dm-stripe.c revision 65988525abde0b0a5833c4e20f32967184a5dcf0 
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
29	/* The size of this target / num. stripes */
30	sector_t stripe_width;
31
32	/* stripe chunk size */
33	uint32_t chunk_shift;
34	sector_t chunk_mask;
35
36	/* Needed for handling events */
37	struct dm_target *ti;
38
39	/* Work struct used for triggering events*/
40	struct work_struct kstriped_ws;
41
42	struct stripe stripe[0];
43};
44
45static struct workqueue_struct *kstriped;
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, kstriped_ws);
54
55	dm_table_event(sc->ti->table);
56
57}
58
59static inline struct stripe_c *alloc_context(unsigned int stripes)
60{
61	size_t len;
62
63	if (dm_array_too_big(sizeof(struct stripe_c), sizeof(struct stripe),
64			     stripes))
65		return NULL;
66
67	len = sizeof(struct stripe_c) + (sizeof(struct stripe) * stripes);
68
69	return kmalloc(len, GFP_KERNEL);
70}
71
72/*
73 * Parse a single <dev> <sector> pair
74 */
75static int get_stripe(struct dm_target *ti, struct stripe_c *sc,
76		      unsigned int stripe, char **argv)
77{
78	unsigned long long start;
79
80	if (sscanf(argv[1], "%llu", &start) != 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	char *end;
103	int r;
104	unsigned int i;
105
106	if (argc < 2) {
107		ti->error = "Not enough arguments";
108		return -EINVAL;
109	}
110
111	stripes = simple_strtoul(argv[0], &end, 10);
112	if (!stripes || *end) {
113		ti->error = "Invalid stripe count";
114		return -EINVAL;
115	}
116
117	chunk_size = simple_strtoul(argv[1], &end, 10);
118	if (*end) {
119		ti->error = "Invalid chunk_size";
120		return -EINVAL;
121	}
122
123	/*
124	 * chunk_size is a power of two
125	 */
126	if (!is_power_of_2(chunk_size) ||
127	    (chunk_size < (PAGE_SIZE >> SECTOR_SHIFT))) {
128		ti->error = "Invalid chunk size";
129		return -EINVAL;
130	}
131
132	if (ti->len & (chunk_size - 1)) {
133		ti->error = "Target length not divisible by "
134		    "chunk size";
135		return -EINVAL;
136	}
137
138	width = ti->len;
139	if (sector_div(width, stripes)) {
140		ti->error = "Target length not divisible by "
141		    "number of stripes";
142		return -EINVAL;
143	}
144
145	/*
146	 * Do we have enough arguments for that many stripes ?
147	 */
148	if (argc != (2 + 2 * stripes)) {
149		ti->error = "Not enough destinations "
150			"specified";
151		return -EINVAL;
152	}
153
154	sc = alloc_context(stripes);
155	if (!sc) {
156		ti->error = "Memory allocation for striped context "
157		    "failed";
158		return -ENOMEM;
159	}
160
161	INIT_WORK(&sc->kstriped_ws, trigger_event);
162
163	/* Set pointer to dm target; used in trigger_event */
164	sc->ti = ti;
165
166	sc->stripes = stripes;
167	sc->stripe_width = width;
168	ti->split_io = chunk_size;
169	ti->num_flush_requests = stripes;
170
171	sc->chunk_mask = ((sector_t) chunk_size) - 1;
172	for (sc->chunk_shift = 0; chunk_size; sc->chunk_shift++)
173		chunk_size >>= 1;
174	sc->chunk_shift--;
175
176	/*
177	 * Get the stripe destinations.
178	 */
179	for (i = 0; i < stripes; i++) {
180		argv += 2;
181
182		r = get_stripe(ti, sc, i, argv);
183		if (r < 0) {
184			ti->error = "Couldn't parse stripe destination";
185			while (i--)
186				dm_put_device(ti, sc->stripe[i].dev);
187			kfree(sc);
188			return r;
189		}
190		atomic_set(&(sc->stripe[i].error_count), 0);
191	}
192
193	ti->private = sc;
194
195	return 0;
196}
197
198static void stripe_dtr(struct dm_target *ti)
199{
200	unsigned int i;
201	struct stripe_c *sc = (struct stripe_c *) ti->private;
202
203	for (i = 0; i < sc->stripes; i++)
204		dm_put_device(ti, sc->stripe[i].dev);
205
206	flush_workqueue(kstriped);
207	kfree(sc);
208}
209
210static void stripe_map_sector(struct stripe_c *sc, sector_t sector,
211			      uint32_t *stripe, sector_t *result)
212{
213	sector_t offset = dm_target_offset(sc->ti, sector);
214	sector_t chunk = offset >> sc->chunk_shift;
215
216	*stripe = sector_div(chunk, sc->stripes);
217	*result = (chunk << sc->chunk_shift) | (offset & sc->chunk_mask);
218}
219
220static int stripe_map(struct dm_target *ti, struct bio *bio,
221		      union map_info *map_context)
222{
223	struct stripe_c *sc = ti->private;
224	uint32_t stripe;
225	unsigned target_request_nr;
226
227	if (unlikely(bio_empty_barrier(bio))) {
228		target_request_nr = map_context->target_request_nr;
229		BUG_ON(target_request_nr >= sc->stripes);
230		bio->bi_bdev = sc->stripe[target_request_nr].dev->bdev;
231		return DM_MAPIO_REMAPPED;
232	}
233
234	stripe_map_sector(sc, bio->bi_sector, &stripe, &bio->bi_sector);
235
236	bio->bi_sector += sc->stripe[stripe].physical_start;
237	bio->bi_bdev = sc->stripe[stripe].dev->bdev;
238
239	return DM_MAPIO_REMAPPED;
240}
241
242/*
243 * Stripe status:
244 *
245 * INFO
246 * #stripes [stripe_name <stripe_name>] [group word count]
247 * [error count 'A|D' <error count 'A|D'>]
248 *
249 * TABLE
250 * #stripes [stripe chunk size]
251 * [stripe_name physical_start <stripe_name physical_start>]
252 *
253 */
254
255static int stripe_status(struct dm_target *ti,
256			 status_type_t type, char *result, unsigned int maxlen)
257{
258	struct stripe_c *sc = (struct stripe_c *) ti->private;
259	char buffer[sc->stripes + 1];
260	unsigned int sz = 0;
261	unsigned int i;
262
263	switch (type) {
264	case STATUSTYPE_INFO:
265		DMEMIT("%d ", sc->stripes);
266		for (i = 0; i < sc->stripes; i++)  {
267			DMEMIT("%s ", sc->stripe[i].dev->name);
268			buffer[i] = atomic_read(&(sc->stripe[i].error_count)) ?
