dm raid: handle failed devices during start up
[~shefty/rdma-dev.git] / drivers / md / dm-raid.c
1 /*
2  * Copyright (C) 2010-2011 Neil Brown
3  * Copyright (C) 2010-2011 Red Hat, Inc. All rights reserved.
4  *
5  * This file is released under the GPL.
6  */
7
8 #include <linux/slab.h>
9 #include <linux/module.h>
10
11 #include "md.h"
12 #include "raid1.h"
13 #include "raid5.h"
14 #include "bitmap.h"
15
16 #include <linux/device-mapper.h>
17
18 #define DM_MSG_PREFIX "raid"
19
20 /*
21  * The following flags are used by dm-raid.c to set up the array state.
22  * They must be cleared before md_run is called.
23  */
24 #define FirstUse 10             /* rdev flag */
25
26 struct raid_dev {
27         /*
28          * Two DM devices, one to hold metadata and one to hold the
29          * actual data/parity.  The reason for this is to not confuse
30          * ti->len and give more flexibility in altering size and
31          * characteristics.
32          *
33          * While it is possible for this device to be associated
34          * with a different physical device than the data_dev, it
35          * is intended for it to be the same.
36          *    |--------- Physical Device ---------|
37          *    |- meta_dev -|------ data_dev ------|
38          */
39         struct dm_dev *meta_dev;
40         struct dm_dev *data_dev;
41         struct md_rdev rdev;
42 };
43
44 /*
45  * Flags for rs->print_flags field.
46  */
47 #define DMPF_SYNC              0x1
48 #define DMPF_NOSYNC            0x2
49 #define DMPF_REBUILD           0x4
50 #define DMPF_DAEMON_SLEEP      0x8
51 #define DMPF_MIN_RECOVERY_RATE 0x10
52 #define DMPF_MAX_RECOVERY_RATE 0x20
53 #define DMPF_MAX_WRITE_BEHIND  0x40
54 #define DMPF_STRIPE_CACHE      0x80
55 #define DMPF_REGION_SIZE       0X100
56 struct raid_set {
57         struct dm_target *ti;
58
59         uint32_t bitmap_loaded;
60         uint32_t print_flags;
61
62         struct mddev md;
63         struct raid_type *raid_type;
64         struct dm_target_callbacks callbacks;
65
66         struct raid_dev dev[0];
67 };
68
69 /* Supported raid types and properties. */
70 static struct raid_type {
71         const char *name;               /* RAID algorithm. */
72         const char *descr;              /* Descriptor text for logging. */
73         const unsigned parity_devs;     /* # of parity devices. */
74         const unsigned minimal_devs;    /* minimal # of devices in set. */
75         const unsigned level;           /* RAID level. */
76         const unsigned algorithm;       /* RAID algorithm. */
77 } raid_types[] = {
78         {"raid1",    "RAID1 (mirroring)",               0, 2, 1, 0 /* NONE */},
79         {"raid4",    "RAID4 (dedicated parity disk)",   1, 2, 5, ALGORITHM_PARITY_0},
80         {"raid5_la", "RAID5 (left asymmetric)",         1, 2, 5, ALGORITHM_LEFT_ASYMMETRIC},
81         {"raid5_ra", "RAID5 (right asymmetric)",        1, 2, 5, ALGORITHM_RIGHT_ASYMMETRIC},
82         {"raid5_ls", "RAID5 (left symmetric)",          1, 2, 5, ALGORITHM_LEFT_SYMMETRIC},
83         {"raid5_rs", "RAID5 (right symmetric)",         1, 2, 5, ALGORITHM_RIGHT_SYMMETRIC},
84         {"raid6_zr", "RAID6 (zero restart)",            2, 4, 6, ALGORITHM_ROTATING_ZERO_RESTART},
85         {"raid6_nr", "RAID6 (N restart)",               2, 4, 6, ALGORITHM_ROTATING_N_RESTART},
86         {"raid6_nc", "RAID6 (N continue)",              2, 4, 6, ALGORITHM_ROTATING_N_CONTINUE}
87 };
88
89 static struct raid_type *get_raid_type(char *name)
90 {
91         int i;
92
93         for (i = 0; i < ARRAY_SIZE(raid_types); i++)
94                 if (!strcmp(raid_types[i].name, name))
95                         return &raid_types[i];
96
97         return NULL;
98 }
99
100 static struct raid_set *context_alloc(struct dm_target *ti, struct raid_type *raid_type, unsigned raid_devs)
101 {
102         unsigned i;
103         struct raid_set *rs;
104         sector_t sectors_per_dev;
105
106         if (raid_devs <= raid_type->parity_devs) {
107                 ti->error = "Insufficient number of devices";
108                 return ERR_PTR(-EINVAL);
109         }
110
111         sectors_per_dev = ti->len;
112         if ((raid_type->level > 1) &&
113             sector_div(sectors_per_dev, (raid_devs - raid_type->parity_devs))) {
114                 ti->error = "Target length not divisible by number of data devices";
115                 return ERR_PTR(-EINVAL);
116         }
117
118         rs = kzalloc(sizeof(*rs) + raid_devs * sizeof(rs->dev[0]), GFP_KERNEL);
119         if (!rs) {
120                 ti->error = "Cannot allocate raid context";
121                 return ERR_PTR(-ENOMEM);
122         }
123
124         mddev_init(&rs->md);
125
126         rs->ti = ti;
127         rs->raid_type = raid_type;
128         rs->md.raid_disks = raid_devs;
129         rs->md.level = raid_type->level;
130         rs->md.new_level = rs->md.level;
131         rs->md.dev_sectors = sectors_per_dev;
132         rs->md.layout = raid_type->algorithm;
133         rs->md.new_layout = rs->md.layout;
134         rs->md.delta_disks = 0;
135         rs->md.recovery_cp = 0;
136
137         for (i = 0; i < raid_devs; i++)
138                 md_rdev_init(&rs->dev[i].rdev);
139
140         /*
141          * Remaining items to be initialized by further RAID params:
142          *  rs->md.persistent
143          *  rs->md.external
144          *  rs->md.chunk_sectors
145          *  rs->md.new_chunk_sectors
146          */
147
148         return rs;
149 }
150
151 static void context_free(struct raid_set *rs)
152 {
153         int i;
154
155         for (i = 0; i < rs->md.raid_disks; i++) {
156                 if (rs->dev[i].meta_dev)
157                         dm_put_device(rs->ti, rs->dev[i].meta_dev);
158                 if (rs->dev[i].rdev.sb_page)
159                         put_page(rs->dev[i].rdev.sb_page);
160                 rs->dev[i].rdev.sb_page = NULL;
161                 rs->dev[i].rdev.sb_loaded = 0;
162                 if (rs->dev[i].data_dev)
163                         dm_put_device(rs->ti, rs->dev[i].data_dev);
164         }
165
166         kfree(rs);
167 }
168
169 /*
170  * For every device we have two words
171  *  <meta_dev>: meta device name or '-' if missing
172  *  <data_dev>: data device name or '-' if missing
173  *
174  * The following are permitted:
175  *    - -
176  *    - <data_dev>
177  *    <meta_dev> <data_dev>
178  *
179  * The following is not allowed:
180  *    <meta_dev> -
181  *
182  * This code parses those words.  If there is a failure,
183  * the caller must use context_free to unwind the operations.
