1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 raid0.c : Multiple Devices driver for Linux
4 Copyright (C) 1994-96 Marc ZYNGIER
5 <zyngier@ufr-info-p7.ibp.fr> or
6 <maz@gloups.fdn.fr>
7 Copyright (C) 1999, 2000 Ingo Molnar, Red Hat
8
9 RAID-0 management functions.
10
11 */
12
13 #include <linux/blkdev.h>
14 #include <linux/seq_file.h>
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/string_choices.h>
18 #include <trace/events/block.h>
19 #include "md.h"
20 #include "raid0.h"
21 #include "raid5.h"
22
23 static int default_layout = 0;
24 module_param(default_layout, int, 0644);
25
26 #define UNSUPPORTED_MDDEV_FLAGS \
27 ((1L << MD_HAS_JOURNAL) | \
28 (1L << MD_JOURNAL_CLEAN) | \
29 (1L << MD_FAILFAST_SUPPORTED) |\
30 (1L << MD_HAS_PPL) | \
31 (1L << MD_HAS_MULTIPLE_PPLS) | \
32 (1L << MD_FAILLAST_DEV) | \
33 (1L << MD_SERIALIZE_POLICY))
34
35 /*
36 * inform the user of the raid configuration
37 */
dump_zones(struct mddev * mddev)38 static void dump_zones(struct mddev *mddev)
39 {
40 int j, k;
41 sector_t zone_size = 0;
42 sector_t zone_start = 0;
43 struct r0conf *conf = mddev->private;
44 int raid_disks = conf->strip_zone[0].nb_dev;
45 pr_debug("md: RAID0 configuration for %s - %d zone%s\n",
46 mdname(mddev),
47 conf->nr_strip_zones, str_plural(conf->nr_strip_zones));
48 for (j = 0; j < conf->nr_strip_zones; j++) {
49 char line[200];
50 int len = 0;
51
52 for (k = 0; k < conf->strip_zone[j].nb_dev; k++)
53 len += scnprintf(line+len, 200-len, "%s%pg", k?"/":"",
54 conf->devlist[j * raid_disks + k]->bdev);
55 pr_debug("md: zone%d=[%s]\n", j, line);
56
57 zone_size = conf->strip_zone[j].zone_end - zone_start;
58 pr_debug(" zone-offset=%10lluKB, device-offset=%10lluKB, size=%10lluKB\n",
59 (unsigned long long)zone_start>>1,
60 (unsigned long long)conf->strip_zone[j].dev_start>>1,
61 (unsigned long long)zone_size>>1);
62 zone_start = conf->strip_zone[j].zone_end;
63 }
64 }
65
create_strip_zones(struct mddev * mddev,struct r0conf ** private_conf)66 static int create_strip_zones(struct mddev *mddev, struct r0conf **private_conf)
67 {
68 int i, c, err;
69 sector_t curr_zone_end, sectors;
70 struct md_rdev *smallest, *rdev1, *rdev2, *rdev, **dev;
71 struct strip_zone *zone;
72 int cnt;
73 struct r0conf *conf = kzalloc_obj(*conf);
74 unsigned int blksize = 512;
75
76 if (!mddev_is_dm(mddev))
77 blksize = queue_logical_block_size(mddev->gendisk->queue);
78
79 *private_conf = ERR_PTR(-ENOMEM);
80 if (!conf)
81 return -ENOMEM;
82 rdev_for_each(rdev1, mddev) {
83 pr_debug("md/raid0:%s: looking at %pg\n",
84 mdname(mddev),
85 rdev1->bdev);
86 c = 0;
87
88 /* round size to chunk_size */
89 sectors = rdev1->sectors;
90 sector_div(sectors, mddev->chunk_sectors);
91 rdev1->sectors = sectors * mddev->chunk_sectors;
92
93 if (mddev_is_dm(mddev))
94 blksize = max(blksize, queue_logical_block_size(
95 rdev1->bdev->bd_disk->queue));
96
97 rdev_for_each(rdev2, mddev) {
98 pr_debug("md/raid0:%s: comparing %pg(%llu)"
99 " with %pg(%llu)\n",
100 mdname(mddev),
101 rdev1->bdev,
102 (unsigned long long)rdev1->sectors,
103 rdev2->bdev,
104 (unsigned long long)rdev2->sectors);
105 if (rdev2 == rdev1) {
106 pr_debug("md/raid0:%s: END\n",
107 mdname(mddev));
108 break;
109 }
