dm-stripe.c revision fd7c092e711ebab55b2688d3859d95dfd0301f73 
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;
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, 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	ti->num_write_same_requests = stripes;
166
167	sc->chunk_size = chunk_size;
168	if (chunk_size & (chunk_size - 1))
169		sc->chunk_size_shift = -1;
170	else
171		sc->chunk_size_shift = __ffs(chunk_size);
172
173	/*
174	 * Get the stripe destinations.
175	 */
176	for (i = 0; i < stripes; i++) {
177		argv += 2;
178
179		r = get_stripe(ti, sc, i, argv);
180		if (r < 0) {
181			ti->error = "Couldn't parse stripe destination";
182			while (i--)
183				dm_put_device(ti, sc->stripe[i].dev);
184			kfree(sc);
185			return r;
186		}
187		atomic_set(&(sc->stripe[i].error_count), 0);
188	}
189
190	ti->private = sc;
191
192	return 0;
193}
194
195static void stripe_dtr(struct dm_target *ti)
196{
197	unsigned int i;
198	struct stripe_c *sc = (struct stripe_c *) ti->private;
199
200	for (i = 0; i < sc->stripes; i++)
201		dm_put_device(ti, sc->stripe[i].dev);
202
203	flush_work(&sc->trigger_event);
204	kfree(sc);
205}
206
207static void stripe_map_sector(struct stripe_c *sc, sector_t sector,
208			      uint32_t *stripe, sector_t *result)
209{
210	sector_t chunk = dm_target_offset(sc->ti, sector);
211	sector_t chunk_offset;
212
213	if (sc->chunk_size_shift < 0)
214		chunk_offset = sector_div(chunk, sc->chunk_size);
215	else {
216		chunk_offset = chunk & (sc->chunk_size - 1);
217		chunk >>= sc->chunk_size_shift;
218	}
219
220	if (sc->stripes_shift < 0)
221		*stripe = sector_div(chunk, sc->stripes);
222	else {
223		*stripe = chunk & (sc->stripes - 1);
224		chunk >>= sc->stripes_shift;
225	}
226
227	if (sc->chunk_size_shift < 0)
228		chunk *= sc->chunk_size;
229	else
230		chunk <<= sc->chunk_size_shift;
231
232	*result = chunk + chunk_offset;
233}
234
235static void stripe_map_range_sector(struct stripe_c *sc, sector_t sector,
236				    uint32_t target_stripe, sector_t *result)
237{
238	uint32_t stripe;
239
240	stripe_map_sector(sc, sector, &stripe, result);
241	if (stripe == target_stripe)
242		return;
243
244	/* round down */
245	sector = *result;
246	if (sc->chunk_size_shift < 0)
247		*result -= sector_div(sector, sc->chunk_size);
248	else
249		*result = sector & ~(sector_t)(sc->chunk_size - 1);
250
251	if (target_stripe < stripe)
252		*result += sc->chunk_size;		/* next chunk */
253}
254
255static int stripe_map_range(struct stripe_c *sc, struct bio *bio,
256			    uint32_t target_stripe)
257{
258	sector_t begin, end;
259
260	stripe_map_range_sector(sc, bio->bi_sector, target_stripe, &begin);
261	stripe_map_range_sector(sc, bio->bi_sector + bio_sectors(bio),
262				target_stripe, &end);
263	if (begin < end) {
264		bio->bi_bdev = sc->stripe[target_stripe].dev->bdev;
265		bio->bi_sector = begin + sc->stripe[target_stripe].physical_start;
266		bio->bi_size = to_bytes(end - begin);
267		return DM_MAPIO_REMAPPED;
268	} else {
269		/* The range doesn't map to the target stripe */
270		bio_endio(bio, 0);
271		return DM_MAPIO_SUBMITTED;
272	}
273}
274
275static int stripe_map(struct dm_target *ti, struct bio *bio)
276{
277	struct stripe_c *sc = ti->private;
278	uint32_t stripe;
279	unsigned target_request_nr;
280
281	if (bio->bi_rw & REQ_FLUSH) {
282		target_request_nr = dm_bio_get_target_request_nr(bio);
283		BUG_ON(target_request_nr >= sc->stripes);
284		bio->bi_bdev = sc->stripe[target_request_nr].dev->bdev;
285		return DM_MAPIO_REMAPPED;
286	}
287	if (unlikely(bio->bi_rw & REQ_DISCARD) ||
288	    unlikely(bio->bi_rw & REQ_WRITE_SAME)) {
289		target_request_nr = dm_bio_get_target_request_nr(bio);
290		BUG_ON(target_request_nr >= sc->stripes);
291		return stripe_map_range(sc, bio, target_request_nr);
292	}
293
294	stripe_map_sector(sc, bio->bi_sector, &stripe, &bio->bi_sector);
295
296	bio->bi_sector += sc->stripe[stripe].physical_start;
297	bio->bi_bdev = sc->stripe[stripe].dev->bdev;
298
299	return DM_MAPIO_REMAPPED;
300}
301
302/*
303 * Stripe status:
304 *
305 * INFO
306 * #stripes [stripe_name <stripe_name>] [group word count]
307 * [error count 'A|D' <error count 'A|D'>]
308 *
309 * TABLE
310 * #stripes [stripe chunk size]
311 * [stripe_name physical_start <stripe_name physical_start>]
312 *
313 */
314
315static void stripe_status(struct dm_target *ti, status_type_t type,
316			  unsigned status_flags, char *result, unsigned maxlen)
317{
318	struct stripe_c *sc = (struct stripe_c *) ti->private;
319	char buffer[sc->stripes + 1];
320	unsigned int sz = 0;
321	unsigned int i;
322
323	switch (type) {
324	case STATUSTYPE_INFO:
325		DMEMIT("%d ", sc->stripes);
326		for (i = 0; i < sc->stripes; i++)  {
327			DMEMIT("%s ", sc->stripe[i].dev->name);
328			buffer[i] = atomic_read(&(sc->stripe[i].error_count)) ?
