draid_xlate(vdev_t *cvd, const range_seg64_t *logical_rs, 2235 range_seg64_t *physical_rs, range_seg64_t *remain_rs) 2236 { 2237 vdev_t *raidvd = cvd->vdev_parent; 2238 ASSERT(raidvd->vdev_ops == &vdev_draid_ops); 2239 2240 vdev_draid_config_t *vdc = raidvd->vdev_tsd; 2241 uint64_t ashift = raidvd->vdev_top->vdev_ashift; 2242 2243 /* Make sure the offsets are block-aligned */ 2244 ASSERT0(logical_rs->rs_start % (1 << ashift)); 2245 ASSERT0(logical_rs->rs_end % (1 << ashift)); 2246 2247 uint64_t logical_start = logical_rs->rs_start; 2248 uint64_t logical_end = logical_rs->rs_end; 2249 2250 /* 2251 * Unaligned ranges must be skipped. All metaslabs are correctly 2252 * aligned so this should not happen, but this case is handled in 2253 * case it's needed by future callers. 2254 */ 2255 uint64_t astart = vdev_draid_get_astart(raidvd, logical_start); 2256 if (astart != logical_start) { 2257 physical_rs->rs_start = logical_start; 2258 physical_rs->rs_end = logical_start; 2259 remain_rs->rs_start = MIN(astart, logical_end); 2260 remain_rs->rs_end = logical_end; 2261 return; 2262 } 2263 2264 /* 2265 * Unlike with mirrors and raidz a dRAID logical range can map 2266 * to multiple non-contiguous physical ranges. This is handled by 2267 * limiting the size of the logical range to a single group and 2268 * setting the remain argument such that it describes the remaining 2269 * unmapped logical range. This is stricter than absolutely 2270 * necessary but helps simplify the logic below. 2271 */ 2272 uint64_t group = vdev_draid_offset_to_group(raidvd, logical_start); 2273 uint64_t nextstart = vdev_draid_group_to_offset(raidvd, group + 1); 2274 if (logical_end > nextstart) 2275 logical_end = nextstart; 2276 2277 /* Find the starting offset for each vdev in the group */ 2278 uint64_t perm, groupstart; 2279 uint64_t start = vdev_draid_logical_to_physical(raidvd, 2280 logical_start, &perm, &groupstart); 2281 uint64_t end = start; 2282 2283 uint8_t *base; 2284 uint64_t iter, id; 2285 vdev_draid_get_perm(vdc, perm, &base, &iter); 2286 2287 /* 2288 * Check if the passed child falls within the group. If it does 2289 * update the start and end to reflect the physical range. 2290 * Otherwise, leave them unmodified which will result in an empty 2291 * (zero-length) physical range being returned. 2292 */ 2293 for (uint64_t i = 0; i < vdc->vdc_groupwidth; i++) { 2294 uint64_t c = (groupstart + i) % vdc->vdc_ndisks; 2295 2296 if (c == 0 && i != 0) { 2297 /* the group wrapped, increment the start */ 2298 start += VDEV_DRAID_ROWHEIGHT; 2299 end = start; 2300 } 2301 2302 id = vdev_draid_permute_id(vdc, base, iter, c); 2303 if (id == cvd->vdev_id) { 2304 uint64_t b_size = (logical_end >> ashift) - 2305 (logical_start >> ashift); 2306 ASSERT3U(b_size, >, 0); 2307 end = start + ((((b_size - 1) / 2308 vdc->vdc_groupwidth) + 1) << ashift); 2309 break; 2310 } 2311 } 2312 physical_rs->rs_start = start; 2313 physical_rs->rs_end = end; 2314 2315 /* 2316 * Only top-level vdevs are allowed to set remain_rs because 2317 * when .vdev_op_xlate() is called for their children the full 2318 * logical range is not provided by vdev_xlate(). 