dm-stripe.c revision 29915202006c2e7bafe81348eb498ff9a724ac61 
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
79	if (sscanf(argv[1], "%llu", &start) != 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	char *end;
102	int r;
103	unsigned int i;
104
105	if (argc < 2) {
106		ti->error = "Not enough arguments";
107		return -EINVAL;
108	}
109
110	stripes = simple_strtoul(argv[0], &end, 10);
111	if (!stripes || *end) {
112		ti->error = "Invalid stripe count";
113		return -EINVAL;
114	}
115
116	chunk_size = simple_strtoul(argv[1], &end, 10);
117	if (*end) {
118		ti->error = "Invalid chunk_size";
119		return -EINVAL;
120	}
121
122	/*
123	 * chunk_size is a power of two
124	 */
125	if (!is_power_of_2(chunk_size) ||
126	    (chunk_size < (PAGE_SIZE >> SECTOR_SHIFT))) {
127		ti->error = "Invalid chunk size";
128		return -EINVAL;
129	}
130
131	if (ti->len & (chunk_size - 1)) {
132		ti->error = "Target length not divisible by "
133		    "chunk size";
134		return -EINVAL;
135	}
136
137	width = ti->len;
138	if (sector_div(width, stripes)) {
139		ti->error = "Target length not divisible by "
140		    "number of stripes";
141		return -EINVAL;
142	}
143
144	/*
145	 * Do we have enough arguments for that many stripes ?
146	 */
147	if (argc != (2 + 2 * stripes)) {
148		ti->error = "Not enough destinations "
149			"specified";
150		return -EINVAL;
151	}
152
153	sc = alloc_context(stripes);
154	if (!sc) {
155		ti->error = "Memory allocation for striped context "
156		    "failed";
157		return -ENOMEM;
158	}
159
160	INIT_WORK(&sc->trigger_event, trigger_event);
161
162	/* Set pointer to dm target; used in trigger_event */
163	sc->ti = ti;
164	sc->stripes = stripes;
165	sc->stripe_width = width;
166
167	if (stripes & (stripes - 1))
168		sc->stripes_shift = -1;
169	else {
170		sc->stripes_shift = ffs(stripes) - 1;
171		sc->stripes_mask = ((sector_t) stripes) - 1;
172	}
173
174	ti->split_io = chunk_size;
175	ti->num_flush_requests = stripes;
176	ti->num_discard_requests = stripes;
177
178	sc->chunk_shift = ffs(chunk_size) - 1;
179	sc->chunk_mask = ((sector_t) chunk_size) - 1;
180
181	/*
182	 * Get the stripe destinations.
183	 */
184	for (i = 0; i < stripes; i++) {
185		argv += 2;
186
187		r = get_stripe(ti, sc, i, argv);
188		if (r < 0) {
189			ti->error = "Couldn't parse stripe destination";
190			while (i--)
191				dm_put_device(ti, sc->stripe[i].dev);
192			kfree(sc);
193			return r;
194		}
195		atomic_set(&(sc->stripe[i].error_count), 0);
196	}
197
198	ti->private = sc;
199
200	return 0;
201}
202
203static void stripe_dtr(struct dm_target *ti)
204{
205	unsigned int i;
206	struct stripe_c *sc = (struct stripe_c *) ti->private;
207
208	for (i = 0; i < sc->stripes; i++)
209		dm_put_device(ti, sc->stripe[i].dev);
210
211	flush_work_sync(&sc->trigger_event);
212	kfree(sc);
213}
214
215static void stripe_map_sector(struct stripe_c *sc, sector_t sector,
216			      uint32_t *stripe, sector_t *result)
217{
218	sector_t offset = dm_target_offset(sc->ti, sector);