110 if (rdev2->sectors == rdev1->sectors) {
111 /*
112 * Not unique, don't count it as a new
113 * group
114 */
115 pr_debug("md/raid0:%s: EQUAL\n",
116 mdname(mddev));
117 c = 1;
118 break;
119 }
120 pr_debug("md/raid0:%s: NOT EQUAL\n",
121 mdname(mddev));
122 }
123 if (!c) {
124 pr_debug("md/raid0:%s: ==> UNIQUE\n",
125 mdname(mddev));
126 conf->nr_strip_zones++;
127 pr_debug("md/raid0:%s: %d zones\n",
128 mdname(mddev), conf->nr_strip_zones);
129 }
130 }
131 pr_debug("md/raid0:%s: FINAL %d zones\n",
132 mdname(mddev), conf->nr_strip_zones);
133
134 /*
135 * now since we have the hard sector sizes, we can make sure
136 * chunk size is a multiple of that sector size
137 */
138 if ((mddev->chunk_sectors << 9) % blksize) {
139 pr_warn("md/raid0:%s: chunk_size of %d not multiple of block size %d\n",
140 mdname(mddev),
141 mddev->chunk_sectors << 9, blksize);
142 err = -EINVAL;
143 goto abort;
144 }
145
146 err = -ENOMEM;
147 conf->strip_zone = kvzalloc_objs(struct strip_zone, conf->nr_strip_zones);
148 if (!conf->strip_zone)
149 goto abort;
150 conf->devlist = kvzalloc(array3_size(sizeof(struct md_rdev *),
151 conf->nr_strip_zones,
152 mddev->raid_disks),
153 GFP_KERNEL);
154 if (!conf->devlist)
155 goto abort;
156
157 /* The first zone must contain all devices, so here we check that
158 * there is a proper alignment of slots to devices and find them all
159 */
160 zone = &conf->strip_zone[0];
161 cnt = 0;
162 smallest = NULL;
163 dev = conf->devlist;
164 err = -EINVAL;
165 rdev_for_each(rdev1, mddev) {
166 int j = rdev1->raid_disk;
167
168 if (mddev->level == 10) {
169 /* taking over a raid10-n2 array */
170 j /= 2;
171 rdev1->new_raid_disk = j;
172 }
173
174 if (mddev->level == 1) {
175 /* taiking over a raid1 array-
176 * we have only one active disk
177 */
178 j = 0;
179 rdev1->new_raid_disk = j;
180 }
181
182 if (j < 0) {
183 pr_warn("md/raid0:%s: remove inactive devices before converting to RAID0\n",
184 mdname(mddev));
185 goto abort;
186 }
187 if (j >= mddev->raid_disks) {
188 pr_warn("md/raid0:%s: bad disk number %d - aborting!\n",
189 mdname(mddev), j);
190 goto abort;
191 }
192 if (dev[j]) {
193 pr_warn("md/raid0:%s: multiple devices for %d - aborting!\n",
194 mdname(mddev), j);
195 goto abort;
196 }
197 dev[j] = rdev1;
198
199 if (!smallest || (rdev1->sectors < smallest->sectors))
200 smallest = rdev1;
201 cnt++;
202 }
203 if (cnt != mddev->raid_disks) {
204 pr_warn("md/raid0:%s: too few disks (%d of %d) - aborting!\n",
205 mdname(mddev), cnt, mddev->raid_disks);
206 goto abort;
207 }
208 zone->nb_dev = cnt;
209 zone->zone_end = smallest->sectors * cnt;
210
211 curr_zone_end = zone->zone_end;
212
213 /* now do the other zones */
214 for (i = 1; i < conf->nr_strip_zones; i++)
215 {
216 int j;
217
218 zone = conf->strip_zone + i;
219 dev = conf->devlist + i * mddev->raid_disks;
220
221 pr_debug("md/raid0:%s: zone %d\n", mdname(mddev), i);
222 zone->dev_start = smallest->sectors;
223 smallest = NULL;
224 c = 0;
225
226 for (j=0; j<cnt; j++) {
227 rdev = conf->devlist[j];
228 if (rdev->sectors <= zone->dev_start) {
229 pr_debug("md/raid0:%s: checking %pg ... nope\n",
230 mdname(mddev),
231 rdev->bdev);
232 continue;
233 }
234 pr_debug("md/raid0:%s: checking %pg ..."