2319 */ 2320 remain_rs->rs_start = logical_end; 2321 remain_rs->rs_end = logical_rs->rs_end; 2322 2323 ASSERT3U(physical_rs->rs_start, <=, logical_start); 2324 ASSERT3U(physical_rs->rs_end - physical_rs->rs_start, <=, 2325 logical_end - logical_start); 2326 } 2327 2328 /* 2329 * Add dRAID specific fields to the config nvlist. 2330 */ 2331 static void 2332 vdev_draid_config_generate(vdev_t *vd, nvlist_t *nv) 2333 { 2334 ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops); 2335 vdev_draid_config_t *vdc = vd->vdev_tsd; 2336 2337 fnvlist_add_uint64(nv, ZPOOL_CONFIG_NPARITY, vdc->vdc_nparity); 2338 fnvlist_add_uint64(nv, ZPOOL_CONFIG_DRAID_NDATA, vdc->vdc_ndata); 2339 fnvlist_add_uint64(nv, ZPOOL_CONFIG_DRAID_NSPARES, vdc->vdc_nspares); 2340 fnvlist_add_uint64(nv, ZPOOL_CONFIG_DRAID_NGROUPS, vdc->vdc_ngroups); 2341 } 2342 2343 /* 2344 * Initialize private dRAID specific fields from the nvlist. 2345 */ 2346 static int 2347 vdev_draid_init(spa_t *spa, nvlist_t *nv, void **tsd) 2348 { 2349 uint64_t ndata, nparity, nspares, ngroups; 2350 int error; 2351 2352 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DRAID_NDATA, &ndata)) 2353 return (SET_ERROR(EINVAL)); 2354 2355 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY, &nparity) || 2356 nparity == 0 || nparity > VDEV_DRAID_MAXPARITY) { 2357 return (SET_ERROR(EINVAL)); 2358 } 2359 2360 uint_t children; 2361 nvlist_t **child; 2362 if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN, 2363 &child, &children) != 0 || children == 0 || 2364 children > VDEV_DRAID_MAX_CHILDREN) { 2365 return (SET_ERROR(EINVAL)); 2366 } 2367 2368 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DRAID_NSPARES, &nspares) || 2369 nspares > 100 || nspares > (children - (ndata + nparity))) { 2370 return (SET_ERROR(EINVAL)); 2371 } 2372 2373 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DRAID_NGROUPS, &ngroups) || 2374 ngroups == 0 || ngroups > VDEV_DRAID_MAX_CHILDREN) { 2375 return (SET_ERROR(EINVAL)); 2376 } 2377 2378 /* 2379 * Validate the minimum number of children exist per group for the 2380 * specified parity level (draid1 >= 2, draid2 >= 3, draid3 >= 4). 2381 */ 2382 if (children < (ndata + nparity + nspares)) 2383 return (SET_ERROR(EINVAL)); 2384 2385 /* 2386 * Create the dRAID configuration using the pool nvlist configuration 2387 * and the fixed mapping for the correct number of children. 2388 */ 2389 vdev_draid_config_t *vdc; 2390 const draid_map_t *map; 2391 2392 error = vdev_draid_lookup_map(children, &map); 2393 if (error) 2394 return (SET_ERROR(EINVAL)); 2395 2396 vdc = kmem_zalloc(sizeof (*vdc), KM_SLEEP); 2397 vdc->vdc_ndata = ndata; 2398 vdc->vdc_nparity = nparity; 2399 vdc->vdc_nspares = nspares; 2400 vdc->vdc_children = children; 2401 vdc->vdc_ngroups = ngroups; 2402 vdc->vdc_nperms = map->dm_nperms; 2403 2404 error = vdev_draid_generate_perms(map, &vdc->vdc_perms); 2405 if (error) { 2406 kmem_free(vdc, sizeof (*vdc)); 2407 return (SET_ERROR(EINVAL)); 2408 } 2409 2410 /* 2411 * Derived constants. 