dm-stripe.c revision 1a66a08ae82b16eb40705ad112ff28873981af92 
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	sector_t stripes_mask;
30
31	/* The size of this target / num. stripes */
32	sector_t stripe_width;
33
34	/* stripe chunk size */
35	uint32_t chunk_shift;
36	sector_t chunk_mask;
37
38	/* Needed for handling events */
39	struct dm_target *ti;
40
41	/* Work struct used for triggering events*/
42	struct work_struct trigger_event;
43
44	struct stripe stripe[0];
45};
46
47/*
48 * An event is triggered whenever a drive
49 * drops out of a stripe volume.
50 */
51static void trigger_event(struct work_struct *work)
52{
53	struct stripe_c *sc = container_of(work, struct stripe_c,
54					   trigger_event);
55	dm_table_event(sc->ti->table);
56}
57
58static inline struct stripe_c *alloc_context(unsigned int stripes)
59{
60	size_t len;
61
62	if (dm_array_too_big(sizeof(struct stripe_c), sizeof(struct stripe),
63			     stripes))
64		return NULL;
65
66	len = sizeof(struct stripe_c) + (sizeof(struct stripe) * stripes);
67
68	return kmalloc(len, GFP_KERNEL);
69}
70
71/*
72 * Parse a single <dev> <sector> pair
73 */
74static int get_stripe(struct dm_target *ti, struct stripe_c *sc,
75		      unsigned int stripe, char **argv)
76{
77	unsigned long long start;
78	char dummy;
79
80	if (sscanf(argv[1], "%llu%c", &start, &dummy) != 1)
81		return -EINVAL;
82
83	if (dm_get_device(ti, argv[0], dm_table_get_mode(ti->table),
84			  &sc->stripe[stripe].dev))
85		return -ENXIO;
86
87	sc->stripe[stripe].physical_start = start;
88
89	return 0;
90}
91
92/*
93 * Construct a striped mapping.
94 * <number of stripes> <chunk size (2^^n)> [<dev_path> <offset>]+
95 */
96static int stripe_ctr(struct dm_target *ti, unsigned int argc, char **argv)
97{
98	struct stripe_c *sc;
99	sector_t width;
100	uint32_t stripes;
101	uint32_t chunk_size;
102	int r;
103	unsigned int i;
104
105	if (argc < 2) {
106		ti->error = "Not enough arguments";
107		return -EINVAL;
108	}
109
110	if (kstrtouint(argv[0], 10, &stripes) || !stripes) {
111		ti->error = "Invalid stripe count";
112		return -EINVAL;
113	}
114
115	if (kstrtouint(argv[1], 10, &chunk_size)) {
116		ti->error = "Invalid chunk_size";
117		return -EINVAL;
118	}
119
120	/*
121	 * chunk_size is a power of two
122	 */
123	if (!is_power_of_2(chunk_size) ||
124	    (chunk_size < (PAGE_SIZE >> SECTOR_SHIFT))) {
125		ti->error = "Invalid chunk size";
126		return -EINVAL;
127	}
128
129	if (ti->len & (chunk_size - 1)) {
130		ti->error = "Target length not divisible by "
131		    "chunk size";
132		return -EINVAL;
133	}
134
135	width = ti->len;
136	if (sector_div(width, stripes)) {
137		ti->error = "Target length not divisible by "
138		    "number of stripes";
139		return -EINVAL;
140	}
141
142	/*
143	 * Do we have enough arguments for that many stripes ?
