Commit | Line | Data |
---|---|---|
af1a8899 | 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
1da177e4 LT |
2 | /* |
3 | raid0.c : Multiple Devices driver for Linux | |
f72ffdd6 | 4 | Copyright (C) 1994-96 Marc ZYNGIER |
1da177e4 LT |
5 | <zyngier@ufr-info-p7.ibp.fr> or |
6 | <maz@gloups.fdn.fr> | |
f72ffdd6 | 7 | Copyright (C) 1999, 2000 Ingo Molnar, Red Hat |
1da177e4 LT |
8 | |
9 | RAID-0 management functions. | |
10 | ||
1da177e4 LT |
11 | */ |
12 | ||
bff61975 | 13 | #include <linux/blkdev.h> |
bff61975 | 14 | #include <linux/seq_file.h> |
056075c7 | 15 | #include <linux/module.h> |
5a0e3ad6 | 16 | #include <linux/slab.h> |
109e3765 | 17 | #include <trace/events/block.h> |
43b2e5d8 | 18 | #include "md.h" |
ef740c37 | 19 | #include "raid0.h" |
9af204cf | 20 | #include "raid5.h" |
1da177e4 | 21 | |
c84a1372 N |
22 | static int default_layout = 0; |
23 | module_param(default_layout, int, 0644); | |
24 | ||
394ed8e4 SL |
25 | #define UNSUPPORTED_MDDEV_FLAGS \ |
26 | ((1L << MD_HAS_JOURNAL) | \ | |
27 | (1L << MD_JOURNAL_CLEAN) | \ | |
ea0213e0 | 28 | (1L << MD_FAILFAST_SUPPORTED) |\ |
ddc08823 PB |
29 | (1L << MD_HAS_PPL) | \ |
30 | (1L << MD_HAS_MULTIPLE_PPLS)) | |
394ed8e4 | 31 | |
46994191 | 32 | /* |
33 | * inform the user of the raid configuration | |
34 | */ | |
fd01b88c | 35 | static void dump_zones(struct mddev *mddev) |
46994191 | 36 | { |
50de8df4 | 37 | int j, k; |
46994191 | 38 | sector_t zone_size = 0; |
39 | sector_t zone_start = 0; | |
e373ab10 | 40 | struct r0conf *conf = mddev->private; |
84707f38 | 41 | int raid_disks = conf->strip_zone[0].nb_dev; |
76603884 N |
42 | pr_debug("md: RAID0 configuration for %s - %d zone%s\n", |
43 | mdname(mddev), | |
44 | conf->nr_strip_zones, conf->nr_strip_zones==1?"":"s"); | |
46994191 | 45 | for (j = 0; j < conf->nr_strip_zones; j++) { |
76603884 N |
46 | char line[200]; |
47 | int len = 0; | |
48 | ||
46994191 | 49 | for (k = 0; k < conf->strip_zone[j].nb_dev; k++) |
1727fd50 | 50 | len += scnprintf(line+len, 200-len, "%s%pg", k?"/":"", |
913cce5a | 51 | conf->devlist[j * raid_disks + k]->bdev); |
76603884 | 52 | pr_debug("md: zone%d=[%s]\n", j, line); |
46994191 | 53 | |
54 | zone_size = conf->strip_zone[j].zone_end - zone_start; | |
76603884 | 55 | pr_debug(" zone-offset=%10lluKB, device-offset=%10lluKB, size=%10lluKB\n", |
46994191 | 56 | (unsigned long long)zone_start>>1, |
57 | (unsigned long long)conf->strip_zone[j].dev_start>>1, | |
58 | (unsigned long long)zone_size>>1); | |
59 | zone_start = conf->strip_zone[j].zone_end; | |
60 | } | |
46994191 | 61 | } |
62 | ||
e373ab10 | 63 | static int create_strip_zones(struct mddev *mddev, struct r0conf **private_conf) |
1da177e4 | 64 | { |
a9f326eb | 65 | int i, c, err; |
49f357a2 | 66 | sector_t curr_zone_end, sectors; |
3cb03002 | 67 | struct md_rdev *smallest, *rdev1, *rdev2, *rdev, **dev; |
1da177e4 LT |
68 | struct strip_zone *zone; |
69 | int cnt; | |
e373ab10 | 70 | struct r0conf *conf = kzalloc(sizeof(*conf), GFP_KERNEL); |
ad6bf88a | 71 | unsigned blksize = 512; |
ed7b0038 | 72 | |
7dedd15d | 73 | *private_conf = ERR_PTR(-ENOMEM); |
ed7b0038 AN |
74 | if (!conf) |
75 | return -ENOMEM; | |
dafb20fa | 76 | rdev_for_each(rdev1, mddev) { |
913cce5a | 77 | pr_debug("md/raid0:%s: looking at %pg\n", |
50de8df4 | 78 | mdname(mddev), |
913cce5a | 79 | rdev1->bdev); |
1da177e4 | 80 | c = 0; |
13f2682b N |
81 | |
82 | /* round size to chunk_size */ | |
83 | sectors = rdev1->sectors; | |
84 | sector_div(sectors, mddev->chunk_sectors); | |
85 | rdev1->sectors = sectors * mddev->chunk_sectors; | |
86 | ||
199dc6ed N |
87 | blksize = max(blksize, queue_logical_block_size( |
88 | rdev1->bdev->bd_disk->queue)); | |
89 | ||
dafb20fa | 90 | rdev_for_each(rdev2, mddev) { |
913cce5a CH |
91 | pr_debug("md/raid0:%s: comparing %pg(%llu)" |
92 | " with %pg(%llu)\n", | |
50de8df4 | 93 | mdname(mddev), |
913cce5a | 94 | rdev1->bdev, |
50de8df4 | 95 | (unsigned long long)rdev1->sectors, |
913cce5a | 96 | rdev2->bdev, |
50de8df4 | 97 | (unsigned long long)rdev2->sectors); |
1da177e4 | 98 | if (rdev2 == rdev1) { |