2412 */ 2413 vdc->vdc_groupwidth = vdc->vdc_ndata + vdc->vdc_nparity; 2414 vdc->vdc_ndisks = vdc->vdc_children - vdc->vdc_nspares; 2415 vdc->vdc_groupsz = vdc->vdc_groupwidth * VDEV_DRAID_ROWHEIGHT; 2416 vdc->vdc_devslicesz = (vdc->vdc_groupsz * vdc->vdc_ngroups) / 2417 vdc->vdc_ndisks; 2418 2419 ASSERT3U(vdc->vdc_groupwidth, >=, 2); 2420 ASSERT3U(vdc->vdc_groupwidth, <=, vdc->vdc_ndisks); 2421 ASSERT3U(vdc->vdc_groupsz, >=, 2 * VDEV_DRAID_ROWHEIGHT); 2422 ASSERT3U(vdc->vdc_devslicesz, >=, VDEV_DRAID_ROWHEIGHT); 2423 ASSERT3U(vdc->vdc_devslicesz % VDEV_DRAID_ROWHEIGHT, ==, 0); 2424 ASSERT3U((vdc->vdc_groupwidth * vdc->vdc_ngroups) % 2425 vdc->vdc_ndisks, ==, 0); 2426 2427 *tsd = vdc; 2428 2429 return (0); 2430 } 2431 2432 static void 2433 vdev_draid_fini(vdev_t *vd) 2434 { 2435 vdev_draid_config_t *vdc = vd->vdev_tsd; 2436 2437 vmem_free(vdc->vdc_perms, sizeof (uint8_t) * 2438 vdc->vdc_children * vdc->vdc_nperms); 2439 kmem_free(vdc, sizeof (*vdc)); 2440 } 2441 2442 static uint64_t 2443 vdev_draid_nparity(vdev_t *vd) 2444 { 2445 vdev_draid_config_t *vdc = vd->vdev_tsd; 2446 2447 return (vdc->vdc_nparity); 2448 } 2449 2450 static uint64_t 2451 vdev_draid_ndisks(vdev_t *vd) 2452 { 2453 vdev_draid_config_t *vdc = vd->vdev_tsd; 2454 2455 return (vdc->vdc_ndisks); 2456 } 2457 2458 vdev_ops_t vdev_draid_ops = { 2459 .vdev_op_init = vdev_draid_init, 2460 .vdev_op_fini = vdev_draid_fini, 2461 .vdev_op_open = vdev_draid_open, 2462 .vdev_op_close = vdev_draid_close, 2463 .vdev_op_asize = vdev_draid_asize, 2464 .vdev_op_min_asize = vdev_draid_min_asize, 2465 .vdev_op_min_alloc = vdev_draid_min_alloc, 2466 .vdev_op_io_start = vdev_draid_io_start, 2467 .vdev_op_io_done = vdev_draid_io_done, 2468 .vdev_op_state_change = vdev_draid_state_change, 2469 .vdev_op_need_resilver = vdev_draid_need_resilver, 2470 .vdev_op_hold = NULL, 2471 .vdev_op_rele = NULL, 2472 .vdev_op_remap = NULL, 2473 .vdev_op_xlate = vdev_draid_xlate, 2474 .vdev_op_rebuild_asize = vdev_draid_rebuild_asize, 2475 .vdev_op_metaslab_init = vdev_draid_metaslab_init, 2476 .vdev_op_config_generate = vdev_draid_config_generate, 2477 .vdev_op_nparity = vdev_draid_nparity, 2478 .vdev_op_ndisks = vdev_draid_ndisks, 2479 .vdev_op_type = VDEV_TYPE_DRAID, 2480 .vdev_op_leaf = B_FALSE, 2481 }; 2482 2483 2484 /* 2485 * A dRAID distributed spare is a virtual leaf vdev which is included in the 2486 * parent dRAID configuration. The last N columns of the dRAID permutation 2487 * table are used to determine on which dRAID children a specific offset 2488 * should be written. These spare leaf vdevs can only be used to replace 2489 * faulted children in the same dRAID configuration. 2490 */ 2491 2492 /* 2493 * Distributed spare state. All fields are set when the distributed spare is 2494 * first opened and are immutable. 