184  */
185 static int dev_parms(struct raid_set *rs, char **argv)
186 {
187         int i;
188         int rebuild = 0;
189         int metadata_available = 0;
190         int ret = 0;
191
192         for (i = 0; i < rs->md.raid_disks; i++, argv += 2) {
193                 rs->dev[i].rdev.raid_disk = i;
194
195                 rs->dev[i].meta_dev = NULL;
196                 rs->dev[i].data_dev = NULL;
197
198                 /*
199                  * There are no offsets, since there is a separate device
200                  * for data and metadata.
201                  */
202                 rs->dev[i].rdev.data_offset = 0;
203                 rs->dev[i].rdev.mddev = &rs->md;
204
205                 if (strcmp(argv[0], "-")) {
206                         ret = dm_get_device(rs->ti, argv[0],
207                                             dm_table_get_mode(rs->ti->table),
208                                             &rs->dev[i].meta_dev);
209                         rs->ti->error = "RAID metadata device lookup failure";
210                         if (ret)
211                                 return ret;
212
213                         rs->dev[i].rdev.sb_page = alloc_page(GFP_KERNEL);
214                         if (!rs->dev[i].rdev.sb_page)
215                                 return -ENOMEM;
216                 }
217
218                 if (!strcmp(argv[1], "-")) {
219                         if (!test_bit(In_sync, &rs->dev[i].rdev.flags) &&
220                             (!rs->dev[i].rdev.recovery_offset)) {
221                                 rs->ti->error = "Drive designated for rebuild not specified";
222                                 return -EINVAL;
223                         }
224
225                         rs->ti->error = "No data device supplied with metadata device";
226                         if (rs->dev[i].meta_dev)
227                                 return -EINVAL;
228
229                         continue;
230                 }
231
232                 ret = dm_get_device(rs->ti, argv[1],
233                                     dm_table_get_mode(rs->ti->table),
234                                     &rs->dev[i].data_dev);
235                 if (ret) {
236                         rs->ti->error = "RAID device lookup failure";
237                         return ret;
238                 }
239
240                 if (rs->dev[i].meta_dev) {
241                         metadata_available = 1;
242                         rs->dev[i].rdev.meta_bdev = rs->dev[i].meta_dev->bdev;
243                 }
244                 rs->dev[i].rdev.bdev = rs->dev[i].data_dev->bdev;
245                 list_add(&rs->dev[i].rdev.same_set, &rs->md.disks);
246                 if (!test_bit(In_sync, &rs->dev[i].rdev.flags))
247                         rebuild++;
248         }
249
250         if (metadata_available) {
251                 rs->md.external = 0;
252                 rs->md.persistent = 1;
253                 rs->md.major_version = 2;
254         } else if (rebuild && !rs->md.recovery_cp) {
255                 /*
256                  * Without metadata, we will not be able to tell if the array
257                  * is in-sync or not - we must assume it is not.  Therefore,
258                  * it is impossible to rebuild a drive.
259                  *
260                  * Even if there is metadata, the on-disk information may
261                  * indicate that the array is not in-sync and it will then
262                  * fail at that time.
263                  *
264                  * User could specify 'nosync' option if desperate.
265                  */
266                 DMERR("Unable to rebuild drive while array is not in-sync");
267                 rs->ti->error = "RAID device lookup failure";
268                 return -EINVAL;
269         }
270
271         return 0;
272 }
273
274 /*
275  * validate_region_size
276  * @rs
277  * @region_size:  region size in sectors.  If 0, pick a size (4MiB default).
278  *
279  * Set rs->md.bitmap_info.chunksize (which really refers to 'region size').
280  * Ensure that (ti->len/region_size < 2^21) - required by MD bitmap.
281  *
282  * Returns: 0 on success, -EINVAL on failure.
283  */
284 static int validate_region_size(struct raid_set *rs, unsigned long region_size)
285 {
286         unsigned long min_region_size = rs->ti->len / (1 << 21);
287
288         if (!region_size) {
289                 /*
290                  * Choose a reasonable default.  All figures in sectors.
291                  */
292                 if (min_region_size > (1 << 13)) {
293                         DMINFO("Choosing default region size of %lu sectors",
294                                region_size);
295                         region_size = min_region_size;
296                 } else {
297                         DMINFO("Choosing default region size of 4MiB");
298                         region_size = 1 << 13; /* sectors */
299                 }
300         } else {
301                 /*
302                  * Validate user-supplied value.
303                  */
304                 if (region_size > rs->ti->len) {
305                         rs->ti->error = "Supplied region size is too large";
306                         return -EINVAL;
307                 }
308
309                 if (region_size < min_region_size) {
310                         DMERR("Supplied region_size (%lu sectors) below minimum (%lu)",
311                               region_size, min_region_size);
312                         rs->ti->error = "Supplied region size is too small";
313                         return -EINVAL;
314                 }
315
316                 if (!is_power_of_2(region_size)) {
317                         rs->ti->error = "Region size is not a power of 2";
318                         return -EINVAL;
319                 }
320
321                 if (region_size < rs->md.chunk_sectors) {
322                         rs->ti->error = "Region size is smaller than the chunk size";
323                         return -EINVAL;
324                 }
325         }
326
327         /*
328          * Convert sectors to bytes.
