xref: /freebsd/sys/contrib/openzfs/module/zfs/vdev_draid.c (revision b197d4b893974c9eb4d7b38704c6d5c486235d6f)
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 https://opensource.org/licenses/CDDL-1.0.
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  * Full stripe writes.  When writing, all columns (D+P) are required.  Parity
637  * is calculated over all the columns, including empty zero filled sectors,
638  * and each is written to disk.  While only the data columns are needed for
639  * a normal read, all of the columns are required for reconstruction when
640  * performing a sequential resilver.
641  *
642  * For "big columns" it's sufficient to map the correct range of the zio ABD.
643  * Partial columns require allocating a gang ABD in order to zero fill the
644  * empty sectors.  When the column is empty a zero filled sector must be
645  * mapped.  In all cases the data ABDs must be the same size as the parity
646  * ABDs (e.g. rc->rc_size == parity_size).
647  */
648 static void
649 vdev_draid_map_alloc_write(zio_t *zio, uint64_t abd_offset, raidz_row_t *rr)
650 {
651 	uint64_t skip_size = 1ULL << zio->io_vd->vdev_top->vdev_ashift;
652 	uint64_t parity_size = rr->rr_col[0].rc_size;
653 	uint64_t abd_off = abd_offset;
654 
655 	ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
656 	ASSERT3U(parity_size, ==, abd_get_size(rr->rr_col[0].rc_abd));
657 
658 	for (uint64_t c = rr->rr_firstdatacol; c < rr->rr_cols; c++) {
659 		raidz_col_t *rc = &rr->rr_col[c];
660 
661 		if (rc->rc_size == 0) {
662 			/* empty data column (small write), add a skip sector */
663 			ASSERT3U(skip_size, ==, parity_size);
664 			rc->rc_abd = abd_get_zeros(skip_size);
665 		} else if (rc->rc_size == parity_size) {
666 			/* this is a "big column" */
667 			rc->rc_abd = abd_get_offset_struct(&rc->rc_abdstruct,
668 			    zio->io_abd, abd_off, rc->rc_size);
669 		} else {
670 			/* short data column, add a skip sector */
671 			ASSERT3U(rc->rc_size + skip_size, ==, parity_size);
672 			rc->rc_abd = abd_alloc_gang();
673 			abd_gang_add(rc->rc_abd, abd_get_offset_size(
674 			    zio->io_abd, abd_off, rc->rc_size), B_TRUE);
675 			abd_gang_add(rc->rc_abd, abd_get_zeros(skip_size),
676 			    B_TRUE);
677 		}
678 
679 		ASSERT3U(abd_get_size(rc->rc_abd), ==, parity_size);
680 
681 		abd_off += rc->rc_size;
682 		rc->rc_size = parity_size;
683 	}
684 
685 	IMPLY(abd_offset != 0, abd_off == zio->io_size);
686 }
687 
688 /*
689  * Scrub/resilver reads.  In order to store the contents of the skip sectors
690  * an additional ABD is allocated.  The columns are handled in the same way
691  * as a full stripe write except instead of using the zero ABD the newly
692  * allocated skip ABD is used to back the skip sectors.  In all cases the
693  * data ABD must be the same size as the parity ABDs.
694  */
695 static void
696 vdev_draid_map_alloc_scrub(zio_t *zio, uint64_t abd_offset, raidz_row_t *rr)
697 {
698 	uint64_t skip_size = 1ULL << zio->io_vd->vdev_top->vdev_ashift;
699 	uint64_t parity_size = rr->rr_col[0].rc_size;
700 	uint64_t abd_off = abd_offset;
701 	uint64_t skip_off = 0;
702 
703 	ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ);
704 	ASSERT3P(rr->rr_abd_empty, ==, NULL);
705 
706 	if (rr->rr_nempty > 0) {
707 		rr->rr_abd_empty = abd_alloc_linear(rr->rr_nempty * skip_size,
708 		    B_FALSE);
709 	}
710 
711 	for (uint64_t c = rr->rr_firstdatacol; c < rr->rr_cols; c++) {
712 		raidz_col_t *rc = &rr->rr_col[c];
713 
714 		if (rc->rc_size == 0) {
715 			/* empty data column (small read), add a skip sector */
716 			ASSERT3U(skip_size, ==, parity_size);
717 			ASSERT3U(rr->rr_nempty, !=, 0);
718 			rc->rc_abd = abd_get_offset_size(rr->rr_abd_empty,
719 			    skip_off, skip_size);
720 			skip_off += skip_size;
721 		} else if (rc->rc_size == parity_size) {
722 			/* this is a "big column" */
723 			rc->rc_abd = abd_get_offset_struct(&rc->rc_abdstruct,
724 			    zio->io_abd, abd_off, rc->rc_size);
725 		} else {
726 			/* short data column, add a skip sector */
727 			ASSERT3U(rc->rc_size + skip_size, ==, parity_size);
728 			ASSERT3U(rr->rr_nempty, !=, 0);
729 			rc->rc_abd = abd_alloc_gang();
730 			abd_gang_add(rc->rc_abd, abd_get_offset_size(
731 			    zio->io_abd, abd_off, rc->rc_size), B_TRUE);
732 			abd_gang_add(rc->rc_abd, abd_get_offset_size(
733 			    rr->rr_abd_empty, skip_off, skip_size), B_TRUE);
734 			skip_off += skip_size;
735 		}
736 
737 		uint64_t abd_size = abd_get_size(rc->rc_abd);
738 		ASSERT3U(abd_size, ==, abd_get_size(rr->rr_col[0].rc_abd));
739 
740 		/*
741 		 * Increase rc_size so the skip ABD is included in subsequent
742 		 * parity calculations.
743 		 */
744 		abd_off += rc->rc_size;
745 		rc->rc_size = abd_size;
746 	}
747 
748 	IMPLY(abd_offset != 0, abd_off == zio->io_size);
749 	ASSERT3U(skip_off, ==, rr->rr_nempty * skip_size);
750 }
751 
752 /*
753  * Normal reads.  In this common case only the columns containing data
754  * are read in to the zio ABDs.  Neither the parity columns or empty skip
755  * sectors are read unless the checksum fails verification.  In which case
756  * vdev_raidz_read_all() will call vdev_draid_map_alloc_empty() to expand
757  * the raid map in order to allow reconstruction using the parity data and
758  * skip sectors.
759  */
760 static void
761 vdev_draid_map_alloc_read(zio_t *zio, uint64_t abd_offset, raidz_row_t *rr)
762 {
763 	uint64_t abd_off = abd_offset;
764 
765 	ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ);
766 
767 	for (uint64_t c = rr->rr_firstdatacol; c < rr->rr_cols; c++) {
768 		raidz_col_t *rc = &rr->rr_col[c];
769 
770 		if (rc->rc_size > 0) {
771 			rc->rc_abd = abd_get_offset_struct(&rc->rc_abdstruct,
772 			    zio->io_abd, abd_off, rc->rc_size);
773 			abd_off += rc->rc_size;
774 		}
775 	}
776 
777 	IMPLY(abd_offset != 0, abd_off == zio->io_size);
778 }
779 
780 /*
781  * Converts a normal "read" raidz_row_t to a "scrub" raidz_row_t. The key
782  * difference is that an ABD is allocated to back skip sectors so they may
783  * be read in to memory, verified, and repaired if needed.
784  */
785 void
786 vdev_draid_map_alloc_empty(zio_t *zio, raidz_row_t *rr)
787 {
788 	uint64_t skip_size = 1ULL << zio->io_vd->vdev_top->vdev_ashift;
789 	uint64_t parity_size = rr->rr_col[0].rc_size;
790 	uint64_t skip_off = 0;
791 
792 	ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ);
793 	ASSERT3P(rr->rr_abd_empty, ==, NULL);
794 
795 	if (rr->rr_nempty > 0) {
796 		rr->rr_abd_empty = abd_alloc_linear(rr->rr_nempty * skip_size,
797 		    B_FALSE);
798 	}
799 
800 	for (uint64_t c = rr->rr_firstdatacol; c < rr->rr_cols; c++) {
801 		raidz_col_t *rc = &rr->rr_col[c];
802 
803 		if (rc->rc_size == 0) {
804 			/* empty data column (small read), add a skip sector */
805 			ASSERT3U(skip_size, ==, parity_size);
806 			ASSERT3U(rr->rr_nempty, !=, 0);
807 			ASSERT3P(rc->rc_abd, ==, NULL);
808 			rc->rc_abd = abd_get_offset_size(rr->rr_abd_empty,
809 			    skip_off, skip_size);
810 			skip_off += skip_size;
811 		} else if (rc->rc_size == parity_size) {
812 			/* this is a "big column", nothing to add */
813 			ASSERT3P(rc->rc_abd, !=, NULL);
814 		} else {
815 			/*
816 			 * short data column, add a skip sector and clear
817 			 * rc_tried to force the entire column to be re-read
818 			 * thereby including the missing skip sector data
819 			 * which is needed for reconstruction.
820 			 */
821 			ASSERT3U(rc->rc_size + skip_size, ==, parity_size);
822 			ASSERT3U(rr->rr_nempty, !=, 0);
823 			ASSERT3P(rc->rc_abd, !=, NULL);
824 			ASSERT(!abd_is_gang(rc->rc_abd));
825 			abd_t *read_abd = rc->rc_abd;
826 			rc->rc_abd = abd_alloc_gang();
827 			abd_gang_add(rc->rc_abd, read_abd, B_TRUE);
828 			abd_gang_add(rc->rc_abd, abd_get_offset_size(
829 			    rr->rr_abd_empty, skip_off, skip_size), B_TRUE);
830 			skip_off += skip_size;
831 			rc->rc_tried = 0;
832 		}
833 
834 		/*
835 		 * Increase rc_size so the empty ABD is included in subsequent
836 		 * parity calculations.
837 		 */
838 		rc->rc_size = parity_size;
839 	}
840 
841 	ASSERT3U(skip_off, ==, rr->rr_nempty * skip_size);
842 }
843 
844 /*
845  * Verify that all empty sectors are zero filled before using them to
846  * calculate parity.  Otherwise, silent corruption in an empty sector will
847  * result in bad parity being generated.  That bad parity will then be
848  * considered authoritative and overwrite the good parity on disk.  This
849  * is possible because the checksum is only calculated over the data,
850  * thus it cannot be used to detect damage in empty sectors.
851  */
852 int
853 vdev_draid_map_verify_empty(zio_t *zio, raidz_row_t *rr)
854 {
855 	uint64_t skip_size = 1ULL << zio->io_vd->vdev_top->vdev_ashift;
856 	uint64_t parity_size = rr->rr_col[0].rc_size;
857 	uint64_t skip_off = parity_size - skip_size;
858 	uint64_t empty_off = 0;
859 	int ret = 0;
860 
861 	ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ);
862 	ASSERT3P(rr->rr_abd_empty, !=, NULL);
863 	ASSERT3U(rr->rr_bigcols, >, 0);
864 
865 	void *zero_buf = kmem_zalloc(skip_size, KM_SLEEP);
866 
867 	for (int c = rr->rr_bigcols; c < rr->rr_cols; c++) {
868 		raidz_col_t *rc = &rr->rr_col[c];
869 
870 		ASSERT3P(rc->rc_abd, !=, NULL);
871 		ASSERT3U(rc->rc_size, ==, parity_size);
872 
873 		if (abd_cmp_buf_off(rc->rc_abd, zero_buf, skip_off,
874 		    skip_size) != 0) {
875 			vdev_raidz_checksum_error(zio, rc, rc->rc_abd);
876 			abd_zero_off(rc->rc_abd, skip_off, skip_size);
877 			rc->rc_error = SET_ERROR(ECKSUM);
878 			ret++;
879 		}
880 
881 		empty_off += skip_size;
882 	}
883 
884 	ASSERT3U(empty_off, ==, abd_get_size(rr->rr_abd_empty));
885 
886 	kmem_free(zero_buf, skip_size);
887 
888 	return (ret);
889 }
890 
891 /*
892  * Given a logical address within a dRAID configuration, return the physical
893  * address on the first drive in the group that this address maps to
894  * (at position 'start' in permutation number 'perm').
