dm-io.c revision bf17ce3a604d943f29bf1bc1a66a4e0d2ad4ec96 
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-io.h"
9
10#include <linux/bio.h>
11#include <linux/mempool.h>
12#include <linux/module.h>
13#include <linux/sched.h>
14#include <linux/slab.h>
15
16struct dm_io_client {
17	mempool_t *pool;
18	struct bio_set *bios;
19};
20
21/* FIXME: can we shrink this ? */
22struct io {
23	unsigned long error;
24	atomic_t count;
25	struct task_struct *sleeper;
26	struct dm_io_client *client;
27	io_notify_fn callback;
28	void *context;
29};
30
31/*
32 * io contexts are only dynamically allocated for asynchronous
33 * io.  Since async io is likely to be the majority of io we'll
34 * have the same number of io contexts as bios! (FIXME: must reduce this).
35 */
36
37static unsigned int pages_to_ios(unsigned int pages)
38{
39	return 4 * pages;	/* too many ? */
40}
41
42/*
43 * Create a client with mempool and bioset.
44 */
45struct dm_io_client *dm_io_client_create(unsigned num_pages)
46{
47	unsigned ios = pages_to_ios(num_pages);
48	struct dm_io_client *client;
49
50	client = kmalloc(sizeof(*client), GFP_KERNEL);
51	if (!client)
52		return ERR_PTR(-ENOMEM);
53
54	client->pool = mempool_create_kmalloc_pool(ios, sizeof(struct io));
55	if (!client->pool)
56		goto bad;
57
58	client->bios = bioset_create(16, 16);
59	if (!client->bios)
60		goto bad;
61
62	return client;
63
64   bad:
65	if (client->pool)
66		mempool_destroy(client->pool);
67	kfree(client);
68	return ERR_PTR(-ENOMEM);
69}
70EXPORT_SYMBOL(dm_io_client_create);
71
72int dm_io_client_resize(unsigned num_pages, struct dm_io_client *client)
73{
74	return mempool_resize(client->pool, pages_to_ios(num_pages),
75			      GFP_KERNEL);
76}
77EXPORT_SYMBOL(dm_io_client_resize);
78
79void dm_io_client_destroy(struct dm_io_client *client)
80{
81	mempool_destroy(client->pool);
82	bioset_free(client->bios);
83	kfree(client);
84}
85EXPORT_SYMBOL(dm_io_client_destroy);
86
87/*-----------------------------------------------------------------
88 * We need to keep track of which region a bio is doing io for.
89 * In order to save a memory allocation we store this the last
90 * bvec which we know is unused (blech).
91 * XXX This is ugly and can OOPS with some configs... find another way.
92 *---------------------------------------------------------------*/
93static inline void bio_set_region(struct bio *bio, unsigned region)
94{
95	bio->bi_io_vec[bio->bi_max_vecs].bv_len = region;
96}
97
98static inline unsigned bio_get_region(struct bio *bio)
99{
100	return bio->bi_io_vec[bio->bi_max_vecs].bv_len;
101}
102
103/*-----------------------------------------------------------------
104 * We need an io object to keep track of the number of bios that
105 * have been dispatched for a particular io.
106 *---------------------------------------------------------------*/
107static void dec_count(struct io *io, unsigned int region, int error)
108{
109	if (error)
110		set_bit(region, &io->error);
111
112	if (atomic_dec_and_test(&io->count)) {
113		if (io->sleeper)
114			wake_up_process(io->sleeper);
115
116		else {
117			int r = io->error;
118			io_notify_fn fn = io->callback;
119			void *context = io->context;
120
121			mempool_free(io, io->client->pool);
122			fn(r, context);
123		}
124	}
125}
126
127static int endio(struct bio *bio, unsigned int done, int error)
128{
129	struct io *io;
130	unsigned region;
131
132	/* keep going until we've finished */
133	if (bio->bi_size)
134		return 1;
135
136	if (error && bio_data_dir(bio) == READ)
137		zero_fill_bio(bio);
138
139	/*
140	 * The bio destructor in bio_put() may use the io object.
141	 */
142	io = bio->bi_private;
143	region = bio_get_region(bio);
144
145	bio->bi_max_vecs++;
146	bio_put(bio);
147
148	dec_count(io, region, error);
149
150	return 0;
151}
152
153/*-----------------------------------------------------------------
154 * These little objects provide an abstraction for getting a new
155 * destination page for io.
156 *---------------------------------------------------------------*/
157struct dpages {
158	void (*get_page)(struct dpages *dp,
159			 struct page **p, unsigned long *len, unsigned *offset);
160	void (*next_page)(struct dpages *dp);
161
162	unsigned context_u;
163	void *context_ptr;
164};
165
166/*
167 * Functions for getting the pages from a list.