219	sector_t chunk = offset >> sc->chunk_shift;
220
221	if (sc->stripes_shift < 0)
222		*stripe = sector_div(chunk, sc->stripes);
223	else {
224		*stripe = chunk & sc->stripes_mask;
225		chunk >>= sc->stripes_shift;
226	}
227
228	*result = (chunk << sc->chunk_shift) | (offset & sc->chunk_mask);
229}
230
231static void stripe_map_range_sector(struct stripe_c *sc, sector_t sector,
232				    uint32_t target_stripe, sector_t *result)
233{
234	uint32_t stripe;
235
236	stripe_map_sector(sc, sector, &stripe, result);
237	if (stripe == target_stripe)
238		return;
239	*result &= ~sc->chunk_mask;			/* round down */
240	if (target_stripe < stripe)
241		*result += sc->chunk_mask + 1;		/* next chunk */
242}
243
244static int stripe_map_discard(struct stripe_c *sc, struct bio *bio,
245			      uint32_t target_stripe)
246{
247	sector_t begin, end;
248
249	stripe_map_range_sector(sc, bio->bi_sector, target_stripe, &begin);
250	stripe_map_range_sector(sc, bio->bi_sector + bio_sectors(bio),
251				target_stripe, &end);
252	if (begin < end) {
253		bio->bi_bdev = sc->stripe[target_stripe].dev->bdev;
254		bio->bi_sector = begin + sc->stripe[target_stripe].physical_start;
255		bio->bi_size = to_bytes(end - begin);
256		return DM_MAPIO_REMAPPED;
257	} else {
258		/* The range doesn't map to the target stripe */
259		bio_endio(bio, 0);
260		return DM_MAPIO_SUBMITTED;
261	}
262}
263
264static int stripe_map(struct dm_target *ti, struct bio *bio,
265		      union map_info *map_context)
266{
267	struct stripe_c *sc = ti->private;
268	uint32_t stripe;
269	unsigned target_request_nr;
270
271	if (bio->bi_rw & REQ_FLUSH) {
272		target_request_nr = map_context->target_request_nr;
273		BUG_ON(target_request_nr >= sc->stripes);
274		bio->bi_bdev = sc->stripe[target_request_nr].dev->bdev;
275		return DM_MAPIO_REMAPPED;
276	}
277	if (unlikely(bio->bi_rw & REQ_DISCARD)) {
278		target_request_nr = map_context->target_request_nr;
279		BUG_ON(target_request_nr >= sc->stripes);
280		return stripe_map_discard(sc, bio, target_request_nr);
281	}
282
283	stripe_map_sector(sc, bio->bi_sector, &stripe, &bio->bi_sector);
284
285	bio->bi_sector += sc->stripe[stripe].physical_start;
286	bio->bi_bdev = sc->stripe[stripe].dev->bdev;
287
288	return DM_MAPIO_REMAPPED;
289}
290
291/*
292 * Stripe status:
293 *
294 * INFO
295 * #stripes [stripe_name <stripe_name>] [group word count]
296 * [error count 'A|D' <error count 'A|D'>]
297 *
298 * TABLE
299 * #stripes [stripe chunk size]
300 * [stripe_name physical_start <stripe_name physical_start>]
301 *
302 */
303
304static int stripe_status(struct dm_target *ti,
305			 status_type_t type, char *result, unsigned int maxlen)
306{
307	struct stripe_c *sc = (struct stripe_c *) ti->private;
308	char buffer[sc->stripes + 1];
309	unsigned int sz = 0;
310	unsigned int i;
311
312	switch (type) {
313	case STATUSTYPE_INFO:
314		DMEMIT("%d ", sc->stripes);
315		for (i = 0; i < sc->stripes; i++)  {
316			DMEMIT("%s ", sc->stripe[i].dev->name);
317			buffer[i] = atomic_read(&(sc->stripe[i].error_count)) ?