895  */
896 static uint64_t
897 vdev_draid_logical_to_physical(vdev_t *vd, uint64_t logical_offset,
898     uint64_t *perm, uint64_t *start)
899 {
900 	vdev_draid_config_t *vdc = vd->vdev_tsd;
901 
902 	/* b is the dRAID (parent) sector offset. */
903 	uint64_t ashift = vd->vdev_top->vdev_ashift;
904 	uint64_t b_offset = logical_offset >> ashift;
905 
906 	/*
907 	 * The height of a row in units of the vdev's minimum sector size.
908 	 * This is the amount of data written to each disk of each group
909 	 * in a given permutation.
910 	 */
911 	uint64_t rowheight_sectors = VDEV_DRAID_ROWHEIGHT >> ashift;
912 
913 	/*
914 	 * We cycle through a disk permutation every groupsz * ngroups chunk
915 	 * of address space. Note that ngroups * groupsz must be a multiple
916 	 * of the number of data drives (ndisks) in order to guarantee
917 	 * alignment. So, for example, if our row height is 16MB, our group
918 	 * size is 10, and there are 13 data drives in the draid, then ngroups
919 	 * will be 13, we will change permutation every 2.08GB and each
920 	 * disk will have 160MB of data per chunk.
921 	 */
922 	uint64_t groupwidth = vdc->vdc_groupwidth;
923 	uint64_t ngroups = vdc->vdc_ngroups;
924 	uint64_t ndisks = vdc->vdc_ndisks;
925 
926 	/*
927 	 * groupstart is where the group this IO will land in "starts" in
928 	 * the permutation array.
929 	 */
930 	uint64_t group = logical_offset / vdc->vdc_groupsz;
931 	uint64_t groupstart = (group * groupwidth) % ndisks;
932 	ASSERT3U(groupstart + groupwidth, <=, ndisks + groupstart);
933 	*start = groupstart;
934 
935 	/* b_offset is the sector offset within a group chunk */
936 	b_offset = b_offset % (rowheight_sectors * groupwidth);
937 	ASSERT0(b_offset % groupwidth);
938 
939 	/*
940 	 * Find the starting byte offset on each child vdev:
941 	 * - within a permutation there are ngroups groups spread over the
942 	 *   rows, where each row covers a slice portion of the disk
943 	 * - each permutation has (groupwidth * ngroups) / ndisks rows
944 	 * - so each permutation covers rows * slice portion of the disk
945 	 * - so we need to find the row where this IO group target begins
946 	 */
947 	*perm = group / ngroups;
948 	uint64_t row = (*perm * ((groupwidth * ngroups) / ndisks)) +
949 	    (((group % ngroups) * groupwidth) / ndisks);
950 
951 	return (((rowheight_sectors * row) +
952 	    (b_offset / groupwidth)) << ashift);
953 }
954 
955 static uint64_t
956 vdev_draid_map_alloc_row(zio_t *zio, raidz_row_t **rrp, uint64_t io_offset,
957     uint64_t abd_offset, uint64_t abd_size)
958 {
959 	vdev_t *vd = zio->io_vd;
960 	vdev_draid_config_t *vdc = vd->vdev_tsd;
961 	uint64_t ashift = vd->vdev_top->vdev_ashift;
962 	uint64_t io_size = abd_size;
963 	uint64_t io_asize = vdev_draid_asize(vd, io_size);
964 	uint64_t group = vdev_draid_offset_to_group(vd, io_offset);
965 	uint64_t start_offset = vdev_draid_group_to_offset(vd, group + 1);
966 
967 	/*
968 	 * Limit the io_size to the space remaining in the group.  A second
969 	 * row in the raidz_map_t is created for the remainder.
970 	 */
971 	if (io_offset + io_asize > start_offset) {
972 		io_size = vdev_draid_asize_to_psize(vd,
973 		    start_offset - io_offset);
974 	}
975 
976 	/*
977 	 * At most a block may span the logical end of one group and the start
978 	 * of the next group. Therefore, at the end of a group the io_size must
979 	 * span the group width evenly and the remainder must be aligned to the
980 	 * start of the next group.
981 	 */
982 	IMPLY(abd_offset == 0 && io_size < zio->io_size,
983 	    (io_asize >> ashift) % vdc->vdc_groupwidth == 0);
984 	IMPLY(abd_offset != 0,
985 	    vdev_draid_group_to_offset(vd, group) == io_offset);
986 
987 	/* Lookup starting byte offset on each child vdev */
988 	uint64_t groupstart, perm;
989 	uint64_t physical_offset = vdev_draid_logical_to_physical(vd,
990 	    io_offset, &perm, &groupstart);
991 
992 	/*
993 	 * If there is less than groupwidth drives available after the group
994 	 * start, the group is going to wrap onto the next row. 'wrap' is the
995 	 * group disk number that starts on the next row.
996 	 */
997 	uint64_t ndisks = vdc->vdc_ndisks;
998 	uint64_t groupwidth = vdc->vdc_groupwidth;
999 	uint64_t wrap = groupwidth;
1000 
1001 	if (groupstart + groupwidth > ndisks)
1002 		wrap = ndisks - groupstart;
1003 
1004 	/* The io size in units of the vdev's minimum sector size. */
1005 	const uint64_t psize = io_size >> ashift;
1006 
1007 	/*
1008 	 * "Quotient": The number of data sectors for this stripe on all but
1009 	 * the "big column" child vdevs that also contain "remainder" data.
1010 	 */
1011 	uint64_t q = psize / vdc->vdc_ndata;
1012 
1013 	/*
1014 	 * "Remainder": The number of partial stripe data sectors in this I/O.
1015 	 * This will add a sector to some, but not all, child vdevs.
1016 	 */
1017 	uint64_t r = psize - q * vdc->vdc_ndata;
1018 
1019 	/* The number of "big columns" - those which contain remainder data. */
1020 	uint64_t bc = (r == 0 ? 0 : r + vdc->vdc_nparity);
1021 	ASSERT3U(bc, <, groupwidth);
1022 
1023 	/* The total number of data and parity sectors for this I/O. */
1024 	uint64_t tot = psize + (vdc->vdc_nparity * (q + (r == 0 ? 0 : 1)));
1025 
1026 	raidz_row_t *rr;
1027 	rr = kmem_alloc(offsetof(raidz_row_t, rr_col[groupwidth]), KM_SLEEP);
1028 	rr->rr_cols = groupwidth;
1029 	rr->rr_scols = groupwidth;
1030 	rr->rr_bigcols = bc;
1031 	rr->rr_missingdata = 0;
1032 	rr->rr_missingparity = 0;
1033 	rr->rr_firstdatacol = vdc->vdc_nparity;
1034 	rr->rr_abd_empty = NULL;
1035 #ifdef ZFS_DEBUG
1036 	rr->rr_offset = io_offset;
1037 	rr->rr_size = io_size;
1038 #endif
1039 	*rrp = rr;
1040 
1041 	uint8_t *base;
1042 	uint64_t iter, asize = 0;
1043 	vdev_draid_get_perm(vdc, perm, &base, &iter);
1044 	for (uint64_t i = 0; i < groupwidth; i++) {
1045 		raidz_col_t *rc = &rr->rr_col[i];
1046 		uint64_t c = (groupstart + i) % ndisks;
1047 
1048 		/* increment the offset if we wrap to the next row */
1049 		if (i == wrap)
1050 			physical_offset += VDEV_DRAID_ROWHEIGHT;
1051 
1052 		rc->rc_devidx = vdev_draid_permute_id(vdc, base, iter, c);
1053 		rc->rc_offset = physical_offset;
1054 		rc->rc_abd = NULL;
1055 		rc->rc_orig_data = NULL;
1056 		rc->rc_error = 0;
1057 		rc->rc_tried = 0;
1058 		rc->rc_skipped = 0;
1059 		rc->rc_force_repair = 0;
1060 		rc->rc_allow_repair = 1;
1061 		rc->rc_need_orig_restore = B_FALSE;
1062 
1063 		if (q == 0 && i >= bc)
1064 			rc->rc_size = 0;
1065 		else if (i < bc)
1066 			rc->rc_size = (q + 1) << ashift;
1067 		else
1068 			rc->rc_size = q << ashift;
1069 
1070 		asize += rc->rc_size;
1071 	}
1072 
1073 	ASSERT3U(asize, ==, tot << ashift);
1074 	rr->rr_nempty = roundup(tot, groupwidth) - tot;
1075 	IMPLY(bc > 0, rr->rr_nempty == groupwidth - bc);
1076 
1077 	/* Allocate buffers for the parity columns */
1078 	for (uint64_t c = 0; c < rr->rr_firstdatacol; c++) {
1079 		raidz_col_t *rc = &rr->rr_col[c];
1080 		rc->rc_abd = abd_alloc_linear(rc->rc_size, B_FALSE);
1081 	}
1082 
1083 	/*
1084 	 * Map buffers for data columns and allocate/map buffers for skip
1085 	 * sectors.  There are three distinct cases for dRAID which are
1086 	 * required to support sequential rebuild.
1087 	 */
1088 	if (zio->io_type == ZIO_TYPE_WRITE) {
1089 		vdev_draid_map_alloc_write(zio, abd_offset, rr);
1090 	} else if ((rr->rr_nempty > 0) &&
1091 	    (zio->io_flags & (ZIO_FLAG_SCRUB | ZIO_FLAG_RESILVER))) {
1092 		vdev_draid_map_alloc_scrub(zio, abd_offset, rr);
1093 	} else {
1094 		ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ);
1095 		vdev_draid_map_alloc_read(zio, abd_offset, rr);
1096 	}
1097 
1098 	return (io_size);
1099 }
1100 
1101 /*
1102  * Allocate the raidz mapping to be applied to the dRAID I/O.  The parity
1103  * calculations for dRAID are identical to raidz however there are a few
1104  * differences in the layout.
1105  *
1106  * - dRAID always allocates a full stripe width. Any extra sectors due
1107  *   this padding are zero filled and written to disk. They will be read
1108  *   back during a scrub or repair operation since they are included in
1109  *   the parity calculation. This property enables sequential resilvering.
1110  *
1111  * - When the block at the logical offset spans redundancy groups then two
1112  *   rows are allocated in the raidz_map_t. One row resides at the end of
1113  *   the first group and the other at the start of the following group.
1114  */
1115 static raidz_map_t *
1116 vdev_draid_map_alloc(zio_t *zio)
1117 {
1118 	raidz_row_t *rr[2];
1119 	uint64_t abd_offset = 0;
1120 	uint64_t abd_size = zio->io_size;
1121 	uint64_t io_offset = zio->io_offset;
1122 	uint64_t size;
1123 	int nrows = 1;
1124 
1125 	size = vdev_draid_map_alloc_row(zio, &rr[0], io_offset,
1126 	    abd_offset, abd_size);
1127 	if (size < abd_size) {
1128 		vdev_t *vd = zio->io_vd;
1129 
1130 		io_offset += vdev_draid_asize(vd, size);
1131 		abd_offset += size;
1132 		abd_size -= size;
1133 		nrows++;
1134 
1135 		ASSERT3U(io_offset, ==, vdev_draid_group_to_offset(
1136 		    vd, vdev_draid_offset_to_group(vd, io_offset)));
1137 		ASSERT3U(abd_offset, <, zio->io_size);
1138 		ASSERT3U(abd_size, !=, 0);
1139 
1140 		size = vdev_draid_map_alloc_row(zio, &rr[1],
1141 		    io_offset, abd_offset, abd_size);
1142 		VERIFY3U(size, ==, abd_size);
1143 	}
1144 
1145 	raidz_map_t *rm;
1146 	rm = kmem_zalloc(offsetof(raidz_map_t, rm_row[nrows]), KM_SLEEP);
1147 	rm->rm_ops = vdev_raidz_math_get_ops();
1148 	rm->rm_nrows = nrows;
1149 	rm->rm_row[0] = rr[0];
1150 	if (nrows == 2)
1151 		rm->rm_row[1] = rr[1];
1152 
1153 	return (rm);
1154 }
1155 
1156 /*
1157  * Given an offset into a dRAID return the next group width aligned offset
1158  * which can be used to start an allocation.
