]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * Copyright (C) 2001-2003 Sistina Software (UK) Limited. | |
3 | * | |
4 | * This file is released under the GPL. | |
5 | */ | |
6 | ||
7fff5e8f | 7 | #include "dm.h" |
586e80e6 | 8 | #include <linux/device-mapper.h> |
1da177e4 LT |
9 | |
10 | #include <linux/module.h> | |
11 | #include <linux/init.h> | |
12 | #include <linux/blkdev.h> | |
13 | #include <linux/bio.h> | |
14 | #include <linux/slab.h> | |
6f3c3f0a | 15 | #include <linux/log2.h> |
1da177e4 | 16 | |
72d94861 | 17 | #define DM_MSG_PREFIX "striped" |
a25eb944 | 18 | #define DM_IO_ERROR_THRESHOLD 15 |
72d94861 | 19 | |
1da177e4 LT |
20 | struct stripe { |
21 | struct dm_dev *dev; | |
22 | sector_t physical_start; | |
a25eb944 BW |
23 | |
24 | atomic_t error_count; | |
1da177e4 LT |
25 | }; |
26 | ||
27 | struct stripe_c { | |
28 | uint32_t stripes; | |
c96053b7 | 29 | int stripes_shift; |
1da177e4 LT |
30 | |
31 | /* The size of this target / num. stripes */ | |
32 | sector_t stripe_width; | |
33 | ||
eb850de6 | 34 | uint32_t chunk_size; |
33d07c0d | 35 | int chunk_size_shift; |
1da177e4 | 36 | |
a25eb944 BW |
37 | /* Needed for handling events */ |
38 | struct dm_target *ti; | |
39 | ||
40 | /* Work struct used for triggering events*/ | |
f521f074 | 41 | struct work_struct trigger_event; |
a25eb944 | 42 | |
1da177e4 LT |
43 | struct stripe stripe[0]; |
44 | }; | |
45 | ||
a25eb944 BW |
46 | /* |
47 | * An event is triggered whenever a drive | |
48 | * drops out of a stripe volume. | |
49 | */ | |
50 | static void trigger_event(struct work_struct *work) | |
51 | { | |
f521f074 TH |
52 | struct stripe_c *sc = container_of(work, struct stripe_c, |
53 | trigger_event); | |
a25eb944 | 54 | dm_table_event(sc->ti->table); |
a25eb944 BW |
55 | } |
56 | ||
1da177e4 LT |
57 | static inline struct stripe_c *alloc_context(unsigned int stripes) |
58 | { | |
59 | size_t len; | |
60 | ||
d63a5ce3 MP |
61 | if (dm_array_too_big(sizeof(struct stripe_c), sizeof(struct stripe), |
62 | stripes)) | |
1da177e4 LT |
63 | return NULL; |
64 | ||
65 | len = sizeof(struct stripe_c) + (sizeof(struct stripe) * stripes); | |
66 | ||
67 | return kmalloc(len, GFP_KERNEL); | |
68 | } | |
69 | ||
70 | /* | |
71 | * Parse a single <dev> <sector> pair | |
72 | */ | |
73 | static int get_stripe(struct dm_target *ti, struct stripe_c *sc, | |
74 | unsigned int stripe, char **argv) | |
75 | { | |
4ee218cd | 76 | unsigned long long start; |
31998ef1 | 77 | char dummy; |
e80d1c80 | 78 | int ret; |
1da177e4 | 79 | |
31998ef1 | 80 | if (sscanf(argv[1], "%llu%c", &start, &dummy) != 1) |
1da177e4 LT |
81 | return -EINVAL; |
82 | ||
e80d1c80 VG |
83 | ret = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), |
84 | &sc->stripe[stripe].dev); | |
85 | if (ret) | |
86 | return ret; | |
1da177e4 LT |
87 | |
88 | sc->stripe[stripe].physical_start = start; | |
a25eb944 | 89 | |
1da177e4 LT |
90 | return 0; |
91 | } | |
92 | ||
93 | /* | |
94 | * Construct a striped mapping. | |
eb850de6 | 95 | * <number of stripes> <chunk size> [<dev_path> <offset>]+ |
1da177e4 LT |
96 | */ |