269				'D' : 'A';
270		}
271		buffer[i] = '\0';
272		DMEMIT("1 %s", buffer);
273		break;
274
275	case STATUSTYPE_TABLE:
276		DMEMIT("%d %llu", sc->stripes,
277			(unsigned long long)sc->chunk_mask + 1);
278		for (i = 0; i < sc->stripes; i++)
279			DMEMIT(" %s %llu", sc->stripe[i].dev->name,
280			    (unsigned long long)sc->stripe[i].physical_start);
281		break;
282	}
283	return 0;
284}
285
286static int stripe_end_io(struct dm_target *ti, struct bio *bio,
287			 int error, union map_info *map_context)
288{
289	unsigned i;
290	char major_minor[16];
291	struct stripe_c *sc = ti->private;
292
293	if (!error)
294		return 0; /* I/O complete */
295
296	if ((error == -EWOULDBLOCK) && (bio->bi_rw & REQ_RAHEAD))
297		return error;
298
299	if (error == -EOPNOTSUPP)
300		return error;
301
302	memset(major_minor, 0, sizeof(major_minor));
303	sprintf(major_minor, "%d:%d",
304		MAJOR(disk_devt(bio->bi_bdev->bd_disk)),
305		MINOR(disk_devt(bio->bi_bdev->bd_disk)));
306
307	/*
308	 * Test to see which stripe drive triggered the event
309	 * and increment error count for all stripes on that device.
310	 * If the error count for a given device exceeds the threshold
311	 * value we will no longer trigger any further events.
312	 */
313	for (i = 0; i < sc->stripes; i++)
314		if (!strcmp(sc->stripe[i].dev->name, major_minor)) {
315			atomic_inc(&(sc->stripe[i].error_count));
316			if (atomic_read(&(sc->stripe[i].error_count)) <
317			    DM_IO_ERROR_THRESHOLD)
318				queue_work(kstriped, &sc->kstriped_ws);
319		}
320
321	return error;
322}
323
324static int stripe_iterate_devices(struct dm_target *ti,
325				  iterate_devices_callout_fn fn, void *data)
326{
327	struct stripe_c *sc = ti->private;
328	int ret = 0;
329	unsigned i = 0;
330
331	do {
332		ret = fn(ti, sc->stripe[i].dev,
333			 sc->stripe[i].physical_start,
334			 sc->stripe_width, data);
335	} while (!ret && ++i < sc->stripes);
336
337	return ret;
338}
339
340static void stripe_io_hints(struct dm_target *ti,
341			    struct queue_limits *limits)
342{
343	struct stripe_c *sc = ti->private;
344	unsigned chunk_size = (sc->chunk_mask + 1) << 9;
345
346	blk_limits_io_min(limits, chunk_size);
347	blk_limits_io_opt(limits, chunk_size * sc->stripes);
348}
349
350static struct target_type stripe_target = {
351	.name   = "striped",
352	.version = {1, 3, 0},
353	.module = THIS_MODULE,
354	.ctr    = stripe_ctr,
355	.dtr    = stripe_dtr,
356	.map    = stripe_map,
357	.end_io = stripe_end_io,
358	.status = stripe_status,
359	.iterate_devices = stripe_iterate_devices,
360	.io_hints = stripe_io_hints,
361};
362
363int __init dm_stripe_init(void)
364{
365	int r;
366
367	r = dm_register_target(&stripe_target);
368	if (r < 0) {
369		DMWARN("target registration failed");
370		return r;
371	}
372
373	kstriped = create_singlethread_workqueue("kstriped");
374	if (!kstriped) {
375		DMERR("failed to create workqueue kstriped");
376		dm_unregister_target(&stripe_target);
377		return -ENOMEM;
378	}
379
380	return r;
381}
382
383void dm_stripe_exit(void)
384{
385	dm_unregister_target(&stripe_target);
386	destroy_workqueue(kstriped);
387
388	return;
389}
390