1/*
2 * Copyright (C) 2003 Sistina Software
3 * Copyright (C) 2006 Red Hat GmbH
4 *
5 * This file is released under the GPL.
6 */
7
8#include "dm.h"
9
10#include <linux/device-mapper.h>
11
12#include <linux/bio.h>
13#include <linux/completion.h>
14#include <linux/mempool.h>
15#include <linux/module.h>
16#include <linux/sched.h>
17#include <linux/slab.h>
18#include <linux/dm-io.h>
19
20#define DM_MSG_PREFIX "io"
21
22#define DM_IO_MAX_REGIONS	BITS_PER_LONG
23
24struct dm_io_client {
25	mempool_t *pool;
26	struct bio_set *bios;
27};
28
29/*
30 * Aligning 'struct io' reduces the number of bits required to store
31 * its address.  Refer to store_io_and_region_in_bio() below.
32 */
33struct io {
34	unsigned long error_bits;
35	atomic_t count;
36	struct dm_io_client *client;
37	io_notify_fn callback;
38	void *context;
39	void *vma_invalidate_address;
40	unsigned long vma_invalidate_size;
41} __attribute__((aligned(DM_IO_MAX_REGIONS)));
42
43static struct kmem_cache *_dm_io_cache;
44
45/*
46 * Create a client with mempool and bioset.
47 */
48struct dm_io_client *dm_io_client_create(void)
49{
50	struct dm_io_client *client;
51	unsigned min_ios = dm_get_reserved_bio_based_ios();
52
53	client = kmalloc(sizeof(*client), GFP_KERNEL);
54	if (!client)
55		return ERR_PTR(-ENOMEM);
56
57	client->pool = mempool_create_slab_pool(min_ios, _dm_io_cache);
58	if (!client->pool)
59		goto bad;
60
61	client->bios = bioset_create(min_ios, 0);
62	if (!client->bios)
63		goto bad;
64
65	return client;
66
67   bad:
68	mempool_destroy(client->pool);
69	kfree(client);
70	return ERR_PTR(-ENOMEM);
71}
72EXPORT_SYMBOL(dm_io_client_create);
73
74void dm_io_client_destroy(struct dm_io_client *client)
75{
76	mempool_destroy(client->pool);
77	bioset_free(client->bios);
78	kfree(client);
79}
80EXPORT_SYMBOL(dm_io_client_destroy);
81
82/*-----------------------------------------------------------------
83 * We need to keep track of which region a bio is doing io for.
84 * To avoid a memory allocation to store just 5 or 6 bits, we
85 * ensure the 'struct io' pointer is aligned so enough low bits are
86 * always zero and then combine it with the region number directly in
87 * bi_private.
88 *---------------------------------------------------------------*/
89static void store_io_and_region_in_bio(struct bio *bio, struct io *io,
90				       unsigned region)
91{
92	if (unlikely(!IS_ALIGNED((unsigned long)io, DM_IO_MAX_REGIONS))) {
93		DMCRIT("Unaligned struct io pointer %p", io);
94		BUG();
95	}
96
97	bio->bi_private = (void *)((unsigned long)io | region);
98}
99
100static void retrieve_io_and_region_from_bio(struct bio *bio, struct io **io,
101				       unsigned *region)
102{
103	unsigned long val = (unsigned long)bio->bi_private;
104
105	*io = (void *)(val & -(unsigned long)DM_IO_MAX_REGIONS);
106	*region = val & (DM_IO_MAX_REGIONS - 1);
107}
108
109/*-----------------------------------------------------------------
110 * We need an io object to keep track of the number of bios that
111 * have been dispatched for a particular io.
112 *---------------------------------------------------------------*/
113static void complete_io(struct io *io)
114{
115	unsigned long error_bits = io->error_bits;
116	io_notify_fn fn = io->callback;
117	void *context = io->context;
118
119	if (io->vma_invalidate_size)
120		invalidate_kernel_vmap_range(io->vma_invalidate_address,
121					     io->vma_invalidate_size);
122
123	mempool_free(io, io->client->pool);
124	fn(error_bits, context);
125}
126
127static void dec_count(struct io *io, unsigned int region, int error)
128{
129	if (error)
130		set_bit(region, &io->error_bits);
131
132	if (atomic_dec_and_test(&io->count))
133		complete_io(io);
134}
135
136static void endio(struct bio *bio)
137{
138	struct io *io;
139	unsigned region;
140	int error;
141
142	if (bio->bi_error && bio_data_dir(bio) == READ)
143		zero_fill_bio(bio);
144
145	/*
146	 * The bio destructor in bio_put() may use the io object.