97 | static int stripe_ctr(struct dm_target *ti, unsigned int argc, char **argv) | |
98 | { | |
99 | struct stripe_c *sc; | |
d793e684 | 100 | sector_t width, tmp_len; |
1da177e4 LT |
101 | uint32_t stripes; |
102 | uint32_t chunk_size; | |
1da177e4 LT |
103 | int r; |
104 | unsigned int i; | |
105 | ||
106 | if (argc < 2) { | |
72d94861 | 107 | ti->error = "Not enough arguments"; |
1da177e4 LT |
108 | return -EINVAL; |
109 | } | |
110 | ||
1a66a08a | 111 | if (kstrtouint(argv[0], 10, &stripes) || !stripes) { |
72d94861 | 112 | ti->error = "Invalid stripe count"; |
1da177e4 LT |
113 | return -EINVAL; |
114 | } | |
115 | ||
8f069b41 | 116 | if (kstrtouint(argv[1], 10, &chunk_size) || !chunk_size) { |
eb850de6 | 117 | ti->error = "Invalid chunk_size"; |
1da177e4 LT |
118 | return -EINVAL; |
119 | } | |
120 | ||
eb850de6 | 121 | width = ti->len; |
d793e684 | 122 | if (sector_div(width, stripes)) { |
72d94861 | 123 | ti->error = "Target length not divisible by " |
d793e684 | 124 | "number of stripes"; |
8ba32fde KC |
125 | return -EINVAL; |
126 | } | |
127 | ||
d793e684 MS |
128 | tmp_len = width; |
129 | if (sector_div(tmp_len, chunk_size)) { | |
72d94861 | 130 | ti->error = "Target length not divisible by " |
d793e684 | 131 | "chunk size"; |
1da177e4 LT |
132 | return -EINVAL; |
133 | } | |
134 | ||
135 | /* | |
136 | * Do we have enough arguments for that many stripes ? | |
137 | */ | |
138 | if (argc != (2 + 2 * stripes)) { | |
72d94861 | 139 | ti->error = "Not enough destinations " |
1da177e4 LT |
140 | "specified"; |
141 | return -EINVAL; | |
142 | } | |
143 | ||
144 | sc = alloc_context(stripes); | |
145 | if (!sc) { | |
72d94861 | 146 | ti->error = "Memory allocation for striped context " |
1da177e4 LT |
147 | "failed"; |
148 | return -ENOMEM; | |
149 | } | |
150 | ||
f521f074 | 151 | INIT_WORK(&sc->trigger_event, trigger_event); |
a25eb944 BW |
152 | |
153 | /* Set pointer to dm target; used in trigger_event */ | |
154 | sc->ti = ti; | |
1da177e4 LT |
155 | sc->stripes = stripes; |
156 | sc->stripe_width = width; | |
c96053b7 MP |
157 | |
158 | if (stripes & (stripes - 1)) | |
159 | sc->stripes_shift = -1; | |
1df05483 MP |
160 | else |
161 | sc->stripes_shift = __ffs(stripes); | |
c96053b7 | 162 | |
542f9038 | 163 | r = dm_set_target_max_io_len(ti, chunk_size); |
a3f2af25 PK |
164 | if (r) { |
165 | kfree(sc); | |
542f9038 | 166 | return r; |
a3f2af25 | 167 | } |
542f9038 | 168 | |
55a62eef AK |
169 | ti->num_flush_bios = stripes; |
170 | ti->num_discard_bios = stripes; | |
171 | ti->num_write_same_bios = stripes; | |
1da177e4 | 172 | |
eb850de6 | 173 | sc->chunk_size = chunk_size; |
33d07c0d MP |
174 | if (chunk_size & (chunk_size - 1)) |
175 | sc->chunk_size_shift = -1; | |
176 | else | |
177 | sc->chunk_size_shift = __ffs(chunk_size); | |
1da177e4 LT |
178 | |
179 | /* | |
180 | * Get the stripe destinations. | |
181 | */ | |
182 | for (i = 0; i < stripes; i++) { | |
183 | argv += 2; | |
184 | ||
185 | r = get_stripe(ti, sc, i, argv); | |
186 | if (r < 0) { | |
72d94861 | 187 | ti->error = "Couldn't parse stripe destination"; |
1da177e4 LT |
188 | while (i--) |