2495 */ 2496 typedef struct { 2497 vdev_t *vds_draid_vdev; /* top-level parent dRAID vdev */ 2498 uint64_t vds_top_guid; /* top-level parent dRAID guid */ 2499 uint64_t vds_spare_id; /* spare id (0 - vdc->vdc_nspares-1) */ 2500 } vdev_draid_spare_t; 2501 2502 /* 2503 * Returns the parent dRAID vdev to which the distributed spare belongs. 2504 * This may be safely called even when the vdev is not open. 2505 */ 2506 vdev_t * 2507 vdev_draid_spare_get_parent(vdev_t *vd) 2508 { 2509 vdev_draid_spare_t *vds = vd->vdev_tsd; 2510 2511 ASSERT3P(vd->vdev_ops, ==, &vdev_draid_spare_ops); 2512 2513 if (vds->vds_draid_vdev != NULL) 2514 return (vds->vds_draid_vdev); 2515 2516 return (vdev_lookup_by_guid(vd->vdev_spa->spa_root_vdev, 2517 vds->vds_top_guid)); 2518 } 2519 2520 /* 2521 * A dRAID space is active when it's the child of a vdev using the 2522 * vdev_spare_ops, vdev_replacing_ops or vdev_draid_ops. 2523 */ 2524 static boolean_t 2525 vdev_draid_spare_is_active(vdev_t *vd) 2526 { 2527 vdev_t *pvd = vd->vdev_parent; 2528 2529 if (pvd != NULL && (pvd->vdev_ops == &vdev_spare_ops || 2530 pvd->vdev_ops == &vdev_replacing_ops || 2531 pvd->vdev_ops == &vdev_draid_ops)) { 2532 return (B_TRUE); 2533 } else { 2534 return (B_FALSE); 2535 } 2536 } 2537 2538 /* 2539 * Given a dRAID distribute spare vdev, returns the physical child vdev 2540 * on which the provided offset resides. This may involve recursing through 2541 * multiple layers of distributed spares. Note that offset is relative to 2542 * this vdev. 2543 */ 2544 vdev_t * 2545 vdev_draid_spare_get_child(vdev_t *vd, uint64_t physical_offset) 2546 { 2547 vdev_draid_spare_t *vds = vd->vdev_tsd; 2548 2549 ASSERT3P(vd->vdev_ops, ==, &vdev_draid_spare_ops); 2550 2551 /* The vdev is closed */ 2552 if (vds->vds_draid_vdev == NULL) 2553 return (NULL); 2554 2555 vdev_t *tvd = vds->vds_draid_vdev; 2556 vdev_draid_config_t *vdc = tvd->vdev_tsd; 2557 2558 ASSERT3P(tvd->vdev_ops, ==, &vdev_draid_ops); 2559 ASSERT3U(vds->vds_spare_id, <, vdc->vdc_nspares); 2560 2561 uint8_t *base; 2562 uint64_t iter; 2563 uint64_t perm = physical_offset / vdc->vdc_devslicesz; 2564 2565 vdev_draid_get_perm(vdc, perm, &base, &iter); 2566 2567 uint64_t cid = vdev_draid_permute_id(vdc, base, iter, 2568 (tvd->vdev_children - 1) - vds->vds_spare_id); 2569 vdev_t *cvd = tvd->vdev_child[cid]; 2570 2571 if (cvd->vdev_ops == &vdev_draid_spare_ops) 2572 return (vdev_draid_spare_get_child(cvd, physical_offset)); 2573 2574 return (cvd); 2575 } 2576 2577 /* ARGSUSED */ 2578 static void 2579 vdev_draid_spare_close(vdev_t *vd) 2580 { 2581 vdev_draid_spare_t *vds = vd->vdev_tsd; 2582 vds->vds_draid_vdev = NULL; 2583 } 2584 2585 /* 2586 * Opening a dRAID spare device is done by looking up the associated dRAID 2587 * top-level vdev guid from the spare configuration. 