1159  */
1160 static uint64_t
1161 vdev_draid_get_astart(vdev_t *vd, const uint64_t start)
1162 {
1163 	vdev_draid_config_t *vdc = vd->vdev_tsd;
1164 
1165 	ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops);
1166 
1167 	return (roundup(start, vdc->vdc_groupwidth << vd->vdev_ashift));
1168 }
1169 
1170 /*
1171  * Allocatable space for dRAID is (children - nspares) * sizeof(smallest child)
1172  * rounded down to the last full slice.  So each child must provide at least
1173  * 1 / (children - nspares) of its asize.
1174  */
1175 static uint64_t
1176 vdev_draid_min_asize(vdev_t *vd)
1177 {
1178 	vdev_draid_config_t *vdc = vd->vdev_tsd;
1179 
1180 	ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops);
1181 
1182 	return (VDEV_DRAID_REFLOW_RESERVE +
1183 	    (vd->vdev_min_asize + vdc->vdc_ndisks - 1) / (vdc->vdc_ndisks));
1184 }
1185 
1186 /*
1187  * When using dRAID the minimum allocation size is determined by the number
1188  * of data disks in the redundancy group.  Full stripes are always used.
1189  */
1190 static uint64_t
1191 vdev_draid_min_alloc(vdev_t *vd)
1192 {
1193 	vdev_draid_config_t *vdc = vd->vdev_tsd;
1194 
1195 	ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops);
1196 
1197 	return (vdc->vdc_ndata << vd->vdev_ashift);
1198 }
1199 
1200 /*
1201  * Returns true if the txg range does not exist on any leaf vdev.
1202  *
1203  * A dRAID spare does not fit into the DTL model. While it has child vdevs
1204  * there is no redundancy among them, and the effective child vdev is
1205  * determined by offset. Essentially we do a vdev_dtl_reassess() on the
1206  * fly by replacing a dRAID spare with the child vdev under the offset.
1207  * Note that it is a recursive process because the child vdev can be
1208  * another dRAID spare and so on.
1209  */
1210 boolean_t
1211 vdev_draid_missing(vdev_t *vd, uint64_t physical_offset, uint64_t txg,
1212     uint64_t size)
1213 {
1214 	if (vd->vdev_ops == &vdev_spare_ops ||
1215 	    vd->vdev_ops == &vdev_replacing_ops) {
1216 		/*
1217 		 * Check all of the readable children, if any child
1218 		 * contains the txg range the data it is not missing.
1219 		 */
1220 		for (int c = 0; c < vd->vdev_children; c++) {
1221 			vdev_t *cvd = vd->vdev_child[c];
1222 
1223 			if (!vdev_readable(cvd))
1224 				continue;
1225 
1226 			if (!vdev_draid_missing(cvd, physical_offset,
1227 			    txg, size))
1228 				return (B_FALSE);
1229 		}
1230 
1231 		return (B_TRUE);
1232 	}
1233 
1234 	if (vd->vdev_ops == &vdev_draid_spare_ops) {
1235 		/*
1236 		 * When sequentially resilvering we don't have a proper
1237 		 * txg range so instead we must presume all txgs are
1238 		 * missing on this vdev until the resilver completes.
1239 		 */
1240 		if (vd->vdev_rebuild_txg != 0)
1241 			return (B_TRUE);
1242 
1243 		/*
1244 		 * DTL_MISSING is set for all prior txgs when a resilver
1245 		 * is started in spa_vdev_attach().
1246 		 */
1247 		if (vdev_dtl_contains(vd, DTL_MISSING, txg, size))
1248 			return (B_TRUE);
1249 
1250 		/*
1251 		 * Consult the DTL on the relevant vdev. Either a vdev
1252 		 * leaf or spare/replace mirror child may be returned so
1253 		 * we must recursively call vdev_draid_missing_impl().
1254 		 */
1255 		vd = vdev_draid_spare_get_child(vd, physical_offset);
1256 		if (vd == NULL)
1257 			return (B_TRUE);
1258 
1259 		return (vdev_draid_missing(vd, physical_offset,
1260 		    txg, size));
1261 	}
1262 
1263 	return (vdev_dtl_contains(vd, DTL_MISSING, txg, size));
1264 }
1265 
1266 /*
1267  * Returns true if the txg is only partially replicated on the leaf vdevs.
1268  */
1269 static boolean_t
1270 vdev_draid_partial(vdev_t *vd, uint64_t physical_offset, uint64_t txg,
1271     uint64_t size)
1272 {
1273 	if (vd->vdev_ops == &vdev_spare_ops ||
1274 	    vd->vdev_ops == &vdev_replacing_ops) {
1275 		/*
1276 		 * Check all of the readable children, if any child is
1277 		 * missing the txg range then it is partially replicated.
1278 		 */
1279 		for (int c = 0; c < vd->vdev_children; c++) {
1280 			vdev_t *cvd = vd->vdev_child[c];
1281 
1282 			if (!vdev_readable(cvd))
1283 				continue;
1284 
1285 			if (vdev_draid_partial(cvd, physical_offset, txg, size))
1286 				return (B_TRUE);
1287 		}
1288 
1289 		return (B_FALSE);
1290 	}
1291 
1292 	if (vd->vdev_ops == &vdev_draid_spare_ops) {
1293 		/*
1294 		 * When sequentially resilvering we don't have a proper
1295 		 * txg range so instead we must presume all txgs are
1296 		 * missing on this vdev until the resilver completes.
1297 		 */
1298 		if (vd->vdev_rebuild_txg != 0)
1299 			return (B_TRUE);
1300 
1301 		/*
1302 		 * DTL_MISSING is set for all prior txgs when a resilver
1303 		 * is started in spa_vdev_attach().
1304 		 */
1305 		if (vdev_dtl_contains(vd, DTL_MISSING, txg, size))
1306 			return (B_TRUE);
1307 
1308 		/*
1309 		 * Consult the DTL on the relevant vdev. Either a vdev
1310 		 * leaf or spare/replace mirror child may be returned so
1311 		 * we must recursively call vdev_draid_missing_impl().
1312 		 */
1313 		vd = vdev_draid_spare_get_child(vd, physical_offset);
1314 		if (vd == NULL)
1315 			return (B_TRUE);
1316 
1317 		return (vdev_draid_partial(vd, physical_offset, txg, size));
1318 	}
1319 
1320 	return (vdev_dtl_contains(vd, DTL_MISSING, txg, size));
1321 }
1322 
1323 /*
1324  * Determine if the vdev is readable at the given offset.
1325  */
1326 boolean_t
1327 vdev_draid_readable(vdev_t *vd, uint64_t physical_offset)
1328 {
1329 	if (vd->vdev_ops == &vdev_draid_spare_ops) {
1330 		vd = vdev_draid_spare_get_child(vd, physical_offset);
1331 		if (vd == NULL)
1332 			return (B_FALSE);
1333 	}
1334 
1335 	if (vd->vdev_ops == &vdev_spare_ops ||
1336 	    vd->vdev_ops == &vdev_replacing_ops) {
1337 
1338 		for (int c = 0; c < vd->vdev_children; c++) {
1339 			vdev_t *cvd = vd->vdev_child[c];
1340 
1341 			if (!vdev_readable(cvd))
1342 				continue;
1343 
1344 			if (vdev_draid_readable(cvd, physical_offset))
1345 				return (B_TRUE);
1346 		}
1347 
1348 		return (B_FALSE);
1349 	}
1350 
1351 	return (vdev_readable(vd));
1352 }
1353 
1354 /*
1355  * Returns the first distributed spare found under the provided vdev tree.
1356  */
1357 static vdev_t *
1358 vdev_draid_find_spare(vdev_t *vd)
1359 {
1360 	if (vd->vdev_ops == &vdev_draid_spare_ops)
1361 		return (vd);
1362 
1363 	for (int c = 0; c < vd->vdev_children; c++) {
1364 		vdev_t *svd = vdev_draid_find_spare(vd->vdev_child[c]);
1365 		if (svd != NULL)
1366 			return (svd);
1367 	}
1368 
1369 	return (NULL);
1370 }
1371 
1372 /*
1373  * Returns B_TRUE if the passed in vdev is currently "faulted".
1374  * Faulted, in this context, means that the vdev represents a
1375  * replacing or sparing vdev tree.
1376  */
1377 static boolean_t
1378 vdev_draid_faulted(vdev_t *vd, uint64_t physical_offset)
1379 {
1380 	if (vd->vdev_ops == &vdev_draid_spare_ops) {
1381 		vd = vdev_draid_spare_get_child(vd, physical_offset);
1382 		if (vd == NULL)
1383 			return (B_FALSE);
1384 
1385 		/*
1386 		 * After resolving the distributed spare to a leaf vdev
1387 		 * check the parent to determine if it's "faulted".
1388 		 */
1389 		vd = vd->vdev_parent;
1390 	}
1391 
1392 	return (vd->vdev_ops == &vdev_replacing_ops ||
1393 	    vd->vdev_ops == &vdev_spare_ops);
1394 }
1395 
1396 /*
1397  * Determine if the dRAID block at the logical offset is degraded.
1398  * Used by sequential resilver.
1399  */
1400 static boolean_t
1401 vdev_draid_group_degraded(vdev_t *vd, uint64_t offset)
1402 {
1403 	vdev_draid_config_t *vdc = vd->vdev_tsd;
1404 
1405 	ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops);
1406 	ASSERT3U(vdev_draid_get_astart(vd, offset), ==, offset);
1407 
1408 	uint64_t groupstart, perm;
1409 	uint64_t physical_offset = vdev_draid_logical_to_physical(vd,
1410 	    offset, &perm, &groupstart);
1411 
1412 	uint8_t *base;
1413 	uint64_t iter;
1414 	vdev_draid_get_perm(vdc, perm, &base, &iter);
1415 
1416 	for (uint64_t i = 0; i < vdc->vdc_groupwidth; i++) {
1417 		uint64_t c = (groupstart + i) % vdc->vdc_ndisks;
1418 		uint64_t cid = vdev_draid_permute_id(vdc, base, iter, c);
1419 		vdev_t *cvd = vd->vdev_child[cid];
1420 
1421 		/* Group contains a faulted vdev. */
1422 		if (vdev_draid_faulted(cvd, physical_offset))
1423 			return (B_TRUE);
1424 
1425 		/*
1426 		 * Always check groups with active distributed spares
1427 		 * because any vdev failure in the pool will affect them.
1428 		 */
1429 		if (vdev_draid_find_spare(cvd) != NULL)
1430 			return (B_TRUE);
1431 	}
1432 
1433 	return (B_FALSE);
1434 }
1435 
1436 /*
1437  * Determine if the txg is missing.  Used by healing resilver.
1438  */
1439 static boolean_t
1440 vdev_draid_group_missing(vdev_t *vd, uint64_t offset, uint64_t txg,
1441     uint64_t size)
1442 {
1443 	vdev_draid_config_t *vdc = vd->vdev_tsd;
1444 
1445 	ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops);
1446 	ASSERT3U(vdev_draid_get_astart(vd, offset), ==, offset);
1447 
1448 	uint64_t groupstart, perm;
1449 	uint64_t physical_offset = vdev_draid_logical_to_physical(vd,
1450 	    offset, &perm, &groupstart);
1451 
1452 	uint8_t *base;
1453 	uint64_t iter;
1454 	vdev_draid_get_perm(vdc, perm, &base, &iter);
1455 
1456 	for (uint64_t i = 0; i < vdc->vdc_groupwidth; i++) {
1457 		uint64_t c = (groupstart + i) % vdc->vdc_ndisks;
1458 		uint64_t cid = vdev_draid_permute_id(vdc, base, iter, c);
1459 		vdev_t *cvd = vd->vdev_child[cid];
1460 
1461 		/* Transaction group is known to be partially replicated. */
1462 		if (vdev_draid_partial(cvd, physical_offset, txg, size))
1463 			return (B_TRUE);
1464 
1465 		/*
1466 		 * Always check groups with active distributed spares
1467 		 * because any vdev failure in the pool will affect them.
