dm-stripe.c revision 8f069b41bce79b0c4b6076acd0f3d15df0a232ed 
1/*
2 * Copyright (C) 2001-2003 Sistina Software (UK) Limited.
3 *
4 * This file is released under the GPL.
5 */
6
7#include <linux/device-mapper.h>
8
9#include <linux/module.h>
10#include <linux/init.h>
11#include <linux/blkdev.h>
12#include <linux/bio.h>
13#include <linux/slab.h>
14#include <linux/log2.h>
15
16#define DM_MSG_PREFIX "striped"
17#define DM_IO_ERROR_THRESHOLD 15
18
19struct stripe {
20	struct dm_dev *dev;
21	sector_t physical_start;
22
23	atomic_t error_count;
24};
25
26struct stripe_c {
27	uint32_t stripes;
28	int stripes_shift;
29
30	/* The size of this target / num. stripes */
31	sector_t stripe_width;
32
33	uint32_t chunk_size;
34
35	/* Needed for handling events */
36	struct dm_target *ti;
37
38	/* Work struct used for triggering events*/
39	struct work_struct trigger_event;
40
41	struct stripe stripe[0];
42};
43
44/*
45 * An event is triggered whenever a drive
46 * drops out of a stripe volume.
47 */
48static void trigger_event(struct work_struct *work)
49{
50	struct stripe_c *sc = container_of(work, struct stripe_c,
51					   trigger_event);
52	dm_table_event(sc->ti->table);
53}
54
55static inline struct stripe_c *alloc_context(unsigned int stripes)
56{
57	size_t len;
58
59	if (dm_array_too_big(sizeof(struct stripe_c), sizeof(struct stripe),
60			     stripes))
61		return NULL;
62
63	len = sizeof(struct stripe_c) + (sizeof(struct stripe) * stripes);
64
65	return kmalloc(len, GFP_KERNEL);
66}
67
68/*
69 * Parse a single <dev> <sector> pair
70 */
71static int get_stripe(struct dm_target *ti, struct stripe_c *sc,
72		      unsigned int stripe, char **argv)
73{
74	unsigned long long start;
75	char dummy;
76
77	if (sscanf(argv[1], "%llu%c", &start, &dummy) != 1)
78		return -EINVAL;
79
80	if (dm_get_device(ti, argv[0], dm_table_get_mode(ti->table),
81			  &sc->stripe[stripe].dev))
82		return -ENXIO;
83
84	sc->stripe[stripe].physical_start = start;
85
86	return 0;
87}
88
89/*
90 * Construct a striped mapping.
91 * <number of stripes> <chunk size> [<dev_path> <offset>]+
92 */
93static int stripe_ctr(struct dm_target *ti, unsigned int argc, char **argv)
94{
95	struct stripe_c *sc;
96	sector_t width;
97	uint32_t stripes;
98	uint32_t chunk_size;
99	int r;
100	unsigned int i;
101
102	if (argc < 2) {
103		ti->error = "Not enough arguments";
104		return -EINVAL;
105	}
106
107	if (kstrtouint(argv[0], 10, &stripes) || !stripes) {
108		ti->error = "Invalid stripe count";
109		return -EINVAL;
110	}
111
112	if (kstrtouint(argv[1], 10, &chunk_size) || !chunk_size) {
113		ti->error = "Invalid chunk_size";
114		return -EINVAL;
115	}
116
117	width = ti->len;
118	if (sector_div(width, chunk_size)) {
119		ti->error = "Target length not divisible by "
120		    "chunk size";
121		return -EINVAL;
122	}
123
124	if (sector_div(width, stripes)) {
125		ti->error = "Target length not divisible by "
126		    "number of stripes";
127		return -EINVAL;
128	}
129
130	/*
131	 * Do we have enough arguments for that many stripes ?
132	 */
133	if (argc != (2 + 2 * stripes)) {
134		ti->error = "Not enough destinations "
135			"specified";
136		return -EINVAL;
137	}
138
139	sc = alloc_context(stripes);
140	if (!sc) {
141		ti->error = "Memory allocation for striped context "
142		    "failed";
143		return -ENOMEM;
144	}
145
146	INIT_WORK(&sc->trigger_event, trigger_event);
147
148	/* Set pointer to dm target; used in trigger_event */
149	sc->ti = ti;
150	sc->stripes = stripes;
151	sc->stripe_width = width;
152
153	if (stripes & (stripes - 1))
154		sc->stripes_shift = -1;
155	else
156		sc->stripes_shift = __ffs(stripes);
157
158	r = dm_set_target_max_io_len(ti, chunk_size);
159	if (r)
160		return r;
161
162	ti->num_flush_requests = stripes;
163	ti->num_discard_requests = stripes;
164
165	sc->chunk_size = chunk_size;
166
167	/*
168	 * Get the stripe destinations.