144	 */
145	if (argc != (2 + 2 * stripes)) {
146		ti->error = "Not enough destinations "
147			"specified";
148		return -EINVAL;
149	}
150
151	sc = alloc_context(stripes);
152	if (!sc) {
153		ti->error = "Memory allocation for striped context "
154		    "failed";
155		return -ENOMEM;
156	}
157
158	INIT_WORK(&sc->trigger_event, trigger_event);
159
160	/* Set pointer to dm target; used in trigger_event */
161	sc->ti = ti;
162	sc->stripes = stripes;
163	sc->stripe_width = width;
164
165	if (stripes & (stripes - 1))
166		sc->stripes_shift = -1;
167	else {
168		sc->stripes_shift = ffs(stripes) - 1;
169		sc->stripes_mask = ((sector_t) stripes) - 1;
170	}
171
172	ti->split_io = chunk_size;
173	ti->num_flush_requests = stripes;
174	ti->num_discard_requests = stripes;
175
176	sc->chunk_shift = ffs(chunk_size) - 1;
177	sc->chunk_mask = ((sector_t) chunk_size) - 1;
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) {
187			ti->error = "Couldn't parse stripe destination";
188			while (i--)
189				dm_put_device(ti, sc->stripe[i].dev);
190			kfree(sc);
191			return r;
192		}
193		atomic_set(&(sc->stripe[i].error_count), 0);
194	}
195
196	ti->private = sc;
197
198	return 0;
199}
200
201static 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
209	flush_work_sync(&sc->trigger_event);
210	kfree(sc);
211}
212
213static void stripe_map_sector(struct stripe_c *sc, sector_t sector,
214			      uint32_t *stripe, sector_t *result)
215{
216	sector_t offset = dm_target_offset(sc->ti, sector);
217	sector_t chunk = offset >> sc->chunk_shift;
218
219	if (sc->stripes_shift < 0)
220		*stripe = sector_div(chunk, sc->stripes);
221	else {
222		*stripe = chunk & sc->stripes_mask;
223		chunk >>= sc->stripes_shift;
224	}
225
226	*result = (chunk << sc->chunk_shift) | (offset & sc->chunk_mask);
227}
228
229static void stripe_map_range_sector(struct stripe_c *sc, sector_t sector,
230				    uint32_t target_stripe, sector_t *result)
231{
232	uint32_t stripe;
233
234	stripe_map_sector(sc, sector, &stripe, result);
235	if (stripe == target_stripe)
236		return;
237	*result &= ~sc->chunk_mask;			/* round down */
238	if (target_stripe < stripe)
239		*result += sc->chunk_mask + 1;		/* next chunk */
240}
241
242static int stripe_map_discard(struct stripe_c *sc, struct bio *bio,
243			      uint32_t target_stripe)
244{
245	sector_t begin, end;
246
247	stripe_map_range_sector(sc, bio->bi_sector, target_stripe, &begin);
248	stripe_map_range_sector(sc, bio->bi_sector + bio_sectors(bio),
249				target_stripe, &end);
250	if (begin < end) {
251		bio->bi_bdev = sc->stripe[target_stripe].dev->bdev;
252		bio->bi_sector = begin + sc->stripe[target_stripe].physical_start;
253		bio->bi_size = to_bytes(end - begin);
254		return DM_MAPIO_REMAPPED;
255	} else {
256		/* The range doesn't map to the target stripe */
257		bio_endio(bio, 0);
258		return DM_MAPIO_SUBMITTED;
259	}
260}
261
262static int stripe_map(struct dm_target *ti, struct bio *bio,
263		      union map_info *map_context)
264{
265	struct stripe_c *sc = ti->private;
266	uint32_t stripe;
267	unsigned target_request_nr;
268
269	if (bio->bi_rw & REQ_FLUSH) {
270		target_request_nr = map_context->target_request_nr;
271		BUG_ON(target_request_nr >= sc->stripes);
272		bio->bi_bdev = sc->stripe[target_request_nr].dev->bdev;
273		return DM_MAPIO_REMAPPED;
274	}
275	if (unlikely(bio->bi_rw & REQ_DISCARD)) {
276		target_request_nr = map_context->target_request_nr;
277		BUG_ON(target_request_nr >= sc->stripes);
278		return stripe_map_discard(sc, bio, target_request_nr);
279	}
280
281	stripe_map_sector(sc, bio->bi_sector, &stripe, &bio->bi_sector);
282
283	bio->bi_sector += sc->stripe[stripe].physical_start;
284	bio->bi_bdev = sc->stripe[stripe].dev->bdev;
285
286	return DM_MAPIO_REMAPPED;
287}
288
289/*
290 * Stripe status:
291 *
292 * INFO
293 * #stripes [stripe_name <stripe_name>] [group word count]
294 * [error count 'A|D' <error count 'A|D'>]
295 *
296 * TABLE
297 * #stripes [stripe chunk size]
298 * [stripe_name physical_start <stripe_name physical_start>]
299 *
300 */
301
302static int stripe_status(struct dm_target *ti,
303			 status_type_t type, char *result, unsigned int maxlen)
304{
305	struct stripe_c *sc = (struct stripe_c *) ti->private;
306	char buffer[sc->stripes + 1];
307	unsigned int sz = 0;
308	unsigned int i;
309
310	switch (type) {
311	case STATUSTYPE_INFO:
312		DMEMIT("%d ", sc->stripes);
313		for (i = 0; i < sc->stripes; i++)  {
314			DMEMIT("%s ", sc->stripe[i].dev->name);
315			buffer[i] = atomic_read(&(sc->stripe[i].error_count)) ?