1468 		 */
1469 		if (vdev_draid_find_spare(cvd) != NULL)
1470 			return (B_TRUE);
1471 	}
1472 
1473 	return (B_FALSE);
1474 }
1475 
1476 /*
1477  * Find the smallest child asize and largest sector size to calculate the
1478  * available capacity.  Distributed spares are ignored since their capacity
1479  * is also based of the minimum child size in the top-level dRAID.
1480  */
1481 static void
1482 vdev_draid_calculate_asize(vdev_t *vd, uint64_t *asizep, uint64_t *max_asizep,
1483     uint64_t *logical_ashiftp, uint64_t *physical_ashiftp)
1484 {
1485 	uint64_t logical_ashift = 0, physical_ashift = 0;
1486 	uint64_t asize = 0, max_asize = 0;
1487 
1488 	ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops);
1489 
1490 	for (int c = 0; c < vd->vdev_children; c++) {
1491 		vdev_t *cvd = vd->vdev_child[c];
1492 
1493 		if (cvd->vdev_ops == &vdev_draid_spare_ops)
1494 			continue;
1495 
1496 		asize = MIN(asize - 1, cvd->vdev_asize - 1) + 1;
1497 		max_asize = MIN(max_asize - 1, cvd->vdev_max_asize - 1) + 1;
1498 		logical_ashift = MAX(logical_ashift, cvd->vdev_ashift);
1499 		physical_ashift = MAX(physical_ashift,
1500 		    cvd->vdev_physical_ashift);
1501 	}
1502 
1503 	*asizep = asize;
1504 	*max_asizep = max_asize;
1505 	*logical_ashiftp = logical_ashift;
1506 	*physical_ashiftp = physical_ashift;
1507 }
1508 
1509 /*
1510  * Open spare vdevs.
1511  */
1512 static boolean_t
1513 vdev_draid_open_spares(vdev_t *vd)
1514 {
1515 	return (vd->vdev_ops == &vdev_draid_spare_ops ||
1516 	    vd->vdev_ops == &vdev_replacing_ops ||
1517 	    vd->vdev_ops == &vdev_spare_ops);
1518 }
1519 
1520 /*
1521  * Open all children, excluding spares.
1522  */
1523 static boolean_t
1524 vdev_draid_open_children(vdev_t *vd)
1525 {
1526 	return (!vdev_draid_open_spares(vd));
1527 }
1528 
1529 /*
1530  * Open a top-level dRAID vdev.
1531  */
1532 static int
1533 vdev_draid_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize,
1534     uint64_t *logical_ashift, uint64_t *physical_ashift)
1535 {
1536 	vdev_draid_config_t *vdc =  vd->vdev_tsd;
1537 	uint64_t nparity = vdc->vdc_nparity;
1538 	int open_errors = 0;
1539 
1540 	if (nparity > VDEV_DRAID_MAXPARITY ||
1541 	    vd->vdev_children < nparity + 1) {
1542 		vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
1543 		return (SET_ERROR(EINVAL));
1544 	}
1545 
1546 	/*
1547 	 * First open the normal children then the distributed spares.  This
1548 	 * ordering is important to ensure the distributed spares calculate
1549 	 * the correct psize in the event that the dRAID vdevs were expanded.
1550 	 */
1551 	vdev_open_children_subset(vd, vdev_draid_open_children);
1552 	vdev_open_children_subset(vd, vdev_draid_open_spares);
1553 
1554 	/* Verify enough of the children are available to continue. */
1555 	for (int c = 0; c < vd->vdev_children; c++) {
1556 		if (vd->vdev_child[c]->vdev_open_error != 0) {
1557 			if ((++open_errors) > nparity) {
1558 				vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS;
1559 				return (SET_ERROR(ENXIO));
1560 			}
1561 		}
1562 	}
1563 
1564 	/*
1565 	 * Allocatable capacity is the sum of the space on all children less
1566 	 * the number of distributed spares rounded down to last full row
1567 	 * and then to the last full group. An additional 32MB of scratch
1568 	 * space is reserved at the end of each child for use by the dRAID
1569 	 * expansion feature.
1570 	 */
1571 	uint64_t child_asize, child_max_asize;
1572 	vdev_draid_calculate_asize(vd, &child_asize, &child_max_asize,
1573 	    logical_ashift, physical_ashift);
1574 
1575 	/*
1576 	 * Should be unreachable since the minimum child size is 64MB, but
1577 	 * we want to make sure an underflow absolutely cannot occur here.
1578 	 */
1579 	if (child_asize < VDEV_DRAID_REFLOW_RESERVE ||
1580 	    child_max_asize < VDEV_DRAID_REFLOW_RESERVE) {
1581 		return (SET_ERROR(ENXIO));
1582 	}
1583 
1584 	child_asize = ((child_asize - VDEV_DRAID_REFLOW_RESERVE) /
1585 	    VDEV_DRAID_ROWHEIGHT) * VDEV_DRAID_ROWHEIGHT;
1586 	child_max_asize = ((child_max_asize - VDEV_DRAID_REFLOW_RESERVE) /
1587 	    VDEV_DRAID_ROWHEIGHT) * VDEV_DRAID_ROWHEIGHT;
1588 
1589 	*asize = (((child_asize * vdc->vdc_ndisks) / vdc->vdc_groupsz) *
1590 	    vdc->vdc_groupsz);
1591 	*max_asize = (((child_max_asize * vdc->vdc_ndisks) / vdc->vdc_groupsz) *
1592 	    vdc->vdc_groupsz);
1593 
1594 	return (0);
1595 }
1596 
1597 /*
1598  * Close a top-level dRAID vdev.
1599  */
1600 static void
1601 vdev_draid_close(vdev_t *vd)
1602 {
1603 	for (int c = 0; c < vd->vdev_children; c++) {
1604 		if (vd->vdev_child[c] != NULL)
1605 			vdev_close(vd->vdev_child[c]);
1606 	}
1607 }
1608 
1609 /*
1610  * Return the maximum asize for a rebuild zio in the provided range
1611  * given the following constraints.  A dRAID chunks may not:
1612  *
1613  * - Exceed the maximum allowed block size (SPA_MAXBLOCKSIZE), or
1614  * - Span dRAID redundancy groups.
1615  */
1616 static uint64_t
1617 vdev_draid_rebuild_asize(vdev_t *vd, uint64_t start, uint64_t asize,
1618     uint64_t max_segment)
1619 {
1620 	vdev_draid_config_t *vdc = vd->vdev_tsd;
1621 
1622 	ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops);
1623 
1624 	uint64_t ashift = vd->vdev_ashift;
1625 	uint64_t ndata = vdc->vdc_ndata;
1626 	uint64_t psize = MIN(P2ROUNDUP(max_segment * ndata, 1 << ashift),
1627 	    SPA_MAXBLOCKSIZE);
1628 
1629 	ASSERT3U(vdev_draid_get_astart(vd, start), ==, start);
1630 	ASSERT3U(asize % (vdc->vdc_groupwidth << ashift), ==, 0);
1631 
1632 	/* Chunks must evenly span all data columns in the group. */
1633 	psize = (((psize >> ashift) / ndata) * ndata) << ashift;
1634 	uint64_t chunk_size = MIN(asize, vdev_psize_to_asize(vd, psize));
1635 
1636 	/* Reduce the chunk size to the group space remaining. */
1637 	uint64_t group = vdev_draid_offset_to_group(vd, start);
1638 	uint64_t left = vdev_draid_group_to_offset(vd, group + 1) - start;
1639 	chunk_size = MIN(chunk_size, left);
1640 
1641 	ASSERT3U(chunk_size % (vdc->vdc_groupwidth << ashift), ==, 0);
1642 	ASSERT3U(vdev_draid_offset_to_group(vd, start), ==,
1643 	    vdev_draid_offset_to_group(vd, start + chunk_size - 1));
1644 
1645 	return (chunk_size);
1646 }
1647 
1648 /*
1649  * Align the start of the metaslab to the group width and slightly reduce
1650  * its size to a multiple of the group width.  Since full stripe writes are
1651  * required by dRAID this space is unallocable.  Furthermore, aligning the
1652  * metaslab start is important for vdev initialize and TRIM which both operate
1653  * on metaslab boundaries which vdev_xlate() expects to be aligned.
1654  */
1655 static void
1656 vdev_draid_metaslab_init(vdev_t *vd, uint64_t *ms_start, uint64_t *ms_size)
1657 {
1658 	vdev_draid_config_t *vdc = vd->vdev_tsd;
1659 
1660 	ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops);
1661 
1662 	uint64_t sz = vdc->vdc_groupwidth << vd->vdev_ashift;
1663 	uint64_t astart = vdev_draid_get_astart(vd, *ms_start);
1664 	uint64_t asize = ((*ms_size - (astart - *ms_start)) / sz) * sz;
1665 
1666 	*ms_start = astart;
1667 	*ms_size = asize;
1668 
1669 	ASSERT0(*ms_start % sz);
1670 	ASSERT0(*ms_size % sz);
1671 }
1672 
1673 /*
1674  * Add virtual dRAID spares to the list of valid spares. In order to accomplish
1675  * this the existing array must be freed and reallocated with the additional
1676  * entries.
1677  */
1678 int
1679 vdev_draid_spare_create(nvlist_t *nvroot, vdev_t *vd, uint64_t *ndraidp,
1680     uint64_t next_vdev_id)
1681 {
1682 	uint64_t draid_nspares = 0;
1683 	uint64_t ndraid = 0;
1684 	int error;
1685 
1686 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
1687 		vdev_t *cvd = vd->vdev_child[i];
1688 
1689 		if (cvd->vdev_ops == &vdev_draid_ops) {
1690 			vdev_draid_config_t *vdc = cvd->vdev_tsd;
1691 			draid_nspares += vdc->vdc_nspares;
1692 			ndraid++;
1693 		}
1694 	}
1695 
1696 	if (draid_nspares == 0) {
1697 		*ndraidp = ndraid;
1698 		return (0);
1699 	}
1700 
1701 	nvlist_t **old_spares, **new_spares;
1702 	uint_t old_nspares;
1703 	error = nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
1704 	    &old_spares, &old_nspares);
1705 	if (error)
1706 		old_nspares = 0;
1707 
1708 	/* Allocate memory and copy of the existing spares. */
1709 	new_spares = kmem_alloc(sizeof (nvlist_t *) *
1710 	    (draid_nspares + old_nspares), KM_SLEEP);
1711 	for (uint_t i = 0; i < old_nspares; i++)
1712 		new_spares[i] = fnvlist_dup(old_spares[i]);
1713 
1714 	/* Add new distributed spares to ZPOOL_CONFIG_SPARES. */
1715 	uint64_t n = old_nspares;
1716 	for (uint64_t vdev_id = 0; vdev_id < vd->vdev_children; vdev_id++) {
1717 		vdev_t *cvd = vd->vdev_child[vdev_id];
1718 		char path[64];
1719 
1720 		if (cvd->vdev_ops != &vdev_draid_ops)
1721 			continue;
1722 
1723 		vdev_draid_config_t *vdc = cvd->vdev_tsd;
1724 		uint64_t nspares = vdc->vdc_nspares;
1725 		uint64_t nparity = vdc->vdc_nparity;
1726 
1727 		for (uint64_t spare_id = 0; spare_id < nspares; spare_id++) {
1728 			memset(path, 0, sizeof (path));
1729 			(void) snprintf(path, sizeof (path) - 1,
1730 			    "%s%llu-%llu-%llu", VDEV_TYPE_DRAID,
1731 			    (u_longlong_t)nparity,
1732 			    (u_longlong_t)next_vdev_id + vdev_id,
1733 			    (u_longlong_t)spare_id);
1734 
1735 			nvlist_t *spare = fnvlist_alloc();
1736 			fnvlist_add_string(spare, ZPOOL_CONFIG_PATH, path);
1737 			fnvlist_add_string(spare, ZPOOL_CONFIG_TYPE,
1738 			    VDEV_TYPE_DRAID_SPARE);
1739 			fnvlist_add_uint64(spare, ZPOOL_CONFIG_TOP_GUID,
1740 			    cvd->vdev_guid);
1741 			fnvlist_add_uint64(spare, ZPOOL_CONFIG_SPARE_ID,
1742 			    spare_id);
1743 			fnvlist_add_uint64(spare, ZPOOL_CONFIG_IS_LOG, 0);
1744 			fnvlist_add_uint64(spare, ZPOOL_CONFIG_IS_SPARE, 1);
1745 			fnvlist_add_uint64(spare, ZPOOL_CONFIG_WHOLE_DISK, 1);
1746 			fnvlist_add_uint64(spare, ZPOOL_CONFIG_ASHIFT,
1747 			    cvd->vdev_ashift);
1748 
1749 			new_spares[n] = spare;
1750 			n++;
1751 		}
1752 	}
1753 
1754 	if (n > 0) {
1755 		(void) nvlist_remove_all(nvroot, ZPOOL_CONFIG_SPARES);
1756 		fnvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
1757 		    (const nvlist_t **)new_spares, n);
1758 	}
1759 
1760 	for (int i = 0; i < n; i++)
1761 		nvlist_free(new_spares[i]);
1762 
1763 	kmem_free(new_spares, sizeof (*new_spares) * n);
1764 	*ndraidp = ndraid;
1765 
1766 	return (0);
1767 }
1768 
1769 /*
1770  * Determine if any portion of the provided block resides on a child vdev
1771  * with a dirty DTL and therefore needs to be resilvered.