329          */
330         rs->md.bitmap_info.chunksize = (region_size << 9);
331
332         return 0;
333 }
334
335 /*
336  * Possible arguments are...
337  *      <chunk_size> [optional_args]
338  *
339  * Argument definitions
340  *    <chunk_size>                      The number of sectors per disk that
341  *                                      will form the "stripe"
342  *    [[no]sync]                        Force or prevent recovery of the
343  *                                      entire array
344  *    [rebuild <idx>]                   Rebuild the drive indicated by the index
345  *    [daemon_sleep <ms>]               Time between bitmap daemon work to
346  *                                      clear bits
347  *    [min_recovery_rate <kB/sec/disk>] Throttle RAID initialization
348  *    [max_recovery_rate <kB/sec/disk>] Throttle RAID initialization
349  *    [write_mostly <idx>]              Indicate a write mostly drive via index
350  *    [max_write_behind <sectors>]      See '-write-behind=' (man mdadm)
351  *    [stripe_cache <sectors>]          Stripe cache size for higher RAIDs
352  *    [region_size <sectors>]           Defines granularity of bitmap
353  */
354 static int parse_raid_params(struct raid_set *rs, char **argv,
355                              unsigned num_raid_params)
356 {
357         unsigned i, rebuild_cnt = 0;
358         unsigned long value, region_size = 0;
359         char *key;
360
361         /*
362          * First, parse the in-order required arguments
363          * "chunk_size" is the only argument of this type.
364          */
365         if ((strict_strtoul(argv[0], 10, &value) < 0)) {
366                 rs->ti->error = "Bad chunk size";
367                 return -EINVAL;
368         } else if (rs->raid_type->level == 1) {
369                 if (value)
370                         DMERR("Ignoring chunk size parameter for RAID 1");
371                 value = 0;
372         } else if (!is_power_of_2(value)) {
373                 rs->ti->error = "Chunk size must be a power of 2";
374                 return -EINVAL;
375         } else if (value < 8) {
376                 rs->ti->error = "Chunk size value is too small";
377                 return -EINVAL;
378         }
379
380         rs->md.new_chunk_sectors = rs->md.chunk_sectors = value;
381         argv++;
382         num_raid_params--;
383
384         /*
385          * We set each individual device as In_sync with a completed
386          * 'recovery_offset'.  If there has been a device failure or
387          * replacement then one of the following cases applies:
388          *
389          *   1) User specifies 'rebuild'.
390          *      - Device is reset when param is read.
391          *   2) A new device is supplied.
392          *      - No matching superblock found, resets device.
393          *   3) Device failure was transient and returns on reload.
394          *      - Failure noticed, resets device for bitmap replay.
395          *   4) Device hadn't completed recovery after previous failure.
396          *      - Superblock is read and overrides recovery_offset.
397          *
398          * What is found in the superblocks of the devices is always
399          * authoritative, unless 'rebuild' or '[no]sync' was specified.
400          */
401         for (i = 0; i < rs->md.raid_disks; i++) {
402                 set_bit(In_sync, &rs->dev[i].rdev.flags);
403                 rs->dev[i].rdev.recovery_offset = MaxSector;
404         }
405
406         /*
407          * Second, parse the unordered optional arguments
408          */
409         for (i = 0; i < num_raid_params; i++) {
410                 if (!strcasecmp(argv[i], "nosync")) {
411                         rs->md.recovery_cp = MaxSector;
412                         rs->print_flags |= DMPF_NOSYNC;
413                         continue;
414                 }
415                 if (!strcasecmp(argv[i], "sync")) {
416                         rs->md.recovery_cp = 0;
417                         rs->print_flags |= DMPF_SYNC;
418                         continue;
419                 }
420
421                 /* The rest of the optional arguments come in key/value pairs */
422                 if ((i + 1) >= num_raid_params) {
423                         rs->ti->error = "Wrong number of raid parameters given";
424                         return -EINVAL;
425                 }
426
427                 key = argv[i++];
428                 if (strict_strtoul(argv[i], 10, &value) < 0) {
429                         rs->ti->error = "Bad numerical argument given in raid params";
430                         return -EINVAL;
431                 }
432
433                 if (!strcasecmp(key, "rebuild")) {
434                         rebuild_cnt++;
435                         if (((rs->raid_type->level != 1) &&
436                              (rebuild_cnt > rs->raid_type->parity_devs)) ||
437                             ((rs->raid_type->level == 1) &&
438                              (rebuild_cnt > (rs->md.raid_disks - 1)))) {
439                                 rs->ti->error = "Too many rebuild devices specified for given RAID type";
440                                 return -EINVAL;
441                         }
442                         if (value > rs->md.raid_disks) {
443                                 rs->ti->error = "Invalid rebuild index given";
444                                 return -EINVAL;
445                         }
446                         clear_bit(In_sync, &rs->dev[value].rdev.flags);
447                         rs->dev[value].rdev.recovery_offset = 0;
448                         rs->print_flags |= DMPF_REBUILD;
449                 } else if (!strcasecmp(key, "write_mostly")) {
450                         if (rs->raid_type->level != 1) {
451                                 rs->ti->error = "write_mostly option is only valid for RAID1";
452                                 return -EINVAL;
453                         }
454                         if (value >= rs->md.raid_disks) {
455                                 rs->ti->error = "Invalid write_mostly drive index given";
456                                 return -EINVAL;
457                         }
458                         set_bit(WriteMostly, &rs->dev[value].rdev.flags);
459                 } else if (!strcasecmp(key, "max_write_behind")) {
460                         if (rs->raid_type->level != 1) {
461                                 rs->ti->error = "max_write_behind option is only valid for RAID1";
462                                 return -EINVAL;
463                         }
464                         rs->print_flags |= DMPF_MAX_WRITE_BEHIND;
465
466                         /*
467                          * In device-mapper, we specify things in sectors, but
468                          * MD records this value in kB
469                          */
470                         value /= 2;
471                         if (value > COUNTER_MAX) {
472                                 rs->ti->error = "Max write-behind limit out of range";
473                                 return -EINVAL;
474                         }
