dm-stripe.c

  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * Copyright (C) 2001-2003 Sistina Software (UK) Limited.
  4 *
  5 * This file is released under the GPL.
  6 */
  7
  8#include "dm.h"
  9#include <linux/device-mapper.h>
 10
 11#include <linux/module.h>
 12#include <linux/init.h>
 13#include <linux/blkdev.h>
 14#include <linux/bio.h>
 15#include <linux/dax.h>
 16#include <linux/slab.h>
 17#include <linux/log2.h>
 18
 19static struct workqueue_struct *dm_stripe_wq;
 20
 21#define DM_MSG_PREFIX "striped"
 22#define DM_IO_ERROR_THRESHOLD 15
 23
 24struct stripe {
 25	struct dm_dev *dev;
 26	sector_t physical_start;
 27
 28	atomic_t error_count;
 29};
 30
 31struct stripe_c {
 32	uint32_t stripes;
 33	int stripes_shift;
 34
 35	/* The size of this target / num. stripes */
 36	sector_t stripe_width;
 37
 38	uint32_t chunk_size;
 39	int chunk_size_shift;
 40
 41	/* Needed for handling events */
 42	struct dm_target *ti;
 43
 44	/* Work struct used for triggering events*/
 45	struct work_struct trigger_event;
 46
 47	struct stripe stripe[] __counted_by(stripes);
 48};
 49
 50/*
 51 * An event is triggered whenever a drive
 52 * drops out of a stripe volume.
 53 */
 54static void trigger_event(struct work_struct *work)
 55{
 56	struct stripe_c *sc = container_of(work, struct stripe_c,
 57					   trigger_event);
 58	dm_table_event(sc->ti->table);
 59}
 60
 61/*
 62 * Parse a single <dev> <sector> pair
 63 */
 64static int get_stripe(struct dm_target *ti, struct stripe_c *sc,
 65		      unsigned int stripe, char **argv)
 66{
 67	unsigned long long start;
 68	char dummy;
 69	int ret;
 70
 71	if (sscanf(argv[1], "%llu%c", &start, &dummy) != 1)
 72		return -EINVAL;
 73
 74	ret = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table),
 75			    &sc->stripe[stripe].dev);
 76	if (ret)
 77		return ret;
 78
 79	sc->stripe[stripe].physical_start = start;
 80
 81	return 0;
 82}
 83
 84/*
 85 * Construct a striped mapping.
 86 * <number of stripes> <chunk size> [<dev_path> <offset>]+
 87 */
 88static int stripe_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 89{
 90	struct stripe_c *sc;
 91	sector_t width, tmp_len;
 92	uint32_t stripes;
 93	uint32_t chunk_size;
 94	int r;
 95	unsigned int i;
 96
 97	if (argc < 2) {
 98		ti->error = "Not enough arguments";
 99		return -EINVAL;
100	}
101
102	if (kstrtouint(argv[0], 10, &stripes) || !stripes) {
103		ti->error = "Invalid stripe count";
104		return -EINVAL;
105	}
106
107	if (kstrtouint(argv[1], 10, &chunk_size) || !chunk_size) {
108		ti->error = "Invalid chunk_size";
109		return -EINVAL;
110	}
111
112	width = ti->len;
113	if (sector_div(width, stripes)) {
114		ti->error = "Target length not divisible by number of stripes";
115		return -EINVAL;
116	}
117
118	tmp_len = width;
119	if (sector_div(tmp_len, chunk_size)) {
120		ti->error = "Target length not divisible by chunk size";
121		return -EINVAL;
122	}
123
124	/*
125	 * Do we have enough arguments for that many stripes ?
126	 */
127	if (argc != (2 + 2 * stripes)) {
128		ti->error = "Not enough destinations specified";
129		return -EINVAL;
130	}
131
132	sc = kmalloc(struct_size(sc, stripe, stripes), GFP_KERNEL);
133	if (!sc) {
134		ti->error = "Memory allocation for striped context failed";
135		return -ENOMEM;
136	}
137
138	INIT_WORK(&sc->trigger_event, trigger_event);
139
140	/* Set pointer to dm target; used in trigger_event */
141	sc->ti = ti;
142	sc->stripes = stripes;
143	sc->stripe_width = width;
144
145	if (stripes & (stripes - 1))
146		sc->stripes_shift = -1;
147	else
148		sc->stripes_shift = __ffs(stripes);
149
150	r = dm_set_target_max_io_len(ti, chunk_size);
151	if (r) {
152		kfree(sc);
153		return r;
154	}
155
156	ti->num_flush_bios = stripes;
157	ti->num_discard_bios = stripes;
158	ti->num_secure_erase_bios = stripes;
159	ti->num_write_zeroes_bios = stripes;
160	ti->flush_bypasses_map = true;
161
162	sc->chunk_size = chunk_size;
163	if (chunk_size & (chunk_size - 1))
164		sc->chunk_size_shift = -1;
165	else
166		sc->chunk_size_shift = __ffs(chunk_size);
167
168	/*
169	 * Get the stripe destinations.
