dm-stripe.c revision f14fa693c93078444b5e95d7cad78ead0383ad50 
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	if (sector_div(width, stripes)) {
136		ti->error = "Target length not divisible by "
137		    "number of stripes";
138		return -EINVAL;
139	}
140
141	/*
142	 * Do we have enough arguments for that many stripes ?
143	 */
144	if (argc != (2 + 2 * stripes)) {
145		ti->error = "Not enough destinations "
146			"specified";
147		return -EINVAL;
148	}
149
150	sc = alloc_context(stripes);
151	if (!sc) {
152		ti->error = "Memory allocation for striped context "
153		    "failed";
154		return -ENOMEM;
155	}
156
157	INIT_WORK(&sc->trigger_event, trigger_event);
158
159	/* Set pointer to dm target; used in trigger_event */
160	sc->ti = ti;
161	sc->stripes = stripes;
162	sc->stripe_width = width;
163
164	if (stripes & (stripes - 1))
165		sc->stripes_shift = -1;
166	else {
167		sc->stripes_shift = ffs(stripes) - 1;
168		sc->stripes_mask = ((sector_t) stripes) - 1;
169	}
170
171	ti->split_io = chunk_size;
172	ti->num_flush_requests = stripes;
173	ti->num_discard_requests = stripes;
174
175	sc->chunk_shift = ffs(chunk_size) - 1;
176	sc->chunk_mask = ((sector_t) chunk_size) - 1;
177
178	/*
179	 * Get the stripe destinations.
180	 */
181	for (i = 0; i < stripes; i++) {
182		argv += 2;
183
184		r = get_stripe(ti, sc, i, argv);
185		if (r < 0) {
186			ti->error = "Couldn't parse stripe destination";
187			while (i--)
188				dm_put_device(ti, sc->stripe[i].dev);
189			kfree(sc);
190			return r;
191		}
192		atomic_set(&(sc->stripe[i].error_count), 0);
193	}
194
195	ti->private = sc;
196
197	return 0;
198}
199
200static void stripe_dtr(struct dm_target *ti)
201{
202	unsigned int i;
203	struct stripe_c *sc = (struct stripe_c *) ti->private;
204
205	for (i = 0; i < sc->stripes; i++)
206		dm_put_device(ti, sc->stripe[i].dev);
207
208	flush_work_sync(&sc->trigger_event);
209	kfree(sc);
210}
211
212static void stripe_map_sector(struct stripe_c *sc, sector_t sector,
213			      uint32_t *stripe, sector_t *result)
214{
215	sector_t offset = dm_target_offset(sc->ti, sector);
216	sector_t chunk = offset >> sc->chunk_shift;
217
218	if (sc->stripes_shift < 0)
219		*stripe = sector_div(chunk, sc->stripes);
220	else {
221		*stripe = chunk & sc->stripes_mask;
222		chunk >>= sc->stripes_shift;
223	}
224
225	*result = (chunk << sc->chunk_shift) | (offset & sc->chunk_mask);
226}
227
228static void stripe_map_range_sector(struct stripe_c *sc, sector_t sector,
229				    uint32_t target_stripe, sector_t *result)
230{
231	uint32_t stripe;
232
233	stripe_map_sector(sc, sector, &stripe, result);
234	if (stripe == target_stripe)
235		return;
236	*result &= ~sc->chunk_mask;			/* round down */
237	if (target_stripe < stripe)
238		*result += sc->chunk_mask + 1;		/* next chunk */
239}
240
241static int stripe_map_discard(struct stripe_c *sc, struct bio *bio,
242			      uint32_t target_stripe)
243{
244	sector_t begin, end;
245
246	stripe_map_range_sector(sc, bio->bi_sector, target_stripe, &begin);
247	stripe_map_range_sector(sc, bio->bi_sector + bio_sectors(bio),
248				target_stripe, &end);
249	if (begin < end) {
250		bio->bi_bdev = sc->stripe[target_stripe].dev->bdev;
251		bio->bi_sector = begin + sc->stripe[target_stripe].physical_start;
252		bio->bi_size = to_bytes(end - begin);
253		return DM_MAPIO_REMAPPED;
254	} else {
255		/* The range doesn't map to the target stripe */
256		bio_endio(bio, 0);
257		return DM_MAPIO_SUBMITTED;
258	}
259}
260
261static int stripe_map(struct dm_target *ti, struct bio *bio,
262		      union map_info *map_context)
263{
264	struct stripe_c *sc = ti->private;
265	uint32_t stripe;
266	unsigned target_request_nr;
267
268	if (bio->bi_rw & REQ_FLUSH) {
269		target_request_nr = map_context->target_request_nr;
270		BUG_ON(target_request_nr >= sc->stripes);
271		bio->bi_bdev = sc->stripe[target_request_nr].dev->bdev;
272		return DM_MAPIO_REMAPPED;
273	}
274	if (unlikely(bio->bi_rw & REQ_DISCARD)) {
275		target_request_nr = map_context->target_request_nr;
276		BUG_ON(target_request_nr >= sc->stripes);
277		return stripe_map_discard(sc, bio, target_request_nr);
278	}
279
280	stripe_map_sector(sc, bio->bi_sector, &stripe, &bio->bi_sector);
281
282	bio->bi_sector += sc->stripe[stripe].physical_start;
283	bio->bi_bdev = sc->stripe[stripe].dev->bdev;
284
285	return DM_MAPIO_REMAPPED;
286}
287
288/*
289 * Stripe status:
290 *
291 * INFO
292 * #stripes [stripe_name <stripe_name>] [group word count]
293 * [error count 'A|D' <error count 'A|D'>]
294 *
295 * TABLE
296 * #stripes [stripe chunk size]
297 * [stripe_name physical_start <stripe_name physical_start>]
298 *
299 */
300
301static int stripe_status(struct dm_target *ti,
302			 status_type_t type, char *result, unsigned int maxlen)
303{
304	struct stripe_c *sc = (struct stripe_c *) ti->private;
305	char buffer[sc->stripes + 1];
306	unsigned int sz = 0;
307	unsigned int i;
308
309	switch (type) {
310	case STATUSTYPE_INFO:
311		DMEMIT("%d ", sc->stripes);
312		for (i = 0; i < sc->stripes; i++)  {
313			DMEMIT("%s ", sc->stripe[i].dev->name);
314			buffer[i] = atomic_read(&(sc->stripe[i].error_count)) ?