50de8df4 N |
99 | pr_debug("md/raid0:%s: END\n", |
100 | mdname(mddev)); | |
1da177e4 LT |
101 | break; |
102 | } | |
dd8ac336 | 103 | if (rdev2->sectors == rdev1->sectors) { |
1da177e4 LT |
104 | /* |
105 | * Not unique, don't count it as a new | |
106 | * group | |
107 | */ | |
50de8df4 N |
108 | pr_debug("md/raid0:%s: EQUAL\n", |
109 | mdname(mddev)); | |
1da177e4 LT |
110 | c = 1; |
111 | break; | |
112 | } | |
50de8df4 N |
113 | pr_debug("md/raid0:%s: NOT EQUAL\n", |
114 | mdname(mddev)); | |
1da177e4 LT |
115 | } |
116 | if (!c) { | |
50de8df4 N |
117 | pr_debug("md/raid0:%s: ==> UNIQUE\n", |
118 | mdname(mddev)); | |
1da177e4 | 119 | conf->nr_strip_zones++; |
50de8df4 N |
120 | pr_debug("md/raid0:%s: %d zones\n", |
121 | mdname(mddev), conf->nr_strip_zones); | |
1da177e4 LT |
122 | } |
123 | } | |
50de8df4 N |
124 | pr_debug("md/raid0:%s: FINAL %d zones\n", |
125 | mdname(mddev), conf->nr_strip_zones); | |
c84a1372 | 126 | |
199dc6ed N |
127 | /* |
128 | * now since we have the hard sector sizes, we can make sure | |
129 | * chunk size is a multiple of that sector size | |
130 | */ | |
131 | if ((mddev->chunk_sectors << 9) % blksize) { | |
76603884 N |
132 | pr_warn("md/raid0:%s: chunk_size of %d not multiple of block size %d\n", |
133 | mdname(mddev), | |
134 | mddev->chunk_sectors << 9, blksize); | |
199dc6ed N |
135 | err = -EINVAL; |
136 | goto abort; | |
137 | } | |
138 | ||
ed7b0038 | 139 | err = -ENOMEM; |
6396bb22 KC |
140 | conf->strip_zone = kcalloc(conf->nr_strip_zones, |
141 | sizeof(struct strip_zone), | |
142 | GFP_KERNEL); | |
1da177e4 | 143 | if (!conf->strip_zone) |
ed7b0038 | 144 | goto abort; |
6396bb22 KC |
145 | conf->devlist = kzalloc(array3_size(sizeof(struct md_rdev *), |
146 | conf->nr_strip_zones, | |
147 | mddev->raid_disks), | |
1da177e4 LT |
148 | GFP_KERNEL); |
149 | if (!conf->devlist) | |
ed7b0038 | 150 | goto abort; |
1da177e4 | 151 | |
1da177e4 LT |
152 | /* The first zone must contain all devices, so here we check that |
153 | * there is a proper alignment of slots to devices and find them all | |
154 | */ | |
155 | zone = &conf->strip_zone[0]; | |
156 | cnt = 0; | |
157 | smallest = NULL; | |
b414579f | 158 | dev = conf->devlist; |
ed7b0038 | 159 | err = -EINVAL; |
dafb20fa | 160 | rdev_for_each(rdev1, mddev) { |
1da177e4 LT |
161 | int j = rdev1->raid_disk; |
162 | ||
e93f68a1 | 163 | if (mddev->level == 10) { |
9af204cf TM |
164 | /* taking over a raid10-n2 array */ |
165 | j /= 2; | |
e93f68a1 N |
166 | rdev1->new_raid_disk = j; |
167 | } | |
9af204cf | 168 | |
fc3a08b8 KW |
169 | if (mddev->level == 1) { |
170 | /* taiking over a raid1 array- | |
171 | * we have only one active disk | |
172 | */ | |
173 | j = 0; | |
174 | rdev1->new_raid_disk = j; | |
175 | } | |
176 | ||
f96c9f30 | 177 | if (j < 0) { |
76603884 N |
178 | pr_warn("md/raid0:%s: remove inactive devices before converting to RAID0\n", |
179 | mdname(mddev)); | |
f96c9f30 N |
180 | goto abort; |
181 | } | |
182 | if (j >= mddev->raid_disks) { | |
76603884 N |
183 | pr_warn("md/raid0:%s: bad disk number %d - aborting!\n", |
184 | mdname(mddev), j); | |
1da177e4 LT |
185 | goto abort; |
186 | } | |
b414579f | 187 | if (dev[j]) { |
76603884 N |
188 | pr_warn("md/raid0:%s: multiple devices for %d - aborting!\n", |
189 | mdname(mddev), j); | |
1da177e4 LT |
190 | goto abort; |
191 | } | |
b414579f | 192 | dev[j] = rdev1; |
1da177e4 | 193 | |
dd8ac336 | 194 | if (!smallest || (rdev1->sectors < smallest->sectors)) |
1da177e4 LT |
195 | smallest = rdev1; |
196 | cnt++; | |
197 | } | |
198 | if (cnt != mddev->raid_disks) { | |
76603884 N |
199 | pr_warn("md/raid0:%s: too few disks (%d of %d) - aborting!\n", |
200 | mdname(mddev), cnt, mddev->raid_disks); | |
1da177e4 LT |
201 | goto abort; |
202 | } | |
203 | zone->nb_dev = cnt; | |
49f357a2 | 204 | zone->zone_end = smallest->sectors * cnt; |
1da177e4 | 205 | |
49f357a2 | 206 | curr_zone_end = zone->zone_end; |
1da177e4 LT |
207 | |
208 | /* now do the other zones */ | |
209 | for (i = 1; i < conf->nr_strip_zones; i++) | |
210 | { | |
a9f326eb N |
