dm-stripe.c revision 1df05483d758ea43abc375869fbe06be506ba827 
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	/* stripe chunk size */
34	uint32_t chunk_shift;
35	sector_t chunk_mask;
36
37	/* Needed for handling events */
38	struct dm_target *ti;
39
40	/* Work struct used for triggering events*/
41	struct work_struct trigger_event;
42
43	struct stripe stripe[0];
44};
45
46/*
47 * An event is triggered whenever a drive
48 * drops out of a stripe volume.
49 */
50static void trigger_event(struct work_struct *work)
51{
52	struct stripe_c *sc = container_of(work, struct stripe_c,
53					   trigger_event);
54	dm_table_event(sc->ti->table);
55}
56
57static inline struct stripe_c *alloc_context(unsigned int stripes)
58{
59	size_t len;
60
61	if (dm_array_too_big(sizeof(struct stripe_c), sizeof(struct stripe),
62			     stripes))
63		return NULL;
64
65	len = sizeof(struct stripe_c) + (sizeof(struct stripe) * stripes);
66
67	return kmalloc(len, GFP_KERNEL);
68}
69
70/*
71 * Parse a single <dev> <sector> pair
72 */
73static int get_stripe(struct dm_target *ti, struct stripe_c *sc,
74		      unsigned int stripe, char **argv)
75{
76	unsigned long long start;
77	char dummy;
78
79	if (sscanf(argv[1], "%llu%c", &start, &dummy) != 1)
80		return -EINVAL;
81
82	if (dm_get_device(ti, argv[0], dm_table_get_mode(ti->table),
83			  &sc->stripe[stripe].dev))
84		return -ENXIO;
85
86	sc->stripe[stripe].physical_start = start;
87
88	return 0;
89}
90
91/*
92 * Construct a striped mapping.
93 * <number of stripes> <chunk size (2^^n)> [<dev_path> <offset>]+
94 */
95static int stripe_ctr(struct dm_target *ti, unsigned int argc, char **argv)
96{
97	struct stripe_c *sc;
98	sector_t width;
99	uint32_t stripes;
100	uint32_t chunk_size;
101	int r;
102	unsigned int i;
103
104	if (argc < 2) {
105		ti->error = "Not enough arguments";
106		return -EINVAL;
107	}
108
109	if (kstrtouint(argv[0], 10, &stripes) || !stripes) {
110		ti->error = "Invalid stripe count";
111		return -EINVAL;
112	}
113
114	if (kstrtouint(argv[1], 10, &chunk_size)) {
115		ti->error = "Invalid chunk_size";
116		return -EINVAL;
117	}
118
119	/*
120	 * chunk_size is a power of two
121	 */
122	if (!is_power_of_2(chunk_size) ||
123	    (chunk_size < (PAGE_SIZE >> SECTOR_SHIFT))) {
124		ti->error = "Invalid chunk size";
125		return -EINVAL;
126	}
127
128	if (ti->len & (chunk_size - 1)) {
129		ti->error = "Target length not divisible by "
130		    "chunk size";
131		return -EINVAL;
132	}
133
134	if (sector_div(width, stripes)) {
135		ti->error = "Target length not divisible by "
136		    "number of stripes";
137		return -EINVAL;
138	}
139
140	/*
141	 * Do we have enough arguments for that many stripes ?
142	 */
143	if (argc != (2 + 2 * stripes)) {
144		ti->error = "Not enough destinations "
145			"specified";
146		return -EINVAL;
147	}
148
149	sc = alloc_context(stripes);
150	if (!sc) {
151		ti->error = "Memory allocation for striped context "
152		    "failed";
153		return -ENOMEM;
154	}
155
156	INIT_WORK(&sc->trigger_event, trigger_event);
157
158	/* Set pointer to dm target; used in trigger_event */
159	sc->ti = ti;
160	sc->stripes = stripes;
161	sc->stripe_width = width;
162
163	if (stripes & (stripes - 1))
164		sc->stripes_shift = -1;
165	else
166		sc->stripes_shift = __ffs(stripes);
167
168	ti->split_io = chunk_size;
169	ti->num_flush_requests = stripes;
170	ti->num_discard_requests = stripes;
171
172	sc->chunk_shift = ffs(chunk_size) - 1;
173	sc->chunk_mask = ((sector_t) chunk_size) - 1;
174
175	/*
176	 * Get the stripe destinations.