315				'D' : 'A';
316		}
317		buffer[i] = '\0';
318		DMEMIT("1 %s", buffer);
319		break;
320
321	case STATUSTYPE_TABLE:
322		DMEMIT("%d %llu", sc->stripes,
323			(unsigned long long)sc->chunk_mask + 1);
324		for (i = 0; i < sc->stripes; i++)
325			DMEMIT(" %s %llu", sc->stripe[i].dev->name,
326			    (unsigned long long)sc->stripe[i].physical_start);
327		break;
328	}
329	return 0;
330}
331
332static int stripe_end_io(struct dm_target *ti, struct bio *bio,
333			 int error, union map_info *map_context)
334{
335	unsigned i;
336	char major_minor[16];
337	struct stripe_c *sc = ti->private;
338
339	if (!error)
340		return 0; /* I/O complete */
341
342	if ((error == -EWOULDBLOCK) && (bio->bi_rw & REQ_RAHEAD))
343		return error;
344
345	if (error == -EOPNOTSUPP)
346		return error;
347
348	memset(major_minor, 0, sizeof(major_minor));
349	sprintf(major_minor, "%d:%d",
350		MAJOR(disk_devt(bio->bi_bdev->bd_disk)),
351		MINOR(disk_devt(bio->bi_bdev->bd_disk)));
352
353	/*
354	 * Test to see which stripe drive triggered the event
355	 * and increment error count for all stripes on that device.
356	 * If the error count for a given device exceeds the threshold
357	 * value we will no longer trigger any further events.
358	 */
359	for (i = 0; i < sc->stripes; i++)
360		if (!strcmp(sc->stripe[i].dev->name, major_minor)) {
361			atomic_inc(&(sc->stripe[i].error_count));
362			if (atomic_read(&(sc->stripe[i].error_count)) <
363			    DM_IO_ERROR_THRESHOLD)
364				schedule_work(&sc->trigger_event);
365		}
366
367	return error;
368}
369
370static int stripe_iterate_devices(struct dm_target *ti,
371				  iterate_devices_callout_fn fn, void *data)
372{
373	struct stripe_c *sc = ti->private;
374	int ret = 0;
375	unsigned i = 0;
376
377	do {
378		ret = fn(ti, sc->stripe[i].dev,
379			 sc->stripe[i].physical_start,
380			 sc->stripe_width, data);
381	} while (!ret && ++i < sc->stripes);
382
383	return ret;
384}
385
386static void stripe_io_hints(struct dm_target *ti,
387			    struct queue_limits *limits)
388{
389	struct stripe_c *sc = ti->private;
390	unsigned chunk_size = (sc->chunk_mask + 1) << 9;
391
392	blk_limits_io_min(limits, chunk_size);
393	blk_limits_io_opt(limits, chunk_size * sc->stripes);
394}
395
396static int stripe_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
397			struct bio_vec *biovec, int max_size)
398{
399	struct stripe_c *sc = ti->private;
400	sector_t bvm_sector = bvm->bi_sector;
401	uint32_t stripe;
402	struct request_queue *q;
403
404	stripe_map_sector(sc, bvm_sector, &stripe, &bvm_sector);
405
406	q = bdev_get_queue(sc->stripe[stripe].dev->bdev);
407	if (!q->merge_bvec_fn)
408		return max_size;
409
410	bvm->bi_bdev = sc->stripe[stripe].dev->bdev;
411	bvm->bi_sector = sc->stripe[stripe].physical_start + bvm_sector;
412
413	return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
414}
415
416static struct target_type stripe_target = {
417	.name   = "striped",
418	.version = {1, 4, 0},
419	.module = THIS_MODULE,
420	.ctr    = stripe_ctr,
421	.dtr    = stripe_dtr,
422	.map    = stripe_map,
423	.end_io = stripe_end_io,
424	.status = stripe_status,
425	.iterate_devices = stripe_iterate_devices,
426	.io_hints = stripe_io_hints,
427	.merge  = stripe_merge,
428};
429
430int __init dm_stripe_init(void)
431{
432	int r;
433
434	r = dm_register_target(&stripe_target);
435	if (r < 0) {
436		DMWARN("target registration failed");
437		return r;
438	}
439
440	return r;
441}
442
443void dm_stripe_exit(void)
444{
445	dm_unregister_target(&stripe_target);
446}
447