475                         rs->md.bitmap_info.max_write_behind = value;
476                 } else if (!strcasecmp(key, "daemon_sleep")) {
477                         rs->print_flags |= DMPF_DAEMON_SLEEP;
478                         if (!value || (value > MAX_SCHEDULE_TIMEOUT)) {
479                                 rs->ti->error = "daemon sleep period out of range";
480                                 return -EINVAL;
481                         }
482                         rs->md.bitmap_info.daemon_sleep = value;
483                 } else if (!strcasecmp(key, "stripe_cache")) {
484                         rs->print_flags |= DMPF_STRIPE_CACHE;
485
486                         /*
487                          * In device-mapper, we specify things in sectors, but
488                          * MD records this value in kB
489                          */
490                         value /= 2;
491
492                         if (rs->raid_type->level < 5) {
493                                 rs->ti->error = "Inappropriate argument: stripe_cache";
494                                 return -EINVAL;
495                         }
496                         if (raid5_set_cache_size(&rs->md, (int)value)) {
497                                 rs->ti->error = "Bad stripe_cache size";
498                                 return -EINVAL;
499                         }
500                 } else if (!strcasecmp(key, "min_recovery_rate")) {
501                         rs->print_flags |= DMPF_MIN_RECOVERY_RATE;
502                         if (value > INT_MAX) {
503                                 rs->ti->error = "min_recovery_rate out of range";
504                                 return -EINVAL;
505                         }
506                         rs->md.sync_speed_min = (int)value;
507                 } else if (!strcasecmp(key, "max_recovery_rate")) {
508                         rs->print_flags |= DMPF_MAX_RECOVERY_RATE;
509                         if (value > INT_MAX) {
510                                 rs->ti->error = "max_recovery_rate out of range";
511                                 return -EINVAL;
512                         }
513                         rs->md.sync_speed_max = (int)value;
514                 } else if (!strcasecmp(key, "region_size")) {
515                         rs->print_flags |= DMPF_REGION_SIZE;
516                         region_size = value;
517                 } else {
518                         DMERR("Unable to parse RAID parameter: %s", key);
519                         rs->ti->error = "Unable to parse RAID parameters";
520                         return -EINVAL;
521                 }
522         }
523
524         if (validate_region_size(rs, region_size))
525                 return -EINVAL;
526
527         if (rs->md.chunk_sectors)
528                 rs->ti->split_io = rs->md.chunk_sectors;
529         else
530                 rs->ti->split_io = region_size;
531
532         if (rs->md.chunk_sectors)
533                 rs->ti->split_io = rs->md.chunk_sectors;
534         else
535                 rs->ti->split_io = region_size;
536
537         /* Assume there are no metadata devices until the drives are parsed */
538         rs->md.persistent = 0;
539         rs->md.external = 1;
540
541         return 0;
542 }
543
544 static void do_table_event(struct work_struct *ws)
545 {
546         struct raid_set *rs = container_of(ws, struct raid_set, md.event_work);
547
548         dm_table_event(rs->ti->table);
549 }
550
551 static int raid_is_congested(struct dm_target_callbacks *cb, int bits)
552 {
553         struct raid_set *rs = container_of(cb, struct raid_set, callbacks);
554
555         if (rs->raid_type->level == 1)
556                 return md_raid1_congested(&rs->md, bits);
557
558         return md_raid5_congested(&rs->md, bits);
559 }
560
561 /*
562  * This structure is never routinely used by userspace, unlike md superblocks.
563  * Devices with this superblock should only ever be accessed via device-mapper.
564  */
565 #define DM_RAID_MAGIC 0x64526D44
566 struct dm_raid_superblock {
567         __le32 magic;           /* "DmRd" */
568         __le32 features;        /* Used to indicate possible future changes */
569
570         __le32 num_devices;     /* Number of devices in this array. (Max 64) */
571         __le32 array_position;  /* The position of this drive in the array */
572
573         __le64 events;          /* Incremented by md when superblock updated */
574         __le64 failed_devices;  /* Bit field of devices to indicate failures */
575
576         /*
577          * This offset tracks the progress of the repair or replacement of
578          * an individual drive.
579          */
580         __le64 disk_recovery_offset;
581
582         /*
583          * This offset tracks the progress of the initial array
584          * synchronisation/parity calculation.
585          */
586         __le64 array_resync_offset;
587
588         /*
589          * RAID characteristics
590          */
591         __le32 level;
592         __le32 layout;
593         __le32 stripe_sectors;
594
595         __u8 pad[452];          /* Round struct to 512 bytes. */
596                                 /* Always set to 0 when writing. */
597 } __packed;
598
599 static int read_disk_sb(struct md_rdev *rdev, int size)
600 {
601         BUG_ON(!rdev->sb_page);
602
603         if (rdev->sb_loaded)
604                 return 0;
605
606         if (!sync_page_io(rdev, 0, size, rdev->sb_page, READ, 1)) {
607                 DMERR("Failed to read superblock of device at position %d",
608                       rdev->raid_disk);
609                 set_bit(Faulty, &rdev->flags);
610                 return -EINVAL;
611         }
612
613         rdev->sb_loaded = 1;
614
615         return 0;
616 }
617
618 static void super_sync(struct mddev *mddev, struct md_rdev *rdev)
619 {
620         struct md_rdev *r;
621         uint64_t failed_devices;
622         struct dm_raid_superblock *sb;
623
624         sb = page_address(rdev->sb_page);
625         failed_devices = le64_to_cpu(sb->failed_devices);
626
627         rdev_for_each(r, mddev)
628                 if ((r->raid_disk >= 0) && test_bit(Faulty, &r->flags))
629                         failed_devices |= (1ULL << r->raid_disk);
630
631         memset(sb, 0, sizeof(*sb));
632
633         sb->magic = cpu_to_le32(DM_RAID_MAGIC);
634         sb->features = cpu_to_le32(0);  /* No features yet */
635
636         sb->num_devices = cpu_to_le32(mddev->raid_disks);
637         sb->array_position = cpu_to_le32(rdev->raid_disk);
638
639         sb->events = cpu_to_le64(mddev->events);
640         sb->failed_devices = cpu_to_le64(failed_devices);
641
642         sb->disk_recovery_offset = cpu_to_le64(rdev->recovery_offset);
643         sb->array_resync_offset = cpu_to_le64(mddev->recovery_cp);
644
645         sb->level = cpu_to_le32(mddev->level);
646         sb->layout = cpu_to_le32(mddev->layout);
647         sb->stripe_sectors = cpu_to_le32(mddev->chunk_sectors);
648 }
649
650 /*
651  * super_load
652  *
653  * This function creates a superblock if one is not found on the device
654  * and will decide which superblock to use if there's a choice.
