dm-stripe.c revision eb850de608cc22e0199b3797cd5c0076bae6cda0 
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) ||
113	    (chunk_size < (PAGE_SIZE >> SECTOR_SHIFT))) {
114		ti->error = "Invalid chunk_size";
115		return -EINVAL;
116	}
117
118	width = ti->len;
119	if (sector_div(width, chunk_size)) {
120		ti->error = "Target length not divisible by "
121		    "chunk size";
122		return -EINVAL;
123	}
124
125	if (sector_div(width, stripes)) {
126		ti->error = "Target length not divisible by "
127		    "number of stripes";
128		return -EINVAL;
129	}
130
131	/*
132	 * Do we have enough arguments for that many stripes ?
133	 */
134	if (argc != (2 + 2 * stripes)) {
135		ti->error = "Not enough destinations "
136			"specified";
137		return -EINVAL;
138	}
139
140	sc = alloc_context(stripes);
141	if (!sc) {
142		ti->error = "Memory allocation for striped context "
143		    "failed";
144		return -ENOMEM;
145	}
146
147	INIT_WORK(&sc->trigger_event, trigger_event);
148
149	/* Set pointer to dm target; used in trigger_event */
150	sc->ti = ti;
151	sc->stripes = stripes;
152	sc->stripe_width = width;
153
154	if (stripes & (stripes - 1))
155		sc->stripes_shift = -1;
156	else
157		sc->stripes_shift = __ffs(stripes);
158
159	r = dm_set_target_max_io_len(ti, chunk_size);
160	if (r)
161		return r;
162
163	ti->num_flush_requests = stripes;
164	ti->num_discard_requests = stripes;
165
166	sc->chunk_size = chunk_size;
167
168	/*
169	 * Get the stripe destinations.
170	 */
171	for (i = 0; i < stripes; i++) {
172		argv += 2;
173
174		r = get_stripe(ti, sc, i, argv);
175		if (r < 0) {
176			ti->error = "Couldn't parse stripe destination";
177			while (i--)
178				dm_put_device(ti, sc->stripe[i].dev);
179			kfree(sc);
180			return r;
181		}
182		atomic_set(&(sc->stripe[i].error_count), 0);
183	}
184
185	ti->private = sc;
186
187	return 0;
188}
189
190static void stripe_dtr(struct dm_target *ti)
191{
192	unsigned int i;
193	struct stripe_c *sc = (struct stripe_c *) ti->private;
194
195	for (i = 0; i < sc->stripes; i++)
196		dm_put_device(ti, sc->stripe[i].dev);
197
198	flush_work_sync(&sc->trigger_event);
199	kfree(sc);
200}
201
202static void stripe_map_sector(struct stripe_c *sc, sector_t sector,
203			      uint32_t *stripe, sector_t *result)
204{
205	sector_t chunk = dm_target_offset(sc->ti, sector);
206	sector_t chunk_offset = sector_div(chunk, sc->chunk_size);
207
208	if (sc->stripes_shift < 0)
209		*stripe = sector_div(chunk, sc->stripes);
210	else {
211		*stripe = chunk & (sc->stripes - 1);
212		chunk >>= sc->stripes_shift;
213	}
214
215	*result = (chunk * sc->chunk_size) + chunk_offset;
216}
217
218static void stripe_map_range_sector(struct stripe_c *sc, sector_t sector,
219				    uint32_t target_stripe, sector_t *result)
220{
221	uint32_t stripe;
222
223	stripe_map_sector(sc, sector, &stripe, result);
224	if (stripe == target_stripe)
225		return;
226
227	/* round down */
228	sector = *result;
229	*result -= sector_div(sector, sc->chunk_size);
230
231	if (target_stripe < stripe)
232		*result += sc->chunk_size;		/* next chunk */
233}
234
235static int stripe_map_discard(struct stripe_c *sc, struct bio *bio,
236			      uint32_t target_stripe)
237{
238	sector_t begin, end;
239
240	stripe_map_range_sector(sc, bio->bi_sector, target_stripe, &begin);
241	stripe_map_range_sector(sc, bio->bi_sector + bio_sectors(bio),
242				target_stripe, &end);
243	if (begin < end) {
244		bio->bi_bdev = sc->stripe[target_stripe].dev->bdev;
245		bio->bi_sector = begin + sc->stripe[target_stripe].physical_start;
246		bio->bi_size = to_bytes(end - begin);
247		return DM_MAPIO_REMAPPED;
248	} else {
249		/* The range doesn't map to the target stripe */
250		bio_endio(bio, 0);
251		return DM_MAPIO_SUBMITTED;
252	}
253}
254
255static int stripe_map(struct dm_target *ti, struct bio *bio,
256		      union map_info *map_context)
257{
258	struct stripe_c *sc = ti->private;
259	uint32_t stripe;
260	unsigned target_request_nr;
261
262	if (bio->bi_rw & REQ_FLUSH) {
263		target_request_nr = map_context->target_request_nr;
264		BUG_ON(target_request_nr >= sc->stripes);
265		bio->bi_bdev = sc->stripe[target_request_nr].dev->bdev;
266		return DM_MAPIO_REMAPPED;
267	}
268	if (unlikely(bio->bi_rw & REQ_DISCARD)) {
269		target_request_nr = map_context->target_request_nr;
270		BUG_ON(target_request_nr >= sc->stripes);
271		return stripe_map_discard(sc, bio, target_request_nr);
272	}
273
274	stripe_map_sector(sc, bio->bi_sector, &stripe, &bio->bi_sector);
275
276	bio->bi_sector += sc->stripe[stripe].physical_start;
277	bio->bi_bdev = sc->stripe[stripe].dev->bdev;
278
279	return DM_MAPIO_REMAPPED;
280}
281
282/*
283 * Stripe status:
284 *
285 * INFO
286 * #stripes [stripe_name <stripe_name>] [group word count]
287 * [error count 'A|D' <error count 'A|D'>]
288 *
289 * TABLE
290 * #stripes [stripe chunk size]
291 * [stripe_name physical_start <stripe_name physical_start>]
292 *
293 */
294
295static int stripe_status(struct dm_target *ti,
296			 status_type_t type, char *result, unsigned int maxlen)
297{
298	struct stripe_c *sc = (struct stripe_c *) ti->private;
299	char buffer[sc->stripes + 1];
300	unsigned int sz = 0;
301	unsigned int i;
302
303	switch (type) {
304	case STATUSTYPE_INFO:
305		DMEMIT("%d ", sc->stripes);
306		for (i = 0; i < sc->stripes; i++)  {
307			DMEMIT("%s ", sc->stripe[i].dev->name);
308			buffer[i] = atomic_read(&(sc->stripe[i].error_count)) ?