235 " contained as device %d\n",
236 mdname(mddev),
237 rdev->bdev, c);
238 dev[c] = rdev;
239 c++;
240 if (!smallest || rdev->sectors < smallest->sectors) {
241 smallest = rdev;
242 pr_debug("md/raid0:%s: (%llu) is smallest!.\n",
243 mdname(mddev),
244 (unsigned long long)rdev->sectors);
245 }
246 }
247
248 zone->nb_dev = c;
249 sectors = (smallest->sectors - zone->dev_start) * c;
250 pr_debug("md/raid0:%s: zone->nb_dev: %d, sectors: %llu\n",
251 mdname(mddev),
252 zone->nb_dev, (unsigned long long)sectors);
253
254 curr_zone_end += sectors;
255 zone->zone_end = curr_zone_end;
256
257 pr_debug("md/raid0:%s: current zone start: %llu\n",
258 mdname(mddev),
259 (unsigned long long)smallest->sectors);
260 }
261
262 if (conf->nr_strip_zones == 1 || conf->strip_zone[1].nb_dev == 1) {
263 conf->layout = RAID0_ORIG_LAYOUT;
264 } else if (mddev->layout == RAID0_ORIG_LAYOUT ||
265 mddev->layout == RAID0_ALT_MULTIZONE_LAYOUT) {
266 conf->layout = mddev->layout;
267 } else if (default_layout == RAID0_ORIG_LAYOUT ||
268 default_layout == RAID0_ALT_MULTIZONE_LAYOUT) {
269 conf->layout = default_layout;
270 } else {
271 pr_err("md/raid0:%s: cannot assemble multi-zone RAID0 with default_layout setting\n",
272 mdname(mddev));
273 pr_err("md/raid0: please set raid0.default_layout to 1 or 2\n");
274 err = -EOPNOTSUPP;
275 goto abort;
276 }
277
278 if (conf->layout == RAID0_ORIG_LAYOUT) {
279 for (i = 1; i < conf->nr_strip_zones; i++) {
280 sector_t first_sector = conf->strip_zone[i-1].zone_end;
281
282 sector_div(first_sector, mddev->chunk_sectors);
283 zone = conf->strip_zone + i;
284 /* disk_shift is first disk index used in the zone */
285 zone->disk_shift = sector_div(first_sector,
286 zone->nb_dev);
287 }
288 }
289
290 pr_debug("md/raid0:%s: done.\n", mdname(mddev));
291 *private_conf = conf;
292
293 return 0;
294 abort:
295 kvfree(conf->strip_zone);
296 kvfree(conf->devlist);
297 kfree(conf);
298 *private_conf = ERR_PTR(err);
299 return err;
300 }
301
302 /* Find the zone which holds a particular offset
303 * Update *sectorp to be an offset in that zone
304 */
find_zone(struct r0conf * conf,sector_t * sectorp)305 static struct strip_zone *find_zone(struct r0conf *conf,
306 sector_t *sectorp)
307 {
308 int i;
309 struct strip_zone *z = conf->strip_zone;
310 sector_t sector = *sectorp;
311
312 for (i = 0; i < conf->nr_strip_zones; i++)
313 if (sector < z[i].zone_end) {
314 if (i)
315 *sectorp = sector - z[i-1].zone_end;
316 return z + i;
317 }
318 BUG();
319 }
320
321 /*
322 * remaps the bio to the target device. we separate two flows.