655  *
656  * Return: 1 if use rdev, 0 if use refdev, -Exxx otherwise
657  */
658 static int super_load(struct md_rdev *rdev, struct md_rdev *refdev)
659 {
660         int ret;
661         struct dm_raid_superblock *sb;
662         struct dm_raid_superblock *refsb;
663         uint64_t events_sb, events_refsb;
664
665         rdev->sb_start = 0;
666         rdev->sb_size = sizeof(*sb);
667
668         ret = read_disk_sb(rdev, rdev->sb_size);
669         if (ret)
670                 return ret;
671
672         sb = page_address(rdev->sb_page);
673
674         /*
675          * Two cases that we want to write new superblocks and rebuild:
676          * 1) New device (no matching magic number)
677          * 2) Device specified for rebuild (!In_sync w/ offset == 0)
678          */
679         if ((sb->magic != cpu_to_le32(DM_RAID_MAGIC)) ||
680             (!test_bit(In_sync, &rdev->flags) && !rdev->recovery_offset)) {
681                 super_sync(rdev->mddev, rdev);
682
683                 set_bit(FirstUse, &rdev->flags);
684
685                 /* Force writing of superblocks to disk */
686                 set_bit(MD_CHANGE_DEVS, &rdev->mddev->flags);
687
688                 /* Any superblock is better than none, choose that if given */
689                 return refdev ? 0 : 1;
690         }
691
692         if (!refdev)
693                 return 1;
694
695         events_sb = le64_to_cpu(sb->events);
696
697         refsb = page_address(refdev->sb_page);
698         events_refsb = le64_to_cpu(refsb->events);
699
700         return (events_sb > events_refsb) ? 1 : 0;
701 }
702
703 static int super_init_validation(struct mddev *mddev, struct md_rdev *rdev)
704 {
705         int role;
706         struct raid_set *rs = container_of(mddev, struct raid_set, md);
707         uint64_t events_sb;
708         uint64_t failed_devices;
709         struct dm_raid_superblock *sb;
710         uint32_t new_devs = 0;
711         uint32_t rebuilds = 0;
712         struct md_rdev *r;
713         struct dm_raid_superblock *sb2;
714
715         sb = page_address(rdev->sb_page);
716         events_sb = le64_to_cpu(sb->events);
717         failed_devices = le64_to_cpu(sb->failed_devices);
718
719         /*
720          * Initialise to 1 if this is a new superblock.
721          */
722         mddev->events = events_sb ? : 1;
723
724         /*
725          * Reshaping is not currently allowed
726          */
727         if ((le32_to_cpu(sb->level) != mddev->level) ||
728             (le32_to_cpu(sb->layout) != mddev->layout) ||
729             (le32_to_cpu(sb->stripe_sectors) != mddev->chunk_sectors)) {
730                 DMERR("Reshaping arrays not yet supported.");
731                 return -EINVAL;
732         }
733
734         /* We can only change the number of devices in RAID1 right now */
735         if ((rs->raid_type->level != 1) &&
736             (le32_to_cpu(sb->num_devices) != mddev->raid_disks)) {
737                 DMERR("Reshaping arrays not yet supported.");
738                 return -EINVAL;
739         }
740
741         if (!(rs->print_flags & (DMPF_SYNC | DMPF_NOSYNC)))
742                 mddev->recovery_cp = le64_to_cpu(sb->array_resync_offset);
743
744         /*
745          * During load, we set FirstUse if a new superblock was written.
746          * There are two reasons we might not have a superblock:
747          * 1) The array is brand new - in which case, all of the
748          *    devices must have their In_sync bit set.  Also,
749          *    recovery_cp must be 0, unless forced.
750          * 2) This is a new device being added to an old array
751          *    and the new device needs to be rebuilt - in which
752          *    case the In_sync bit will /not/ be set and
753          *    recovery_cp must be MaxSector.
754          */
755         rdev_for_each(r, mddev) {
756                 if (!test_bit(In_sync, &r->flags)) {
757                         DMINFO("Device %d specified for rebuild: "
758                                "Clearing superblock", r->raid_disk);
759                         rebuilds++;
760                 } else if (test_bit(FirstUse, &r->flags))
761                         new_devs++;
762         }
763
764         if (!rebuilds) {
765                 if (new_devs == mddev->raid_disks) {
766                         DMINFO("Superblocks created for new array");
767                         set_bit(MD_ARRAY_FIRST_USE, &mddev->flags);
768                 } else if (new_devs) {
769                         DMERR("New device injected "
770                               "into existing array without 'rebuild' "
771                               "parameter specified");
772                         return -EINVAL;
773                 }
774         } else if (new_devs) {
775                 DMERR("'rebuild' devices cannot be "
776                       "injected into an array with other first-time devices");
777                 return -EINVAL;
778         } else if (mddev->recovery_cp != MaxSector) {
779                 DMERR("'rebuild' specified while array is not in-sync");
780                 return -EINVAL;
781         }
782
783         /*
784          * Now we set the Faulty bit for those devices that are
785          * recorded in the superblock as failed.
786          */
787         rdev_for_each(r, mddev) {
788                 if (!r->sb_page)
789                         continue;
790                 sb2 = page_address(r->sb_page);
791                 sb2->failed_devices = 0;
792
793                 /*
794                  * Check for any device re-ordering.