318				'D' : 'A';
319		}
320		buffer[i] = '\0';
321		DMEMIT("1 %s", buffer);
322		break;
323
324	case STATUSTYPE_TABLE:
325		DMEMIT("%d %llu", sc->stripes,
326			(unsigned long long)sc->chunk_mask + 1);
327		for (i = 0; i < sc->stripes; i++)
328			DMEMIT(" %s %llu", sc->stripe[i].dev->name,
329			    (unsigned long long)sc->stripe[i].physical_start);
330		break;
331	}
332	return 0;
333}
334
335static int stripe_end_io(struct dm_target *ti, struct bio *bio,
336			 int error, union map_info *map_context)
337{
338	unsigned i;
339	char major_minor[16];
340	struct stripe_c *sc = ti->private;
341
342	if (!error)
343		return 0; /* I/O complete */
344
345	if ((error == -EWOULDBLOCK) && (bio->bi_rw & REQ_RAHEAD))
346		return error;
347
348	if (error == -EOPNOTSUPP)
349		return error;
350
351	memset(major_minor, 0, sizeof(major_minor));
352	sprintf(major_minor, "%d:%d",
353		MAJOR(disk_devt(bio->bi_bdev->bd_disk)),
354		MINOR(disk_devt(bio->bi_bdev->bd_disk)));
355
356	/*
357	 * Test to see which stripe drive triggered the event
358	 * and increment error count for all stripes on that device.
359	 * If the error count for a given device exceeds the threshold
360	 * value we will no longer trigger any further events.
361	 */
362	for (i = 0; i < sc->stripes; i++)
363		if (!strcmp(sc->stripe[i].dev->name, major_minor)) {
364			atomic_inc(&(sc->stripe[i].error_count));
365			if (atomic_read(&(sc->stripe[i].error_count)) <
366			    DM_IO_ERROR_THRESHOLD)
367				schedule_work(&sc->trigger_event);
368		}
369
370	return error;
371}
372
373static int stripe_iterate_devices(struct dm_target *ti,
374				  iterate_devices_callout_fn fn, void *data)
375{
376	struct stripe_c *sc = ti->private;
377	int ret = 0;
378	unsigned i = 0;
379
380	do {
381		ret = fn(ti, sc->stripe[i].dev,
382			 sc->stripe[i].physical_start,
383			 sc->stripe_width, data);
384	} while (!ret && ++i < sc->stripes);
385
386	return ret;
387}
388
389static void stripe_io_hints(struct dm_target *ti,
390			    struct queue_limits *limits)
391{
392	struct stripe_c *sc = ti->private;
393	unsigned chunk_size = (sc->chunk_mask + 1) << 9;
394
395	blk_limits_io_min(limits, chunk_size);
396	blk_limits_io_opt(limits, chunk_size * sc->stripes);
397}
398
399static int stripe_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
400			struct bio_vec *biovec, int max_size)
401{
402	struct stripe_c *sc = ti->private;
403	sector_t bvm_sector = bvm->bi_sector;
404	uint32_t stripe;
405	struct request_queue *q;
406
407	stripe_map_sector(sc, bvm_sector, &stripe, &bvm_sector);
408
409	q = bdev_get_queue(sc->stripe[stripe].dev->bdev);
410	if (!q->merge_bvec_fn)
411		return max_size;
412
413	bvm->bi_bdev = sc->stripe[stripe].dev->bdev;
414	bvm->bi_sector = sc->stripe[stripe].physical_start + bvm_sector;
415
416	return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
417}
418
419static struct target_type stripe_target = {
420	.name   = "striped",
421	.version = {1, 4, 0},
422	.module = THIS_MODULE,
423	.ctr    = stripe_ctr,
424	.dtr    = stripe_dtr,
425	.map    = stripe_map,
426	.end_io = stripe_end_io,
427	.status = stripe_status,
428	.iterate_devices = stripe_iterate_devices,
429	.io_hints = stripe_io_hints,
430	.merge  = stripe_merge,
431};
432
433int __init dm_stripe_init(void)
434{
435	int r;
436
437	r = dm_register_target(&stripe_target);
438	if (r < 0) {
439		DMWARN("target registration failed");
440		return r;
441	}
442
443	return r;
444}
445
446void dm_stripe_exit(void)
447{
448	dm_unregister_target(&stripe_target);
449}
450