dm-stripe.c revision d793e684277124d55c5d2444007e224635821346 
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	int chunk_size_shift;
35
36	/* Needed for handling events */
37	struct dm_target *ti;
38
39	/* Work struct used for triggering events*/
40	struct work_struct trigger_event;
41
42	struct stripe stripe[0];
43};
44
45/*
46 * An event is triggered whenever a drive
47 * drops out of a stripe volume.
48 */
49static void trigger_event(struct work_struct *work)
50{
51	struct stripe_c *sc = container_of(work, struct stripe_c,
52					   trigger_event);
53	dm_table_event(sc->ti->table);
54}
55
56static inline struct stripe_c *alloc_context(unsigned int stripes)
57{
58	size_t len;
59
60	if (dm_array_too_big(sizeof(struct stripe_c), sizeof(struct stripe),
61			     stripes))
62		return NULL;
63
64	len = sizeof(struct stripe_c) + (sizeof(struct stripe) * stripes);
65
66	return kmalloc(len, GFP_KERNEL);
67}
68
69/*
70 * Parse a single <dev> <sector> pair
71 */
72static int get_stripe(struct dm_target *ti, struct stripe_c *sc,
73		      unsigned int stripe, char **argv)
74{
75	unsigned long long start;
76	char dummy;
77
78	if (sscanf(argv[1], "%llu%c", &start, &dummy) != 1)
79		return -EINVAL;
80
81	if (dm_get_device(ti, argv[0], dm_table_get_mode(ti->table),
82			  &sc->stripe[stripe].dev))
83		return -ENXIO;
84
85	sc->stripe[stripe].physical_start = start;
86
87	return 0;
88}
89
90/*
91 * Construct a striped mapping.
92 * <number of stripes> <chunk size> [<dev_path> <offset>]+
93 */
94static int stripe_ctr(struct dm_target *ti, unsigned int argc, char **argv)
95{
96	struct stripe_c *sc;
97	sector_t width, tmp_len;
98	uint32_t stripes;
99	uint32_t chunk_size;
100	int r;
101	unsigned int i;
102
103	if (argc < 2) {
104		ti->error = "Not enough arguments";
105		return -EINVAL;
106	}
107
108	if (kstrtouint(argv[0], 10, &stripes) || !stripes) {
109		ti->error = "Invalid stripe count";
110		return -EINVAL;
111	}
112
113	if (kstrtouint(argv[1], 10, &chunk_size) || !chunk_size) {
114		ti->error = "Invalid chunk_size";
115		return -EINVAL;
116	}
117
118	width = ti->len;
119	if (sector_div(width, stripes)) {
120		ti->error = "Target length not divisible by "
121		    "number of stripes";
122		return -EINVAL;
123	}
124
125	tmp_len = width;
126	if (sector_div(tmp_len, chunk_size)) {
127		ti->error = "Target length not divisible by "
128		    "chunk size";
129		return -EINVAL;
130	}
131
132	/*
133	 * Do we have enough arguments for that many stripes ?
134	 */
135	if (argc != (2 + 2 * stripes)) {
136		ti->error = "Not enough destinations "
137			"specified";
138		return -EINVAL;
139	}
140
141	sc = alloc_context(stripes);
142	if (!sc) {
143		ti->error = "Memory allocation for striped context "
144		    "failed";
145		return -ENOMEM;
146	}
147
148	INIT_WORK(&sc->trigger_event, trigger_event);
149
150	/* Set pointer to dm target; used in trigger_event */
151	sc->ti = ti;
152	sc->stripes = stripes;
153	sc->stripe_width = width;
154
155	if (stripes & (stripes - 1))
156		sc->stripes_shift = -1;
157	else
158		sc->stripes_shift = __ffs(stripes);
159
160	r = dm_set_target_max_io_len(ti, chunk_size);
161	if (r)
162		return r;
163
164	ti->num_flush_bios = stripes;
165	ti->num_discard_bios = stripes;
166	ti->num_write_same_bios = stripes;
167
168	sc->chunk_size = chunk_size;
169	if (chunk_size & (chunk_size - 1))
170		sc->chunk_size_shift = -1;
171	else
172		sc->chunk_size_shift = __ffs(chunk_size);
173
174	/*
175	 * Get the stripe destinations.