211 | int j; |
212 | ||
1da177e4 | 213 | zone = conf->strip_zone + i; |
b414579f | 214 | dev = conf->devlist + i * mddev->raid_disks; |
1da177e4 | 215 | |
50de8df4 | 216 | pr_debug("md/raid0:%s: zone %d\n", mdname(mddev), i); |
d27a43ab | 217 | zone->dev_start = smallest->sectors; |
1da177e4 LT |
218 | smallest = NULL; |
219 | c = 0; | |
220 | ||
221 | for (j=0; j<cnt; j++) { | |
b414579f | 222 | rdev = conf->devlist[j]; |
d27a43ab | 223 | if (rdev->sectors <= zone->dev_start) { |
913cce5a | 224 | pr_debug("md/raid0:%s: checking %pg ... nope\n", |
50de8df4 | 225 | mdname(mddev), |
913cce5a | 226 | rdev->bdev); |
dd8ac336 AN |
227 | continue; |
228 | } | |
913cce5a | 229 | pr_debug("md/raid0:%s: checking %pg ..." |
50de8df4 N |
230 | " contained as device %d\n", |
231 | mdname(mddev), | |
913cce5a | 232 | rdev->bdev, c); |
b414579f | 233 | dev[c] = rdev; |
dd8ac336 AN |
234 | c++; |
235 | if (!smallest || rdev->sectors < smallest->sectors) { | |
236 | smallest = rdev; | |
50de8df4 N |
237 | pr_debug("md/raid0:%s: (%llu) is smallest!.\n", |
238 | mdname(mddev), | |
239 | (unsigned long long)rdev->sectors); | |
dd8ac336 | 240 | } |
1da177e4 LT |
241 | } |
242 | ||
243 | zone->nb_dev = c; | |
49f357a2 | 244 | sectors = (smallest->sectors - zone->dev_start) * c; |
50de8df4 N |
245 | pr_debug("md/raid0:%s: zone->nb_dev: %d, sectors: %llu\n", |
246 | mdname(mddev), | |
247 | zone->nb_dev, (unsigned long long)sectors); | |
1da177e4 | 248 | |
49f357a2 | 249 | curr_zone_end += sectors; |
d27a43ab | 250 | zone->zone_end = curr_zone_end; |
1da177e4 | 251 | |
50de8df4 N |
252 | pr_debug("md/raid0:%s: current zone start: %llu\n", |
253 | mdname(mddev), | |
254 | (unsigned long long)smallest->sectors); | |
1da177e4 | 255 | } |
1da177e4 | 256 | |
ea23994e PH |
257 | if (conf->nr_strip_zones == 1 || conf->strip_zone[1].nb_dev == 1) { |
258 | conf->layout = RAID0_ORIG_LAYOUT; | |
259 | } else if (mddev->layout == RAID0_ORIG_LAYOUT || | |
260 | mddev->layout == RAID0_ALT_MULTIZONE_LAYOUT) { | |
261 | conf->layout = mddev->layout; | |
262 | } else if (default_layout == RAID0_ORIG_LAYOUT || | |
263 | default_layout == RAID0_ALT_MULTIZONE_LAYOUT) { | |
264 | conf->layout = default_layout; | |
265 | } else { | |
266 | pr_err("md/raid0:%s: cannot assemble multi-zone RAID0 with default_layout setting\n", | |
267 | mdname(mddev)); | |
268 | pr_err("md/raid0: please set raid0.default_layout to 1 or 2\n"); | |
269 | err = -EOPNOTSUPP; | |
270 | goto abort; | |
271 | } | |
272 | ||
e8360070 JB |
273 | if (conf->layout == RAID0_ORIG_LAYOUT) { |
274 | for (i = 1; i < conf->nr_strip_zones; i++) { | |
275 | sector_t first_sector = conf->strip_zone[i-1].zone_end; | |
276 | ||
277 | sector_div(first_sector, mddev->chunk_sectors); | |
278 | zone = conf->strip_zone + i; | |
279 | /* disk_shift is first disk index used in the zone */ | |
280 | zone->disk_shift = sector_div(first_sector, | |
281 | zone->nb_dev); | |
282 | } | |
283 | } | |
284 | ||
50de8df4 | 285 | pr_debug("md/raid0:%s: done.\n", mdname(mddev)); |
9af204cf TM |
286 | *private_conf = conf; |
287 | ||
1da177e4 | 288 | return 0; |
5568a603 | 289 | abort: |
ed7b0038 AN |
290 | kfree(conf->strip_zone); |
291 | kfree(conf->devlist); | |
292 | kfree(conf); | |
58ebb34c | 293 | *private_conf = ERR_PTR(err); |
ed7b0038 | 294 | return err; |
1da177e4 LT |
295 | } |
296 | ||
ba13da47 N |
297 | /* Find the zone which holds a particular offset |
298 | * Update *sectorp to be an offset in that zone | |
299 | */ | |
300 | static struct strip_zone *find_zone(struct r0conf *conf, | |
301 | sector_t *sectorp) | |
302 | { | |
303 | int i; | |
304 | struct strip_zone *z = conf->strip_zone; | |
305 | sector_t sector = *sectorp; | |
306 | ||
307 | for (i = 0; i < conf->nr_strip_zones; i++) | |
308 | if (sector < z[i].zone_end) { | |
309 | if (i) | |
310 | *sectorp = sector - z[i-1].zone_end; | |
311 | return z + i; | |
312 | } | |
313 | BUG(); | |
314 | } | |
315 | ||
316 | /* | |
317 | * remaps the bio to the target device. we separate two flows. | |
47d68979 | 318 | * power 2 flow and a general flow for the sake of performance |
