dm-io.c revision 6712ecf8f648118c3363c142196418f89a510b90 
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 void endio(struct bio *bio, int error)
128{
129	struct io *io;
130	unsigned region;
131
132	if (error && bio_data_dir(bio) == READ)
133		zero_fill_bio(bio);
134
135	/*
136	 * The bio destructor in bio_put() may use the io object.
137	 */
138	io = bio->bi_private;
139	region = bio_get_region(bio);
140
141	bio->bi_max_vecs++;
142	bio_put(bio);
143
144	dec_count(io, region, error);
145}
146
147/*-----------------------------------------------------------------
148 * These little objects provide an abstraction for getting a new
149 * destination page for io.
150 *---------------------------------------------------------------*/
151struct dpages {
152	void (*get_page)(struct dpages *dp,
153			 struct page **p, unsigned long *len, unsigned *offset);
154	void (*next_page)(struct dpages *dp);
155
156	unsigned context_u;
157	void *context_ptr;
158};
159
160/*
161 * Functions for getting the pages from a list.
162 */
163static void list_get_page(struct dpages *dp,
164		  struct page **p, unsigned long *len, unsigned *offset)
165{
166	unsigned o = dp->context_u;
167	struct page_list *pl = (struct page_list *) dp->context_ptr;
168
169	*p = pl->page;
170	*len = PAGE_SIZE - o;
171	*offset = o;
172}
173
174static void list_next_page(struct dpages *dp)
175{
176	struct page_list *pl = (struct page_list *) dp->context_ptr;
177	dp->context_ptr = pl->next;
178	dp->context_u = 0;
179}
180
181static void list_dp_init(struct dpages *dp, struct page_list *pl, unsigned offset)
182{
183	dp->get_page = list_get_page;
184	dp->next_page = list_next_page;
185	dp->context_u = offset;
186	dp->context_ptr = pl;
187}
188
189/*
190 * Functions for getting the pages from a bvec.
191 */
192static void bvec_get_page(struct dpages *dp,
193		  struct page **p, unsigned long *len, unsigned *offset)
194{
195	struct bio_vec *bvec = (struct bio_vec *) dp->context_ptr;
196	*p = bvec->bv_page;
197	*len = bvec->bv_len;
198	*offset = bvec->bv_offset;
199}
200
201static void bvec_next_page(struct dpages *dp)
202{
203	struct bio_vec *bvec = (struct bio_vec *) dp->context_ptr;
204	dp->context_ptr = bvec + 1;
205}
206
207static void bvec_dp_init(struct dpages *dp, struct bio_vec *bvec)
208{
209	dp->get_page = bvec_get_page;
210	dp->next_page = bvec_next_page;
211	dp->context_ptr = bvec;
212}
213
214/*
215 * Functions for getting the pages from a VMA.
216 */
217static void vm_get_page(struct dpages *dp,
218		 struct page **p, unsigned long *len, unsigned *offset)
219{
220	*p = vmalloc_to_page(dp->context_ptr);
221	*offset = dp->context_u;
222	*len = PAGE_SIZE - dp->context_u;
223}
224
225static void vm_next_page(struct dpages *dp)
226{
227	dp->context_ptr += PAGE_SIZE - dp->context_u;
228	dp->context_u = 0;
229}
230
231static void vm_dp_init(struct dpages *dp, void *data)
232{
233	dp->get_page = vm_get_page;
234	dp->next_page = vm_next_page;
235	dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1);
236	dp->context_ptr = data;
237}
238
239static void dm_bio_destructor(struct bio *bio)
240{
241	struct io *io = bio->bi_private;
242
243	bio_free(bio, io->client->bios);
244}
245
246/*
247 * Functions for getting the pages from kernel memory.
248 */
249static void km_get_page(struct dpages *dp, struct page **p, unsigned long *len,
250			unsigned *offset)
251{
252	*p = virt_to_page(dp->context_ptr);
253	*offset = dp->context_u;
254	*len = PAGE_SIZE - dp->context_u;
255}
256
257static void km_next_page(struct dpages *dp)
258{
259	dp->context_ptr += PAGE_SIZE - dp->context_u;
260	dp->context_u = 0;
261}
262
263static void km_dp_init(struct dpages *dp, void *data)
264{
265	dp->get_page = km_get_page;
266	dp->next_page = km_next_page;
267	dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1);
268	dp->context_ptr = data;
269}
270
271/*-----------------------------------------------------------------
272 * IO routines that accept a list of pages.
273 *---------------------------------------------------------------*/
274static void do_region(int rw, unsigned int region, struct io_region *where,
275		      struct dpages *dp, struct io *io)
276{
277	struct bio *bio;
278	struct page *page;
279	unsigned long len;
280	unsigned offset;
281	unsigned num_bvecs;
282	sector_t remaining = where->count;
283
284	while (remaining) {
285		/*
286		 * Allocate a suitably sized-bio: we add an extra
287		 * bvec for bio_get/set_region() and decrement bi_max_vecs
288		 * to hide it from bio_add_page().