2588 */ 2589 static int 2590 vdev_draid_spare_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize, 2591 uint64_t *logical_ashift, uint64_t *physical_ashift) 2592 { 2593 vdev_draid_spare_t *vds = vd->vdev_tsd; 2594 vdev_t *rvd = vd->vdev_spa->spa_root_vdev; 2595 uint64_t asize, max_asize; 2596 2597 vdev_t *tvd = vdev_lookup_by_guid(rvd, vds->vds_top_guid); 2598 if (tvd == NULL) { 2599 /* 2600 * When spa_vdev_add() is labeling new spares the 2601 * associated dRAID is not attached to the root vdev 2602 * nor does this spare have a parent. Simulate a valid 2603 * device in order to allow the label to be initialized 2604 * and the distributed spare added to the configuration. 2605 */ 2606 if (vd->vdev_parent == NULL) { 2607 *psize = *max_psize = SPA_MINDEVSIZE; 2608 *logical_ashift = *physical_ashift = ASHIFT_MIN; 2609 return (0); 2610 } 2611 2612 return (SET_ERROR(EINVAL)); 2613 } 2614 2615 vdev_draid_config_t *vdc = tvd->vdev_tsd; 2616 if (tvd->vdev_ops != &vdev_draid_ops || vdc == NULL) 2617 return (SET_ERROR(EINVAL)); 2618 2619 if (vds->vds_spare_id >= vdc->vdc_nspares) 2620 return (SET_ERROR(EINVAL)); 2621 2622 /* 2623 * Neither tvd->vdev_asize or tvd->vdev_max_asize can be used here 2624 * because the caller may be vdev_draid_open() in which case the 2625 * values are stale as they haven't yet been updated by vdev_open(). 2626 * To avoid this always recalculate the dRAID asize and max_asize. 2627 */ 2628 vdev_draid_calculate_asize(tvd, &asize, &max_asize, 2629 logical_ashift, physical_ashift); 2630 2631 *psize = asize + VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE; 2632 *max_psize = max_asize + VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE; 2633 2634 vds->vds_draid_vdev = tvd; 2635 2636 return (0); 2637 } 2638 2639 /* 2640 * Completed distributed spare IO. Store the result in the parent zio 2641 * as if it had performed the operation itself. Only the first error is 2642 * preserved if there are multiple errors. 2643 */ 2644 static void 2645 vdev_draid_spare_child_done(zio_t *zio) 2646 { 2647 zio_t *pio = zio->io_private; 2648 2649 /* 2650 * IOs are issued to non-writable vdevs in order to keep their 2651 * DTLs accurate. However, we don't want to propagate the 2652 * error in to the distributed spare's DTL. When resilvering 2653 * vdev_draid_need_resilver() will consult the relevant DTL 2654 * to determine if the data is missing and must be repaired. 2655 */ 2656 if (!vdev_writeable(zio->io_vd)) 2657 return; 2658 2659 if (pio->io_error == 0) 2660 pio->io_error = zio->io_error; 2661 } 2662 2663 /* 2664 * Returns a valid label nvlist for the distributed spare vdev. This is 2665 * used to bypass the IO pipeline to avoid the complexity of constructing 2666 * a complete label with valid checksum to return when read. 2667 */ 2668 nvlist_t * 2669 vdev_draid_read_config_spare(vdev_t *vd) 2670 { 2671 spa_t *spa = vd->vdev_spa; 2672 spa_aux_vdev_t *sav = &spa->spa_spares; 2673 uint64_t guid = vd->vdev_guid; 2674 2675 nvlist_t *nv = fnvlist_alloc(); 2676 fnvlist_add_uint64(nv, ZPOOL_CONFIG_IS_SPARE, 1); 2677 fnvlist_add_uint64(nv, ZPOOL_CONFIG_CREATE_TXG, vd->vdev_crtxg); 2678 fnvlist_add_uint64(nv, ZPOOL_CONFIG_VERSION, spa_version(spa)); 2679 fnvlist_add_string(nv, ZPOOL_CONFIG_POOL_NAME, spa_name(spa)); 2680 fnvlist_add_uint64(nv, ZPOOL_CONFIG_POOL_GUID, spa_guid(spa)); 2681 fnvlist_add_uint64(nv, ZPOOL_CONFIG_POOL_TXG, spa->spa_config_txg); 2682 fnvlist_add_uint64(nv, ZPOOL_CONFIG_TOP_GUID, vd->vdev_top->vdev_guid); 2683 fnvlist_add_uint64(nv, ZPOOL_CONFIG_POOL_STATE, 2684 vdev_draid_spare_is_active(vd) ? 