1772  */
1773 static boolean_t
1774 vdev_draid_need_resilver(vdev_t *vd, const dva_t *dva, size_t psize,
1775     uint64_t phys_birth)
1776 {
1777 	uint64_t offset = DVA_GET_OFFSET(dva);
1778 	uint64_t asize = vdev_draid_asize(vd, psize);
1779 
1780 	if (phys_birth == TXG_UNKNOWN) {
1781 		/*
1782 		 * Sequential resilver.  There is no meaningful phys_birth
1783 		 * for this block, we can only determine if block resides
1784 		 * in a degraded group in which case it must be resilvered.
1785 		 */
1786 		ASSERT3U(vdev_draid_offset_to_group(vd, offset), ==,
1787 		    vdev_draid_offset_to_group(vd, offset + asize - 1));
1788 
1789 		return (vdev_draid_group_degraded(vd, offset));
1790 	} else {
1791 		/*
1792 		 * Healing resilver.  TXGs not in DTL_PARTIAL are intact,
1793 		 * as are blocks in non-degraded groups.
1794 		 */
1795 		if (!vdev_dtl_contains(vd, DTL_PARTIAL, phys_birth, 1))
1796 			return (B_FALSE);
1797 
1798 		if (vdev_draid_group_missing(vd, offset, phys_birth, 1))
1799 			return (B_TRUE);
1800 
1801 		/* The block may span groups in which case check both. */
1802 		if (vdev_draid_offset_to_group(vd, offset) !=
1803 		    vdev_draid_offset_to_group(vd, offset + asize - 1)) {
1804 			if (vdev_draid_group_missing(vd,
1805 			    offset + asize, phys_birth, 1))
1806 				return (B_TRUE);
1807 		}
1808 
1809 		return (B_FALSE);
1810 	}
1811 }
1812 
1813 static boolean_t
1814 vdev_draid_rebuilding(vdev_t *vd)
1815 {
1816 	if (vd->vdev_ops->vdev_op_leaf && vd->vdev_rebuild_txg)
1817 		return (B_TRUE);
1818 
1819 	for (int i = 0; i < vd->vdev_children; i++) {
1820 		if (vdev_draid_rebuilding(vd->vdev_child[i])) {
1821 			return (B_TRUE);
1822 		}
1823 	}
1824 
1825 	return (B_FALSE);
1826 }
1827 
1828 static void
1829 vdev_draid_io_verify(vdev_t *vd, raidz_row_t *rr, int col)
1830 {
1831 #ifdef ZFS_DEBUG
1832 	range_seg64_t logical_rs, physical_rs, remain_rs;
1833 	logical_rs.rs_start = rr->rr_offset;
1834 	logical_rs.rs_end = logical_rs.rs_start +
1835 	    vdev_draid_asize(vd, rr->rr_size);
1836 
1837 	raidz_col_t *rc = &rr->rr_col[col];
1838 	vdev_t *cvd = vd->vdev_child[rc->rc_devidx];
1839 
1840 	vdev_xlate(cvd, &logical_rs, &physical_rs, &remain_rs);
1841 	ASSERT(vdev_xlate_is_empty(&remain_rs));
1842 	ASSERT3U(rc->rc_offset, ==, physical_rs.rs_start);
1843 	ASSERT3U(rc->rc_offset, <, physical_rs.rs_end);
1844 	ASSERT3U(rc->rc_offset + rc->rc_size, ==, physical_rs.rs_end);
1845 #endif
1846 }
1847 
1848 /*
1849  * For write operations:
1850  * 1. Generate the parity data
1851  * 2. Create child zio write operations to each column's vdev, for both
1852  *    data and parity.  A gang ABD is allocated by vdev_draid_map_alloc()
1853  *    if a skip sector needs to be added to a column.
1854  */
1855 static void
1856 vdev_draid_io_start_write(zio_t *zio, raidz_row_t *rr)
1857 {
1858 	vdev_t *vd = zio->io_vd;
1859 	raidz_map_t *rm = zio->io_vsd;
1860 
1861 	vdev_raidz_generate_parity_row(rm, rr);
1862 
1863 	for (int c = 0; c < rr->rr_cols; c++) {
1864 		raidz_col_t *rc = &rr->rr_col[c];
1865 
1866 		/*
1867 		 * Empty columns are zero filled and included in the parity
1868 		 * calculation and therefore must be written.
1869 		 */
1870 		ASSERT3U(rc->rc_size, !=, 0);
1871 
1872 		/* Verify physical to logical translation */
1873 		vdev_draid_io_verify(vd, rr, c);
1874 
1875 		zio_nowait(zio_vdev_child_io(zio, NULL,
1876 		    vd->vdev_child[rc->rc_devidx], rc->rc_offset,
1877 		    rc->rc_abd, rc->rc_size, zio->io_type, zio->io_priority,
1878 		    0, vdev_raidz_child_done, rc));
1879 	}
1880 }
1881 
1882 /*
1883  * For read operations:
1884  * 1. The vdev_draid_map_alloc() function will create a minimal raidz
1885  *    mapping for the read based on the zio->io_flags.  There are two
1886  *    possible mappings either 1) a normal read, or 2) a scrub/resilver.
1887  * 2. Create the zio read operations.  This will include all parity
1888  *    columns and skip sectors for a scrub/resilver.
1889  */
1890 static void
1891 vdev_draid_io_start_read(zio_t *zio, raidz_row_t *rr)
1892 {
1893 	vdev_t *vd = zio->io_vd;
1894 
1895 	/* Sequential rebuild must do IO at redundancy group boundary. */
1896 	IMPLY(zio->io_priority == ZIO_PRIORITY_REBUILD, rr->rr_nempty == 0);
1897 
1898 	/*
1899 	 * Iterate over the columns in reverse order so that we hit the parity
1900 	 * last.  Any errors along the way will force us to read the parity.
1901 	 * For scrub/resilver IOs which verify skip sectors, a gang ABD will
1902 	 * have been allocated to store them and rc->rc_size is increased.
1903 	 */
1904 	for (int c = rr->rr_cols - 1; c >= 0; c--) {
1905 		raidz_col_t *rc = &rr->rr_col[c];
1906 		vdev_t *cvd = vd->vdev_child[rc->rc_devidx];
1907 
1908 		if (!vdev_draid_readable(cvd, rc->rc_offset)) {
1909 			if (c >= rr->rr_firstdatacol)
1910 				rr->rr_missingdata++;
1911 			else
1912 				rr->rr_missingparity++;
1913 			rc->rc_error = SET_ERROR(ENXIO);
1914 			rc->rc_tried = 1;
1915 			rc->rc_skipped = 1;
1916 			continue;
1917 		}
1918 
1919 		if (vdev_draid_missing(cvd, rc->rc_offset, zio->io_txg, 1)) {
1920 			if (c >= rr->rr_firstdatacol)
1921 				rr->rr_missingdata++;
1922 			else
1923 				rr->rr_missingparity++;
1924 			rc->rc_error = SET_ERROR(ESTALE);
1925 			rc->rc_skipped = 1;
1926 			continue;
1927 		}
1928 
1929 		/*
1930 		 * Empty columns may be read during vdev_draid_io_done().
1931 		 * Only skip them after the readable and missing checks
1932 		 * verify they are available.
1933 		 */
1934 		if (rc->rc_size == 0) {
1935 			rc->rc_skipped = 1;
1936 			continue;
1937 		}
1938 
1939 		if (zio->io_flags & ZIO_FLAG_RESILVER) {
1940 			vdev_t *svd;
1941 
1942 			/*
1943 			 * Sequential rebuilds need to always consider the data
1944 			 * on the child being rebuilt to be stale.  This is
1945 			 * important when all columns are available to aid
1946 			 * known reconstruction in identifing which columns
1947 			 * contain incorrect data.
1948 			 *
1949 			 * Furthermore, all repairs need to be constrained to
1950 			 * the devices being rebuilt because without a checksum
1951 			 * we cannot verify the data is actually correct and
1952 			 * performing an incorrect repair could result in
1953 			 * locking in damage and making the data unrecoverable.
1954 			 */
1955 			if (zio->io_priority == ZIO_PRIORITY_REBUILD) {
1956 				if (vdev_draid_rebuilding(cvd)) {
1957 					if (c >= rr->rr_firstdatacol)
1958 						rr->rr_missingdata++;
1959 					else
1960 						rr->rr_missingparity++;
1961 					rc->rc_error = SET_ERROR(ESTALE);
1962 					rc->rc_skipped = 1;
1963 					rc->rc_allow_repair = 1;
1964 					continue;
1965 				} else {
1966 					rc->rc_allow_repair = 0;
1967 				}
1968 			} else {
1969 				rc->rc_allow_repair = 1;
1970 			}
1971 
1972 			/*
1973 			 * If this child is a distributed spare then the
1974 			 * offset might reside on the vdev being replaced.
1975 			 * In which case this data must be written to the
1976 			 * new device.  Failure to do so would result in
1977 			 * checksum errors when the old device is detached
1978 			 * and the pool is scrubbed.
1979 			 */
1980 			if ((svd = vdev_draid_find_spare(cvd)) != NULL) {
1981 				svd = vdev_draid_spare_get_child(svd,
1982 				    rc->rc_offset);
1983 				if (svd && (svd->vdev_ops == &vdev_spare_ops ||
1984 				    svd->vdev_ops == &vdev_replacing_ops)) {
1985 					rc->rc_force_repair = 1;
1986 
1987 					if (vdev_draid_rebuilding(svd))
1988 						rc->rc_allow_repair = 1;
1989 				}
1990 			}
1991 
1992 			/*
1993 			 * Always issue a repair IO to this child when its
1994 			 * a spare or replacing vdev with an active rebuild.
1995 			 */
1996 			if ((cvd->vdev_ops == &vdev_spare_ops ||
1997 			    cvd->vdev_ops == &vdev_replacing_ops) &&
1998 			    vdev_draid_rebuilding(cvd)) {
1999 				rc->rc_force_repair = 1;
2000 				rc->rc_allow_repair = 1;
2001 			}
2002 		}
2003 	}
2004 
2005 	/*
2006 	 * Either a parity or data column is missing this means a repair
2007 	 * may be attempted by vdev_draid_io_done().  Expand the raid map
2008 	 * to read in empty columns which are needed along with the parity
2009 	 * during reconstruction.