323 * power 2 flow and a general flow for the sake of performance
324 */
map_sector(struct mddev * mddev,struct strip_zone * zone,sector_t sector,sector_t * sector_offset)325 static struct md_rdev *map_sector(struct mddev *mddev, struct strip_zone *zone,
326 sector_t sector, sector_t *sector_offset)
327 {
328 unsigned int sect_in_chunk;
329 sector_t chunk;
330 struct r0conf *conf = mddev->private;
331 int raid_disks = conf->strip_zone[0].nb_dev;
332 unsigned int chunk_sects = mddev->chunk_sectors;
333
334 if (is_power_of_2(chunk_sects)) {
335 int chunksect_bits = ffz(~chunk_sects);
336 /* find the sector offset inside the chunk */
337 sect_in_chunk = sector & (chunk_sects - 1);
338 sector >>= chunksect_bits;
339 /* chunk in zone */
340 chunk = *sector_offset;
341 /* quotient is the chunk in real device*/
342 sector_div(chunk, zone->nb_dev << chunksect_bits);
343 } else{
344 sect_in_chunk = sector_div(sector, chunk_sects);
345 chunk = *sector_offset;
346 sector_div(chunk, chunk_sects * zone->nb_dev);
347 }
348 /*
349 * position the bio over the real device
350 * real sector = chunk in device + starting of zone
351 * + the position in the chunk
352 */
353 *sector_offset = (chunk * chunk_sects) + sect_in_chunk;
354 return conf->devlist[(zone - conf->strip_zone)*raid_disks
355 + sector_div(sector, zone->nb_dev)];
356 }
357
raid0_size(struct mddev * mddev,sector_t sectors,int raid_disks)358 static sector_t raid0_size(struct mddev *mddev, sector_t sectors, int raid_disks)
359 {
360 sector_t array_sectors = 0;
361 struct md_rdev *rdev;
362
363 WARN_ONCE(sectors || raid_disks,
364 "%s does not support generic reshape\n", __func__);
365
366 rdev_for_each(rdev, mddev)
367 array_sectors += (rdev->sectors &
368 ~(sector_t)(mddev->chunk_sectors-1));
369
370 return array_sectors;
371 }
372
raid0_free(struct mddev * mddev,void * priv)373 static void raid0_free(struct mddev *mddev, void *priv)
374 {
375 struct r0conf *conf = priv;
376
377 kvfree(conf->strip_zone);
378 kvfree(conf->devlist);
379 kfree(conf);
380 }
381
raid0_set_limits(struct mddev * mddev)382 static int raid0_set_limits(struct mddev *mddev)
383 {
384 struct queue_limits lim;
385 int err;
386
387 md_init_stacking_limits(&lim);
388 lim.features |= BLK_FEAT_NOWAIT;
389 lim.max_hw_sectors = mddev->chunk_sectors;
390 lim.max_write_zeroes_sectors = mddev->chunk_sectors;
391 lim.max_hw_wzeroes_unmap_sectors = mddev->chunk_sectors;
392 lim.logical_block_size = mddev->logical_block_size;
393 lim.io_min = mddev->chunk_sectors << 9;
394 lim.io_opt = lim.io_min * mddev->raid_disks;
395 lim.chunk_sectors = mddev->chunk_sectors;
396 lim.features |= BLK_FEAT_ATOMIC_WRITES;
397 lim.features |= BLK_FEAT_PCI_P2PDMA;
398 err = mddev_stack_rdev_limits(mddev, &lim, MDDEV_STACK_INTEGRITY);
399 if (err)
400 return err;
401 return queue_limits_set(mddev->gendisk->queue, &lim);
402 }
403
raid0_run(struct mddev * mddev)404 static int raid0_run(struct mddev *mddev)
405 {
406 struct r0conf *conf;
407 int ret;
408
409 if (mddev->chunk_sectors == 0) {
410 pr_warn("md/raid0:%s: chunk size must be set.\n", mdname(mddev));
411 return -EINVAL;
412 }
413 if (md_check_no_bitmap(mddev))
414 return -EINVAL;
415
416 if (!mddev_is_dm(mddev)) {
417 ret = raid0_set_limits(mddev);
418 if (ret)
419 return ret;
420 }
421
422 /* if private is not null, we are here after takeover */
423 if (mddev->private == NULL) {
424 ret = create_strip_zones(mddev, &conf);
425 if (ret < 0)
426 return ret;
427 mddev->private = conf;
428 }
429 conf = mddev->private;
430
431 /* calculate array device size */
432 md_set_array_sectors(mddev, raid0_size(mddev, 0, 0));
433
434 pr_debug("md/raid0:%s: md_size is %llu sectors.\n",
435 mdname(mddev),
436 (unsigned long long)mddev->array_sectors);
437
438 dump_zones(mddev);
439
440 return md_integrity_register(mddev);
441 }
442
443 /*
444 * Convert disk_index to the disk order in which it is read/written.
