dm-stripe.c revision 7fff5e8f727285cf54e6aba10f31b196f207b98a 
1/*
2 * Copyright (C) 2001-2003 Sistina Software (UK) Limited.
3 *
4 * This file is released under the GPL.
5 */
6
7#include "dm.h"
8#include <linux/device-mapper.h>
9
10#include <linux/module.h>
11#include <linux/init.h>
12#include <linux/blkdev.h>
13#include <linux/bio.h>
14#include <linux/slab.h>
15#include <linux/log2.h>
16
17#define DM_MSG_PREFIX "striped"
18#define DM_IO_ERROR_THRESHOLD 15
19
20struct stripe {
21	struct dm_dev *dev;
22	sector_t physical_start;
23
24	atomic_t error_count;
25};
26
27struct stripe_c {
28	uint32_t stripes;
29	int stripes_shift;
30
31	/* The size of this target / num. stripes */
32	sector_t stripe_width;
33
34	uint32_t chunk_size;
35	int chunk_size_shift;
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> [<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, tmp_len;
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) || !chunk_size) {
115		ti->error = "Invalid chunk_size";
116		return -EINVAL;
117	}
118
119	width = ti->len;
120	if (sector_div(width, stripes)) {
121		ti->error = "Target length not divisible by "
122		    "number of stripes";
123		return -EINVAL;
124	}
125
126	tmp_len = width;
127	if (sector_div(tmp_len, chunk_size)) {
128		ti->error = "Target length not divisible by "
129		    "chunk size";
130		return -EINVAL;
131	}
132
133	/*
134	 * Do we have enough arguments for that many stripes ?
135	 */
136	if (argc != (2 + 2 * stripes)) {
137		ti->error = "Not enough destinations "
138			"specified";
139		return -EINVAL;
140	}
141
142	sc = alloc_context(stripes);
143	if (!sc) {
144		ti->error = "Memory allocation for striped context "
145		    "failed";
146		return -ENOMEM;
147	}
148
149	INIT_WORK(&sc->trigger_event, trigger_event);
150
151	/* Set pointer to dm target; used in trigger_event */
152	sc->ti = ti;
153	sc->stripes = stripes;
154	sc->stripe_width = width;
155
156	if (stripes & (stripes - 1))
157		sc->stripes_shift = -1;
158	else
159		sc->stripes_shift = __ffs(stripes);
160
161	r = dm_set_target_max_io_len(ti, chunk_size);
162	if (r)
163		return r;
164
165	ti->num_flush_bios = stripes;
166	ti->num_discard_bios = stripes;
167	ti->num_write_same_bios = stripes;
168
169	sc->chunk_size = chunk_size;
170	if (chunk_size & (chunk_size - 1))
171		sc->chunk_size_shift = -1;
172	else
173		sc->chunk_size_shift = __ffs(chunk_size);
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(&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 chunk = dm_target_offset(sc->ti, sector);
213	sector_t chunk_offset;
214
215	if (sc->chunk_size_shift < 0)
216		chunk_offset = sector_div(chunk, sc->chunk_size);
217	else {
218		chunk_offset = chunk & (sc->chunk_size - 1);
219		chunk >>= sc->chunk_size_shift;
220	}
221
222	if (sc->stripes_shift < 0)
223		*stripe = sector_div(chunk, sc->stripes);
224	else {
225		*stripe = chunk & (sc->stripes - 1);
226		chunk >>= sc->stripes_shift;
227	}
228
229	if (sc->chunk_size_shift < 0)