2010 	 */
2011 	if ((rr->rr_missingdata > 0 || rr->rr_missingparity > 0) &&
2012 	    rr->rr_nempty > 0 && rr->rr_abd_empty == NULL) {
2013 		vdev_draid_map_alloc_empty(zio, rr);
2014 	}
2015 
2016 	for (int c = rr->rr_cols - 1; c >= 0; c--) {
2017 		raidz_col_t *rc = &rr->rr_col[c];
2018 		vdev_t *cvd = vd->vdev_child[rc->rc_devidx];
2019 
2020 		if (rc->rc_error || rc->rc_size == 0)
2021 			continue;
2022 
2023 		if (c >= rr->rr_firstdatacol || rr->rr_missingdata > 0 ||
2024 		    (zio->io_flags & (ZIO_FLAG_SCRUB | ZIO_FLAG_RESILVER))) {
2025 			zio_nowait(zio_vdev_child_io(zio, NULL, cvd,
2026 			    rc->rc_offset, rc->rc_abd, rc->rc_size,
2027 			    zio->io_type, zio->io_priority, 0,
2028 			    vdev_raidz_child_done, rc));
2029 		}
2030 	}
2031 }
2032 
2033 /*
2034  * Start an IO operation to a dRAID vdev.
2035  */
2036 static void
2037 vdev_draid_io_start(zio_t *zio)
2038 {
2039 	vdev_t *vd __maybe_unused = zio->io_vd;
2040 
2041 	ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops);
2042 	ASSERT3U(zio->io_offset, ==, vdev_draid_get_astart(vd, zio->io_offset));
2043 
2044 	raidz_map_t *rm = vdev_draid_map_alloc(zio);
2045 	zio->io_vsd = rm;
2046 	zio->io_vsd_ops = &vdev_raidz_vsd_ops;
2047 
2048 	if (zio->io_type == ZIO_TYPE_WRITE) {
2049 		for (int i = 0; i < rm->rm_nrows; i++) {
2050 			vdev_draid_io_start_write(zio, rm->rm_row[i]);
2051 		}
2052 	} else {
2053 		ASSERT(zio->io_type == ZIO_TYPE_READ);
2054 
2055 		for (int i = 0; i < rm->rm_nrows; i++) {
2056 			vdev_draid_io_start_read(zio, rm->rm_row[i]);
2057 		}
2058 	}
2059 
2060 	zio_execute(zio);
2061 }
2062 
2063 /*
2064  * Complete an IO operation on a dRAID vdev.  The raidz logic can be applied
2065  * to dRAID since the layout is fully described by the raidz_map_t.
2066  */
2067 static void
2068 vdev_draid_io_done(zio_t *zio)
2069 {
2070 	vdev_raidz_io_done(zio);
2071 }
2072 
2073 static void
2074 vdev_draid_state_change(vdev_t *vd, int faulted, int degraded)
2075 {
2076 	vdev_draid_config_t *vdc = vd->vdev_tsd;
2077 	ASSERT(vd->vdev_ops == &vdev_draid_ops);
2078 
2079 	if (faulted > vdc->vdc_nparity)
2080 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2081 		    VDEV_AUX_NO_REPLICAS);
2082 	else if (degraded + faulted != 0)
2083 		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE);
2084 	else
2085 		vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE);
2086 }
2087 
2088 static void
2089 vdev_draid_xlate(vdev_t *cvd, const range_seg64_t *logical_rs,
2090     range_seg64_t *physical_rs, range_seg64_t *remain_rs)
2091 {
2092 	vdev_t *raidvd = cvd->vdev_parent;
2093 	ASSERT(raidvd->vdev_ops == &vdev_draid_ops);
2094 
2095 	vdev_draid_config_t *vdc = raidvd->vdev_tsd;
2096 	uint64_t ashift = raidvd->vdev_top->vdev_ashift;
2097 
2098 	/* Make sure the offsets are block-aligned */
2099 	ASSERT0(logical_rs->rs_start % (1 << ashift));
2100 	ASSERT0(logical_rs->rs_end % (1 << ashift));
2101 
2102 	uint64_t logical_start = logical_rs->rs_start;
2103 	uint64_t logical_end = logical_rs->rs_end;
2104 
2105 	/*
2106 	 * Unaligned ranges must be skipped. All metaslabs are correctly
2107 	 * aligned so this should not happen, but this case is handled in
2108 	 * case it's needed by future callers.
2109 	 */
2110 	uint64_t astart = vdev_draid_get_astart(raidvd, logical_start);
2111 	if (astart != logical_start) {
2112 		physical_rs->rs_start = logical_start;
2113 		physical_rs->rs_end = logical_start;
2114 		remain_rs->rs_start = MIN(astart, logical_end);
2115 		remain_rs->rs_end = logical_end;
2116 		return;
2117 	}
2118 
2119 	/*
2120 	 * Unlike with mirrors and raidz a dRAID logical range can map
2121 	 * to multiple non-contiguous physical ranges. This is handled by
2122 	 * limiting the size of the logical range to a single group and
2123 	 * setting the remain argument such that it describes the remaining
2124 	 * unmapped logical range. This is stricter than absolutely
2125 	 * necessary but helps simplify the logic below.
2126 	 */
2127 	uint64_t group = vdev_draid_offset_to_group(raidvd, logical_start);
2128 	uint64_t nextstart = vdev_draid_group_to_offset(raidvd, group + 1);
2129 	if (logical_end > nextstart)
2130 		logical_end = nextstart;
2131 
2132 	/* Find the starting offset for each vdev in the group */
2133 	uint64_t perm, groupstart;
2134 	uint64_t start = vdev_draid_logical_to_physical(raidvd,
2135 	    logical_start, &perm, &groupstart);
2136 	uint64_t end = start;
2137 
2138 	uint8_t *base;
2139 	uint64_t iter, id;
2140 	vdev_draid_get_perm(vdc, perm, &base, &iter);
2141 
2142 	/*
2143 	 * Check if the passed child falls within the group.  If it does
2144 	 * update the start and end to reflect the physical range.
2145 	 * Otherwise, leave them unmodified which will result in an empty
2146 	 * (zero-length) physical range being returned.
2147 	 */
2148 	for (uint64_t i = 0; i < vdc->vdc_groupwidth; i++) {
2149 		uint64_t c = (groupstart + i) % vdc->vdc_ndisks;
2150 
2151 		if (c == 0 && i != 0) {
2152 			/* the group wrapped, increment the start */
2153 			start += VDEV_DRAID_ROWHEIGHT;
2154 			end = start;
2155 		}
2156 
2157 		id = vdev_draid_permute_id(vdc, base, iter, c);
2158 		if (id == cvd->vdev_id) {
2159 			uint64_t b_size = (logical_end >> ashift) -
2160 			    (logical_start >> ashift);
2161 			ASSERT3U(b_size, >, 0);
2162 			end = start + ((((b_size - 1) /
2163 			    vdc->vdc_groupwidth) + 1) << ashift);
2164 			break;
2165 		}
2166 	}
2167 	physical_rs->rs_start = start;
2168 	physical_rs->rs_end = end;
2169 
2170 	/*
2171 	 * Only top-level vdevs are allowed to set remain_rs because
2172 	 * when .vdev_op_xlate() is called for their children the full
2173 	 * logical range is not provided by vdev_xlate().
2174 	 */
2175 	remain_rs->rs_start = logical_end;
2176 	remain_rs->rs_end = logical_rs->rs_end;
2177 
2178 	ASSERT3U(physical_rs->rs_start, <=, logical_start);
2179 	ASSERT3U(physical_rs->rs_end - physical_rs->rs_start, <=,
2180 	    logical_end - logical_start);
2181 }
2182 
2183 /*
2184  * Add dRAID specific fields to the config nvlist.
2185  */
2186 static void
2187 vdev_draid_config_generate(vdev_t *vd, nvlist_t *nv)
2188 {
2189 	ASSERT3P(vd->vdev_ops, ==, &vdev_draid_ops);
2190 	vdev_draid_config_t *vdc = vd->vdev_tsd;
2191 
2192 	fnvlist_add_uint64(nv, ZPOOL_CONFIG_NPARITY, vdc->vdc_nparity);
2193 	fnvlist_add_uint64(nv, ZPOOL_CONFIG_DRAID_NDATA, vdc->vdc_ndata);
2194 	fnvlist_add_uint64(nv, ZPOOL_CONFIG_DRAID_NSPARES, vdc->vdc_nspares);
2195 	fnvlist_add_uint64(nv, ZPOOL_CONFIG_DRAID_NGROUPS, vdc->vdc_ngroups);
2196 }
2197 
2198 /*
2199  * Initialize private dRAID specific fields from the nvlist.
2200  */
2201 static int
2202 vdev_draid_init(spa_t *spa, nvlist_t *nv, void **tsd)
2203 {
2204 	(void) spa;
2205 	uint64_t ndata, nparity, nspares, ngroups;
2206 	int error;
2207 
2208 	if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DRAID_NDATA, &ndata))
2209 		return (SET_ERROR(EINVAL));
2210 
2211 	if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY, &nparity) ||
2212 	    nparity == 0 || nparity > VDEV_DRAID_MAXPARITY) {
2213 		return (SET_ERROR(EINVAL));
2214 	}
2215 
2216 	uint_t children;
2217 	nvlist_t **child;
2218 	if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
2219 	    &child, &children) != 0 || children == 0 ||
2220 	    children > VDEV_DRAID_MAX_CHILDREN) {
2221 		return (SET_ERROR(EINVAL));
2222 	}
2223 
2224 	if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DRAID_NSPARES, &nspares) ||
2225 	    nspares > 100 || nspares > (children - (ndata + nparity))) {
2226 		return (SET_ERROR(EINVAL));
2227 	}
2228 
2229 	if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DRAID_NGROUPS, &ngroups) ||
2230 	    ngroups == 0 || ngroups > VDEV_DRAID_MAX_CHILDREN) {
2231 		return (SET_ERROR(EINVAL));
2232 	}
2233 
2234 	/*
2235 	 * Validate the minimum number of children exist per group for the
2236 	 * specified parity level (draid1 >= 2, draid2 >= 3, draid3 >= 4).
2237 	 */
2238 	if (children < (ndata + nparity + nspares))
2239 		return (SET_ERROR(EINVAL));
2240 
2241 	/*
2242 	 * Create the dRAID configuration using the pool nvlist configuration
2243 	 * and the fixed mapping for the correct number of children.
2244 	 */
2245 	vdev_draid_config_t *vdc;
2246 	const draid_map_t *map;
2247 
2248 	error = vdev_draid_lookup_map(children, &map);
2249 	if (error)
2250 		return (SET_ERROR(EINVAL));
2251 
2252 	vdc = kmem_zalloc(sizeof (*vdc), KM_SLEEP);
2253 	vdc->vdc_ndata = ndata;
2254 	vdc->vdc_nparity = nparity;
2255 	vdc->vdc_nspares = nspares;
2256 	vdc->vdc_children = children;
2257 	vdc->vdc_ngroups = ngroups;
2258 	vdc->vdc_nperms = map->dm_nperms;
2259 
2260 	error = vdev_draid_generate_perms(map, &vdc->vdc_perms);
2261 	if (error) {
2262 		kmem_free(vdc, sizeof (*vdc));
2263 		return (SET_ERROR(EINVAL));
2264 	}
2265 
2266 	/*
2267 	 * Derived constants.