ba13da47 N |
319 | */ |
320 | static struct md_rdev *map_sector(struct mddev *mddev, struct strip_zone *zone, | |
321 | sector_t sector, sector_t *sector_offset) | |
322 | { | |
323 | unsigned int sect_in_chunk; | |
324 | sector_t chunk; | |
325 | struct r0conf *conf = mddev->private; | |
326 | int raid_disks = conf->strip_zone[0].nb_dev; | |
327 | unsigned int chunk_sects = mddev->chunk_sectors; | |
328 | ||
329 | if (is_power_of_2(chunk_sects)) { | |
330 | int chunksect_bits = ffz(~chunk_sects); | |
331 | /* find the sector offset inside the chunk */ | |
332 | sect_in_chunk = sector & (chunk_sects - 1); | |
333 | sector >>= chunksect_bits; | |
334 | /* chunk in zone */ | |
335 | chunk = *sector_offset; | |
336 | /* quotient is the chunk in real device*/ | |
337 | sector_div(chunk, zone->nb_dev << chunksect_bits); | |
338 | } else{ | |
339 | sect_in_chunk = sector_div(sector, chunk_sects); | |
340 | chunk = *sector_offset; | |
341 | sector_div(chunk, chunk_sects * zone->nb_dev); | |
342 | } | |
343 | /* | |
344 | * position the bio over the real device | |
345 | * real sector = chunk in device + starting of zone | |
346 | * + the position in the chunk | |
347 | */ | |
348 | *sector_offset = (chunk * chunk_sects) + sect_in_chunk; | |
349 | return conf->devlist[(zone - conf->strip_zone)*raid_disks | |
350 | + sector_div(sector, zone->nb_dev)]; | |
351 | } | |
352 | ||
fd01b88c | 353 | static sector_t raid0_size(struct mddev *mddev, sector_t sectors, int raid_disks) |
80c3a6ce DW |
354 | { |
355 | sector_t array_sectors = 0; | |
3cb03002 | 356 | struct md_rdev *rdev; |
80c3a6ce DW |
357 | |
358 | WARN_ONCE(sectors || raid_disks, | |
359 | "%s does not support generic reshape\n", __func__); | |
360 | ||
dafb20fa | 361 | rdev_for_each(rdev, mddev) |
a6468539 N |
362 | array_sectors += (rdev->sectors & |
363 | ~(sector_t)(mddev->chunk_sectors-1)); | |
80c3a6ce DW |
364 | |
365 | return array_sectors; | |
366 | } | |
367 | ||
0c031fd3 XN |
368 | static void free_conf(struct mddev *mddev, struct r0conf *conf) |
369 | { | |
370 | kfree(conf->strip_zone); | |
371 | kfree(conf->devlist); | |
372 | kfree(conf); | |
0c031fd3 XN |
373 | } |
374 | ||
375 | static void raid0_free(struct mddev *mddev, void *priv) | |
376 | { | |
377 | struct r0conf *conf = priv; | |
378 | ||
379 | free_conf(mddev, conf); | |
0c031fd3 | 380 | } |
0366ef84 | 381 | |
fd01b88c | 382 | static int raid0_run(struct mddev *mddev) |
1da177e4 | 383 | { |
e373ab10 | 384 | struct r0conf *conf; |
5568a603 | 385 | int ret; |
1da177e4 | 386 | |
9d8f0363 | 387 | if (mddev->chunk_sectors == 0) { |
76603884 | 388 | pr_warn("md/raid0:%s: chunk size must be set.\n", mdname(mddev)); |
2604b703 N |
389 | return -EINVAL; |
390 | } | |
0894cc30 AN |
391 | if (md_check_no_bitmap(mddev)) |
392 | return -EINVAL; | |
753f2856 | 393 | |
9af204cf TM |
394 | /* if private is not null, we are here after takeover */ |
395 | if (mddev->private == NULL) { | |
396 | ret = create_strip_zones(mddev, &conf); | |
397 | if (ret < 0) | |
c567c86b | 398 | return ret; |
9af204cf TM |
399 | mddev->private = conf; |
400 | } | |
401 | conf = mddev->private; | |
199dc6ed N |
402 | if (mddev->queue) { |
403 | struct md_rdev *rdev; | |
199dc6ed | 404 | |
199dc6ed | 405 | blk_queue_max_hw_sectors(mddev->queue, mddev->chunk_sectors); |
3deff1a7 | 406 | blk_queue_max_write_zeroes_sectors(mddev->queue, mddev->chunk_sectors); |
199dc6ed N |
407 | |
408 | blk_queue_io_min(mddev->queue, mddev->chunk_sectors << 9); | |
409 | blk_queue_io_opt(mddev->queue, | |
410 | (mddev->chunk_sectors << 9) * mddev->raid_disks); | |
411 | ||
66eefe5d N |
412 | rdev_for_each(rdev, mddev) { |
413 | disk_stack_limits(mddev->gendisk, rdev->bdev, | |
414 | rdev->data_offset << 9); | |
66eefe5d | 415 | } |
199dc6ed | 416 | } |
1da177e4 LT |
417 | |
418 | /* calculate array device size */ | |
1f403624 | 419 | md_set_array_sectors(mddev, raid0_size(mddev, 0, 0)); |
1da177e4 | 420 | |
76603884 N |
421 | pr_debug("md/raid0:%s: md_size is %llu sectors.\n", |
422 | mdname(mddev), | |