189 | dm_put_device(ti, sc->stripe[i].dev); | |
190 | kfree(sc); | |
191 | return r; | |
192 | } | |
a25eb944 | 193 | atomic_set(&(sc->stripe[i].error_count), 0); |
1da177e4 LT |
194 | } |
195 | ||
196 | ti->private = sc; | |
a25eb944 | 197 | |
1da177e4 LT |
198 | return 0; |
199 | } | |
200 | ||
201 | static void stripe_dtr(struct dm_target *ti) | |
202 | { | |
203 | unsigned int i; | |
204 | struct stripe_c *sc = (struct stripe_c *) ti->private; | |
205 | ||
206 | for (i = 0; i < sc->stripes; i++) | |
207 | dm_put_device(ti, sc->stripe[i].dev); | |
208 | ||
43829731 | 209 | flush_work(&sc->trigger_event); |
1da177e4 LT |
210 | kfree(sc); |
211 | } | |
212 | ||
65988525 MP |
213 | static void stripe_map_sector(struct stripe_c *sc, sector_t sector, |
214 | uint32_t *stripe, sector_t *result) | |
215 | { | |
eb850de6 | 216 | sector_t chunk = dm_target_offset(sc->ti, sector); |
33d07c0d MP |
217 | sector_t chunk_offset; |
218 | ||
219 | if (sc->chunk_size_shift < 0) | |
220 | chunk_offset = sector_div(chunk, sc->chunk_size); | |
221 | else { | |
222 | chunk_offset = chunk & (sc->chunk_size - 1); | |
223 | chunk >>= sc->chunk_size_shift; | |
224 | } | |
65988525 | 225 | |
c96053b7 MP |
226 | if (sc->stripes_shift < 0) |
227 | *stripe = sector_div(chunk, sc->stripes); | |
228 | else { | |
1df05483 | 229 | *stripe = chunk & (sc->stripes - 1); |
c96053b7 MP |
230 | chunk >>= sc->stripes_shift; |
231 | } | |
232 | ||
33d07c0d MP |
233 | if (sc->chunk_size_shift < 0) |
234 | chunk *= sc->chunk_size; | |
235 | else | |
236 | chunk <<= sc->chunk_size_shift; | |
237 | ||
238 | *result = chunk + chunk_offset; | |
65988525 MP |
239 | } |
240 | ||
7b76ec11 MP |
241 | static void stripe_map_range_sector(struct stripe_c *sc, sector_t sector, |
242 | uint32_t target_stripe, sector_t *result) | |
243 | { | |
244 | uint32_t stripe; | |
245 | ||
246 | stripe_map_sector(sc, sector, &stripe, result); | |
247 | if (stripe == target_stripe) | |
248 | return; | |
eb850de6 MS |
249 | |
250 | /* round down */ | |
251 | sector = *result; | |
33d07c0d MP |
252 | if (sc->chunk_size_shift < 0) |
253 | *result -= sector_div(sector, sc->chunk_size); | |
254 | else | |
255 | *result = sector & ~(sector_t)(sc->chunk_size - 1); | |
eb850de6 | 256 | |
7b76ec11 | 257 | if (target_stripe < stripe) |
eb850de6 | 258 | *result += sc->chunk_size; /* next chunk */ |
7b76ec11 MP |
259 | } |
260 | ||
45e621d4 MS |
261 | static int stripe_map_range(struct stripe_c *sc, struct bio *bio, |
262 | uint32_t target_stripe) | |
7b76ec11 MP |
263 | { |
264 | sector_t begin, end; | |
265 | ||
4f024f37 KO |
266 | stripe_map_range_sector(sc, bio->bi_iter.bi_sector, |
267 | target_stripe, &begin); | |
f73a1c7d | 268 | stripe_map_range_sector(sc, bio_end_sector(bio), |
7b76ec11 MP |
269 | target_stripe, &end); |
270 | if (begin < end) { | |
271 | bio->bi_bdev = sc->stripe[target_stripe].dev->bdev; | |
4f024f37 KO |
272 | bio->bi_iter.bi_sector = begin + |
273 | sc->stripe[target_stripe].physical_start; | |
274 | bio->bi_iter.bi_size = to_bytes(end - begin); | |
7b76ec11 MP |
275 | return DM_MAPIO_REMAPPED; |