2685 POOL_STATE_ACTIVE : POOL_STATE_SPARE); 2686 2687 /* Set the vdev guid based on the vdev list in sav_count. */ 2688 for (int i = 0; i < sav->sav_count; i++) { 2689 if (sav->sav_vdevs[i]->vdev_ops == &vdev_draid_spare_ops && 2690 strcmp(sav->sav_vdevs[i]->vdev_path, vd->vdev_path) == 0) { 2691 guid = sav->sav_vdevs[i]->vdev_guid; 2692 break; 2693 } 2694 } 2695 2696 fnvlist_add_uint64(nv, ZPOOL_CONFIG_GUID, guid); 2697 2698 return (nv); 2699 } 2700 2701 /* 2702 * Handle any ioctl requested of the distributed spare. Only flushes 2703 * are supported in which case all children must be flushed. 2704 */ 2705 static int 2706 vdev_draid_spare_ioctl(zio_t *zio) 2707 { 2708 vdev_t *vd = zio->io_vd; 2709 int error = 0; 2710 2711 if (zio->io_cmd == DKIOCFLUSHWRITECACHE) { 2712 for (int c = 0; c < vd->vdev_children; c++) { 2713 zio_nowait(zio_vdev_child_io(zio, NULL, 2714 vd->vdev_child[c], zio->io_offset, zio->io_abd, 2715 zio->io_size, zio->io_type, zio->io_priority, 0, 2716 vdev_draid_spare_child_done, zio)); 2717 } 2718 } else { 2719 error = SET_ERROR(ENOTSUP); 2720 } 2721 2722 return (error); 2723 } 2724 2725 /* 2726 * Initiate an IO to the distributed spare. For normal IOs this entails using 2727 * the zio->io_offset and permutation table to calculate which child dRAID vdev 2728 * is responsible for the data. Then passing along the zio to that child to 2729 * perform the actual IO. The label ranges are not stored on disk and require 2730 * some special handling which is described below. 2731 */ 2732 static void 2733 vdev_draid_spare_io_start(zio_t *zio) 2734 { 2735 vdev_t *cvd = NULL, *vd = zio->io_vd; 2736 vdev_draid_spare_t *vds = vd->vdev_tsd; 2737 uint64_t offset = zio->io_offset - VDEV_LABEL_START_SIZE; 2738 2739 /* 2740 * If the vdev is closed, it's likely in the REMOVED or FAULTED state. 2741 * Nothing to be done here but return failure. 2742 */ 2743 if (vds == NULL) { 2744 zio->io_error = ENXIO; 2745 zio_interrupt(zio); 2746 return; 2747 } 2748 2749 switch (zio->io_type) { 2750 case ZIO_TYPE_IOCTL: 2751 zio->io_error = vdev_draid_spare_ioctl(zio); 2752 break; 2753 2754 case ZIO_TYPE_WRITE: 2755 if (VDEV_OFFSET_IS_LABEL(vd, zio->io_offset)) { 2756 /* 2757 * Accept probe IOs and config writers to simulate the 2758 * existence of an on disk label. vdev_label_sync(), 2759 * vdev_uberblock_sync() and vdev_copy_uberblocks() 2760 * skip the distributed spares. This only leaves 2761 * vdev_label_init() which is allowed to succeed to 2762 * avoid adding special cases the function. 