309				'D' : 'A';
310		}
311		buffer[i] = '\0';
312		DMEMIT("1 %s", buffer);
313		break;
314
315	case STATUSTYPE_TABLE:
316		DMEMIT("%d %llu", sc->stripes,
317			(unsigned long long)sc->chunk_size);
318		for (i = 0; i < sc->stripes; i++)
319			DMEMIT(" %s %llu", sc->stripe[i].dev->name,
320			    (unsigned long long)sc->stripe[i].physical_start);
321		break;
322	}
323	return 0;
324}
325
326static int stripe_end_io(struct dm_target *ti, struct bio *bio,
327			 int error, union map_info *map_context)
328{
329	unsigned i;
330	char major_minor[16];
331	struct stripe_c *sc = ti->private;
332
333	if (!error)
334		return 0; /* I/O complete */
335
336	if ((error == -EWOULDBLOCK) && (bio->bi_rw & REQ_RAHEAD))
337		return error;
338
339	if (error == -EOPNOTSUPP)
340		return error;
341
342	memset(major_minor, 0, sizeof(major_minor));
343	sprintf(major_minor, "%d:%d",
344		MAJOR(disk_devt(bio->bi_bdev->bd_disk)),
345		MINOR(disk_devt(bio->bi_bdev->bd_disk)));
346
347	/*
348	 * Test to see which stripe drive triggered the event
349	 * and increment error count for all stripes on that device.
350	 * If the error count for a given device exceeds the threshold
351	 * value we will no longer trigger any further events.
352	 */
353	for (i = 0; i < sc->stripes; i++)
354		if (!strcmp(sc->stripe[i].dev->name, major_minor)) {
355			atomic_inc(&(sc->stripe[i].error_count));
356			if (atomic_read(&(sc->stripe[i].error_count)) <
357			    DM_IO_ERROR_THRESHOLD)
358				schedule_work(&sc->trigger_event);
359		}
360
361	return error;
362}
363
364static int stripe_iterate_devices(struct dm_target *ti,
365				  iterate_devices_callout_fn fn, void *data)
366{
367	struct stripe_c *sc = ti->private;
368	int ret = 0;
369	unsigned i = 0;
370
371	do {
372		ret = fn(ti, sc->stripe[i].dev,
373			 sc->stripe[i].physical_start,
374			 sc->stripe_width, data);
375	} while (!ret && ++i < sc->stripes);
376
377	return ret;
378}
379
380static void stripe_io_hints(struct dm_target *ti,
381			    struct queue_limits *limits)
382{
383	struct stripe_c *sc = ti->private;
384	unsigned chunk_size = sc->chunk_size << SECTOR_SHIFT;
385
386	blk_limits_io_min(limits, chunk_size);
387	blk_limits_io_opt(limits, chunk_size * sc->stripes);
388}
389
390static int stripe_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
391			struct bio_vec *biovec, int max_size)
392{
393	struct stripe_c *sc = ti->private;
394	sector_t bvm_sector = bvm->bi_sector;
395	uint32_t stripe;
396	struct request_queue *q;
397
398	stripe_map_sector(sc, bvm_sector, &stripe, &bvm_sector);
399
400	q = bdev_get_queue(sc->stripe[stripe].dev->bdev);
401	if (!q->merge_bvec_fn)
402		return max_size;
403
404	bvm->bi_bdev = sc->stripe[stripe].dev->bdev;
405	bvm->bi_sector = sc->stripe[stripe].physical_start + bvm_sector;
406
407	return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
408}
409
410static struct target_type stripe_target = {
411	.name   = "striped",
412	.version = {1, 5, 0},
413	.module = THIS_MODULE,
414	.ctr    = stripe_ctr,
415	.dtr    = stripe_dtr,
416	.map    = stripe_map,
417	.end_io = stripe_end_io,
418	.status = stripe_status,
419	.iterate_devices = stripe_iterate_devices,
420	.io_hints = stripe_io_hints,
421	.merge  = stripe_merge,
422};
423
424int __init dm_stripe_init(void)
425{
426	int r;
427
428	r = dm_register_target(&stripe_target);
429	if (r < 0) {
430		DMWARN("target registration failed");
431		return r;
432	}
433
434	return r;
435}
436
437void dm_stripe_exit(void)
438{
439	dm_unregister_target(&stripe_target);
440}
441