795                  */
796                 if (!test_bit(FirstUse, &r->flags) && (r->raid_disk >= 0)) {
797                         role = le32_to_cpu(sb2->array_position);
798                         if (role != r->raid_disk) {
799                                 if (rs->raid_type->level != 1) {
800                                         rs->ti->error = "Cannot change device "
801                                                 "positions in RAID array";
802                                         return -EINVAL;
803                                 }
804                                 DMINFO("RAID1 device #%d now at position #%d",
805                                        role, r->raid_disk);
806                         }
807
808                         /*
809                          * Partial recovery is performed on
810                          * returning failed devices.
811                          */
812                         if (failed_devices & (1 << role))
813                                 set_bit(Faulty, &r->flags);
814                 }
815         }
816
817         return 0;
818 }
819
820 static int super_validate(struct mddev *mddev, struct md_rdev *rdev)
821 {
822         struct dm_raid_superblock *sb = page_address(rdev->sb_page);
823
824         /*
825          * If mddev->events is not set, we know we have not yet initialized
826          * the array.
827          */
828         if (!mddev->events && super_init_validation(mddev, rdev))
829                 return -EINVAL;
830
831         mddev->bitmap_info.offset = 4096 >> 9; /* Enable bitmap creation */
832         rdev->mddev->bitmap_info.default_offset = 4096 >> 9;
833         if (!test_bit(FirstUse, &rdev->flags)) {
834                 rdev->recovery_offset = le64_to_cpu(sb->disk_recovery_offset);
835                 if (rdev->recovery_offset != MaxSector)
836                         clear_bit(In_sync, &rdev->flags);
837         }
838
839         /*
840          * If a device comes back, set it as not In_sync and no longer faulty.
841          */
842         if (test_bit(Faulty, &rdev->flags)) {
843                 clear_bit(Faulty, &rdev->flags);
844                 clear_bit(In_sync, &rdev->flags);
845                 rdev->saved_raid_disk = rdev->raid_disk;
846                 rdev->recovery_offset = 0;
847         }
848
849         clear_bit(FirstUse, &rdev->flags);
850
851         return 0;
852 }
853
854 /*
855  * Analyse superblocks and select the freshest.
856  */
857 static int analyse_superblocks(struct dm_target *ti, struct raid_set *rs)
858 {
859         int ret;
860         unsigned redundancy = 0;
861         struct raid_dev *dev;
862         struct md_rdev *rdev, *freshest;
863         struct mddev *mddev = &rs->md;
864
865         switch (rs->raid_type->level) {
866         case 1:
867                 redundancy = rs->md.raid_disks - 1;
868                 break;
869         case 4:
870         case 5:
871         case 6:
872                 redundancy = rs->raid_type->parity_devs;
873                 break;
874         default:
875                 ti->error = "Unknown RAID type";
876                 return -EINVAL;
877         }
878
879         freshest = NULL;
880         rdev_for_each(rdev, mddev) {
881                 if (!rdev->meta_bdev)
882                         continue;
883
884                 ret = super_load(rdev, freshest);
885
886                 switch (ret) {
887                 case 1:
888                         freshest = rdev;
889                         break;
890                 case 0:
891                         break;
892                 default:
893                         dev = container_of(rdev, struct raid_dev, rdev);
894                         if (redundancy--) {
895                                 if (dev->meta_dev)
896                                         dm_put_device(ti, dev->meta_dev);
897
898                                 dev->meta_dev = NULL;
899                                 rdev->meta_bdev = NULL;
900
901                                 if (rdev->sb_page)
902                                         put_page(rdev->sb_page);
903
904                                 rdev->sb_page = NULL;
905
906                                 rdev->sb_loaded = 0;
907
908                                 /*
909                                  * We might be able to salvage the data device
910                                  * even though the meta device has failed.  For
911                                  * now, we behave as though '- -' had been
912                                  * set for this device in the table.
913                                  */
914                                 if (dev->data_dev)
915                                         dm_put_device(ti, dev->data_dev);
916
917                                 dev->data_dev = NULL;
918                                 rdev->bdev = NULL;
919
920                                 list_del(&rdev->same_set);
921
922                                 continue;
923                         }
924                         ti->error = "Failed to load superblock";
925                         return ret;
926                 }
927         }
928
929         if (!freshest)
930                 return 0;
931
932         /*
933          * Validation of the freshest device provides the source of
934          * validation for the remaining devices.
935          */
936         ti->error = "Unable to assemble array: Invalid superblocks";
937         if (super_validate(mddev, freshest))
938                 return -EINVAL;
939
940         rdev_for_each(rdev, mddev)
941                 if ((rdev != freshest) && super_validate(mddev, rdev))
942                         return -EINVAL;
943
944         return 0;
945 }
946
947 /*
948  * Construct a RAID4/5/6 mapping:
949  * Args:
950  *      <raid_type> <#raid_params> <raid_params>                \
951  *      <#raid_devs> { <meta_dev1> <dev1> .. <meta_devN> <devN> }
952  *
953  * <raid_params> varies by <raid_type>.  See 'parse_raid_params' for
954  * details on possible <raid_params>.