230		chunk *= sc->chunk_size;
231	else
232		chunk <<= sc->chunk_size_shift;
233
234	*result = chunk + chunk_offset;
235}
236
237static void stripe_map_range_sector(struct stripe_c *sc, sector_t sector,
238				    uint32_t target_stripe, sector_t *result)
239{
240	uint32_t stripe;
241
242	stripe_map_sector(sc, sector, &stripe, result);
243	if (stripe == target_stripe)
244		return;
245
246	/* round down */
247	sector = *result;
248	if (sc->chunk_size_shift < 0)
249		*result -= sector_div(sector, sc->chunk_size);
250	else
251		*result = sector & ~(sector_t)(sc->chunk_size - 1);
252
253	if (target_stripe < stripe)
254		*result += sc->chunk_size;		/* next chunk */
255}
256
257static int stripe_map_range(struct stripe_c *sc, struct bio *bio,
258			    uint32_t target_stripe)
259{
260	sector_t begin, end;
261
262	stripe_map_range_sector(sc, bio->bi_sector, target_stripe, &begin);
263	stripe_map_range_sector(sc, bio_end_sector(bio),
264				target_stripe, &end);
265	if (begin < end) {
266		bio->bi_bdev = sc->stripe[target_stripe].dev->bdev;
267		bio->bi_sector = begin + sc->stripe[target_stripe].physical_start;
268		bio->bi_size = to_bytes(end - begin);
269		return DM_MAPIO_REMAPPED;
270	} else {
271		/* The range doesn't map to the target stripe */
272		bio_endio(bio, 0);
273		return DM_MAPIO_SUBMITTED;
274	}
275}
276
277static int stripe_map(struct dm_target *ti, struct bio *bio)
278{
279	struct stripe_c *sc = ti->private;
280	uint32_t stripe;
281	unsigned target_bio_nr;
282
283	if (bio->bi_rw & REQ_FLUSH) {
284		target_bio_nr = dm_bio_get_target_bio_nr(bio);
285		BUG_ON(target_bio_nr >= sc->stripes);
286		bio->bi_bdev = sc->stripe[target_bio_nr].dev->bdev;
287		return DM_MAPIO_REMAPPED;
288	}
289	if (unlikely(bio->bi_rw & REQ_DISCARD) ||
290	    unlikely(bio->bi_rw & REQ_WRITE_SAME)) {
291		target_bio_nr = dm_bio_get_target_bio_nr(bio);
292		BUG_ON(target_bio_nr >= sc->stripes);
293		return stripe_map_range(sc, bio, target_bio_nr);
294	}
295
296	stripe_map_sector(sc, bio->bi_sector, &stripe, &bio->bi_sector);
297
298	bio->bi_sector += sc->stripe[stripe].physical_start;
299	bio->bi_bdev = sc->stripe[stripe].dev->bdev;
300
301	return DM_MAPIO_REMAPPED;
302}
303
304/*
305 * Stripe status:
306 *
307 * INFO
308 * #stripes [stripe_name <stripe_name>] [group word count]
309 * [error count 'A|D' <error count 'A|D'>]
310 *
311 * TABLE
312 * #stripes [stripe chunk size]
313 * [stripe_name physical_start <stripe_name physical_start>]
314 *
315 */
316
317static void stripe_status(struct dm_target *ti, status_type_t type,
318			  unsigned status_flags, char *result, unsigned maxlen)
319{
320	struct stripe_c *sc = (struct stripe_c *) ti->private;
321	char buffer[sc->stripes + 1];
322	unsigned int sz = 0;
323	unsigned int i;
324
325	switch (type) {
326	case STATUSTYPE_INFO:
327		DMEMIT("%d ", sc->stripes);
328		for (i = 0; i < sc->stripes; i++)  {
329			DMEMIT("%s ", sc->stripe[i].dev->name);
330			buffer[i] = atomic_read(&(sc->stripe[i].error_count)) ?