176	 */
177	for (i = 0; i < stripes; i++) {
178		argv += 2;
179
180		r = get_stripe(ti, sc, i, argv);
181		if (r < 0) {
182			ti->error = "Couldn't parse stripe destination";
183			while (i--)
184				dm_put_device(ti, sc->stripe[i].dev);
185			kfree(sc);
186			return r;
187		}
188		atomic_set(&(sc->stripe[i].error_count), 0);
189	}
190
191	ti->private = sc;
192
193	return 0;
194}
195
196static void stripe_dtr(struct dm_target *ti)
197{
198	unsigned int i;
199	struct stripe_c *sc = (struct stripe_c *) ti->private;
200
201	for (i = 0; i < sc->stripes; i++)
202		dm_put_device(ti, sc->stripe[i].dev);
203
204	flush_work(&sc->trigger_event);
205	kfree(sc);
206}
207
208static void stripe_map_sector(struct stripe_c *sc, sector_t sector,
209			      uint32_t *stripe, sector_t *result)
210{
211	sector_t chunk = dm_target_offset(sc->ti, sector);
212	sector_t chunk_offset;
213
214	if (sc->chunk_size_shift < 0)
215		chunk_offset = sector_div(chunk, sc->chunk_size);
216	else {
217		chunk_offset = chunk & (sc->chunk_size - 1);
218		chunk >>= sc->chunk_size_shift;
219	}
220
221	if (sc->stripes_shift < 0)
222		*stripe = sector_div(chunk, sc->stripes);
223	else {
224		*stripe = chunk & (sc->stripes - 1);
225		chunk >>= sc->stripes_shift;
226	}
227
228	if (sc->chunk_size_shift < 0)
229		chunk *= sc->chunk_size;
230	else
231		chunk <<= sc->chunk_size_shift;
232
233	*result = chunk + chunk_offset;
234}
235
236static void stripe_map_range_sector(struct stripe_c *sc, sector_t sector,
237				    uint32_t target_stripe, sector_t *result)
238{
239	uint32_t stripe;
240
241	stripe_map_sector(sc, sector, &stripe, result);
242	if (stripe == target_stripe)
243		return;
244
245	/* round down */
246	sector = *result;
247	if (sc->chunk_size_shift < 0)
248		*result -= sector_div(sector, sc->chunk_size);
249	else
250		*result = sector & ~(sector_t)(sc->chunk_size - 1);
251
252	if (target_stripe < stripe)
253		*result += sc->chunk_size;		/* next chunk */
254}
255
256static int stripe_map_range(struct stripe_c *sc, struct bio *bio,
257			    uint32_t target_stripe)
258{
259	sector_t begin, end;
260
261	stripe_map_range_sector(sc, bio->bi_sector, target_stripe, &begin);
262	stripe_map_range_sector(sc, bio_end_sector(bio),
263				target_stripe, &end);
264	if (begin < end) {
265		bio->bi_bdev = sc->stripe[target_stripe].dev->bdev;
266		bio->bi_sector = begin + sc->stripe[target_stripe].physical_start;
267		bio->bi_size = to_bytes(end - begin);
268		return DM_MAPIO_REMAPPED;
269	} else {
270		/* The range doesn't map to the target stripe */
271		bio_endio(bio, 0);
272		return DM_MAPIO_SUBMITTED;
273	}
274}
275
276static int stripe_map(struct dm_target *ti, struct bio *bio)
277{
278	struct stripe_c *sc = ti->private;
279	uint32_t stripe;
280	unsigned target_bio_nr;
281
282	if (bio->bi_rw & REQ_FLUSH) {
283		target_bio_nr = dm_bio_get_target_bio_nr(bio);
284		BUG_ON(target_bio_nr >= sc->stripes);
285		bio->bi_bdev = sc->stripe[target_bio_nr].dev->bdev;
286		return DM_MAPIO_REMAPPED;
287	}
288	if (unlikely(bio->bi_rw & REQ_DISCARD) ||
289	    unlikely(bio->bi_rw & REQ_WRITE_SAME)) {
290		target_bio_nr = dm_bio_get_target_bio_nr(bio);
291		BUG_ON(target_bio_nr >= sc->stripes);
292		return stripe_map_range(sc, bio, target_bio_nr);
293	}
294
295	stripe_map_sector(sc, bio->bi_sector, &stripe, &bio->bi_sector);
296
297	bio->bi_sector += sc->stripe[stripe].physical_start;
298	bio->bi_bdev = sc->stripe[stripe].dev->bdev;
299
300	return DM_MAPIO_REMAPPED;
301}
302
303/*
304 * Stripe status:
305 *
306 * INFO
307 * #stripes [stripe_name <stripe_name>] [group word count]
308 * [error count 'A|D' <error count 'A|D'>]
309 *
310 * TABLE
311 * #stripes [stripe chunk size]
312 * [stripe_name physical_start <stripe_name physical_start>]
313 *
314 */
315
316static void stripe_status(struct dm_target *ti, status_type_t type,
317			  unsigned status_flags, char *result, unsigned maxlen)
318{
319	struct stripe_c *sc = (struct stripe_c *) ti->private;
320	char buffer[sc->stripes + 1];
321	unsigned int sz = 0;
322	unsigned int i;
323
324	switch (type) {
325	case STATUSTYPE_INFO:
326		DMEMIT("%d ", sc->stripes);
327		for (i = 0; i < sc->stripes; i++)  {
328			DMEMIT("%s ", sc->stripe[i].dev->name);
329			buffer[i] = atomic_read(&(sc->stripe[i].error_count)) ?