955  */
956 static int raid_ctr(struct dm_target *ti, unsigned argc, char **argv)
957 {
958         int ret;
959         struct raid_type *rt;
960         unsigned long num_raid_params, num_raid_devs;
961         struct raid_set *rs = NULL;
962
963         /* Must have at least <raid_type> <#raid_params> */
964         if (argc < 2) {
965                 ti->error = "Too few arguments";
966                 return -EINVAL;
967         }
968
969         /* raid type */
970         rt = get_raid_type(argv[0]);
971         if (!rt) {
972                 ti->error = "Unrecognised raid_type";
973                 return -EINVAL;
974         }
975         argc--;
976         argv++;
977
978         /* number of RAID parameters */
979         if (strict_strtoul(argv[0], 10, &num_raid_params) < 0) {
980                 ti->error = "Cannot understand number of RAID parameters";
981                 return -EINVAL;
982         }
983         argc--;
984         argv++;
985
986         /* Skip over RAID params for now and find out # of devices */
987         if (num_raid_params + 1 > argc) {
988                 ti->error = "Arguments do not agree with counts given";
989                 return -EINVAL;
990         }
991
992         if ((strict_strtoul(argv[num_raid_params], 10, &num_raid_devs) < 0) ||
993             (num_raid_devs >= INT_MAX)) {
994                 ti->error = "Cannot understand number of raid devices";
995                 return -EINVAL;
996         }
997
998         rs = context_alloc(ti, rt, (unsigned)num_raid_devs);
999         if (IS_ERR(rs))
1000                 return PTR_ERR(rs);
1001
1002         ret = parse_raid_params(rs, argv, (unsigned)num_raid_params);
1003         if (ret)
1004                 goto bad;
1005
1006         ret = -EINVAL;
1007
1008         argc -= num_raid_params + 1; /* +1: we already have num_raid_devs */
1009         argv += num_raid_params + 1;
1010
1011         if (argc != (num_raid_devs * 2)) {
1012                 ti->error = "Supplied RAID devices does not match the count given";
1013                 goto bad;
1014         }
1015
1016         ret = dev_parms(rs, argv);
1017         if (ret)
1018                 goto bad;
1019
1020         rs->md.sync_super = super_sync;
1021         ret = analyse_superblocks(ti, rs);
1022         if (ret)
1023                 goto bad;
1024
1025         INIT_WORK(&rs->md.event_work, do_table_event);
1026         ti->private = rs;
1027         ti->num_flush_requests = 1;
1028
1029         mutex_lock(&rs->md.reconfig_mutex);
1030         ret = md_run(&rs->md);
1031         rs->md.in_sync = 0; /* Assume already marked dirty */
1032         mutex_unlock(&rs->md.reconfig_mutex);
1033
1034         if (ret) {
1035                 ti->error = "Fail to run raid array";
1036                 goto bad;
1037         }
1038
1039         rs->callbacks.congested_fn = raid_is_congested;
1040         dm_table_add_target_callbacks(ti->table, &rs->callbacks);
1041
1042         mddev_suspend(&rs->md);
1043         return 0;
1044
1045 bad:
1046         context_free(rs);
1047
1048         return ret;
1049 }
1050
1051 static void raid_dtr(struct dm_target *ti)
1052 {
1053         struct raid_set *rs = ti->private;
1054
1055         list_del_init(&rs->callbacks.list);
1056         md_stop(&rs->md);
1057         context_free(rs);
1058 }
1059
1060 static int raid_map(struct dm_target *ti, struct bio *bio, union map_info *map_context)
1061 {
1062         struct raid_set *rs = ti->private;
1063         struct mddev *mddev = &rs->md;
1064
1065         mddev->pers->make_request(mddev, bio);
1066
1067         return DM_MAPIO_SUBMITTED;
1068 }
1069
1070 static int raid_status(struct dm_target *ti, status_type_t type,
1071                        char *result, unsigned maxlen)
1072 {
1073         struct raid_set *rs = ti->private;
1074         unsigned raid_param_cnt = 1; /* at least 1 for chunksize */
1075         unsigned sz = 0;
1076         int i, array_in_sync = 0;
1077         sector_t sync;
1078
1079         switch (type) {
1080         case STATUSTYPE_INFO:
1081                 DMEMIT("%s %d ", rs->raid_type->name, rs->md.raid_disks);
1082
1083                 if (test_bit(MD_RECOVERY_RUNNING, &rs->md.recovery))
1084                         sync = rs->md.curr_resync_completed;
1085                 else
1086                         sync = rs->md.recovery_cp;
1087
1088                 if (sync >= rs->md.resync_max_sectors) {
1089                         array_in_sync = 1;
1090                         sync = rs->md.resync_max_sectors;
1091                 } else {
1092                         /*
1093                          * The array may be doing an initial sync, or it may
1094                          * be rebuilding individual components.  If all the
1095                          * devices are In_sync, then it is the array that is
1096                          * being initialized.
1097                          */
1098                         for (i = 0; i < rs->md.raid_disks; i++)
1099                                 if (!test_bit(In_sync, &rs->dev[i].rdev.flags))
1100                                         array_in_sync = 1;
1101                 }
1102                 /*
1103                  * Status characters:
1104                  *  'D' = Dead/Failed device
1105                  *  'a' = Alive but not in-sync
1106                  *  'A' = Alive and in-sync
1107                  */
1108                 for (i = 0; i < rs->md.raid_disks; i++) {
1109                         if (test_bit(Faulty, &rs->dev[i].rdev.flags))
1110                                 DMEMIT("D");
1111                         else if (!array_in_sync ||
1112                                  !test_bit(In_sync, &rs->dev[i].rdev.flags))
1113                                 DMEMIT("a");
1114                         else
1115                                 DMEMIT("A");
1116                 }
1117
1118                 /*
1119                  * In-sync ratio:
1120                  *  The in-sync ratio shows the progress of:
1121                  *   - Initializing the array
1122                  *   - Rebuilding a subset of devices of the array
1123                  *  The user can distinguish between the two by referring
1124                  *  to the status characters.