329				'D' : 'A';
330		}
331		buffer[i] = '\0';
332		DMEMIT("1 %s", buffer);
333		break;
334
335	case STATUSTYPE_TABLE:
336		DMEMIT("%d %llu", sc->stripes,
337			(unsigned long long)sc->chunk_size);
338		for (i = 0; i < sc->stripes; i++)
339			DMEMIT(" %s %llu", sc->stripe[i].dev->name,
340			    (unsigned long long)sc->stripe[i].physical_start);
341		break;
342	}
343}
344
345static int stripe_end_io(struct dm_target *ti, struct bio *bio, int error)
346{
347	unsigned i;
348	char major_minor[16];
349	struct stripe_c *sc = ti->private;
350
351	if (!error)
352		return 0; /* I/O complete */
353
354	if ((error == -EWOULDBLOCK) && (bio->bi_rw & REQ_RAHEAD))
355		return error;
356
357	if (error == -EOPNOTSUPP)
358		return error;
359
360	memset(major_minor, 0, sizeof(major_minor));
361	sprintf(major_minor, "%d:%d",
362		MAJOR(disk_devt(bio->bi_bdev->bd_disk)),
363		MINOR(disk_devt(bio->bi_bdev->bd_disk)));
364
365	/*
366	 * Test to see which stripe drive triggered the event
367	 * and increment error count for all stripes on that device.
368	 * If the error count for a given device exceeds the threshold
369	 * value we will no longer trigger any further events.
370	 */
371	for (i = 0; i < sc->stripes; i++)
372		if (!strcmp(sc->stripe[i].dev->name, major_minor)) {
373			atomic_inc(&(sc->stripe[i].error_count));
374			if (atomic_read(&(sc->stripe[i].error_count)) <
375			    DM_IO_ERROR_THRESHOLD)
376				schedule_work(&sc->trigger_event);
377		}
378
379	return error;
380}
381
382static int stripe_iterate_devices(struct dm_target *ti,
383				  iterate_devices_callout_fn fn, void *data)
384{
385	struct stripe_c *sc = ti->private;
386	int ret = 0;
387	unsigned i = 0;
388
389	do {
390		ret = fn(ti, sc->stripe[i].dev,
391			 sc->stripe[i].physical_start,
392			 sc->stripe_width, data);
393	} while (!ret && ++i < sc->stripes);
394
395	return ret;
396}
397
398static void stripe_io_hints(struct dm_target *ti,
399			    struct queue_limits *limits)
400{
401	struct stripe_c *sc = ti->private;
402	unsigned chunk_size = sc->chunk_size << SECTOR_SHIFT;
403
404	blk_limits_io_min(limits, chunk_size);
405	blk_limits_io_opt(limits, chunk_size * sc->stripes);
406}
407
408static int stripe_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
409			struct bio_vec *biovec, int max_size)
410{
411	struct stripe_c *sc = ti->private;
412	sector_t bvm_sector = bvm->bi_sector;
413	uint32_t stripe;
414	struct request_queue *q;
415
416	stripe_map_sector(sc, bvm_sector, &stripe, &bvm_sector);
417
418	q = bdev_get_queue(sc->stripe[stripe].dev->bdev);
419	if (!q->merge_bvec_fn)
420		return max_size;
421
422	bvm->bi_bdev = sc->stripe[stripe].dev->bdev;
423	bvm->bi_sector = sc->stripe[stripe].physical_start + bvm_sector;
424
425	return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
426}
427
428static struct target_type stripe_target = {
429	.name   = "striped",
430	.version = {1, 5, 1},
431	.module = THIS_MODULE,
432	.ctr    = stripe_ctr,
433	.dtr    = stripe_dtr,
434	.map    = stripe_map,
435	.end_io = stripe_end_io,
436	.status = stripe_status,
437	.iterate_devices = stripe_iterate_devices,
438	.io_hints = stripe_io_hints,
439	.merge  = stripe_merge,
440};
441
442int __init dm_stripe_init(void)
443{
444	int r;
445
446	r = dm_register_target(&stripe_target);
447	if (r < 0) {
448		DMWARN("target registration failed");
449		return r;
450	}
451
452	return r;
453}
454
455void dm_stripe_exit(void)
456{
457	dm_unregister_target(&stripe_target);
458}
459