423 | (unsigned long long)mddev->array_sectors); | |
753f2856 | 424 | |
46994191 | 425 | dump_zones(mddev); |
0366ef84 | 426 | |
427 | ret = md_integrity_register(mddev); | |
0c031fd3 | 428 | if (ret) |
c567c86b | 429 | free_conf(mddev, conf); |
0366ef84 | 430 | |
431 | return ret; | |
1da177e4 LT |
432 | } |
433 | ||
e8360070 JB |
434 | /* |
435 | * Convert disk_index to the disk order in which it is read/written. | |
436 | * For example, if we have 4 disks, they are numbered 0,1,2,3. If we | |
437 | * write the disks starting at disk 3, then the read/write order would | |
438 | * be disk 3, then 0, then 1, and then disk 2 and we want map_disk_shift() | |
439 | * to map the disks as follows 0,1,2,3 => 1,2,3,0. So disk 0 would map | |
440 | * to 1, 1 to 2, 2 to 3, and 3 to 0. That way we can compare disks in | |
441 | * that 'output' space to understand the read/write disk ordering. | |
442 | */ | |
443 | static int map_disk_shift(int disk_index, int num_disks, int disk_shift) | |
444 | { | |
445 | return ((disk_index + num_disks - disk_shift) % num_disks); | |
446 | } | |
447 | ||
29efc390 SL |
448 | static void raid0_handle_discard(struct mddev *mddev, struct bio *bio) |
449 | { | |
450 | struct r0conf *conf = mddev->private; | |
451 | struct strip_zone *zone; | |
452 | sector_t start = bio->bi_iter.bi_sector; | |
453 | sector_t end; | |
454 | unsigned int stripe_size; | |
455 | sector_t first_stripe_index, last_stripe_index; | |
456 | sector_t start_disk_offset; | |
457 | unsigned int start_disk_index; | |
458 | sector_t end_disk_offset; | |
459 | unsigned int end_disk_index; | |
460 | unsigned int disk; | |
e8360070 | 461 | sector_t orig_start, orig_end; |
29efc390 | 462 | |
e8360070 | 463 | orig_start = start; |
29efc390 SL |
464 | zone = find_zone(conf, &start); |
465 | ||
466 | if (bio_end_sector(bio) > zone->zone_end) { | |
467 | struct bio *split = bio_split(bio, | |
468 | zone->zone_end - bio->bi_iter.bi_sector, GFP_NOIO, | |
afeee514 | 469 | &mddev->bio_set); |
29efc390 | 470 | bio_chain(split, bio); |
ed00aabd | 471 | submit_bio_noacct(bio); |
29efc390 SL |
472 | bio = split; |
473 | end = zone->zone_end; | |
474 | } else | |
475 | end = bio_end_sector(bio); | |
476 | ||
e8360070 | 477 | orig_end = end; |
29efc390 SL |
478 | if (zone != conf->strip_zone) |
479 | end = end - zone[-1].zone_end; | |
480 | ||
481 | /* Now start and end is the offset in zone */ | |
482 | stripe_size = zone->nb_dev * mddev->chunk_sectors; | |
483 | ||
484 | first_stripe_index = start; | |
485 | sector_div(first_stripe_index, stripe_size); | |
486 | last_stripe_index = end; | |
487 | sector_div(last_stripe_index, stripe_size); | |
488 | ||
e8360070 JB |
489 | /* In the first zone the original and alternate layouts are the same */ |
490 | if ((conf->layout == RAID0_ORIG_LAYOUT) && (zone != conf->strip_zone)) { | |
491 | sector_div(orig_start, mddev->chunk_sectors); | |
492 | start_disk_index = sector_div(orig_start, zone->nb_dev); | |
493 | start_disk_index = map_disk_shift(start_disk_index, | |
494 | zone->nb_dev, | |
495 | zone->disk_shift); | |
496 | sector_div(orig_end, mddev->chunk_sectors); | |
497 | end_disk_index = sector_div(orig_end, zone->nb_dev); | |
498 | end_disk_index = map_disk_shift(end_disk_index, | |
499 | zone->nb_dev, zone->disk_shift); | |
500 | } else { | |
501 | start_disk_index = (int)(start - first_stripe_index * stripe_size) / | |
502 | mddev->chunk_sectors; | |
503 | end_disk_index = (int)(end - last_stripe_index * stripe_size) / | |
504 | mddev->chunk_sectors; | |
505 | } | |
29efc390 SL |
506 | start_disk_offset = ((int)(start - first_stripe_index * stripe_size) % |
507 | mddev->chunk_sectors) + | |
508 | first_stripe_index * mddev->chunk_sectors; | |
29efc390 SL |
509 | end_disk_offset = ((int)(end - last_stripe_index * stripe_size) % |
510 | mddev->chunk_sectors) + | |
511 | last_stripe_index * mddev->chunk_sectors; | |
512 | ||
513 | for (disk = 0; disk < zone->nb_dev; disk++) { | |
514 | sector_t dev_start, dev_end; | |
29efc390 | 515 | struct md_rdev *rdev; |
e8360070 JB |
516 | int compare_disk; |