168 */
169static void list_get_page(struct dpages *dp,
170		  struct page **p, unsigned long *len, unsigned *offset)
171{
172	unsigned o = dp->context_u;
173	struct page_list *pl = (struct page_list *) dp->context_ptr;
174
175	*p = pl->page;
176	*len = PAGE_SIZE - o;
177	*offset = o;
178}
179
180static void list_next_page(struct dpages *dp)
181{
182	struct page_list *pl = (struct page_list *) dp->context_ptr;
183	dp->context_ptr = pl->next;
184	dp->context_u = 0;
185}
186
187static void list_dp_init(struct dpages *dp, struct page_list *pl, unsigned offset)
188{
189	dp->get_page = list_get_page;
190	dp->next_page = list_next_page;
191	dp->context_u = offset;
192	dp->context_ptr = pl;
193}
194
195/*
196 * Functions for getting the pages from a bvec.
197 */
198static void bvec_get_page(struct dpages *dp,
199		  struct page **p, unsigned long *len, unsigned *offset)
200{
201	struct bio_vec *bvec = (struct bio_vec *) dp->context_ptr;
202	*p = bvec->bv_page;
203	*len = bvec->bv_len;
204	*offset = bvec->bv_offset;
205}
206
207static void bvec_next_page(struct dpages *dp)
208{
209	struct bio_vec *bvec = (struct bio_vec *) dp->context_ptr;
210	dp->context_ptr = bvec + 1;
211}
212
213static void bvec_dp_init(struct dpages *dp, struct bio_vec *bvec)
214{
215	dp->get_page = bvec_get_page;
216	dp->next_page = bvec_next_page;
217	dp->context_ptr = bvec;
218}
219
220/*
221 * Functions for getting the pages from a VMA.
222 */
223static void vm_get_page(struct dpages *dp,
224		 struct page **p, unsigned long *len, unsigned *offset)
225{
226	*p = vmalloc_to_page(dp->context_ptr);
227	*offset = dp->context_u;
228	*len = PAGE_SIZE - dp->context_u;
229}
230
231static void vm_next_page(struct dpages *dp)
232{
233	dp->context_ptr += PAGE_SIZE - dp->context_u;
234	dp->context_u = 0;
235}
236
237static void vm_dp_init(struct dpages *dp, void *data)
238{
239	dp->get_page = vm_get_page;
240	dp->next_page = vm_next_page;
241	dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1);
242	dp->context_ptr = data;
243}
244
245static void dm_bio_destructor(struct bio *bio)
246{
247	struct io *io = bio->bi_private;
248
249	bio_free(bio, io->client->bios);
250}
251
252/*
253 * Functions for getting the pages from kernel memory.
254 */
255static void km_get_page(struct dpages *dp, struct page **p, unsigned long *len,
256			unsigned *offset)
257{
258	*p = virt_to_page(dp->context_ptr);
259	*offset = dp->context_u;
260	*len = PAGE_SIZE - dp->context_u;
261}
262
263static void km_next_page(struct dpages *dp)
264{
265	dp->context_ptr += PAGE_SIZE - dp->context_u;
266	dp->context_u = 0;
267}
268
269static void km_dp_init(struct dpages *dp, void *data)
270{
271	dp->get_page = km_get_page;
272	dp->next_page = km_next_page;
273	dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1);
274	dp->context_ptr = data;
275}
276
277/*-----------------------------------------------------------------
278 * IO routines that accept a list of pages.
279 *---------------------------------------------------------------*/
280static void do_region(int rw, unsigned int region, struct io_region *where,
281		      struct dpages *dp, struct io *io)
282{
283	struct bio *bio;
284	struct page *page;
285	unsigned long len;
286	unsigned offset;
287	unsigned num_bvecs;
288	sector_t remaining = where->count;
289
290	while (remaining) {
291		/*
292		 * Allocate a suitably sized-bio: we add an extra
293		 * bvec for bio_get/set_region() and decrement bi_max_vecs
294		 * to hide it from bio_add_page().
295		 */
296		num_bvecs = (remaining / (PAGE_SIZE >> SECTOR_SHIFT)) + 2;
297		bio = bio_alloc_bioset(GFP_NOIO, num_bvecs, io->client->bios);
298		bio->bi_sector = where->sector + (where->count - remaining);
299		bio->bi_bdev = where->bdev;
300		bio->bi_end_io = endio;
301		bio->bi_private = io;
302		bio->bi_destructor = dm_bio_destructor;
303		bio->bi_max_vecs--;
304		bio_set_region(bio, region);
305
306		/*
307		 * Try and add as many pages as possible.