316				'D' : 'A';
317		}
318		buffer[i] = '\0';
319		DMEMIT("1 %s", buffer);
320		break;
321
322	case STATUSTYPE_TABLE:
323		DMEMIT("%d %llu", sc->stripes,
324			(unsigned long long)sc->chunk_mask + 1);
325		for (i = 0; i < sc->stripes; i++)
326			DMEMIT(" %s %llu", sc->stripe[i].dev->name,
327			    (unsigned long long)sc->stripe[i].physical_start);
328		break;
329	}
330	return 0;
331}
332
333static int stripe_end_io(struct dm_target *ti, struct bio *bio,
334			 int error, union map_info *map_context)
335{
336	unsigned i;
337	char major_minor[16];
338	struct stripe_c *sc = ti->private;
339
340	if (!error)
341		return 0; /* I/O complete */
342
343	if ((error == -EWOULDBLOCK) && (bio->bi_rw & REQ_RAHEAD))
344		return error;
345
346	if (error == -EOPNOTSUPP)
347		return error;
348
349	memset(major_minor, 0, sizeof(major_minor));
350	sprintf(major_minor, "%d:%d",
351		MAJOR(disk_devt(bio->bi_bdev->bd_disk)),
352		MINOR(disk_devt(bio->bi_bdev->bd_disk)));
353
354	/*
355	 * Test to see which stripe drive triggered the event
356	 * and increment error count for all stripes on that device.
357	 * If the error count for a given device exceeds the threshold
358	 * value we will no longer trigger any further events.
359	 */
360	for (i = 0; i < sc->stripes; i++)
361		if (!strcmp(sc->stripe[i].dev->name, major_minor)) {
362			atomic_inc(&(sc->stripe[i].error_count));
363			if (atomic_read(&(sc->stripe[i].error_count)) <
364			    DM_IO_ERROR_THRESHOLD)
365				schedule_work(&sc->trigger_event);
366		}
367
368	return error;
369}
370
371static int stripe_iterate_devices(struct dm_target *ti,
372				  iterate_devices_callout_fn fn, void *data)
373{
374	struct stripe_c *sc = ti->private;
375	int ret = 0;
376	unsigned i = 0;
377
378	do {
379		ret = fn(ti, sc->stripe[i].dev,
380			 sc->stripe[i].physical_start,
381			 sc->stripe_width, data);
382	} while (!ret && ++i < sc->stripes);
383
384	return ret;
385}
386
387static void stripe_io_hints(struct dm_target *ti,
388			    struct queue_limits *limits)
389{
390	struct stripe_c *sc = ti->private;
391	unsigned chunk_size = (sc->chunk_mask + 1) << 9;
392
393	blk_limits_io_min(limits, chunk_size);
394	blk_limits_io_opt(limits, chunk_size * sc->stripes);
395}
396
397static int stripe_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
398			struct bio_vec *biovec, int max_size)
399{
400	struct stripe_c *sc = ti->private;
401	sector_t bvm_sector = bvm->bi_sector;
402	uint32_t stripe;
403	struct request_queue *q;
404
405	stripe_map_sector(sc, bvm_sector, &stripe, &bvm_sector);
406
407	q = bdev_get_queue(sc->stripe[stripe].dev->bdev);
408	if (!q->merge_bvec_fn)
409		return max_size;
410
411	bvm->bi_bdev = sc->stripe[stripe].dev->bdev;
412	bvm->bi_sector = sc->stripe[stripe].physical_start + bvm_sector;
413
414	return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
415}
416
417static struct target_type stripe_target = {
418	.name   = "striped",
419	.version = {1, 4, 0},
420	.module = THIS_MODULE,
421	.ctr    = stripe_ctr,
422	.dtr    = stripe_dtr,
423	.map    = stripe_map,
424	.end_io = stripe_end_io,
425	.status = stripe_status,
426	.iterate_devices = stripe_iterate_devices,
427	.io_hints = stripe_io_hints,
428	.merge  = stripe_merge,
429};
430
431int __init dm_stripe_init(void)
432{
433	int r;
434
435	r = dm_register_target(&stripe_target);
436	if (r < 0) {
437		DMWARN("target registration failed");
438		return r;
439	}
440
441	return r;
442}
443
444void dm_stripe_exit(void)
445{
446	dm_unregister_target(&stripe_target);
447}
448