445 * For example, if we have 4 disks, they are numbered 0,1,2,3. If we
446 * write the disks starting at disk 3, then the read/write order would
447 * be disk 3, then 0, then 1, and then disk 2 and we want map_disk_shift()
448 * to map the disks as follows 0,1,2,3 => 1,2,3,0. So disk 0 would map
449 * to 1, 1 to 2, 2 to 3, and 3 to 0. That way we can compare disks in
450 * that 'output' space to understand the read/write disk ordering.
451 */
map_disk_shift(int disk_index,int num_disks,int disk_shift)452 static int map_disk_shift(int disk_index, int num_disks, int disk_shift)
453 {
454 return ((disk_index + num_disks - disk_shift) % num_disks);
455 }
456
raid0_handle_discard(struct mddev * mddev,struct bio * bio)457 static void raid0_handle_discard(struct mddev *mddev, struct bio *bio)
458 {
459 struct r0conf *conf = mddev->private;
460 struct strip_zone *zone;
461 sector_t start = bio->bi_iter.bi_sector;
462 sector_t end;
463 unsigned int stripe_size;
464 sector_t first_stripe_index, last_stripe_index;
465 sector_t start_disk_offset;
466 unsigned int start_disk_index;
467 sector_t end_disk_offset;
468 unsigned int end_disk_index;
469 unsigned int disk;
470 sector_t orig_start, orig_end;
471
472 orig_start = start;
473 zone = find_zone(conf, &start);
474
475 if (bio_end_sector(bio) > zone->zone_end) {
476 bio = bio_submit_split_bioset(bio,
477 zone->zone_end - bio->bi_iter.bi_sector,
478 &mddev->bio_set);
479 if (!bio)
480 return;
481
482 end = zone->zone_end;
483 } else {
484 end = bio_end_sector(bio);
485 }
486
487 orig_end = end;
488 if (zone != conf->strip_zone)
489 end = end - zone[-1].zone_end;
490
491 /* Now start and end is the offset in zone */
492 stripe_size = zone->nb_dev * mddev->chunk_sectors;
493
494 first_stripe_index = start;
495 sector_div(first_stripe_index, stripe_size);
496 last_stripe_index = end;
497 sector_div(last_stripe_index, stripe_size);
498
499 /* In the first zone the original and alternate layouts are the same */
500 if ((conf->layout == RAID0_ORIG_LAYOUT) && (zone != conf->strip_zone)) {
501 sector_div(orig_start, mddev->chunk_sectors);
502 start_disk_index = sector_div(orig_start, zone->nb_dev);
503 start_disk_index = map_disk_shift(start_disk_index,
504 zone->nb_dev,
505 zone->disk_shift);
506 sector_div(orig_end, mddev->chunk_sectors);
507 end_disk_index = sector_div(orig_end, zone->nb_dev);
508 end_disk_index = map_disk_shift(end_disk_index,
509 zone->nb_dev, zone->disk_shift);
510 } else {
511 start_disk_index = (int)(start - first_stripe_index * stripe_size) /
512 mddev->chunk_sectors;
513 end_disk_index = (int)(end - last_stripe_index * stripe_size) /
514 mddev->chunk_sectors;
515 }
516 start_disk_offset = ((int)(start - first_stripe_index * stripe_size) %
517 mddev->chunk_sectors) +
518 first_stripe_index * mddev->chunk_sectors;
519 end_disk_offset = ((int)(end - last_stripe_index * stripe_size) %
520 mddev->chunk_sectors) +
521 last_stripe_index * mddev->chunk_sectors;
522
523 for (disk = 0; disk < zone->nb_dev; disk++) {
524 sector_t dev_start, dev_end;
525 struct md_rdev *rdev;
526 int compare_disk;
527
528 compare_disk = map_disk_shift(disk, zone->nb_dev,