289		 */
290		num_bvecs = dm_sector_div_up(remaining,
291					     (PAGE_SIZE >> SECTOR_SHIFT));
292		num_bvecs = 1 + min_t(int, bio_get_nr_vecs(where->bdev),
293				      num_bvecs);
294		bio = bio_alloc_bioset(GFP_NOIO, num_bvecs, io->client->bios);
295		bio->bi_sector = where->sector + (where->count - remaining);
296		bio->bi_bdev = where->bdev;
297		bio->bi_end_io = endio;
298		bio->bi_private = io;
299		bio->bi_destructor = dm_bio_destructor;
300		bio->bi_max_vecs--;
301		bio_set_region(bio, region);
302
303		/*
304		 * Try and add as many pages as possible.
305		 */
306		while (remaining) {
307			dp->get_page(dp, &page, &len, &offset);
308			len = min(len, to_bytes(remaining));
309			if (!bio_add_page(bio, page, len, offset))
310				break;
311
312			offset = 0;
313			remaining -= to_sector(len);
314			dp->next_page(dp);
315		}
316
317		atomic_inc(&io->count);
318		submit_bio(rw, bio);
319	}
320}
321
322static void dispatch_io(int rw, unsigned int num_regions,
323			struct io_region *where, struct dpages *dp,
324			struct io *io, int sync)
325{
326	int i;
327	struct dpages old_pages = *dp;
328
329	if (sync)
330		rw |= (1 << BIO_RW_SYNC);
331
332	/*
333	 * For multiple regions we need to be careful to rewind
334	 * the dp object for each call to do_region.
335	 */
336	for (i = 0; i < num_regions; i++) {
337		*dp = old_pages;
338		if (where[i].count)
339			do_region(rw, i, where + i, dp, io);
340	}
341
342	/*
343	 * Drop the extra reference that we were holding to avoid
344	 * the io being completed too early.
345	 */
346	dec_count(io, 0, 0);
347}
348
349static int sync_io(struct dm_io_client *client, unsigned int num_regions,
350		   struct io_region *where, int rw, struct dpages *dp,
351		   unsigned long *error_bits)
352{
353	struct io io;
354
355	if (num_regions > 1 && rw != WRITE) {
356		WARN_ON(1);
357		return -EIO;
358	}
359
360	io.error = 0;
361	atomic_set(&io.count, 1); /* see dispatch_io() */
362	io.sleeper = current;
363	io.client = client;
364
365	dispatch_io(rw, num_regions, where, dp, &io, 1);
366
367	while (1) {
368		set_current_state(TASK_UNINTERRUPTIBLE);
369
370		if (!atomic_read(&io.count) || signal_pending(current))
371			break;
372
373		io_schedule();
374	}
375	set_current_state(TASK_RUNNING);
376
377	if (atomic_read(&io.count))
378		return -EINTR;
379
380	if (error_bits)
381		*error_bits = io.error;
382
383	return io.error ? -EIO : 0;
384}
385
386static int async_io(struct dm_io_client *client, unsigned int num_regions,
387		    struct io_region *where, int rw, struct dpages *dp,
388		    io_notify_fn fn, void *context)
389{
390	struct io *io;
391
392	if (num_regions > 1 && rw != WRITE) {
393		WARN_ON(1);
394		fn(1, context);
395		return -EIO;
396	}
397
398	io = mempool_alloc(client->pool, GFP_NOIO);
399	io->error = 0;
400	atomic_set(&io->count, 1); /* see dispatch_io() */
401	io->sleeper = NULL;
402	io->client = client;
403	io->callback = fn;
404	io->context = context;
405
406	dispatch_io(rw, num_regions, where, dp, io, 0);
407	return 0;
408}
409
410static int dp_init(struct dm_io_request *io_req, struct dpages *dp)
411{
412	/* Set up dpages based on memory type */
413	switch (io_req->mem.type) {
414	case DM_IO_PAGE_LIST:
415		list_dp_init(dp, io_req->mem.ptr.pl, io_req->mem.offset);
416		break;
417
418	case DM_IO_BVEC:
419		bvec_dp_init(dp, io_req->mem.ptr.bvec);
420		break;
421
422	case DM_IO_VMA:
423		vm_dp_init(dp, io_req->mem.ptr.vma);
424		break;
425
426	case DM_IO_KMEM:
427		km_dp_init(dp, io_req->mem.ptr.addr);
428		break;
429
430	default:
431		return -EINVAL;
432	}
433
434	return 0;
435}
436
437/*
438 * New collapsed (a)synchronous interface
439 */
440int dm_io(struct dm_io_request *io_req, unsigned num_regions,
441	  struct io_region *where, unsigned long *sync_error_bits)
442{
443	int r;
444	struct dpages dp;
445
446	r = dp_init(io_req, &dp);
447	if (r)
448		return r;
449
450	if (!io_req->notify.fn)
451		return sync_io(io_req->client, num_regions, where,
452			       io_req->bi_rw, &dp, sync_error_bits);
453
454	return async_io(io_req->client, num_regions, where, io_req->bi_rw,
455			&dp, io_req->notify.fn, io_req->notify.context);
456}
457EXPORT_SYMBOL(dm_io);
458