2268 	 */
2269 	vdc->vdc_groupwidth = vdc->vdc_ndata + vdc->vdc_nparity;
2270 	vdc->vdc_ndisks = vdc->vdc_children - vdc->vdc_nspares;
2271 	vdc->vdc_groupsz = vdc->vdc_groupwidth * VDEV_DRAID_ROWHEIGHT;
2272 	vdc->vdc_devslicesz = (vdc->vdc_groupsz * vdc->vdc_ngroups) /
2273 	    vdc->vdc_ndisks;
2274 
2275 	ASSERT3U(vdc->vdc_groupwidth, >=, 2);
2276 	ASSERT3U(vdc->vdc_groupwidth, <=, vdc->vdc_ndisks);
2277 	ASSERT3U(vdc->vdc_groupsz, >=, 2 * VDEV_DRAID_ROWHEIGHT);
2278 	ASSERT3U(vdc->vdc_devslicesz, >=, VDEV_DRAID_ROWHEIGHT);
2279 	ASSERT3U(vdc->vdc_devslicesz % VDEV_DRAID_ROWHEIGHT, ==, 0);
2280 	ASSERT3U((vdc->vdc_groupwidth * vdc->vdc_ngroups) %
2281 	    vdc->vdc_ndisks, ==, 0);
2282 
2283 	*tsd = vdc;
2284 
2285 	return (0);
2286 }
2287 
2288 static void
2289 vdev_draid_fini(vdev_t *vd)
2290 {
2291 	vdev_draid_config_t *vdc = vd->vdev_tsd;
2292 
2293 	vmem_free(vdc->vdc_perms, sizeof (uint8_t) *
2294 	    vdc->vdc_children * vdc->vdc_nperms);
2295 	kmem_free(vdc, sizeof (*vdc));
2296 }
2297 
2298 static uint64_t
2299 vdev_draid_nparity(vdev_t *vd)
2300 {
2301 	vdev_draid_config_t *vdc = vd->vdev_tsd;
2302 
2303 	return (vdc->vdc_nparity);
2304 }
2305 
2306 static uint64_t
2307 vdev_draid_ndisks(vdev_t *vd)
2308 {
2309 	vdev_draid_config_t *vdc = vd->vdev_tsd;
2310 
2311 	return (vdc->vdc_ndisks);
2312 }
2313 
2314 vdev_ops_t vdev_draid_ops = {
2315 	.vdev_op_init = vdev_draid_init,
2316 	.vdev_op_fini = vdev_draid_fini,
2317 	.vdev_op_open = vdev_draid_open,
2318 	.vdev_op_close = vdev_draid_close,
2319 	.vdev_op_asize = vdev_draid_asize,
2320 	.vdev_op_min_asize = vdev_draid_min_asize,
2321 	.vdev_op_min_alloc = vdev_draid_min_alloc,
2322 	.vdev_op_io_start = vdev_draid_io_start,
2323 	.vdev_op_io_done = vdev_draid_io_done,
2324 	.vdev_op_state_change = vdev_draid_state_change,
2325 	.vdev_op_need_resilver = vdev_draid_need_resilver,
2326 	.vdev_op_hold = NULL,
2327 	.vdev_op_rele = NULL,
2328 	.vdev_op_remap = NULL,
2329 	.vdev_op_xlate = vdev_draid_xlate,
2330 	.vdev_op_rebuild_asize = vdev_draid_rebuild_asize,
2331 	.vdev_op_metaslab_init = vdev_draid_metaslab_init,
2332 	.vdev_op_config_generate = vdev_draid_config_generate,
2333 	.vdev_op_nparity = vdev_draid_nparity,
2334 	.vdev_op_ndisks = vdev_draid_ndisks,
2335 	.vdev_op_type = VDEV_TYPE_DRAID,
2336 	.vdev_op_leaf = B_FALSE,
2337 };
2338 
2339 
2340 /*
2341  * A dRAID distributed spare is a virtual leaf vdev which is included in the
2342  * parent dRAID configuration.  The last N columns of the dRAID permutation
2343  * table are used to determine on which dRAID children a specific offset
2344  * should be written.  These spare leaf vdevs can only be used to replace
2345  * faulted children in the same dRAID configuration.
2346  */
2347 
2348 /*
2349  * Distributed spare state.  All fields are set when the distributed spare is
2350  * first opened and are immutable.
2351  */
2352 typedef struct {
2353 	vdev_t *vds_draid_vdev;		/* top-level parent dRAID vdev */
2354 	uint64_t vds_top_guid;		/* top-level parent dRAID guid */
2355 	uint64_t vds_spare_id;		/* spare id (0 - vdc->vdc_nspares-1) */
2356 } vdev_draid_spare_t;
2357 
2358 /*
2359  * Returns the parent dRAID vdev to which the distributed spare belongs.
2360  * This may be safely called even when the vdev is not open.
2361  */
2362 vdev_t *
2363 vdev_draid_spare_get_parent(vdev_t *vd)
2364 {
2365 	vdev_draid_spare_t *vds = vd->vdev_tsd;
2366 
2367 	ASSERT3P(vd->vdev_ops, ==, &vdev_draid_spare_ops);
2368 
2369 	if (vds->vds_draid_vdev != NULL)
2370 		return (vds->vds_draid_vdev);
2371 
2372 	return (vdev_lookup_by_guid(vd->vdev_spa->spa_root_vdev,
2373 	    vds->vds_top_guid));
2374 }
2375 
2376 /*
2377  * A dRAID space is active when it's the child of a vdev using the
2378  * vdev_spare_ops, vdev_replacing_ops or vdev_draid_ops.
2379  */
2380 static boolean_t
2381 vdev_draid_spare_is_active(vdev_t *vd)
2382 {
2383 	vdev_t *pvd = vd->vdev_parent;
2384 
2385 	if (pvd != NULL && (pvd->vdev_ops == &vdev_spare_ops ||
2386 	    pvd->vdev_ops == &vdev_replacing_ops ||
2387 	    pvd->vdev_ops == &vdev_draid_ops)) {
2388 		return (B_TRUE);
2389 	} else {
2390 		return (B_FALSE);
2391 	}
2392 }
2393 
2394 /*
2395  * Given a dRAID distribute spare vdev, returns the physical child vdev
2396  * on which the provided offset resides.  This may involve recursing through
2397  * multiple layers of distributed spares.  Note that offset is relative to
2398  * this vdev.
2399  */
2400 vdev_t *
2401 vdev_draid_spare_get_child(vdev_t *vd, uint64_t physical_offset)
2402 {
2403 	vdev_draid_spare_t *vds = vd->vdev_tsd;
2404 
2405 	ASSERT3P(vd->vdev_ops, ==, &vdev_draid_spare_ops);
2406 
2407 	/* The vdev is closed */
2408 	if (vds->vds_draid_vdev == NULL)
2409 		return (NULL);
2410 
2411 	vdev_t *tvd = vds->vds_draid_vdev;
2412 	vdev_draid_config_t *vdc = tvd->vdev_tsd;
2413 
2414 	ASSERT3P(tvd->vdev_ops, ==, &vdev_draid_ops);
2415 	ASSERT3U(vds->vds_spare_id, <, vdc->vdc_nspares);
2416 
2417 	uint8_t *base;
2418 	uint64_t iter;
2419 	uint64_t perm = physical_offset / vdc->vdc_devslicesz;
2420 
2421 	vdev_draid_get_perm(vdc, perm, &base, &iter);
2422 
2423 	uint64_t cid = vdev_draid_permute_id(vdc, base, iter,
2424 	    (tvd->vdev_children - 1) - vds->vds_spare_id);
2425 	vdev_t *cvd = tvd->vdev_child[cid];
2426 
2427 	if (cvd->vdev_ops == &vdev_draid_spare_ops)
2428 		return (vdev_draid_spare_get_child(cvd, physical_offset));
2429 
2430 	return (cvd);
2431 }
2432 
2433 static void
2434 vdev_draid_spare_close(vdev_t *vd)
2435 {
2436 	vdev_draid_spare_t *vds = vd->vdev_tsd;
2437 	vds->vds_draid_vdev = NULL;
2438 }
2439 
2440 /*
2441  * Opening a dRAID spare device is done by looking up the associated dRAID
2442  * top-level vdev guid from the spare configuration.
2443  */
2444 static int
2445 vdev_draid_spare_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
2446     uint64_t *logical_ashift, uint64_t *physical_ashift)
2447 {
2448 	vdev_draid_spare_t *vds = vd->vdev_tsd;
2449 	vdev_t *rvd = vd->vdev_spa->spa_root_vdev;
2450 	uint64_t asize, max_asize;
2451 
2452 	vdev_t *tvd = vdev_lookup_by_guid(rvd, vds->vds_top_guid);
2453 	if (tvd == NULL) {
2454 		/*
2455 		 * When spa_vdev_add() is labeling new spares the
2456 		 * associated dRAID is not attached to the root vdev
2457 		 * nor does this spare have a parent.  Simulate a valid
2458 		 * device in order to allow the label to be initialized
2459 		 * and the distributed spare added to the configuration.
2460 		 */
2461 		if (vd->vdev_parent == NULL) {
2462 			*psize = *max_psize = SPA_MINDEVSIZE;
2463 			*logical_ashift = *physical_ashift = ASHIFT_MIN;
2464 			return (0);
2465 		}
2466 
2467 		return (SET_ERROR(EINVAL));
2468 	}
2469 
2470 	vdev_draid_config_t *vdc = tvd->vdev_tsd;
2471 	if (tvd->vdev_ops != &vdev_draid_ops || vdc == NULL)
2472 		return (SET_ERROR(EINVAL));
2473 
2474 	if (vds->vds_spare_id >= vdc->vdc_nspares)
2475 		return (SET_ERROR(EINVAL));
2476 
2477 	/*
2478 	 * Neither tvd->vdev_asize or tvd->vdev_max_asize can be used here
2479 	 * because the caller may be vdev_draid_open() in which case the
2480 	 * values are stale as they haven't yet been updated by vdev_open().
2481 	 * To avoid this always recalculate the dRAID asize and max_asize.
2482 	 */
2483 	vdev_draid_calculate_asize(tvd, &asize, &max_asize,
2484 	    logical_ashift, physical_ashift);
2485 
2486 	*psize = asize + VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2487 	*max_psize = max_asize + VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2488 
2489 	vds->vds_draid_vdev = tvd;
2490 
2491 	return (0);
2492 }
2493 
2494 /*
2495  * Completed distributed spare IO.  Store the result in the parent zio
2496  * as if it had performed the operation itself.  Only the first error is
2497  * preserved if there are multiple errors.
2498  */
2499 static void
2500 vdev_draid_spare_child_done(zio_t *zio)
2501 {
2502 	zio_t *pio = zio->io_private;
2503 
2504 	/*
2505 	 * IOs are issued to non-writable vdevs in order to keep their
2506 	 * DTLs accurate.  However, we don't want to propagate the
2507 	 * error in to the distributed spare's DTL.  When resilvering
2508 	 * vdev_draid_need_resilver() will consult the relevant DTL
2509 	 * to determine if the data is missing and must be repaired.
2510 	 */
2511 	if (!vdev_writeable(zio->io_vd))
2512 		return;
2513 
2514 	if (pio->io_error == 0)
2515 		pio->io_error = zio->io_error;
2516 }
2517 
2518 /*
2519  * Returns a valid label nvlist for the distributed spare vdev.  This is
2520  * used to bypass the IO pipeline to avoid the complexity of constructing
2521  * a complete label with valid checksum to return when read.
2522  */
2523 nvlist_t *
2524 vdev_draid_read_config_spare(vdev_t *vd)
2525 {
2526 	spa_t *spa = vd->vdev_spa;
2527 	spa_aux_vdev_t *sav = &spa->spa_spares;
2528 	uint64_t guid = vd->vdev_guid;
2529 
2530 	nvlist_t *nv = fnvlist_alloc();
2531 	fnvlist_add_uint64(nv, ZPOOL_CONFIG_IS_SPARE, 1);
2532 	fnvlist_add_uint64(nv, ZPOOL_CONFIG_CREATE_TXG, vd->vdev_crtxg);
2533 	fnvlist_add_uint64(nv, ZPOOL_CONFIG_VERSION, spa_version(spa));
2534 	fnvlist_add_string(nv, ZPOOL_CONFIG_POOL_NAME, spa_name(spa));
2535 	fnvlist_add_uint64(nv, ZPOOL_CONFIG_POOL_GUID, spa_guid(spa));
2536 	fnvlist_add_uint64(nv, ZPOOL_CONFIG_POOL_TXG, spa->spa_config_txg);
2537 	fnvlist_add_uint64(nv, ZPOOL_CONFIG_TOP_GUID, vd->vdev_top->vdev_guid);
2538 	fnvlist_add_uint64(nv, ZPOOL_CONFIG_POOL_STATE,
2539 	    vdev_draid_spare_is_active(vd) ?