169	 */
170	for (i = 0; i < stripes; i++) {
171		argv += 2;
172
173		r = get_stripe(ti, sc, i, argv);
174		if (r < 0) {
175			ti->error = "Couldn't parse stripe destination";
176			while (i--)
177				dm_put_device(ti, sc->stripe[i].dev);
178			kfree(sc);
179			return r;
180		}
181		atomic_set(&(sc->stripe[i].error_count), 0);
182	}
183
184	ti->private = sc;
185
186	return 0;
187}
188
189static void stripe_dtr(struct dm_target *ti)
190{
191	unsigned int i;
192	struct stripe_c *sc = (struct stripe_c *) ti->private;
193
194	for (i = 0; i < sc->stripes; i++)
195		dm_put_device(ti, sc->stripe[i].dev);
196
197	flush_work_sync(&sc->trigger_event);
198	kfree(sc);
199}
200
201static void stripe_map_sector(struct stripe_c *sc, sector_t sector,
202			      uint32_t *stripe, sector_t *result)
203{
204	sector_t chunk = dm_target_offset(sc->ti, sector);
205	sector_t chunk_offset = sector_div(chunk, sc->chunk_size);
206
207	if (sc->stripes_shift < 0)
208		*stripe = sector_div(chunk, sc->stripes);
209	else {
210		*stripe = chunk & (sc->stripes - 1);
211		chunk >>= sc->stripes_shift;
212	}
213
214	*result = (chunk * sc->chunk_size) + chunk_offset;
215}
216
217static void stripe_map_range_sector(struct stripe_c *sc, sector_t sector,
218				    uint32_t target_stripe, sector_t *result)
219{
220	uint32_t stripe;
221
222	stripe_map_sector(sc, sector, &stripe, result);
223	if (stripe == target_stripe)
224		return;
225
226	/* round down */
227	sector = *result;
228	*result -= sector_div(sector, sc->chunk_size);
229
230	if (target_stripe < stripe)
231		*result += sc->chunk_size;		/* next chunk */
232}
233
234static int stripe_map_discard(struct stripe_c *sc, struct bio *bio,
235			      uint32_t target_stripe)
236{
237	sector_t begin, end;
238
239	stripe_map_range_sector(sc, bio->bi_sector, target_stripe, &begin);
240	stripe_map_range_sector(sc, bio->bi_sector + bio_sectors(bio),
241				target_stripe, &end);
242	if (begin < end) {
243		bio->bi_bdev = sc->stripe[target_stripe].dev->bdev;
244		bio->bi_sector = begin + sc->stripe[target_stripe].physical_start;
245		bio->bi_size = to_bytes(end - begin);
246		return DM_MAPIO_REMAPPED;
247	} else {
248		/* The range doesn't map to the target stripe */
249		bio_endio(bio, 0);
250		return DM_MAPIO_SUBMITTED;
251	}
252}
253
254static int stripe_map(struct dm_target *ti, struct bio *bio,
255		      union map_info *map_context)
256{
257	struct stripe_c *sc = ti->private;
258	uint32_t stripe;
259	unsigned target_request_nr;
260
261	if (bio->bi_rw & REQ_FLUSH) {
262		target_request_nr = map_context->target_request_nr;
263		BUG_ON(target_request_nr >= sc->stripes);
264		bio->bi_bdev = sc->stripe[target_request_nr].dev->bdev;
265		return DM_MAPIO_REMAPPED;
266	}
267	if (unlikely(bio->bi_rw & REQ_DISCARD)) {
268		target_request_nr = map_context->target_request_nr;
269		BUG_ON(target_request_nr >= sc->stripes);
270		return stripe_map_discard(sc, bio, target_request_nr);
271	}
272
273	stripe_map_sector(sc, bio->bi_sector, &stripe, &bio->bi_sector);
274
275	bio->bi_sector += sc->stripe[stripe].physical_start;
276	bio->bi_bdev = sc->stripe[stripe].dev->bdev;
277
278	return DM_MAPIO_REMAPPED;
279}
280
281/*
282 * Stripe status:
283 *
284 * INFO
285 * #stripes [stripe_name <stripe_name>] [group word count]
286 * [error count 'A|D' <error count 'A|D'>]
287 *
288 * TABLE
289 * #stripes [stripe chunk size]
290 * [stripe_name physical_start <stripe_name physical_start>]
291 *
292 */
293
294static int stripe_status(struct dm_target *ti,
295			 status_type_t type, char *result, unsigned int maxlen)
296{
297	struct stripe_c *sc = (struct stripe_c *) ti->private;
298	char buffer[sc->stripes + 1];
299	unsigned int sz = 0;
300	unsigned int i;
301
302	switch (type) {
303	case STATUSTYPE_INFO:
304		DMEMIT("%d ", sc->stripes);
305		for (i = 0; i < sc->stripes; i++)  {
306			DMEMIT("%s ", sc->stripe[i].dev->name);
307			buffer[i] = atomic_read(&(sc->stripe[i].error_count)) ?