147	 */
148	retrieve_io_and_region_from_bio(bio, &io, &region);
149
150	error = bio->bi_error;
151	bio_put(bio);
152
153	dec_count(io, region, error);
154}
155
156/*-----------------------------------------------------------------
157 * These little objects provide an abstraction for getting a new
158 * destination page for io.
159 *---------------------------------------------------------------*/
160struct dpages {
161	void (*get_page)(struct dpages *dp,
162			 struct page **p, unsigned long *len, unsigned *offset);
163	void (*next_page)(struct dpages *dp);
164
165	unsigned context_u;
166	void *context_ptr;
167
168	void *vma_invalidate_address;
169	unsigned long vma_invalidate_size;
170};
171
172/*
173 * Functions for getting the pages from a list.
174 */
175static void list_get_page(struct dpages *dp,
176		  struct page **p, unsigned long *len, unsigned *offset)
177{
178	unsigned o = dp->context_u;
179	struct page_list *pl = (struct page_list *) dp->context_ptr;
180
181	*p = pl->page;
182	*len = PAGE_SIZE - o;
183	*offset = o;
184}
185
186static void list_next_page(struct dpages *dp)
187{
188	struct page_list *pl = (struct page_list *) dp->context_ptr;
189	dp->context_ptr = pl->next;
190	dp->context_u = 0;
191}
192
193static void list_dp_init(struct dpages *dp, struct page_list *pl, unsigned offset)
194{
195	dp->get_page = list_get_page;
196	dp->next_page = list_next_page;
197	dp->context_u = offset;
198	dp->context_ptr = pl;
199}
200
201/*
202 * Functions for getting the pages from a bvec.
203 */
204static void bio_get_page(struct dpages *dp, struct page **p,
205			 unsigned long *len, unsigned *offset)
206{
207	struct bio_vec *bvec = dp->context_ptr;
208	*p = bvec->bv_page;
209	*len = bvec->bv_len - dp->context_u;
210	*offset = bvec->bv_offset + dp->context_u;
211}
212
213static void bio_next_page(struct dpages *dp)
214{
215	struct bio_vec *bvec = dp->context_ptr;
216	dp->context_ptr = bvec + 1;
217	dp->context_u = 0;
218}
219
220static void bio_dp_init(struct dpages *dp, struct bio *bio)
221{
222	dp->get_page = bio_get_page;
223	dp->next_page = bio_next_page;
224	dp->context_ptr = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
225	dp->context_u = bio->bi_iter.bi_bvec_done;
226}
227
228/*
229 * Functions for getting the pages from a VMA.
230 */
231static void vm_get_page(struct dpages *dp,
232		 struct page **p, unsigned long *len, unsigned *offset)
233{
234	*p = vmalloc_to_page(dp->context_ptr);
235	*offset = dp->context_u;
236	*len = PAGE_SIZE - dp->context_u;
237}
238
239static void vm_next_page(struct dpages *dp)
240{
241	dp->context_ptr += PAGE_SIZE - dp->context_u;
242	dp->context_u = 0;
243}
244
245static void vm_dp_init(struct dpages *dp, void *data)
246{
247	dp->get_page = vm_get_page;
248	dp->next_page = vm_next_page;
249	dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1);
250	dp->context_ptr = data;
251}
252
253/*
254 * Functions for getting the pages from kernel memory.
255 */
256static void km_get_page(struct dpages *dp, struct page **p, unsigned long *len,
257			unsigned *offset)
258{
259	*p = virt_to_page(dp->context_ptr);
260	*offset = dp->context_u;
261	*len = PAGE_SIZE - dp->context_u;
262}
263
264static void km_next_page(struct dpages *dp)
265{
266	dp->context_ptr += PAGE_SIZE - dp->context_u;
267	dp->context_u = 0;
268}
269
270static void km_dp_init(struct dpages *dp, void *data)
271{
272	dp->get_page = km_get_page;
273	dp->next_page = km_next_page;
274	dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1);
275	dp->context_ptr = data;
276}
277
278/*-----------------------------------------------------------------
279 * IO routines that accept a list of pages.