2540 	    POOL_STATE_ACTIVE : POOL_STATE_SPARE);
2541 
2542 	/* Set the vdev guid based on the vdev list in sav_count. */
2543 	for (int i = 0; i < sav->sav_count; i++) {
2544 		if (sav->sav_vdevs[i]->vdev_ops == &vdev_draid_spare_ops &&
2545 		    strcmp(sav->sav_vdevs[i]->vdev_path, vd->vdev_path) == 0) {
2546 			guid = sav->sav_vdevs[i]->vdev_guid;
2547 			break;
2548 		}
2549 	}
2550 
2551 	fnvlist_add_uint64(nv, ZPOOL_CONFIG_GUID, guid);
2552 
2553 	return (nv);
2554 }
2555 
2556 /*
2557  * Handle any ioctl requested of the distributed spare.  Only flushes
2558  * are supported in which case all children must be flushed.
2559  */
2560 static int
2561 vdev_draid_spare_ioctl(zio_t *zio)
2562 {
2563 	vdev_t *vd = zio->io_vd;
2564 	int error = 0;
2565 
2566 	if (zio->io_cmd == DKIOCFLUSHWRITECACHE) {
2567 		for (int c = 0; c < vd->vdev_children; c++) {
2568 			zio_nowait(zio_vdev_child_io(zio, NULL,
2569 			    vd->vdev_child[c], zio->io_offset, zio->io_abd,
2570 			    zio->io_size, zio->io_type, zio->io_priority, 0,
2571 			    vdev_draid_spare_child_done, zio));
2572 		}
2573 	} else {
2574 		error = SET_ERROR(ENOTSUP);
2575 	}
2576 
2577 	return (error);
2578 }
2579 
2580 /*
2581  * Initiate an IO to the distributed spare.  For normal IOs this entails using
2582  * the zio->io_offset and permutation table to calculate which child dRAID vdev
2583  * is responsible for the data.  Then passing along the zio to that child to
2584  * perform the actual IO.  The label ranges are not stored on disk and require
2585  * some special handling which is described below.
2586  */
2587 static void
2588 vdev_draid_spare_io_start(zio_t *zio)
2589 {
2590 	vdev_t *cvd = NULL, *vd = zio->io_vd;
2591 	vdev_draid_spare_t *vds = vd->vdev_tsd;
2592 	uint64_t offset = zio->io_offset - VDEV_LABEL_START_SIZE;
2593 
2594 	/*
2595 	 * If the vdev is closed, it's likely in the REMOVED or FAULTED state.
2596 	 * Nothing to be done here but return failure.
2597 	 */
2598 	if (vds == NULL) {
2599 		zio->io_error = ENXIO;
2600 		zio_interrupt(zio);
2601 		return;
2602 	}
2603 
2604 	switch (zio->io_type) {
2605 	case ZIO_TYPE_IOCTL:
2606 		zio->io_error = vdev_draid_spare_ioctl(zio);
2607 		break;
2608 
2609 	case ZIO_TYPE_WRITE:
2610 		if (VDEV_OFFSET_IS_LABEL(vd, zio->io_offset)) {
2611 			/*
2612 			 * Accept probe IOs and config writers to simulate the
2613 			 * existence of an on disk label.  vdev_label_sync(),
2614 			 * vdev_uberblock_sync() and vdev_copy_uberblocks()
2615 			 * skip the distributed spares.  This only leaves
2616 			 * vdev_label_init() which is allowed to succeed to
2617 			 * avoid adding special cases the function.
2618 			 */
2619 			if (zio->io_flags & ZIO_FLAG_PROBE ||
2620 			    zio->io_flags & ZIO_FLAG_CONFIG_WRITER) {
2621 				zio->io_error = 0;
2622 			} else {
2623 				zio->io_error = SET_ERROR(EIO);
2624 			}
2625 		} else {
2626 			cvd = vdev_draid_spare_get_child(vd, offset);
2627 
2628 			if (cvd == NULL) {
2629 				zio->io_error = SET_ERROR(ENXIO);
2630 			} else {
2631 				zio_nowait(zio_vdev_child_io(zio, NULL, cvd,
2632 				    offset, zio->io_abd, zio->io_size,
2633 				    zio->io_type, zio->io_priority, 0,
2634 				    vdev_draid_spare_child_done, zio));
2635 			}
2636 		}
2637 		break;
2638 
2639 	case ZIO_TYPE_READ:
2640 		if (VDEV_OFFSET_IS_LABEL(vd, zio->io_offset)) {
2641 			/*
2642 			 * Accept probe IOs to simulate the existence of a
2643 			 * label.  vdev_label_read_config() bypasses the
2644 			 * pipeline to read the label configuration and
2645 			 * vdev_uberblock_load() skips distributed spares
2646 			 * when attempting to locate the best uberblock.
2647 			 */
2648 			if (zio->io_flags & ZIO_FLAG_PROBE) {
2649 				zio->io_error = 0;
2650 			} else {
2651 				zio->io_error = SET_ERROR(EIO);
2652 			}
2653 		} else {
2654 			cvd = vdev_draid_spare_get_child(vd, offset);
2655 
2656 			if (cvd == NULL || !vdev_readable(cvd)) {
2657 				zio->io_error = SET_ERROR(ENXIO);
2658 			} else {
2659 				zio_nowait(zio_vdev_child_io(zio, NULL, cvd,
2660 				    offset, zio->io_abd, zio->io_size,
2661 				    zio->io_type, zio->io_priority, 0,
2662 				    vdev_draid_spare_child_done, zio));
2663 			}
2664 		}
2665 		break;
2666 
2667 	case ZIO_TYPE_TRIM:
2668 		/* The vdev label ranges are never trimmed */
2669 		ASSERT0(VDEV_OFFSET_IS_LABEL(vd, zio->io_offset));
2670 
2671 		cvd = vdev_draid_spare_get_child(vd, offset);
2672 
2673 		if (cvd == NULL || !cvd->vdev_has_trim) {
2674 			zio->io_error = SET_ERROR(ENXIO);
2675 		} else {
2676 			zio_nowait(zio_vdev_child_io(zio, NULL, cvd,
2677 			    offset, zio->io_abd, zio->io_size,
2678 			    zio->io_type, zio->io_priority, 0,
2679 			    vdev_draid_spare_child_done, zio));
2680 		}
2681 		break;
2682 
2683 	default:
2684 		zio->io_error = SET_ERROR(ENOTSUP);
2685 		break;
2686 	}
2687 
2688 	zio_execute(zio);
2689 }
2690 
2691 static void
2692 vdev_draid_spare_io_done(zio_t *zio)
2693 {
2694 	(void) zio;
2695 }
2696 
2697 /*
2698  * Lookup the full spare config in spa->spa_spares.sav_config and
2699  * return the top_guid and spare_id for the named spare.
2700  */
2701 static int
2702 vdev_draid_spare_lookup(spa_t *spa, nvlist_t *nv, uint64_t *top_guidp,
2703     uint64_t *spare_idp)
2704 {
2705 	nvlist_t **spares;
2706 	uint_t nspares;
2707 	int error;
2708 
2709 	if ((spa->spa_spares.sav_config == NULL) ||
2710 	    (nvlist_lookup_nvlist_array(spa->spa_spares.sav_config,
2711 	    ZPOOL_CONFIG_SPARES, &spares, &nspares) != 0)) {
2712 		return (SET_ERROR(ENOENT));
2713 	}
2714 
2715 	char *spare_name;
2716 	error = nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &spare_name);
2717 	if (error != 0)
2718 		return (SET_ERROR(EINVAL));
2719 
2720 	for (int i = 0; i < nspares; i++) {
2721 		nvlist_t *spare = spares[i];
2722 		uint64_t top_guid, spare_id;
2723 		char *type, *path;
2724 
2725 		/* Skip non-distributed spares */
2726 		error = nvlist_lookup_string(spare, ZPOOL_CONFIG_TYPE, &type);
2727 		if (error != 0 || strcmp(type, VDEV_TYPE_DRAID_SPARE) != 0)
2728 			continue;
2729 
2730 		/* Skip spares with the wrong name */
2731 		error = nvlist_lookup_string(spare, ZPOOL_CONFIG_PATH, &path);
2732 		if (error != 0 || strcmp(path, spare_name) != 0)
2733 			continue;
2734 
2735 		/* Found the matching spare */
2736 		error = nvlist_lookup_uint64(spare,
2737 		    ZPOOL_CONFIG_TOP_GUID, &top_guid);
2738 		if (error == 0) {
2739 			error = nvlist_lookup_uint64(spare,
2740 			    ZPOOL_CONFIG_SPARE_ID, &spare_id);
2741 		}
2742 
2743 		if (error != 0) {
2744 			return (SET_ERROR(EINVAL));
2745 		} else {
2746 			*top_guidp = top_guid;
2747 			*spare_idp = spare_id;
2748 			return (0);
2749 		}
2750 	}
2751 
2752 	return (SET_ERROR(ENOENT));
2753 }
2754 
2755 /*
2756  * Initialize private dRAID spare specific fields from the nvlist.
2757  */
2758 static int
2759 vdev_draid_spare_init(spa_t *spa, nvlist_t *nv, void **tsd)
2760 {
2761 	vdev_draid_spare_t *vds;
2762 	uint64_t top_guid = 0;
2763 	uint64_t spare_id;
2764 
2765 	/*
2766 	 * In the normal case check the list of spares stored in the spa
2767 	 * to lookup the top_guid and spare_id for provided spare config.
2768 	 * When creating a new pool or adding vdevs the spare list is not
2769 	 * yet populated and the values are provided in the passed config.
2770 	 */
2771 	if (vdev_draid_spare_lookup(spa, nv, &top_guid, &spare_id) != 0) {
2772 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_TOP_GUID,
2773 		    &top_guid) != 0)
2774 			return (SET_ERROR(EINVAL));
2775 
2776 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_SPARE_ID,
2777 		    &spare_id) != 0)
2778 			return (SET_ERROR(EINVAL));
2779 	}
2780 
2781 	vds = kmem_alloc(sizeof (vdev_draid_spare_t), KM_SLEEP);
2782 	vds->vds_draid_vdev = NULL;
2783 	vds->vds_top_guid = top_guid;
2784 	vds->vds_spare_id = spare_id;
2785 
2786 	*tsd = vds;
2787 
2788 	return (0);
2789 }
2790 
2791 static void
2792 vdev_draid_spare_fini(vdev_t *vd)
2793 {
2794 	kmem_free(vd->vdev_tsd, sizeof (vdev_draid_spare_t));
2795 }
2796 
2797 static void
2798 vdev_draid_spare_config_generate(vdev_t *vd, nvlist_t *nv)
2799 {
2800 	vdev_draid_spare_t *vds = vd->vdev_tsd;
2801 
2802 	ASSERT3P(vd->vdev_ops, ==, &vdev_draid_spare_ops);
2803 
2804 	fnvlist_add_uint64(nv, ZPOOL_CONFIG_TOP_GUID, vds->vds_top_guid);
2805 	fnvlist_add_uint64(nv, ZPOOL_CONFIG_SPARE_ID, vds->vds_spare_id);
2806 }
2807 
2808 vdev_ops_t vdev_draid_spare_ops = {
2809 	.vdev_op_init = vdev_draid_spare_init,
2810 	.vdev_op_fini = vdev_draid_spare_fini,
2811 	.vdev_op_open = vdev_draid_spare_open,
2812 	.vdev_op_close = vdev_draid_spare_close,
2813 	.vdev_op_asize = vdev_default_asize,
2814 	.vdev_op_min_asize = vdev_default_min_asize,
2815 	.vdev_op_min_alloc = NULL,
2816 	.vdev_op_io_start = vdev_draid_spare_io_start,
2817 	.vdev_op_io_done = vdev_draid_spare_io_done,
2818 	.vdev_op_state_change = NULL,
2819 	.vdev_op_need_resilver = NULL,
2820 	.vdev_op_hold = NULL,
2821 	.vdev_op_rele = NULL,
2822 	.vdev_op_remap = NULL,
2823 	.vdev_op_xlate = vdev_default_xlate,
2824 	.vdev_op_rebuild_asize = NULL,
2825 	.vdev_op_metaslab_init = NULL,
2826 	.vdev_op_config_generate = vdev_draid_spare_config_generate,
2827 	.vdev_op_nparity = NULL,
2828 	.vdev_op_ndisks = NULL,
2829 	.vdev_op_type = VDEV_TYPE_DRAID_SPARE,
2830 	.vdev_op_leaf = B_TRUE,
2831 };
2832