170	 */
171	for (i = 0; i < stripes; i++) {
172		argv += 2;
173
174		r = get_stripe(ti, sc, i, argv);
175		if (r < 0) {
176			ti->error = "Couldn't parse stripe destination";
177			while (i--)
178				dm_put_device(ti, sc->stripe[i].dev);
179			kfree(sc);
180			return r;
181		}
182		atomic_set(&(sc->stripe[i].error_count), 0);
183	}
184
185	ti->private = sc;
186
187	return 0;
188}
189
190static void stripe_dtr(struct dm_target *ti)
191{
192	unsigned int i;
193	struct stripe_c *sc = ti->private;
194
195	for (i = 0; i < sc->stripes; i++)
196		dm_put_device(ti, sc->stripe[i].dev);
197
198	flush_work(&sc->trigger_event);
199	kfree(sc);
200}
201
202static void stripe_map_sector(struct stripe_c *sc, sector_t sector,
203			      uint32_t *stripe, sector_t *result)
204{
205	sector_t chunk = dm_target_offset(sc->ti, sector);
206	sector_t chunk_offset;
207
208	if (sc->chunk_size_shift < 0)
209		chunk_offset = sector_div(chunk, sc->chunk_size);
210	else {
211		chunk_offset = chunk & (sc->chunk_size - 1);
212		chunk >>= sc->chunk_size_shift;
213	}
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	if (sc->chunk_size_shift < 0)
223		chunk *= sc->chunk_size;
224	else
225		chunk <<= sc->chunk_size_shift;
226
227	*result = chunk + chunk_offset;
228}
229
230static void stripe_map_range_sector(struct stripe_c *sc, sector_t sector,
231				    uint32_t target_stripe, sector_t *result)
232{
233	uint32_t stripe;
234
235	stripe_map_sector(sc, sector, &stripe, result);
236	if (stripe == target_stripe)
237		return;
238
239	/* round down */
240	sector = *result;
241	if (sc->chunk_size_shift < 0)
242		*result -= sector_div(sector, sc->chunk_size);
243	else
244		*result = sector & ~(sector_t)(sc->chunk_size - 1);
245
246	if (target_stripe < stripe)
247		*result += sc->chunk_size;		/* next chunk */
248}
249
250static int stripe_map_range(struct stripe_c *sc, struct bio *bio,
251			    uint32_t target_stripe)
252{
253	sector_t begin, end;
254
255	stripe_map_range_sector(sc, bio->bi_iter.bi_sector,
256				target_stripe, &begin);
257	stripe_map_range_sector(sc, bio_end_sector(bio),
258				target_stripe, &end);
259	if (begin < end) {
260		bio_set_dev(bio, sc->stripe[target_stripe].dev->bdev);
261		bio->bi_iter.bi_sector = begin +
262			sc->stripe[target_stripe].physical_start;
263		bio->bi_iter.bi_size = to_bytes(end - begin);
264		return DM_MAPIO_REMAPPED;
265	}
266
267	/* The range doesn't map to the target stripe */
268	bio_endio(bio);
269	return DM_MAPIO_SUBMITTED;
270}
271
272int stripe_map(struct dm_target *ti, struct bio *bio)
273{
274	struct stripe_c *sc = ti->private;
275	uint32_t stripe;
276	unsigned int target_bio_nr;
277
278	if (bio->bi_opf & REQ_PREFLUSH) {
279		target_bio_nr = dm_bio_get_target_bio_nr(bio);
280		BUG_ON(target_bio_nr >= sc->stripes);
281		bio_set_dev(bio, sc->stripe[target_bio_nr].dev->bdev);
282		return DM_MAPIO_REMAPPED;
283	}
284	if (unlikely(bio_op(bio) == REQ_OP_DISCARD) ||
285	    unlikely(bio_op(bio) == REQ_OP_SECURE_ERASE) ||
286	    unlikely(bio_op(bio) == REQ_OP_WRITE_ZEROES)) {
287		target_bio_nr = dm_bio_get_target_bio_nr(bio);
288		BUG_ON(target_bio_nr >= sc->stripes);
289		return stripe_map_range(sc, bio, target_bio_nr);