517 | ||
518 | compare_disk = map_disk_shift(disk, zone->nb_dev, | |
519 | zone->disk_shift); | |
29efc390 | 520 | |
e8360070 | 521 | if (compare_disk < start_disk_index) |
29efc390 SL |
522 | dev_start = (first_stripe_index + 1) * |
523 | mddev->chunk_sectors; | |
e8360070 | 524 | else if (compare_disk > start_disk_index) |
29efc390 SL |
525 | dev_start = first_stripe_index * mddev->chunk_sectors; |
526 | else | |
527 | dev_start = start_disk_offset; | |
528 | ||
e8360070 | 529 | if (compare_disk < end_disk_index) |
29efc390 | 530 | dev_end = (last_stripe_index + 1) * mddev->chunk_sectors; |
e8360070 | 531 | else if (compare_disk > end_disk_index) |
29efc390 SL |
532 | dev_end = last_stripe_index * mddev->chunk_sectors; |
533 | else | |
534 | dev_end = end_disk_offset; | |
535 | ||
536 | if (dev_end <= dev_start) | |
537 | continue; | |
538 | ||
539 | rdev = conf->devlist[(zone - conf->strip_zone) * | |
540 | conf->strip_zone[0].nb_dev + disk]; | |
cf78408f | 541 | md_submit_discard_bio(mddev, rdev, bio, |
29efc390 | 542 | dev_start + zone->dev_start + rdev->data_offset, |
cf78408f | 543 | dev_end - dev_start); |
29efc390 SL |
544 | } |
545 | bio_endio(bio); | |
546 | } | |
547 | ||
af50e20a | 548 | static void raid0_map_submit_bio(struct mddev *mddev, struct bio *bio) |
fbb704ef | 549 | { |
c84a1372 | 550 | struct r0conf *conf = mddev->private; |
1da177e4 | 551 | struct strip_zone *zone; |
3cb03002 | 552 | struct md_rdev *tmp_dev; |
af50e20a JK |
553 | sector_t bio_sector = bio->bi_iter.bi_sector; |
554 | sector_t sector = bio_sector; | |
1da177e4 | 555 | |
cc22b540 | 556 | md_account_bio(mddev, &bio); |
10764815 | 557 | |
f00d7c85 | 558 | zone = find_zone(mddev->private, §or); |
c84a1372 N |
559 | switch (conf->layout) { |
560 | case RAID0_ORIG_LAYOUT: | |
af50e20a | 561 | tmp_dev = map_sector(mddev, zone, bio_sector, §or); |
c84a1372 N |
562 | break; |
563 | case RAID0_ALT_MULTIZONE_LAYOUT: | |
564 | tmp_dev = map_sector(mddev, zone, sector, §or); | |
565 | break; | |
566 | default: | |
e3fc3f3d | 567 | WARN(1, "md/raid0:%s: Invalid layout\n", mdname(mddev)); |
c84a1372 | 568 | bio_io_error(bio); |
af50e20a | 569 | return; |
c84a1372 | 570 | } |
62f7b198 | 571 | |
c31fea2f | 572 | if (unlikely(is_rdev_broken(tmp_dev))) { |
62f7b198 | 573 | bio_io_error(bio); |
c31fea2f | 574 | md_error(mddev, tmp_dev); |
af50e20a | 575 | return; |
62f7b198 GP |
576 | } |
577 | ||
74d46992 | 578 | bio_set_dev(bio, tmp_dev->bdev); |
f00d7c85 N |
579 | bio->bi_iter.bi_sector = sector + zone->dev_start + |
580 | tmp_dev->data_offset; | |
581 | ||
29efc390 | 582 | if (mddev->gendisk) |
1c02fca6 CH |
583 | trace_block_bio_remap(bio, disk_devt(mddev->gendisk), |
584 | bio_sector); | |
29efc390 | 585 | mddev_check_write_zeroes(mddev, bio); |
ed00aabd | 586 | submit_bio_noacct(bio); |
af50e20a JK |
587 | } |
588 | ||
589 | static bool raid0_make_request(struct mddev *mddev, struct bio *bio) | |
590 | { | |
591 | sector_t sector; | |
592 | unsigned chunk_sects; | |
593 | unsigned sectors; | |
594 | ||
595 | if (unlikely(bio->bi_opf & REQ_PREFLUSH) | |
596 | && md_flush_request(mddev, bio)) | |
597 | return true; | |
598 | ||
599 | if (unlikely((bio_op(bio) == REQ_OP_DISCARD))) { | |
600 | raid0_handle_discard(mddev, bio); | |
601 | return true; | |
602 | } | |
603 | ||
604 | sector = bio->bi_iter.bi_sector; | |
605 | chunk_sects = mddev->chunk_sectors; | |
606 | ||
607 | sectors = chunk_sects - | |
608 | (likely(is_power_of_2(chunk_sects)) | |
609 | ? (sector & (chunk_sects-1)) | |
610 | : sector_div(sector, chunk_sects)); | |
611 | ||
612 | if (sectors < bio_sectors(bio)) { | |
613 | struct bio *split = bio_split(bio, sectors, GFP_NOIO, | |
614 | &mddev->bio_set); | |
615 | bio_chain(split, bio); | |
319ff40a | 616 | raid0_map_submit_bio(mddev, bio); |
af50e20a JK |
617 | bio = split; |
618 | } | |
619 | ||
620 | raid0_map_submit_bio(mddev, bio); | |
cc27b0c7 | 621 | return true; |
1da177e4 | 622 | } |
8299d7f7 | 623 | |
fd01b88c | 624 | static void raid0_status(struct seq_file *seq, struct mddev *mddev) |