330				'D' : 'A';
331		}
332		buffer[i] = '\0';
333		DMEMIT("1 %s", buffer);
334		break;
335
336	case STATUSTYPE_TABLE:
337		DMEMIT("%d %llu", sc->stripes,
338			(unsigned long long)sc->chunk_size);
339		for (i = 0; i < sc->stripes; i++)
340			DMEMIT(" %s %llu", sc->stripe[i].dev->name,
341			    (unsigned long long)sc->stripe[i].physical_start);
342		break;
343	}
344}
345
346static int stripe_end_io(struct dm_target *ti, struct bio *bio, int error)
347{
348	unsigned i;
349	char major_minor[16];
350	struct stripe_c *sc = ti->private;
351
352	if (!error)
353		return 0; /* I/O complete */
354
355	if ((error == -EWOULDBLOCK) && (bio->bi_rw & REQ_RAHEAD))
356		return error;
357
358	if (error == -EOPNOTSUPP)
359		return error;
360
361	memset(major_minor, 0, sizeof(major_minor));
362	sprintf(major_minor, "%d:%d",
363		MAJOR(disk_devt(bio->bi_bdev->bd_disk)),
364		MINOR(disk_devt(bio->bi_bdev->bd_disk)));
365
366	/*
367	 * Test to see which stripe drive triggered the event
368	 * and increment error count for all stripes on that device.
369	 * If the error count for a given device exceeds the threshold
370	 * value we will no longer trigger any further events.
371	 */
372	for (i = 0; i < sc->stripes; i++)
373		if (!strcmp(sc->stripe[i].dev->name, major_minor)) {
374			atomic_inc(&(sc->stripe[i].error_count));
375			if (atomic_read(&(sc->stripe[i].error_count)) <
376			    DM_IO_ERROR_THRESHOLD)
377				schedule_work(&sc->trigger_event);
378		}
379
380	return error;
381}
382
383static int stripe_iterate_devices(struct dm_target *ti,
384				  iterate_devices_callout_fn fn, void *data)
385{
386	struct stripe_c *sc = ti->private;
387	int ret = 0;
388	unsigned i = 0;
389
390	do {
391		ret = fn(ti, sc->stripe[i].dev,
392			 sc->stripe[i].physical_start,
393			 sc->stripe_width, data);
394	} while (!ret && ++i < sc->stripes);
395
396	return ret;
397}
398
399static void stripe_io_hints(struct dm_target *ti,
400			    struct queue_limits *limits)
401{
402	struct stripe_c *sc = ti->private;
403	unsigned chunk_size = sc->chunk_size << SECTOR_SHIFT;
404
405	blk_limits_io_min(limits, chunk_size);
406	blk_limits_io_opt(limits, chunk_size * sc->stripes);
407}
408
409static int stripe_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
410			struct bio_vec *biovec, int max_size)
411{
412	struct stripe_c *sc = ti->private;
413	sector_t bvm_sector = bvm->bi_sector;
414	uint32_t stripe;
415	struct request_queue *q;
416
417	stripe_map_sector(sc, bvm_sector, &stripe, &bvm_sector);
418
419	q = bdev_get_queue(sc->stripe[stripe].dev->bdev);
420	if (!q->merge_bvec_fn)
421		return max_size;
422
423	bvm->bi_bdev = sc->stripe[stripe].dev->bdev;
424	bvm->bi_sector = sc->stripe[stripe].physical_start + bvm_sector;
425
426	return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
427}
428
429static struct target_type stripe_target = {
430	.name   = "striped",
431	.version = {1, 5, 1},
432	.module = THIS_MODULE,
433	.ctr    = stripe_ctr,
434	.dtr    = stripe_dtr,
435	.map    = stripe_map,
436	.end_io = stripe_end_io,
437	.status = stripe_status,
438	.iterate_devices = stripe_iterate_devices,
439	.io_hints = stripe_io_hints,
440	.merge  = stripe_merge,
441};
442
443int __init dm_stripe_init(void)
444{
445	int r;
446
447	r = dm_register_target(&stripe_target);
448	if (r < 0) {
449		DMWARN("target registration failed");
450		return r;
451	}
452
453	return r;
454}
455
456void dm_stripe_exit(void)
457{
458	dm_unregister_target(&stripe_target);
459}
460