276 | } else { | |
277 | /* The range doesn't map to the target stripe */ | |
4246a0b6 | 278 | bio_endio(bio); |
7b76ec11 MP |
279 | return DM_MAPIO_SUBMITTED; |
280 | } | |
281 | } | |
282 | ||
7de3ee57 | 283 | static int stripe_map(struct dm_target *ti, struct bio *bio) |
1da177e4 | 284 | { |
65988525 | 285 | struct stripe_c *sc = ti->private; |
374bf7e7 | 286 | uint32_t stripe; |
55a62eef | 287 | unsigned target_bio_nr; |
1da177e4 | 288 | |
d87f4c14 | 289 | if (bio->bi_rw & REQ_FLUSH) { |
55a62eef AK |
290 | target_bio_nr = dm_bio_get_target_bio_nr(bio); |
291 | BUG_ON(target_bio_nr >= sc->stripes); | |
292 | bio->bi_bdev = sc->stripe[target_bio_nr].dev->bdev; | |
374bf7e7 MP |
293 | return DM_MAPIO_REMAPPED; |
294 | } | |
45e621d4 MS |
295 | if (unlikely(bio->bi_rw & REQ_DISCARD) || |
296 | unlikely(bio->bi_rw & REQ_WRITE_SAME)) { | |
55a62eef AK |
297 | target_bio_nr = dm_bio_get_target_bio_nr(bio); |
298 | BUG_ON(target_bio_nr >= sc->stripes); | |
299 | return stripe_map_range(sc, bio, target_bio_nr); | |
7b76ec11 | 300 | } |
374bf7e7 | 301 | |
4f024f37 KO |
302 | stripe_map_sector(sc, bio->bi_iter.bi_sector, |
303 | &stripe, &bio->bi_iter.bi_sector); | |
1da177e4 | 304 | |
4f024f37 | 305 | bio->bi_iter.bi_sector += sc->stripe[stripe].physical_start; |
1da177e4 | 306 | bio->bi_bdev = sc->stripe[stripe].dev->bdev; |
65988525 | 307 | |
d2a7ad29 | 308 | return DM_MAPIO_REMAPPED; |
1da177e4 LT |
309 | } |
310 | ||
4f7f5c67 BW |
311 | /* |
312 | * Stripe status: | |
313 | * | |
314 | * INFO | |
315 | * #stripes [stripe_name <stripe_name>] [group word count] | |
316 | * [error count 'A|D' <error count 'A|D'>] | |
317 | * | |
318 | * TABLE | |
319 | * #stripes [stripe chunk size] | |
320 | * [stripe_name physical_start <stripe_name physical_start>] | |
321 | * | |
322 | */ | |
323 | ||
fd7c092e MP |
324 | static void stripe_status(struct dm_target *ti, status_type_t type, |
325 | unsigned status_flags, char *result, unsigned maxlen) | |
1da177e4 LT |
326 | { |
327 | struct stripe_c *sc = (struct stripe_c *) ti->private; | |
4f7f5c67 | 328 | char buffer[sc->stripes + 1]; |
1da177e4 LT |
329 | unsigned int sz = 0; |
330 | unsigned int i; | |
331 | ||
332 | switch (type) { | |
333 | case STATUSTYPE_INFO: | |
4f7f5c67 BW |
334 | DMEMIT("%d ", sc->stripes); |
335 | for (i = 0; i < sc->stripes; i++) { | |
336 | DMEMIT("%s ", sc->stripe[i].dev->name); | |
337 | buffer[i] = atomic_read(&(sc->stripe[i].error_count)) ? | |
338 | 'D' : 'A'; | |
339 | } | |
340 | buffer[i] = '\0'; | |
341 | DMEMIT("1 %s", buffer); | |
1da177e4 LT |
342 | break; |
343 | ||
344 | case STATUSTYPE_TABLE: | |
4ee218cd | 345 | DMEMIT("%d %llu", sc->stripes, |
eb850de6 | 346 | (unsigned long long)sc->chunk_size); |
1da177e4 | 347 | for (i = 0; i < sc->stripes; i++) |
4ee218cd AM |
348 | DMEMIT(" %s %llu", sc->stripe[i].dev->name, |
349 | (unsigned long long)sc->stripe[i].physical_start); | |
1da177e4 LT |
350 | break; |
351 | } | |
1da177e4 LT |
352 | } |
353 | ||
7de3ee57 | 354 | static int stripe_end_io(struct dm_target *ti, struct bio *bio, int error) |
a25eb944 BW |
355 | { |