331				'D' : 'A';
332		}
333		buffer[i] = '\0';
334		DMEMIT("1 %s", buffer);
335		break;
336
337	case STATUSTYPE_TABLE:
338		DMEMIT("%d %llu", sc->stripes,
339			(unsigned long long)sc->chunk_size);
340		for (i = 0; i < sc->stripes; i++)
341			DMEMIT(" %s %llu", sc->stripe[i].dev->name,
342			    (unsigned long long)sc->stripe[i].physical_start);
343		break;
344	}
345}
346
347static int stripe_end_io(struct dm_target *ti, struct bio *bio, int error)
348{
349	unsigned i;
350	char major_minor[16];
351	struct stripe_c *sc = ti->private;
352
353	if (!error)
354		return 0; /* I/O complete */
355
356	if ((error == -EWOULDBLOCK) && (bio->bi_rw & REQ_RAHEAD))
357		return error;
358
359	if (error == -EOPNOTSUPP)
360		return error;
361
362	memset(major_minor, 0, sizeof(major_minor));
363	sprintf(major_minor, "%d:%d",
364		MAJOR(disk_devt(bio->bi_bdev->bd_disk)),
365		MINOR(disk_devt(bio->bi_bdev->bd_disk)));
366
367	/*
368	 * Test to see which stripe drive triggered the event
369	 * and increment error count for all stripes on that device.
370	 * If the error count for a given device exceeds the threshold
371	 * value we will no longer trigger any further events.
372	 */
373	for (i = 0; i < sc->stripes; i++)
374		if (!strcmp(sc->stripe[i].dev->name, major_minor)) {
375			atomic_inc(&(sc->stripe[i].error_count));
376			if (atomic_read(&(sc->stripe[i].error_count)) <
377			    DM_IO_ERROR_THRESHOLD)
378				schedule_work(&sc->trigger_event);
379		}
380
381	return error;
382}
383
384static int stripe_iterate_devices(struct dm_target *ti,
385				  iterate_devices_callout_fn fn, void *data)
386{
387	struct stripe_c *sc = ti->private;
388	int ret = 0;
389	unsigned i = 0;
390
391	do {
392		ret = fn(ti, sc->stripe[i].dev,
393			 sc->stripe[i].physical_start,
394			 sc->stripe_width, data);
395	} while (!ret && ++i < sc->stripes);
396
397	return ret;
398}
399
400static void stripe_io_hints(struct dm_target *ti,
401			    struct queue_limits *limits)
402{
403	struct stripe_c *sc = ti->private;
404	unsigned chunk_size = sc->chunk_size << SECTOR_SHIFT;
405
406	blk_limits_io_min(limits, chunk_size);
407	blk_limits_io_opt(limits, chunk_size * sc->stripes);
408}
409
410static int stripe_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
411			struct bio_vec *biovec, int max_size)
412{
413	struct stripe_c *sc = ti->private;
414	sector_t bvm_sector = bvm->bi_sector;
415	uint32_t stripe;
416	struct request_queue *q;
417
418	stripe_map_sector(sc, bvm_sector, &stripe, &bvm_sector);
419
420	q = bdev_get_queue(sc->stripe[stripe].dev->bdev);
421	if (!q->merge_bvec_fn)
422		return max_size;
423
424	bvm->bi_bdev = sc->stripe[stripe].dev->bdev;
425	bvm->bi_sector = sc->stripe[stripe].physical_start + bvm_sector;
426
427	return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
428}
429
430static struct target_type stripe_target = {
431	.name   = "striped",
432	.version = {1, 5, 1},
433	.module = THIS_MODULE,
434	.ctr    = stripe_ctr,
435	.dtr    = stripe_dtr,
436	.map    = stripe_map,
437	.end_io = stripe_end_io,
438	.status = stripe_status,
439	.iterate_devices = stripe_iterate_devices,
440	.io_hints = stripe_io_hints,
441	.merge  = stripe_merge,
442};
443
444int __init dm_stripe_init(void)
445{
446	int r;
447
448	r = dm_register_target(&stripe_target);
449	if (r < 0) {
450		DMWARN("target registration failed");
451		return r;
452	}
453
454	return r;
455}
456
457void dm_stripe_exit(void)
458{
459	dm_unregister_target(&stripe_target);
460}
461