290	}
291
292	stripe_map_sector(sc, bio->bi_iter.bi_sector,
293			  &stripe, &bio->bi_iter.bi_sector);
294
295	bio->bi_iter.bi_sector += sc->stripe[stripe].physical_start;
296	bio_set_dev(bio, sc->stripe[stripe].dev->bdev);
297
298	return DM_MAPIO_REMAPPED;
299}
300
301#if IS_ENABLED(CONFIG_FS_DAX)
302static struct dax_device *stripe_dax_pgoff(struct dm_target *ti, pgoff_t *pgoff)
303{
304	struct stripe_c *sc = ti->private;
305	struct block_device *bdev;
306	sector_t dev_sector;
307	uint32_t stripe;
308
309	stripe_map_sector(sc, *pgoff * PAGE_SECTORS, &stripe, &dev_sector);
310	dev_sector += sc->stripe[stripe].physical_start;
311	bdev = sc->stripe[stripe].dev->bdev;
312
313	*pgoff = (get_start_sect(bdev) + dev_sector) >> PAGE_SECTORS_SHIFT;
314	return sc->stripe[stripe].dev->dax_dev;
315}
316
317static long stripe_dax_direct_access(struct dm_target *ti, pgoff_t pgoff,
318		long nr_pages, enum dax_access_mode mode, void **kaddr,
319		unsigned long *pfn)
320{
321	struct dax_device *dax_dev = stripe_dax_pgoff(ti, &pgoff);
322
323	return dax_direct_access(dax_dev, pgoff, nr_pages, mode, kaddr, pfn);
324}
325
326static int stripe_dax_zero_page_range(struct dm_target *ti, pgoff_t pgoff,
327				      size_t nr_pages)
328{
329	struct dax_device *dax_dev = stripe_dax_pgoff(ti, &pgoff);
330
331	return dax_zero_page_range(dax_dev, pgoff, nr_pages);
332}
333
334static size_t stripe_dax_recovery_write(struct dm_target *ti, pgoff_t pgoff,
335		void *addr, size_t bytes, struct iov_iter *i)
336{
337	struct dax_device *dax_dev = stripe_dax_pgoff(ti, &pgoff);
338
339	return dax_recovery_write(dax_dev, pgoff, addr, bytes, i);
340}
341
342#else
343#define stripe_dax_direct_access NULL
344#define stripe_dax_zero_page_range NULL
345#define stripe_dax_recovery_write NULL
346#endif
347
348/*
349 * Stripe status:
350 *
351 * INFO
352 * #stripes [stripe_name <stripe_name>] [group word count]
353 * [error count 'A|D' <error count 'A|D'>]
354 *
355 * TABLE
356 * #stripes [stripe chunk size]
357 * [stripe_name physical_start <stripe_name physical_start>]
358 *
359 */
360
361static void stripe_status(struct dm_target *ti, status_type_t type,
362			  unsigned int status_flags, char *result, unsigned int maxlen)
363{
364	struct stripe_c *sc = ti->private;
365	unsigned int sz = 0;
366	unsigned int i;
367
368	switch (type) {
369	case STATUSTYPE_INFO:
370		DMEMIT("%d ", sc->stripes);
371		for (i = 0; i < sc->stripes; i++)
372			DMEMIT("%s ", sc->stripe[i].dev->name);
373
374		DMEMIT("1 ");
375		for (i = 0; i < sc->stripes; i++)
376			DMEMIT("%c", atomic_read(&(sc->stripe[i].error_count)) ?  'D' : 'A');
377		break;
378
379	case STATUSTYPE_TABLE:
380		DMEMIT("%d %llu", sc->stripes,
381			(unsigned long long)sc->chunk_size);
382		for (i = 0; i < sc->stripes; i++)
383			DMEMIT(" %s %llu", sc->stripe[i].dev->name,
384			    (unsigned long long)sc->stripe[i].physical_start);
385		break;
386
387	case STATUSTYPE_IMA:
388		DMEMIT_TARGET_NAME_VERSION(ti->type);
389		DMEMIT(",stripes=%d,chunk_size=%llu", sc->stripes,
390		       (unsigned long long)sc->chunk_size);
391
392		for (i = 0; i < sc->stripes; i++) {
393			DMEMIT(",stripe_%d_device_name=%s", i, sc->stripe[i].dev->name);