1125                  */
1126                 DMEMIT(" %llu/%llu",
1127                        (unsigned long long) sync,
1128                        (unsigned long long) rs->md.resync_max_sectors);
1129
1130                 break;
1131         case STATUSTYPE_TABLE:
1132                 /* The string you would use to construct this array */
1133                 for (i = 0; i < rs->md.raid_disks; i++) {
1134                         if ((rs->print_flags & DMPF_REBUILD) &&
1135                             rs->dev[i].data_dev &&
1136                             !test_bit(In_sync, &rs->dev[i].rdev.flags))
1137                                 raid_param_cnt += 2; /* for rebuilds */
1138                         if (rs->dev[i].data_dev &&
1139                             test_bit(WriteMostly, &rs->dev[i].rdev.flags))
1140                                 raid_param_cnt += 2;
1141                 }
1142
1143                 raid_param_cnt += (hweight32(rs->print_flags & ~DMPF_REBUILD) * 2);
1144                 if (rs->print_flags & (DMPF_SYNC | DMPF_NOSYNC))
1145                         raid_param_cnt--;
1146
1147                 DMEMIT("%s %u %u", rs->raid_type->name,
1148                        raid_param_cnt, rs->md.chunk_sectors);
1149
1150                 if ((rs->print_flags & DMPF_SYNC) &&
1151                     (rs->md.recovery_cp == MaxSector))
1152                         DMEMIT(" sync");
1153                 if (rs->print_flags & DMPF_NOSYNC)
1154                         DMEMIT(" nosync");
1155
1156                 for (i = 0; i < rs->md.raid_disks; i++)
1157                         if ((rs->print_flags & DMPF_REBUILD) &&
1158                             rs->dev[i].data_dev &&
1159                             !test_bit(In_sync, &rs->dev[i].rdev.flags))
1160                                 DMEMIT(" rebuild %u", i);
1161
1162                 if (rs->print_flags & DMPF_DAEMON_SLEEP)
1163                         DMEMIT(" daemon_sleep %lu",
1164                                rs->md.bitmap_info.daemon_sleep);
1165
1166                 if (rs->print_flags & DMPF_MIN_RECOVERY_RATE)
1167                         DMEMIT(" min_recovery_rate %d", rs->md.sync_speed_min);
1168
1169                 if (rs->print_flags & DMPF_MAX_RECOVERY_RATE)
1170                         DMEMIT(" max_recovery_rate %d", rs->md.sync_speed_max);
1171
1172                 for (i = 0; i < rs->md.raid_disks; i++)
1173                         if (rs->dev[i].data_dev &&
1174                             test_bit(WriteMostly, &rs->dev[i].rdev.flags))
1175                                 DMEMIT(" write_mostly %u", i);
1176
1177                 if (rs->print_flags & DMPF_MAX_WRITE_BEHIND)
1178                         DMEMIT(" max_write_behind %lu",
1179                                rs->md.bitmap_info.max_write_behind);
1180
1181                 if (rs->print_flags & DMPF_STRIPE_CACHE) {
1182                         struct r5conf *conf = rs->md.private;
1183
1184                         /* convert from kiB to sectors */
1185                         DMEMIT(" stripe_cache %d",
1186                                conf ? conf->max_nr_stripes * 2 : 0);
1187                 }
1188
1189                 if (rs->print_flags & DMPF_REGION_SIZE)
1190                         DMEMIT(" region_size %lu",
1191                                rs->md.bitmap_info.chunksize >> 9);
1192
1193                 DMEMIT(" %d", rs->md.raid_disks);
1194                 for (i = 0; i < rs->md.raid_disks; i++) {
1195                         if (rs->dev[i].meta_dev)
1196                                 DMEMIT(" %s", rs->dev[i].meta_dev->name);
1197                         else
1198                                 DMEMIT(" -");
1199
1200                         if (rs->dev[i].data_dev)
1201                                 DMEMIT(" %s", rs->dev[i].data_dev->name);
1202                         else
1203                                 DMEMIT(" -");
1204                 }
1205         }
1206
1207         return 0;
1208 }
1209
1210 static int raid_iterate_devices(struct dm_target *ti, iterate_devices_callout_fn fn, void *data)
1211 {
1212         struct raid_set *rs = ti->private;
1213         unsigned i;
1214         int ret = 0;
1215
1216         for (i = 0; !ret && i < rs->md.raid_disks; i++)
1217                 if (rs->dev[i].data_dev)
1218                         ret = fn(ti,
1219                                  rs->dev[i].data_dev,
1220                                  0, /* No offset on data devs */
1221                                  rs->md.dev_sectors,
1222                                  data);
1223
1224         return ret;
1225 }
1226
1227 static void raid_io_hints(struct dm_target *ti, struct queue_limits *limits)
1228 {
1229         struct raid_set *rs = ti->private;
1230         unsigned chunk_size = rs->md.chunk_sectors << 9;
1231         struct r5conf *conf = rs->md.private;
1232
1233         blk_limits_io_min(limits, chunk_size);
1234         blk_limits_io_opt(limits, chunk_size * (conf->raid_disks - conf->max_degraded));
1235 }
1236
1237 static void raid_presuspend(struct dm_target *ti)
1238 {
1239         struct raid_set *rs = ti->private;
1240
1241         md_stop_writes(&rs->md);
1242 }
1243
1244 static void raid_postsuspend(struct dm_target *ti)
1245 {
1246         struct raid_set *rs = ti->private;
1247
1248         mddev_suspend(&rs->md);
1249 }
1250
1251 static void raid_resume(struct dm_target *ti)
1252 {
1253         struct raid_set *rs = ti->private;
1254
1255         if (!rs->bitmap_loaded) {
1256                 bitmap_load(&rs->md);
1257                 rs->bitmap_loaded = 1;
1258         } else
1259                 md_wakeup_thread(rs->md.thread);
1260
1261         mddev_resume(&rs->md);
1262 }
1263
1264 static struct target_type raid_target = {
1265         .name = "raid",
1266         .version = {1, 2, 0},
1267         .module = THIS_MODULE,
1268         .ctr = raid_ctr,
1269         .dtr = raid_dtr,
1270         .map = raid_map,
1271         .status = raid_status,
1272         .iterate_devices = raid_iterate_devices,
1273         .io_hints = raid_io_hints,
1274         .presuspend = raid_presuspend,
1275         .postsuspend = raid_postsuspend,
1276         .resume = raid_resume,
1277 };
1278
1279 static int __init dm_raid_init(void)
1280 {
1281         return dm_register_target(&raid_target);
1282 }
1283
1284 static void __exit dm_raid_exit(void)
1285 {
1286         dm_unregister_target(&raid_target);
1287 }
1288
1289 module_init(dm_raid_init);
1290 module_exit(dm_raid_exit);
1291
1292 MODULE_DESCRIPTION(DM_NAME " raid4/5/6 target");
1293 MODULE_ALIAS("dm-raid4");
1294 MODULE_ALIAS("dm-raid5");
1295 MODULE_ALIAS("dm-raid6");
1296 MODULE_AUTHOR("Neil Brown <dm-devel@redhat.com>");
1297 MODULE_LICENSE("GPL");