2763 */ 2764 if (zio->io_flags & ZIO_FLAG_PROBE || 2765 zio->io_flags & ZIO_FLAG_CONFIG_WRITER) { 2766 zio->io_error = 0; 2767 } else { 2768 zio->io_error = SET_ERROR(EIO); 2769 } 2770 } else { 2771 cvd = vdev_draid_spare_get_child(vd, offset); 2772 2773 if (cvd == NULL) { 2774 zio->io_error = SET_ERROR(ENXIO); 2775 } else { 2776 zio_nowait(zio_vdev_child_io(zio, NULL, cvd, 2777 offset, zio->io_abd, zio->io_size, 2778 zio->io_type, zio->io_priority, 0, 2779 vdev_draid_spare_child_done, zio)); 2780 } 2781 } 2782 break; 2783 2784 case ZIO_TYPE_READ: 2785 if (VDEV_OFFSET_IS_LABEL(vd, zio->io_offset)) { 2786 /* 2787 * Accept probe IOs to simulate the existence of a 2788 * label. vdev_label_read_config() bypasses the 2789 * pipeline to read the label configuration and 2790 * vdev_uberblock_load() skips distributed spares 2791 * when attempting to locate the best uberblock. 2792 */ 2793 if (zio->io_flags & ZIO_FLAG_PROBE) { 2794 zio->io_error = 0; 2795 } else { 2796 zio->io_error = SET_ERROR(EIO); 2797 } 2798 } else { 2799 cvd = vdev_draid_spare_get_child(vd, offset); 2800 2801 if (cvd == NULL || !vdev_readable(cvd)) { 2802 zio->io_error = SET_ERROR(ENXIO); 2803 } else { 2804 zio_nowait(zio_vdev_child_io(zio, NULL, cvd, 2805 offset, zio->io_abd, zio->io_size, 2806 zio->io_type, zio->io_priority, 0, 2807 vdev_draid_spare_child_done, zio)); 2808 } 2809 } 2810 break; 2811 2812 case ZIO_TYPE_TRIM: 2813 /* The vdev label ranges are never trimmed */ 2814 ASSERT0(VDEV_OFFSET_IS_LABEL(vd, zio->io_offset)); 2815 2816 cvd = vdev_draid_spare_get_child(vd, offset); 2817 2818 if (cvd == NULL || !cvd->vdev_has_trim) { 2819 zio->io_error = SET_ERROR(ENXIO); 2820 } else { 2821 zio_nowait(zio_vdev_child_io(zio, NULL, cvd, 2822 offset, zio->io_abd, zio->io_size, 2823 zio->io_type, zio->io_priority, 0, 2824 vdev_draid_spare_child_done, zio)); 2825 } 2826 break; 2827 2828 default: 2829 zio->io_error = SET_ERROR(ENOTSUP); 2830 break; 2831 } 2832 2833 zio_execute(zio); 2834 } 2835 2836 /* ARGSUSED */ 2837 static void 2838 vdev_draid_spare_io_done(zio_t *zio) 2839 { 2840 } 2841 2842 /* 2843 * Lookup the full spare config in spa->spa_spares.sav_config and 2844 * return the top_guid and spare_id for the named spare. 2845 */ 2846 static int 2847 vdev_draid_spare_lookup(spa_t *spa, nvlist_t *nv, uint64_t *top_guidp, 2848 uint64_t *spare_idp) 2849 { 2850 nvlist_t **spares; 2851 uint_t nspares; 2852 int error; 2853 2854 if ((spa->spa_spares.sav_config == NULL) || 2855 (nvlist_lookup_nvlist_array(spa->spa_spares.sav_config, 2856 ZPOOL_CONFIG_SPARES, &spares, &nspares) != 0)) { 2857 return (SET_ERROR(ENOENT)); 2858 } 2859 2860 char *spare_name; 2861 error = nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &spare_name); 2862 if (error != 0) 2863 return (SET_ERROR(EINVAL)); 2864 2865 for (int i = 0; i < nspares; i++) { 2866 nvlist_t *spare = spares[i]; 2867 uint64_t top_guid, spare_id; 2868 char *type, *path; 2869 2870 /* Skip non-distributed spares */ 2871 error = nvlist_lookup_string(spare, ZPOOL_CONFIG_TYPE, &type); 2872 if (error != 0 || strcmp(type, VDEV_TYPE_DRAID_SPARE) != 0) 2873 continue; 2874 2875 /* Skip spares with the wrong name */ 2876 error = nvlist_lookup_string(spare, ZPOOL_CONFIG_PATH, &path); 2877 if (error != 