394			DMEMIT(",stripe_%d_physical_start=%llu", i,
395			       (unsigned long long)sc->stripe[i].physical_start);
396			DMEMIT(",stripe_%d_status=%c", i,
397			       atomic_read(&(sc->stripe[i].error_count)) ? 'D' : 'A');
398		}
399		DMEMIT(";");
400		break;
401	}
402}
403
404static int stripe_end_io(struct dm_target *ti, struct bio *bio,
405		blk_status_t *error)
406{
407	unsigned int i;
408	char major_minor[22];
409	struct stripe_c *sc = ti->private;
410
411	if (!*error)
412		return DM_ENDIO_DONE; /* I/O complete */
413
414	if (bio->bi_opf & REQ_RAHEAD)
415		return DM_ENDIO_DONE;
416
417	if (*error == BLK_STS_NOTSUPP)
418		return DM_ENDIO_DONE;
419
420	format_dev_t(major_minor, bio_dev(bio));
421
422	/*
423	 * Test to see which stripe drive triggered the event
424	 * and increment error count for all stripes on that device.
425	 * If the error count for a given device exceeds the threshold
426	 * value we will no longer trigger any further events.
427	 */
428	for (i = 0; i < sc->stripes; i++)
429		if (!strcmp(sc->stripe[i].dev->name, major_minor)) {
430			atomic_inc(&(sc->stripe[i].error_count));
431			if (atomic_read(&(sc->stripe[i].error_count)) <
432			    DM_IO_ERROR_THRESHOLD)
433				queue_work(dm_stripe_wq, &sc->trigger_event);
434		}
435
436	return DM_ENDIO_DONE;
437}
438
439static int stripe_iterate_devices(struct dm_target *ti,
440				  iterate_devices_callout_fn fn, void *data)
441{
442	struct stripe_c *sc = ti->private;
443	int ret = 0;
444	unsigned int i = 0;
445
446	do {
447		ret = fn(ti, sc->stripe[i].dev,
448			 sc->stripe[i].physical_start,
449			 sc->stripe_width, data);
450	} while (!ret && ++i < sc->stripes);
451
452	return ret;
453}
454
455static void stripe_io_hints(struct dm_target *ti,
456			    struct queue_limits *limits)
457{
458	struct stripe_c *sc = ti->private;
459	unsigned int io_min, io_opt;
460
461	limits->chunk_sectors = sc->chunk_size;
462
463	if (!check_shl_overflow(sc->chunk_size, SECTOR_SHIFT, &io_min) &&
464	    !check_mul_overflow(io_min, sc->stripes, &io_opt)) {
465		limits->io_min = io_min;
466		limits->io_opt = io_opt;
467	}
468}
469
470static struct target_type stripe_target = {
471	.name   = "striped",
472	.version = {1, 7, 0},
473	.features = DM_TARGET_PASSES_INTEGRITY | DM_TARGET_NOWAIT |
474		    DM_TARGET_ATOMIC_WRITES | DM_TARGET_PASSES_CRYPTO,
475	.module = THIS_MODULE,
476	.ctr    = stripe_ctr,
477	.dtr    = stripe_dtr,
478	.map    = stripe_map,
479	.end_io = stripe_end_io,
480	.status = stripe_status,
481	.iterate_devices = stripe_iterate_devices,
482	.io_hints = stripe_io_hints,
483	.direct_access = stripe_dax_direct_access,
484	.dax_zero_page_range = stripe_dax_zero_page_range,
485	.dax_recovery_write = stripe_dax_recovery_write,
486};
487
488int __init dm_stripe_init(void)
489{
490	int r;
491
492	dm_stripe_wq = alloc_workqueue("dm_stripe_wq", 0, 0);
493	if (!dm_stripe_wq)
494		return -ENOMEM;
495	r = dm_register_target(&stripe_target);
496	if (r < 0) {
497		destroy_workqueue(dm_stripe_wq);
498		DMWARN("target registration failed");
499	}
500
501	return r;
502}
503
504void dm_stripe_exit(void)
505{
506	dm_unregister_target(&stripe_target);
507	destroy_workqueue(dm_stripe_wq);
508}