529 zone->disk_shift);
530
531 if (compare_disk < start_disk_index)
532 dev_start = (first_stripe_index + 1) *
533 mddev->chunk_sectors;
534 else if (compare_disk > start_disk_index)
535 dev_start = first_stripe_index * mddev->chunk_sectors;
536 else
537 dev_start = start_disk_offset;
538
539 if (compare_disk < end_disk_index)
540 dev_end = (last_stripe_index + 1) * mddev->chunk_sectors;
541 else if (compare_disk > end_disk_index)
542 dev_end = last_stripe_index * mddev->chunk_sectors;
543 else
544 dev_end = end_disk_offset;
545
546 if (dev_end <= dev_start)
547 continue;
548
549 rdev = conf->devlist[(zone - conf->strip_zone) *
550 conf->strip_zone[0].nb_dev + disk];
551 md_submit_discard_bio(mddev, rdev, bio,
552 dev_start + zone->dev_start + rdev->data_offset,
553 dev_end - dev_start);
554 }
555 bio_endio(bio);
556 }
557
raid0_map_submit_bio(struct mddev * mddev,struct bio * bio)558 static void raid0_map_submit_bio(struct mddev *mddev, struct bio *bio)
559 {
560 struct r0conf *conf = mddev->private;
561 struct strip_zone *zone;
562 struct md_rdev *tmp_dev;
563 sector_t bio_sector = bio->bi_iter.bi_sector;
564 sector_t sector = bio_sector;
565
566 md_account_bio(mddev, &bio);
567
568 zone = find_zone(mddev->private, §or);
569 switch (conf->layout) {
570 case RAID0_ORIG_LAYOUT:
571 tmp_dev = map_sector(mddev, zone, bio_sector, §or);
572 break;
573 case RAID0_ALT_MULTIZONE_LAYOUT:
574 tmp_dev = map_sector(mddev, zone, sector, §or);
575 break;
576 default:
577 WARN(1, "md/raid0:%s: Invalid layout\n", mdname(mddev));
578 bio_io_error(bio);
579 return;
580 }
581
582 if (unlikely(is_rdev_broken(tmp_dev))) {
583 bio_io_error(bio);
584 md_error(mddev, tmp_dev);
585 return;
586 }
587
588 bio_set_dev(bio, tmp_dev->bdev);
589 bio->bi_iter.bi_sector = sector + zone->dev_start +
590 tmp_dev->data_offset;
591 mddev_trace_remap(mddev, bio, bio_sector);
592 mddev_check_write_zeroes(mddev, bio);
593 submit_bio_noacct(bio);
594 }
595
raid0_make_request(struct mddev * mddev,struct bio * bio)596 static bool raid0_make_request(struct mddev *mddev, struct bio *bio)
597 {
598 sector_t sector;
599 unsigned chunk_sects;
600 unsigned sectors;
601
602 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
603 && md_flush_request(mddev, bio))
604 return true;
605
606 if (unlikely((bio_op(bio) == REQ_OP_DISCARD))) {
607 raid0_handle_discard(mddev, bio);
608 return true;
609 }
610
611 sector = bio->bi_iter.bi_sector;
612 chunk_sects = mddev->chunk_sectors;
613
614 sectors = chunk_sects -
615 (likely(is_power_of_2(chunk_sects))
616 ? (sector & (chunk_sects-1))
617 : sector_div(sector, chunk_sects));
618
619 if (sectors < bio_sectors(bio)) {
620 bio = bio_submit_split_bioset(bio, sectors,
621 &mddev->bio_set);
622 if (!bio)
623 return true;
624 }
625
626 raid0_map_submit_bio(mddev, bio);
627 return true;
628 }
629
raid0_status(struct seq_file * seq,struct mddev * mddev)630 static void raid0_status(struct seq_file *seq, struct mddev *mddev)
631 {
632 seq_printf(seq, " %dk chunks", mddev->chunk_sectors / 2);
633 return;
634 }
635
raid0_error(struct mddev * mddev,struct md_rdev * rdev)636 static void raid0_error(struct mddev *mddev, struct md_rdev *rdev)