280 *---------------------------------------------------------------*/
281static void do_region(int rw, unsigned region, struct dm_io_region *where,
282		      struct dpages *dp, struct io *io)
283{
284	struct bio *bio;
285	struct page *page;
286	unsigned long len;
287	unsigned offset;
288	unsigned num_bvecs;
289	sector_t remaining = where->count;
290	struct request_queue *q = bdev_get_queue(where->bdev);
291	unsigned short logical_block_size = queue_logical_block_size(q);
292	sector_t num_sectors;
293	unsigned int uninitialized_var(special_cmd_max_sectors);
294
295	/*
296	 * Reject unsupported discard and write same requests.
297	 */
298	if (rw & REQ_DISCARD)
299		special_cmd_max_sectors = q->limits.max_discard_sectors;
300	else if (rw & REQ_WRITE_SAME)
301		special_cmd_max_sectors = q->limits.max_write_same_sectors;
302	if ((rw & (REQ_DISCARD | REQ_WRITE_SAME)) && special_cmd_max_sectors == 0) {
303		dec_count(io, region, -EOPNOTSUPP);
304		return;
305	}
306
307	/*
308	 * where->count may be zero if rw holds a flush and we need to
309	 * send a zero-sized flush.
310	 */
311	do {
312		/*
313		 * Allocate a suitably sized-bio.
314		 */
315		if ((rw & REQ_DISCARD) || (rw & REQ_WRITE_SAME))
316			num_bvecs = 1;
317		else
318			num_bvecs = min_t(int, BIO_MAX_PAGES,
319					  dm_sector_div_up(remaining, (PAGE_SIZE >> SECTOR_SHIFT)));
320
321		bio = bio_alloc_bioset(GFP_NOIO, num_bvecs, io->client->bios);
322		bio->bi_iter.bi_sector = where->sector + (where->count - remaining);
323		bio->bi_bdev = where->bdev;
324		bio->bi_end_io = endio;
325		store_io_and_region_in_bio(bio, io, region);
326
327		if (rw & REQ_DISCARD) {
328			num_sectors = min_t(sector_t, special_cmd_max_sectors, remaining);
329			bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
330			remaining -= num_sectors;
331		} else if (rw & REQ_WRITE_SAME) {
332			/*
333			 * WRITE SAME only uses a single page.
334			 */
335			dp->get_page(dp, &page, &len, &offset);
336			bio_add_page(bio, page, logical_block_size, offset);
337			num_sectors = min_t(sector_t, special_cmd_max_sectors, remaining);
338			bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
339
340			offset = 0;
341			remaining -= num_sectors;
342			dp->next_page(dp);
343		} else while (remaining) {
344			/*
345			 * Try and add as many pages as possible.
346			 */
347			dp->get_page(dp, &page, &len, &offset);
348			len = min(len, to_bytes(remaining));
349			if (!bio_add_page(bio, page, len, offset))
350				break;
351
352			offset = 0;
353			remaining -= to_sector(len);
354			dp->next_page(dp);
355		}
356
357		atomic_inc(&io->count);
358		submit_bio(rw, bio);
359	} while (remaining);
360}
361
362static void dispatch_io(int rw, unsigned int num_regions,
363			struct dm_io_region *where, struct dpages *dp,
364			struct io *io, int sync)
365{
366	int i;
367	struct dpages old_pages = *dp;
368
369	BUG_ON(num_regions > DM_IO_MAX_REGIONS);
370
371	if (sync)
372		rw |= REQ_SYNC;
373
374	/*
375	 * For multiple regions we need to be careful to rewind
376	 * the dp object for each call to do_region.
377	 */
378	for (i = 0; i < num_regions; i++) {
379		*dp = old_pages;
380		if (where[i].count || (rw & REQ_FLUSH))
381			do_region(rw, i, where + i, dp, io);
382	}
383
384	/*
385	 * Drop the extra reference that we were holding to avoid
386	 * the io being completed too early.