356 | unsigned i; | |
357 | char major_minor[16]; | |
358 | struct stripe_c *sc = ti->private; | |
359 | ||
360 | if (!error) | |
361 | return 0; /* I/O complete */ | |
362 | ||
7b6d91da | 363 | if ((error == -EWOULDBLOCK) && (bio->bi_rw & REQ_RAHEAD)) |
a25eb944 BW |
364 | return error; |
365 | ||
366 | if (error == -EOPNOTSUPP) | |
367 | return error; | |
368 | ||
369 | memset(major_minor, 0, sizeof(major_minor)); | |
370 | sprintf(major_minor, "%d:%d", | |
f331c029 TH |
371 | MAJOR(disk_devt(bio->bi_bdev->bd_disk)), |
372 | MINOR(disk_devt(bio->bi_bdev->bd_disk))); | |
a25eb944 BW |
373 | |
374 | /* | |
375 | * Test to see which stripe drive triggered the event | |
376 | * and increment error count for all stripes on that device. | |
377 | * If the error count for a given device exceeds the threshold | |
378 | * value we will no longer trigger any further events. | |
379 | */ | |
380 | for (i = 0; i < sc->stripes; i++) | |
381 | if (!strcmp(sc->stripe[i].dev->name, major_minor)) { | |
382 | atomic_inc(&(sc->stripe[i].error_count)); | |
383 | if (atomic_read(&(sc->stripe[i].error_count)) < | |
384 | DM_IO_ERROR_THRESHOLD) | |
f521f074 | 385 | schedule_work(&sc->trigger_event); |
a25eb944 BW |
386 | } |
387 | ||
388 | return error; | |
389 | } | |
390 | ||
af4874e0 MS |
391 | static int stripe_iterate_devices(struct dm_target *ti, |
392 | iterate_devices_callout_fn fn, void *data) | |
393 | { | |
394 | struct stripe_c *sc = ti->private; | |
395 | int ret = 0; | |
396 | unsigned i = 0; | |
397 | ||
5dea271b | 398 | do { |
af4874e0 | 399 | ret = fn(ti, sc->stripe[i].dev, |
5dea271b MS |
400 | sc->stripe[i].physical_start, |
401 | sc->stripe_width, data); | |
402 | } while (!ret && ++i < sc->stripes); | |
af4874e0 MS |
403 | |
404 | return ret; | |
405 | } | |
406 | ||
40bea431 MS |
407 | static void stripe_io_hints(struct dm_target *ti, |
408 | struct queue_limits *limits) | |
409 | { | |
410 | struct stripe_c *sc = ti->private; | |
eb850de6 | 411 | unsigned chunk_size = sc->chunk_size << SECTOR_SHIFT; |
40bea431 MS |
412 | |
413 | blk_limits_io_min(limits, chunk_size); | |
3c5820c7 | 414 | blk_limits_io_opt(limits, chunk_size * sc->stripes); |
40bea431 MS |
415 | } |
416 | ||
1da177e4 LT |
417 | static struct target_type stripe_target = { |
418 | .name = "striped", | |
fd7c092e | 419 | .version = {1, 5, 1}, |
1da177e4 LT |
420 | .module = THIS_MODULE, |
421 | .ctr = stripe_ctr, | |
422 | .dtr = stripe_dtr, | |
423 | .map = stripe_map, | |
a25eb944 | 424 | .end_io = stripe_end_io, |
1da177e4 | 425 | .status = stripe_status, |
af4874e0 | 426 | .iterate_devices = stripe_iterate_devices, |
40bea431 | 427 | .io_hints = stripe_io_hints, |
1da177e4 LT |
428 | }; |
429 | ||
430 | int __init dm_stripe_init(void) | |
431 | { | |
432 | int r; | |
433 | ||
434 | r = dm_register_target(&stripe_target); | |
e223e1de | 435 | if (r < 0) |
72d94861 | 436 | DMWARN("target registration failed"); |
1da177e4 LT |
437 | |
438 | return r; | |
439 | } | |
440 | ||
441 | void dm_stripe_exit(void) | |
442 | { | |
10d3bd09 | 443 | dm_unregister_target(&stripe_target); |
1da177e4 | 444 | } |