637 {
638 if (!test_and_set_bit(MD_BROKEN, &mddev->flags)) {
639 char *md_name = mdname(mddev);
640
641 pr_crit("md/raid0%s: Disk failure on %pg detected, failing array.\n",
642 md_name, rdev->bdev);
643 }
644 }
645
raid0_takeover_raid45(struct mddev * mddev)646 static void *raid0_takeover_raid45(struct mddev *mddev)
647 {
648 struct md_rdev *rdev;
649 struct r0conf *priv_conf;
650
651 if (mddev->degraded != 1) {
652 pr_warn("md/raid0:%s: raid5 must be degraded! Degraded disks: %d\n",
653 mdname(mddev),
654 mddev->degraded);
655 return ERR_PTR(-EINVAL);
656 }
657
658 rdev_for_each(rdev, mddev) {
659 /* check slot number for a disk */
660 if (rdev->raid_disk == mddev->raid_disks-1) {
661 pr_warn("md/raid0:%s: raid5 must have missing parity disk!\n",
662 mdname(mddev));
663 return ERR_PTR(-EINVAL);
664 }
665 rdev->sectors = mddev->dev_sectors;
666 }
667
668 /* Set new parameters */
669 mddev->new_level = 0;
670 mddev->new_layout = 0;
671 mddev->new_chunk_sectors = mddev->chunk_sectors;
672 mddev->raid_disks--;
673 mddev->delta_disks = -1;
674 /* make sure it will be not marked as dirty */
675 mddev->resync_offset = MaxSector;
676 mddev_clear_unsupported_flags(mddev, UNSUPPORTED_MDDEV_FLAGS);
677
678 create_strip_zones(mddev, &priv_conf);
679
680 return priv_conf;
681 }
682
raid0_takeover_raid10(struct mddev * mddev)683 static void *raid0_takeover_raid10(struct mddev *mddev)
684 {
685 struct r0conf *priv_conf;
686
687 /* Check layout:
688 * - far_copies must be 1
689 * - near_copies must be 2
690 * - disks number must be even
691 * - all mirrors must be already degraded
692 */
693 if (mddev->layout != ((1 << 8) + 2)) {
694 pr_warn("md/raid0:%s:: Raid0 cannot takeover layout: 0x%x\n",
695 mdname(mddev),
696 mddev->layout);
697 return ERR_PTR(-EINVAL);
698 }
699 if (mddev->raid_disks & 1) {
700 pr_warn("md/raid0:%s: Raid0 cannot takeover Raid10 with odd disk number.\n",
701 mdname(mddev));
702 return ERR_PTR(-EINVAL);
703 }
704 if (mddev->degraded != (mddev->raid_disks>>1)) {
705 pr_warn("md/raid0:%s: All mirrors must be already degraded!\n",
706 mdname(mddev));
707 return ERR_PTR(-EINVAL);
708 }
709
710 /* Set new parameters */
711 mddev->new_level = 0;
712 mddev->new_layout = 0;
713 mddev->new_chunk_sectors = mddev->chunk_sectors;
714 mddev->delta_disks = - mddev->raid_disks / 2;
715 mddev->raid_disks += mddev->delta_disks;
716 mddev->degraded = 0;
717 /* make sure it will be not marked as dirty */
718 mddev->resync_offset = MaxSector;
719 mddev_clear_unsupported_flags(mddev, UNSUPPORTED_MDDEV_FLAGS);
720
721 create_strip_zones(mddev, &priv_conf);
722 return priv_conf;
723 }
724
raid0_takeover_raid1(struct mddev * mddev)725 static void *raid0_takeover_raid1(struct mddev *mddev)
726 {
727 struct r0conf *priv_conf;
728 int chunksect;
729
730 /* Check layout:
731 * - (N - 1) mirror drives must be already faulty
732 */
733 if ((mddev->raid_disks - 1) != mddev->degraded) {
734 pr_err("md/raid0:%s: (N - 1) mirrors drives must be already faulty!\n",
735 mdname(mddev));
736 return ERR_PTR(-EINVAL);
737 }
738
739 /*
740 * a raid1 doesn't have the notion of chunk size, so
741 * figure out the largest suitable size we can use.