387	 */
388	dec_count(io, 0, 0);
389}
390
391struct sync_io {
392	unsigned long error_bits;
393	struct completion wait;
394};
395
396static void sync_io_complete(unsigned long error, void *context)
397{
398	struct sync_io *sio = context;
399
400	sio->error_bits = error;
401	complete(&sio->wait);
402}
403
404static int sync_io(struct dm_io_client *client, unsigned int num_regions,
405		   struct dm_io_region *where, int rw, struct dpages *dp,
406		   unsigned long *error_bits)
407{
408	struct io *io;
409	struct sync_io sio;
410
411	if (num_regions > 1 && (rw & RW_MASK) != WRITE) {
412		WARN_ON(1);
413		return -EIO;
414	}
415
416	init_completion(&sio.wait);
417
418	io = mempool_alloc(client->pool, GFP_NOIO);
419	io->error_bits = 0;
420	atomic_set(&io->count, 1); /* see dispatch_io() */
421	io->client = client;
422	io->callback = sync_io_complete;
423	io->context = &sio;
424
425	io->vma_invalidate_address = dp->vma_invalidate_address;
426	io->vma_invalidate_size = dp->vma_invalidate_size;
427
428	dispatch_io(rw, num_regions, where, dp, io, 1);
429
430	wait_for_completion_io(&sio.wait);
431
432	if (error_bits)
433		*error_bits = sio.error_bits;
434
435	return sio.error_bits ? -EIO : 0;
436}
437
438static int async_io(struct dm_io_client *client, unsigned int num_regions,
439		    struct dm_io_region *where, int rw, struct dpages *dp,
440		    io_notify_fn fn, void *context)
441{
442	struct io *io;
443
444	if (num_regions > 1 && (rw & RW_MASK) != WRITE) {
445		WARN_ON(1);
446		fn(1, context);
447		return -EIO;
448	}
449
450	io = mempool_alloc(client->pool, GFP_NOIO);
451	io->error_bits = 0;
452	atomic_set(&io->count, 1); /* see dispatch_io() */
453	io->client = client;
454	io->callback = fn;
455	io->context = context;
456
457	io->vma_invalidate_address = dp->vma_invalidate_address;
458	io->vma_invalidate_size = dp->vma_invalidate_size;
459
460	dispatch_io(rw, num_regions, where, dp, io, 0);
461	return 0;
462}
463
464static int dp_init(struct dm_io_request *io_req, struct dpages *dp,
465		   unsigned long size)
466{
467	/* Set up dpages based on memory type */
468
469	dp->vma_invalidate_address = NULL;
470	dp->vma_invalidate_size = 0;
471
472	switch (io_req->mem.type) {
473	case DM_IO_PAGE_LIST:
474		list_dp_init(dp, io_req->mem.ptr.pl, io_req->mem.offset);
475		break;
476
477	case DM_IO_BIO:
478		bio_dp_init(dp, io_req->mem.ptr.bio);
479		break;
480
481	case DM_IO_VMA:
482		flush_kernel_vmap_range(io_req->mem.ptr.vma, size);
483		if ((io_req->bi_rw & RW_MASK) == READ) {
484			dp->vma_invalidate_address = io_req->mem.ptr.vma;
485			dp->vma_invalidate_size = size;
486		}
487		vm_dp_init(dp, io_req->mem.ptr.vma);
488		break;
489
490	case DM_IO_KMEM:
491		km_dp_init(dp, io_req->mem.ptr.addr);
492		break;
493
494	default:
495		return -EINVAL;
496	}
497
498	return 0;
499}
500
501/*
502 * New collapsed (a)synchronous interface.
503 *
504 * If the IO is asynchronous (i.e. it has notify.fn), you must either unplug
505 * the queue with blk_unplug() some time later or set REQ_SYNC in io_req->bi_rw.
506 * If you fail to do one of these, the IO will be submitted to the disk after
507 * q->unplug_delay, which defaults to 3ms in blk-settings.c.
508 */
509int dm_io(struct dm_io_request *io_req, unsigned num_regions,
510	  struct dm_io_region *where, unsigned long *sync_error_bits)
511{
512	int r;
513	struct dpages dp;
514
515	r = dp_init(io_req, &dp, (unsigned long)where->count << SECTOR_SHIFT);
516	if (r)
517		return r;
518
519	if (!io_req->notify.fn)
520		return sync_io(io_req->client, num_regions, where,
521			       io_req->bi_rw, &dp, sync_error_bits);
522
523	return async_io(io_req->client, num_regions, where, io_req->bi_rw,
524			&dp, io_req->notify.fn, io_req->notify.context);
525}
526EXPORT_SYMBOL(dm_io);
527
528int __init dm_io_init(void)
529{
530	_dm_io_cache = KMEM_CACHE(io, 0);
531	if (!_dm_io_cache)
532		return -ENOMEM;
533
534	return 0;
535}
536
537void dm_io_exit(void)
538{
539	kmem_cache_destroy(_dm_io_cache);
540	_dm_io_cache = NULL;
541}
542