1da177e4 | 625 | { |
9d8f0363 | 626 | seq_printf(seq, " %dk chunks", mddev->chunk_sectors / 2); |
1da177e4 LT |
627 | return; |
628 | } | |
629 | ||
c31fea2f MT |
630 | static void raid0_error(struct mddev *mddev, struct md_rdev *rdev) |
631 | { | |
632 | if (!test_and_set_bit(MD_BROKEN, &mddev->flags)) { | |
633 | char *md_name = mdname(mddev); | |
634 | ||
635 | pr_crit("md/raid0%s: Disk failure on %pg detected, failing array.\n", | |
636 | md_name, rdev->bdev); | |
637 | } | |
638 | } | |
639 | ||
fd01b88c | 640 | static void *raid0_takeover_raid45(struct mddev *mddev) |
9af204cf | 641 | { |
3cb03002 | 642 | struct md_rdev *rdev; |
e373ab10 | 643 | struct r0conf *priv_conf; |
9af204cf TM |
644 | |
645 | if (mddev->degraded != 1) { | |
76603884 N |
646 | pr_warn("md/raid0:%s: raid5 must be degraded! Degraded disks: %d\n", |
647 | mdname(mddev), | |
648 | mddev->degraded); | |
9af204cf TM |
649 | return ERR_PTR(-EINVAL); |
650 | } | |
651 | ||
dafb20fa | 652 | rdev_for_each(rdev, mddev) { |
9af204cf TM |
653 | /* check slot number for a disk */ |
654 | if (rdev->raid_disk == mddev->raid_disks-1) { | |
76603884 N |
655 | pr_warn("md/raid0:%s: raid5 must have missing parity disk!\n", |
656 | mdname(mddev)); | |
9af204cf TM |
657 | return ERR_PTR(-EINVAL); |
658 | } | |
eea136d6 | 659 | rdev->sectors = mddev->dev_sectors; |
9af204cf TM |
660 | } |
661 | ||
662 | /* Set new parameters */ | |
663 | mddev->new_level = 0; | |
001048a3 | 664 | mddev->new_layout = 0; |
9af204cf TM |
665 | mddev->new_chunk_sectors = mddev->chunk_sectors; |
666 | mddev->raid_disks--; | |
667 | mddev->delta_disks = -1; | |
668 | /* make sure it will be not marked as dirty */ | |
669 | mddev->recovery_cp = MaxSector; | |
394ed8e4 | 670 | mddev_clear_unsupported_flags(mddev, UNSUPPORTED_MDDEV_FLAGS); |
9af204cf TM |
671 | |
672 | create_strip_zones(mddev, &priv_conf); | |
6995f0b2 | 673 | |
9af204cf TM |
674 | return priv_conf; |
675 | } | |
676 | ||
fd01b88c | 677 | static void *raid0_takeover_raid10(struct mddev *mddev) |
9af204cf | 678 | { |
e373ab10 | 679 | struct r0conf *priv_conf; |
9af204cf TM |
680 | |
681 | /* Check layout: | |
682 | * - far_copies must be 1 | |
683 | * - near_copies must be 2 | |
684 | * - disks number must be even | |
685 | * - all mirrors must be already degraded | |
686 | */ | |
687 | if (mddev->layout != ((1 << 8) + 2)) { | |
76603884 N |
688 | pr_warn("md/raid0:%s:: Raid0 cannot takeover layout: 0x%x\n", |
689 | mdname(mddev), | |
690 | mddev->layout); | |
9af204cf TM |
691 | return ERR_PTR(-EINVAL); |
692 | } | |
693 | if (mddev->raid_disks & 1) { | |
76603884 N |
694 | pr_warn("md/raid0:%s: Raid0 cannot takeover Raid10 with odd disk number.\n", |
695 | mdname(mddev)); | |
9af204cf TM |
696 | return ERR_PTR(-EINVAL); |
697 | } | |
698 | if (mddev->degraded != (mddev->raid_disks>>1)) { | |
76603884 N |
699 | pr_warn("md/raid0:%s: All mirrors must be already degraded!\n", |
700 | mdname(mddev)); | |
9af204cf TM |
701 | return ERR_PTR(-EINVAL); |
702 | } | |
703 | ||
704 | /* Set new parameters */ | |
705 | mddev->new_level = 0; | |
001048a3 | 706 | mddev->new_layout = 0; |
9af204cf TM |
707 | mddev->new_chunk_sectors = mddev->chunk_sectors; |
708 | mddev->delta_disks = - mddev->raid_disks / 2; | |
709 | mddev->raid_disks += mddev->delta_disks; | |
710 | mddev->degraded = 0; | |
711 | /* make sure it will be not marked as dirty */ | |
712 | mddev->recovery_cp = MaxSector; | |
394ed8e4 | 713 | mddev_clear_unsupported_flags(mddev, UNSUPPORTED_MDDEV_FLAGS); |
9af204cf TM |
714 | |
715 | create_strip_zones(mddev, &priv_conf); | |
9af204cf TM |
716 | return priv_conf; |
717 | } | |
718 | ||
fd01b88c | 719 | static void *raid0_takeover_raid1(struct mddev *mddev) |
fc3a08b8 | 720 | { |
e373ab10 | 721 | struct r0conf *priv_conf; |
24b961f8 | 722 | int chunksect; |
fc3a08b8 KW |
723 | |
724 | /* Check layout: | |
725 | * - (N - 1) mirror drives must be already faulty | |
726 | */ | |
727 | if ((mddev->raid_disks - 1) != mddev->degraded) { | |
76603884 | 728 | pr_err("md/raid0:%s: (N - 1) mirrors drives must be already faulty!\n", |