0 || strcmp(path, spare_name) != 0) 2878 continue; 2879 2880 /* Found the matching spare */ 2881 error = nvlist_lookup_uint64(spare, 2882 ZPOOL_CONFIG_TOP_GUID, &top_guid); 2883 if (error == 0) { 2884 error = nvlist_lookup_uint64(spare, 2885 ZPOOL_CONFIG_SPARE_ID, &spare_id); 2886 } 2887 2888 if (error != 0) { 2889 return (SET_ERROR(EINVAL)); 2890 } else { 2891 *top_guidp = top_guid; 2892 *spare_idp = spare_id; 2893 return (0); 2894 } 2895 } 2896 2897 return (SET_ERROR(ENOENT)); 2898 } 2899 2900 /* 2901 * Initialize private dRAID spare specific fields from the nvlist. 2902 */ 2903 static int 2904 vdev_draid_spare_init(spa_t *spa, nvlist_t *nv, void **tsd) 2905 { 2906 vdev_draid_spare_t *vds; 2907 uint64_t top_guid = 0; 2908 uint64_t spare_id; 2909 2910 /* 2911 * In the normal case check the list of spares stored in the spa 2912 * to lookup the top_guid and spare_id for provided spare config. 2913 * When creating a new pool or adding vdevs the spare list is not 2914 * yet populated and the values are provided in the passed config. 2915 */ 2916 if (vdev_draid_spare_lookup(spa, nv, &top_guid, &spare_id) != 0) { 2917 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_TOP_GUID, 2918 &top_guid) != 0) 2919 return (SET_ERROR(EINVAL)); 2920 2921 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_SPARE_ID, 2922 &spare_id) != 0) 2923 return (SET_ERROR(EINVAL)); 2924 } 2925 2926 vds = kmem_alloc(sizeof (vdev_draid_spare_t), KM_SLEEP); 2927 vds->vds_draid_vdev = NULL; 2928 vds->vds_top_guid = top_guid; 2929 vds->vds_spare_id = spare_id; 2930 2931 *tsd = vds; 2932 2933 return (0); 2934 } 2935 2936 static void 2937 vdev_draid_spare_fini(vdev_t *vd) 2938 { 2939 kmem_free(vd->vdev_tsd, sizeof (vdev_draid_spare_t)); 2940 } 2941 2942 static void 2943 vdev_draid_spare_config_generate(vdev_t *vd, nvlist_t *nv) 2944 { 2945 vdev_draid_spare_t *vds = vd->vdev_tsd; 2946 2947 ASSERT3P(vd->vdev_ops, ==, &vdev_draid_spare_ops); 2948 2949 fnvlist_add_uint64(nv, ZPOOL_CONFIG_TOP_GUID, vds->vds_top_guid); 2950 fnvlist_add_uint64(nv, ZPOOL_CONFIG_SPARE_ID, vds->vds_spare_id); 2951 } 2952 2953 vdev_ops_t vdev_draid_spare_ops = { 2954 .vdev_op_init = vdev_draid_spare_init, 2955 .vdev_op_fini = vdev_draid_spare_fini, 2956 .vdev_op_open = vdev_draid_spare_open, 2957 .vdev_op_close = vdev_draid_spare_close, 2958 .vdev_op_asize = vdev_default_asize, 2959 .vdev_op_min_asize = vdev_default_min_asize, 2960 .vdev_op_min_alloc = NULL, 2961 .vdev_op_io_start = vdev_draid_spare_io_start, 2962 .vdev_op_io_done = vdev_draid_spare_io_done, 2963 .vdev_op_state_change = NULL, 2964 .vdev_op_need_resilver = NULL, 2965 .vdev_op_hold = NULL, 2966 .vdev_op_rele = NULL, 2967 .vdev_op_remap = NULL, 2968 .vdev_op_xlate = vdev_default_xlate, 2969 .vdev_op_rebuild_asize = NULL, 2970 .vdev_op_metaslab_init = NULL, 2971 .vdev_op_config_generate = vdev_draid_spare_config_generate, 2972 .vdev_op_nparity = NULL, 2973 .vdev_op_ndisks = NULL, 2974 .vdev_op_type = VDEV_TYPE_DRAID_SPARE, 2975 .vdev_op_leaf = B_TRUE, 2976 }; 2977