177	 */
178	for (i = 0; i < stripes; i++) {
179		argv += 2;
180
181		r = get_stripe(ti, sc, i, argv);
182		if (r < 0) {
183			ti->error = "Couldn't parse stripe destination";
184			while (i--)
185				dm_put_device(ti, sc->stripe[i].dev);
186			kfree(sc);
187			return r;
188		}
189		atomic_set(&(sc->stripe[i].error_count), 0);
190	}
191
192	ti->private = sc;
193
194	return 0;
195}
196
197static void stripe_dtr(struct dm_target *ti)
198{
199	unsigned int i;
200	struct stripe_c *sc = (struct stripe_c *) ti->private;
201
202	for (i = 0; i < sc->stripes; i++)
203		dm_put_device(ti, sc->stripe[i].dev);
204
205	flush_work_sync(&sc->trigger_event);
206	kfree(sc);
207}
208
209static void stripe_map_sector(struct stripe_c *sc, sector_t sector,
210			      uint32_t *stripe, sector_t *result)
211{
212	sector_t offset = dm_target_offset(sc->ti, sector);
213	sector_t chunk = offset >> sc->chunk_shift;
214
215	if (sc->stripes_shift < 0)
216		*stripe = sector_div(chunk, sc->stripes);
217	else {
218		*stripe = chunk & (sc->stripes - 1);
219		chunk >>= sc->stripes_shift;
220	}
221
222	*result = (chunk << sc->chunk_shift) | (offset & sc->chunk_mask);
223}
224
225static void stripe_map_range_sector(struct stripe_c *sc, sector_t sector,
226				    uint32_t target_stripe, sector_t *result)
227{
228	uint32_t stripe;
229
230	stripe_map_sector(sc, sector, &stripe, result);
231	if (stripe == target_stripe)
232		return;
233	*result &= ~sc->chunk_mask;			/* round down */
234	if (target_stripe < stripe)
235		*result += sc->chunk_mask + 1;		/* next chunk */
236}
237
238static int stripe_map_discard(struct stripe_c *sc, struct bio *bio,
239			      uint32_t target_stripe)
240{
241	sector_t begin, end;
242
243	stripe_map_range_sector(sc, bio->bi_sector, target_stripe, &begin);
244	stripe_map_range_sector(sc, bio->bi_sector + bio_sectors(bio),
245				target_stripe, &end);
246	if (begin < end) {
247		bio->bi_bdev = sc->stripe[target_stripe].dev->bdev;
248		bio->bi_sector = begin + sc->stripe[target_stripe].physical_start;
249		bio->bi_size = to_bytes(end - begin);
250		return DM_MAPIO_REMAPPED;
251	} else {
252		/* The range doesn't map to the target stripe */
253		bio_endio(bio, 0);
254		return DM_MAPIO_SUBMITTED;
255	}
256}
257
258static int stripe_map(struct dm_target *ti, struct bio *bio,
259		      union map_info *map_context)
260{
261	struct stripe_c *sc = ti->private;
262	uint32_t stripe;
263	unsigned target_request_nr;
264
265	if (bio->bi_rw & REQ_FLUSH) {
266		target_request_nr = map_context->target_request_nr;
267		BUG_ON(target_request_nr >= sc->stripes);
268		bio->bi_bdev = sc->stripe[target_request_nr].dev->bdev;
269		return DM_MAPIO_REMAPPED;
270	}
271	if (unlikely(bio->bi_rw & REQ_DISCARD)) {
272		target_request_nr = map_context->target_request_nr;
273		BUG_ON(target_request_nr >= sc->stripes);
274		return stripe_map_discard(sc, bio, target_request_nr);
275	}
276
277	stripe_map_sector(sc, bio->bi_sector, &stripe, &bio->bi_sector);
278
279	bio->bi_sector += sc->stripe[stripe].physical_start;
280	bio->bi_bdev = sc->stripe[stripe].dev->bdev;
281
282	return DM_MAPIO_REMAPPED;
283}
284
285/*
286 * Stripe status:
287 *
288 * INFO
289 * #stripes [stripe_name <stripe_name>] [group word count]
290 * [error count 'A|D' <error count 'A|D'>]
291 *
292 * TABLE
293 * #stripes [stripe chunk size]
294 * [stripe_name physical_start <stripe_name physical_start>]
295 *
296 */
297
298static int stripe_status(struct dm_target *ti,
299			 status_type_t type, char *result, unsigned int maxlen)
300{
301	struct stripe_c *sc = (struct stripe_c *) ti->private;
302	char buffer[sc->stripes + 1];
303	unsigned int sz = 0;
304	unsigned int i;
305
306	switch (type) {
307	case STATUSTYPE_INFO:
308		DMEMIT("%d ", sc->stripes);
309		for (i = 0; i < sc->stripes; i++)  {
310			DMEMIT("%s ", sc->stripe[i].dev->name);
311			buffer[i] = atomic_read(&(sc->stripe[i].error_count)) ?