fc3a08b8 KW |
729 | mdname(mddev)); |
730 | return ERR_PTR(-EINVAL); | |
731 | } | |
732 | ||
24b961f8 JS |
733 | /* |
734 | * a raid1 doesn't have the notion of chunk size, so | |
735 | * figure out the largest suitable size we can use. | |
736 | */ | |
737 | chunksect = 64 * 2; /* 64K by default */ | |
738 | ||
739 | /* The array must be an exact multiple of chunksize */ | |
740 | while (chunksect && (mddev->array_sectors & (chunksect - 1))) | |
741 | chunksect >>= 1; | |
742 | ||
743 | if ((chunksect << 9) < PAGE_SIZE) | |
744 | /* array size does not allow a suitable chunk size */ | |
745 | return ERR_PTR(-EINVAL); | |
746 | ||
fc3a08b8 KW |
747 | /* Set new parameters */ |
748 | mddev->new_level = 0; | |
749 | mddev->new_layout = 0; | |
24b961f8 JS |
750 | mddev->new_chunk_sectors = chunksect; |
751 | mddev->chunk_sectors = chunksect; | |
fc3a08b8 | 752 | mddev->delta_disks = 1 - mddev->raid_disks; |
f7bee809 | 753 | mddev->raid_disks = 1; |
fc3a08b8 KW |
754 | /* make sure it will be not marked as dirty */ |
755 | mddev->recovery_cp = MaxSector; | |
394ed8e4 | 756 | mddev_clear_unsupported_flags(mddev, UNSUPPORTED_MDDEV_FLAGS); |
fc3a08b8 KW |
757 | |
758 | create_strip_zones(mddev, &priv_conf); | |
759 | return priv_conf; | |
760 | } | |
761 | ||
fd01b88c | 762 | static void *raid0_takeover(struct mddev *mddev) |
9af204cf TM |
763 | { |
764 | /* raid0 can take over: | |
049d6c1e | 765 | * raid4 - if all data disks are active. |
9af204cf TM |
766 | * raid5 - providing it is Raid4 layout and one disk is faulty |
767 | * raid10 - assuming we have all necessary active disks | |
fc3a08b8 | 768 | * raid1 - with (N -1) mirror drives faulty |
9af204cf | 769 | */ |
a8461a61 N |
770 | |
771 | if (mddev->bitmap) { | |
76603884 N |
772 | pr_warn("md/raid0: %s: cannot takeover array with bitmap\n", |
773 | mdname(mddev)); | |
a8461a61 N |
774 | return ERR_PTR(-EBUSY); |
775 | } | |
049d6c1e MT |
776 | if (mddev->level == 4) |
777 | return raid0_takeover_raid45(mddev); | |
778 | ||
9af204cf TM |
779 | if (mddev->level == 5) { |
780 | if (mddev->layout == ALGORITHM_PARITY_N) | |
049d6c1e | 781 | return raid0_takeover_raid45(mddev); |
9af204cf | 782 | |
76603884 N |
783 | pr_warn("md/raid0:%s: Raid can only takeover Raid5 with layout: %d\n", |
784 | mdname(mddev), ALGORITHM_PARITY_N); | |
9af204cf TM |
785 | } |
786 | ||
787 | if (mddev->level == 10) | |
788 | return raid0_takeover_raid10(mddev); | |
789 | ||
fc3a08b8 KW |
790 | if (mddev->level == 1) |
791 | return raid0_takeover_raid1(mddev); | |
792 | ||
76603884 | 793 | pr_warn("Takeover from raid%i to raid0 not supported\n", |
fc3a08b8 KW |
794 | mddev->level); |
795 | ||
9af204cf TM |
796 | return ERR_PTR(-EINVAL); |
797 | } | |
798 | ||
b03e0ccb | 799 | static void raid0_quiesce(struct mddev *mddev, int quiesce) |
9af204cf TM |
800 | { |
801 | } | |
802 | ||
84fc4b56 | 803 | static struct md_personality raid0_personality= |
1da177e4 LT |
804 | { |
805 | .name = "raid0", | |
2604b703 | 806 | .level = 0, |
1da177e4 LT |
807 | .owner = THIS_MODULE, |
808 | .make_request = raid0_make_request, | |
809 | .run = raid0_run, | |
afa0f557 | 810 | .free = raid0_free, |
1da177e4 | 811 | .status = raid0_status, |
80c3a6ce | 812 | .size = raid0_size, |
9af204cf TM |
813 | .takeover = raid0_takeover, |
814 | .quiesce = raid0_quiesce, | |
c31fea2f | 815 | .error_handler = raid0_error, |
1da177e4 LT |
816 | }; |
817 | ||
818 | static int __init raid0_init (void) | |
819 | { | |
2604b703 | 820 | return register_md_personality (&raid0_personality); |
1da177e4 LT |
821 | } |
822 | ||
823 | static void raid0_exit (void) | |
824 | { | |
2604b703 | 825 | unregister_md_personality (&raid0_personality); |
1da177e4 LT |
826 | } |
827 | ||
828 | module_init(raid0_init); | |
829 | module_exit(raid0_exit); | |
830 | MODULE_LICENSE("GPL"); | |
0efb9e61 | 831 | MODULE_DESCRIPTION("RAID0 (striping) personality for MD"); |
1da177e4 | 832 | MODULE_ALIAS("md-personality-2"); /* RAID0 */ |
d9d166c2 | 833 | MODULE_ALIAS("md-raid0"); |
2604b703 | 834 | MODULE_ALIAS("md-level-0"); |