308		 */
309		while (remaining) {
310			dp->get_page(dp, &page, &len, &offset);
311			len = min(len, to_bytes(remaining));
312			if (!bio_add_page(bio, page, len, offset))
313				break;
314
315			offset = 0;
316			remaining -= to_sector(len);
317			dp->next_page(dp);
318		}
319
320		atomic_inc(&io->count);
321		submit_bio(rw, bio);
322	}
323}
324
325static void dispatch_io(int rw, unsigned int num_regions,
326			struct io_region *where, struct dpages *dp,
327			struct io *io, int sync)
328{
329	int i;
330	struct dpages old_pages = *dp;
331
332	if (sync)
333		rw |= (1 << BIO_RW_SYNC);
334
335	/*
336	 * For multiple regions we need to be careful to rewind
337	 * the dp object for each call to do_region.
338	 */
339	for (i = 0; i < num_regions; i++) {
340		*dp = old_pages;
341		if (where[i].count)
342			do_region(rw, i, where + i, dp, io);
343	}
344
345	/*
346	 * Drop the extra reference that we were holding to avoid
347	 * the io being completed too early.
348	 */
349	dec_count(io, 0, 0);
350}
351
352static int sync_io(struct dm_io_client *client, unsigned int num_regions,
353		   struct io_region *where, int rw, struct dpages *dp,
354		   unsigned long *error_bits)
355{
356	struct io io;
357
358	if (num_regions > 1 && rw != WRITE) {
359		WARN_ON(1);
360		return -EIO;
361	}
362
363	io.error = 0;
364	atomic_set(&io.count, 1); /* see dispatch_io() */
365	io.sleeper = current;
366	io.client = client;
367
368	dispatch_io(rw, num_regions, where, dp, &io, 1);
369
370	while (1) {
371		set_current_state(TASK_UNINTERRUPTIBLE);
372
373		if (!atomic_read(&io.count) || signal_pending(current))
374			break;
375
376		io_schedule();
377	}
378	set_current_state(TASK_RUNNING);
379
380	if (atomic_read(&io.count))
381		return -EINTR;
382
383	if (error_bits)
384		*error_bits = io.error;
385
386	return io.error ? -EIO : 0;
387}
388
389static int async_io(struct dm_io_client *client, unsigned int num_regions,
390		    struct io_region *where, int rw, struct dpages *dp,
391		    io_notify_fn fn, void *context)
392{
393	struct io *io;
394
395	if (num_regions > 1 && rw != WRITE) {
396		WARN_ON(1);
397		fn(1, context);
398		return -EIO;
399	}
400
401	io = mempool_alloc(client->pool, GFP_NOIO);
402	io->error = 0;
403	atomic_set(&io->count, 1); /* see dispatch_io() */
404	io->sleeper = NULL;
405	io->client = client;
406	io->callback = fn;
407	io->context = context;
408
409	dispatch_io(rw, num_regions, where, dp, io, 0);
410	return 0;
411}
412
413static int dp_init(struct dm_io_request *io_req, struct dpages *dp)
414{
415	/* Set up dpages based on memory type */
416	switch (io_req->mem.type) {
417	case DM_IO_PAGE_LIST:
418		list_dp_init(dp, io_req->mem.ptr.pl, io_req->mem.offset);
419		break;
420
421	case DM_IO_BVEC:
422		bvec_dp_init(dp, io_req->mem.ptr.bvec);
423		break;
424
425	case DM_IO_VMA:
426		vm_dp_init(dp, io_req->mem.ptr.vma);
427		break;
428
429	case DM_IO_KMEM:
430		km_dp_init(dp, io_req->mem.ptr.addr);
431		break;
432
433	default:
434		return -EINVAL;
435	}
436
437	return 0;
438}
439
440/*
441 * New collapsed (a)synchronous interface
442 */
443int dm_io(struct dm_io_request *io_req, unsigned num_regions,
444	  struct io_region *where, unsigned long *sync_error_bits)
445{
446	int r;
447	struct dpages dp;
448
449	r = dp_init(io_req, &dp);
450	if (r)
451		return r;
452
453	if (!io_req->notify.fn)
454		return sync_io(io_req->client, num_regions, where,
455			       io_req->bi_rw, &dp, sync_error_bits);
456
457	return async_io(io_req->client, num_regions, where, io_req->bi_rw,
458			&dp, io_req->notify.fn, io_req->notify.context);
459}
460EXPORT_SYMBOL(dm_io);
461