742 */
743 chunksect = 64 * 2; /* 64K by default */
744
745 /* The array must be an exact multiple of chunksize */
746 while (chunksect && (mddev->array_sectors & (chunksect - 1)))
747 chunksect >>= 1;
748
749 if ((chunksect << 9) < PAGE_SIZE)
750 /* array size does not allow a suitable chunk size */
751 return ERR_PTR(-EINVAL);
752
753 /* Set new parameters */
754 mddev->new_level = 0;
755 mddev->new_layout = 0;
756 mddev->new_chunk_sectors = chunksect;
757 mddev->chunk_sectors = chunksect;
758 mddev->delta_disks = 1 - mddev->raid_disks;
759 mddev->raid_disks = 1;
760 /* make sure it will be not marked as dirty */
761 mddev->resync_offset = MaxSector;
762 mddev_clear_unsupported_flags(mddev, UNSUPPORTED_MDDEV_FLAGS);
763
764 create_strip_zones(mddev, &priv_conf);
765 return priv_conf;
766 }
767
raid0_takeover(struct mddev * mddev)768 static void *raid0_takeover(struct mddev *mddev)
769 {
770 /* raid0 can take over:
771 * raid4 - if all data disks are active.
772 * raid5 - providing it is Raid4 layout and one disk is faulty
773 * raid10 - assuming we have all necessary active disks
774 * raid1 - with (N -1) mirror drives faulty
775 */
776
777 if (mddev->bitmap) {
778 pr_warn("md/raid0: %s: cannot takeover array with bitmap\n",
779 mdname(mddev));
780 return ERR_PTR(-EBUSY);
781 }
782 if (mddev->level == 4)
783 return raid0_takeover_raid45(mddev);
784
785 if (mddev->level == 5) {
786 if (mddev->layout == ALGORITHM_PARITY_N)
787 return raid0_takeover_raid45(mddev);
788
789 pr_warn("md/raid0:%s: Raid can only takeover Raid5 with layout: %d\n",
790 mdname(mddev), ALGORITHM_PARITY_N);
791 }
792
793 if (mddev->level == 10)
794 return raid0_takeover_raid10(mddev);
795
796 if (mddev->level == 1)
797 return raid0_takeover_raid1(mddev);
798
799 pr_warn("Takeover from raid%i to raid0 not supported\n",
800 mddev->level);
801
802 return ERR_PTR(-EINVAL);
803 }
804
raid0_quiesce(struct mddev * mddev,int quiesce)805 static void raid0_quiesce(struct mddev *mddev, int quiesce)
806 {
807 }
808
809 static struct md_personality raid0_personality=
810 {
811 .head = {
812 .type = MD_PERSONALITY,
813 .id = ID_RAID0,
814 .name = "raid0",
815 .owner = THIS_MODULE,
816 },
817
818 .make_request = raid0_make_request,
819 .run = raid0_run,
820 .free = raid0_free,
821 .status = raid0_status,
822 .size = raid0_size,
823 .takeover = raid0_takeover,
824 .quiesce = raid0_quiesce,
825 .error_handler = raid0_error,
826 };
827
raid0_init(void)828 static int __init raid0_init(void)
829 {
830 return register_md_submodule(&raid0_personality.head);
831 }
832
raid0_exit(void)833 static void __exit raid0_exit(void)
834 {
835 unregister_md_submodule(&raid0_personality.head);
836 }
837
838 module_init(raid0_init);
839 module_exit(raid0_exit);
840 MODULE_LICENSE("GPL");
841 MODULE_DESCRIPTION("RAID0 (striping) personality for MD");
842 MODULE_ALIAS("md-personality-2"); /* RAID0 */
843 MODULE_ALIAS("md-raid0");
844 MODULE_ALIAS("md-level-0");
845