312				'D' : 'A';
313		}
314		buffer[i] = '\0';
315		DMEMIT("1 %s", buffer);
316		break;
317
318	case STATUSTYPE_TABLE:
319		DMEMIT("%d %llu", sc->stripes,
320			(unsigned long long)sc->chunk_mask + 1);
321		for (i = 0; i < sc->stripes; i++)
322			DMEMIT(" %s %llu", sc->stripe[i].dev->name,
323			    (unsigned long long)sc->stripe[i].physical_start);
324		break;
325	}
326	return 0;
327}
328
329static int stripe_end_io(struct dm_target *ti, struct bio *bio,
330			 int error, union map_info *map_context)
331{
332	unsigned i;
333	char major_minor[16];
334	struct stripe_c *sc = ti->private;
335
336	if (!error)
337		return 0; /* I/O complete */
338
339	if ((error == -EWOULDBLOCK) && (bio->bi_rw & REQ_RAHEAD))
340		return error;
341
342	if (error == -EOPNOTSUPP)
343		return error;
344
345	memset(major_minor, 0, sizeof(major_minor));
346	sprintf(major_minor, "%d:%d",
347		MAJOR(disk_devt(bio->bi_bdev->bd_disk)),
348		MINOR(disk_devt(bio->bi_bdev->bd_disk)));
349
350	/*
351	 * Test to see which stripe drive triggered the event
352	 * and increment error count for all stripes on that device.
353	 * If the error count for a given device exceeds the threshold
354	 * value we will no longer trigger any further events.
355	 */
356	for (i = 0; i < sc->stripes; i++)
357		if (!strcmp(sc->stripe[i].dev->name, major_minor)) {
358			atomic_inc(&(sc->stripe[i].error_count));
359			if (atomic_read(&(sc->stripe[i].error_count)) <
360			    DM_IO_ERROR_THRESHOLD)
361				schedule_work(&sc->trigger_event);
362		}
363
364	return error;
365}
366
367static int stripe_iterate_devices(struct dm_target *ti,
368				  iterate_devices_callout_fn fn, void *data)
369{
370	struct stripe_c *sc = ti->private;
371	int ret = 0;
372	unsigned i = 0;
373
374	do {
375		ret = fn(ti, sc->stripe[i].dev,
376			 sc->stripe[i].physical_start,
377			 sc->stripe_width, data);
378	} while (!ret && ++i < sc->stripes);
379
380	return ret;
381}
382
383static void stripe_io_hints(struct dm_target *ti,
384			    struct queue_limits *limits)
385{
386	struct stripe_c *sc = ti->private;
387	unsigned chunk_size = (sc->chunk_mask + 1) << 9;
388
389	blk_limits_io_min(limits, chunk_size);
390	blk_limits_io_opt(limits, chunk_size * sc->stripes);
391}
392
393static int stripe_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
394			struct bio_vec *biovec, int max_size)
395{
396	struct stripe_c *sc = ti->private;
397	sector_t bvm_sector = bvm->bi_sector;
398	uint32_t stripe;
399	struct request_queue *q;
400
401	stripe_map_sector(sc, bvm_sector, &stripe, &bvm_sector);
402
403	q = bdev_get_queue(sc->stripe[stripe].dev->bdev);
404	if (!q->merge_bvec_fn)
405		return max_size;
406
407	bvm->bi_bdev = sc->stripe[stripe].dev->bdev;
408	bvm->bi_sector = sc->stripe[stripe].physical_start + bvm_sector;
409
410	return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
411}
412
413static struct target_type stripe_target = {
414	.name   = "striped",
415	.version = {1, 4, 0},
416	.module = THIS_MODULE,
417	.ctr    = stripe_ctr,
418	.dtr    = stripe_dtr,
419	.map    = stripe_map,
420	.end_io = stripe_end_io,
421	.status = stripe_status,
422	.iterate_devices = stripe_iterate_devices,
423	.io_hints = stripe_io_hints,
424	.merge  = stripe_merge,
425};
426
427int __init dm_stripe_init(void)
428{
429	int r;
430
431	r = dm_register_target(&stripe_target);
432	if (r < 0) {
433		DMWARN("target registration failed");
434		return r;
435	}
436
437	return r;
438}
439
440void dm_stripe_exit(void)
441{
442	dm_unregister_target(&stripe_target);
443}
444