308				'D' : 'A';
309		}
310		buffer[i] = '\0';
311		DMEMIT("1 %s", buffer);
312		break;
313
314	case STATUSTYPE_TABLE:
315		DMEMIT("%d %llu", sc->stripes,
316			(unsigned long long)sc->chunk_size);
317		for (i = 0; i < sc->stripes; i++)
318			DMEMIT(" %s %llu", sc->stripe[i].dev->name,
319			    (unsigned long long)sc->stripe[i].physical_start);
320		break;
321	}
322	return 0;
323}
324
325static int stripe_end_io(struct dm_target *ti, struct bio *bio,
326			 int error, union map_info *map_context)
327{
328	unsigned i;
329	char major_minor[16];
330	struct stripe_c *sc = ti->private;
331
332	if (!error)
333		return 0; /* I/O complete */
334
335	if ((error == -EWOULDBLOCK) && (bio->bi_rw & REQ_RAHEAD))
336		return error;
337
338	if (error == -EOPNOTSUPP)
339		return error;
340
341	memset(major_minor, 0, sizeof(major_minor));
342	sprintf(major_minor, "%d:%d",
343		MAJOR(disk_devt(bio->bi_bdev->bd_disk)),
344		MINOR(disk_devt(bio->bi_bdev->bd_disk)));
345
346	/*
347	 * Test to see which stripe drive triggered the event
348	 * and increment error count for all stripes on that device.
349	 * If the error count for a given device exceeds the threshold
350	 * value we will no longer trigger any further events.
351	 */
352	for (i = 0; i < sc->stripes; i++)
353		if (!strcmp(sc->stripe[i].dev->name, major_minor)) {
354			atomic_inc(&(sc->stripe[i].error_count));
355			if (atomic_read(&(sc->stripe[i].error_count)) <
356			    DM_IO_ERROR_THRESHOLD)
357				schedule_work(&sc->trigger_event);
358		}
359
360	return error;
361}
362
363static int stripe_iterate_devices(struct dm_target *ti,
364				  iterate_devices_callout_fn fn, void *data)
365{
366	struct stripe_c *sc = ti->private;
367	int ret = 0;
368	unsigned i = 0;
369
370	do {
371		ret = fn(ti, sc->stripe[i].dev,
372			 sc->stripe[i].physical_start,
373			 sc->stripe_width, data);
374	} while (!ret && ++i < sc->stripes);
375
376	return ret;
377}
378
379static void stripe_io_hints(struct dm_target *ti,
380			    struct queue_limits *limits)
381{
382	struct stripe_c *sc = ti->private;
383	unsigned chunk_size = sc->chunk_size << SECTOR_SHIFT;
384
385	blk_limits_io_min(limits, chunk_size);
386	blk_limits_io_opt(limits, chunk_size * sc->stripes);
387}
388
389static int stripe_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
390			struct bio_vec *biovec, int max_size)
391{
392	struct stripe_c *sc = ti->private;
393	sector_t bvm_sector = bvm->bi_sector;
394	uint32_t stripe;
395	struct request_queue *q;
396
397	stripe_map_sector(sc, bvm_sector, &stripe, &bvm_sector);
398
399	q = bdev_get_queue(sc->stripe[stripe].dev->bdev);
400	if (!q->merge_bvec_fn)
401		return max_size;
402
403	bvm->bi_bdev = sc->stripe[stripe].dev->bdev;
404	bvm->bi_sector = sc->stripe[stripe].physical_start + bvm_sector;
405
406	return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
407}
408
409static struct target_type stripe_target = {
410	.name   = "striped",
411	.version = {1, 5, 0},
412	.module = THIS_MODULE,
413	.ctr    = stripe_ctr,
414	.dtr    = stripe_dtr,
415	.map    = stripe_map,
416	.end_io = stripe_end_io,
417	.status = stripe_status,
418	.iterate_devices = stripe_iterate_devices,
419	.io_hints = stripe_io_hints,
420	.merge  = stripe_merge,
421};
422
423int __init dm_stripe_init(void)
424{
425	int r;
426
427	r = dm_register_target(&stripe_target);
428	if (r < 0) {
429		DMWARN("target registration failed");
430		return r;
431	}
432
433	return r;
434}
435
436void dm_stripe_exit(void)
437{
438	dm_unregister_target(&stripe_target);
439}
440