dm-io.c revision d659e6cc98766a1a61d6bdd283f95d149abd7719 
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 <linux/device-mapper.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#include <linux/dm-io.h>
16
17struct dm_io_client {
18	mempool_t *pool;
19	struct bio_set *bios;
20};
21
22/* FIXME: can we shrink this ? */
23struct io {
24	unsigned long error_bits;
25	atomic_t count;
26	struct task_struct *sleeper;
27	struct dm_io_client *client;
28	io_notify_fn callback;
29	void *context;
30};
31
32/*
33 * io contexts are only dynamically allocated for asynchronous
34 * io.  Since async io is likely to be the majority of io we'll
35 * have the same number of io contexts as bios! (FIXME: must reduce this).
36 */
37
38static unsigned int pages_to_ios(unsigned int pages)
39{
40	return 4 * pages;	/* too many ? */
41}
42
43/*
44 * Create a client with mempool and bioset.
45 */
46struct dm_io_client *dm_io_client_create(unsigned num_pages)
47{
48	unsigned ios = pages_to_ios(num_pages);
49	struct dm_io_client *client;
50
51	client = kmalloc(sizeof(*client), GFP_KERNEL);
52	if (!client)
53		return ERR_PTR(-ENOMEM);
54
55	client->pool = mempool_create_kmalloc_pool(ios, sizeof(struct io));
56	if (!client->pool)
57		goto bad;
58
59	client->bios = bioset_create(16, 0);
60	if (!client->bios)
61		goto bad;
62
63	return client;
64
65   bad:
66	if (client->pool)
67		mempool_destroy(client->pool);
68	kfree(client);
69	return ERR_PTR(-ENOMEM);
70}
71EXPORT_SYMBOL(dm_io_client_create);
72
73int dm_io_client_resize(unsigned num_pages, struct dm_io_client *client)
74{
75	return mempool_resize(client->pool, pages_to_ios(num_pages),
76			      GFP_KERNEL);
77}
78EXPORT_SYMBOL(dm_io_client_resize);
79
80void dm_io_client_destroy(struct dm_io_client *client)
81{
82	mempool_destroy(client->pool);
83	bioset_free(client->bios);
84	kfree(client);
85}
86EXPORT_SYMBOL(dm_io_client_destroy);
87
88/*-----------------------------------------------------------------
89 * We need to keep track of which region a bio is doing io for.
90 * In order to save a memory allocation we store this the last
91 * bvec which we know is unused (blech).
92 * XXX This is ugly and can OOPS with some configs... find another way.
93 *---------------------------------------------------------------*/
94static inline void bio_set_region(struct bio *bio, unsigned region)
95{
96	bio->bi_io_vec[bio->bi_max_vecs].bv_len = region;
97}
98
99static inline unsigned bio_get_region(struct bio *bio)
100{
101	return bio->bi_io_vec[bio->bi_max_vecs].bv_len;
102}
103
104/*-----------------------------------------------------------------
105 * We need an io object to keep track of the number of bios that
106 * have been dispatched for a particular io.
107 *---------------------------------------------------------------*/
108static void dec_count(struct io *io, unsigned int region, int error)
109{
110	if (error)
111		set_bit(region, &io->error_bits);
112
113	if (atomic_dec_and_test(&io->count)) {
114		if (io->sleeper)
115			wake_up_process(io->sleeper);
116
117		else {
118			unsigned long r = io->error_bits;
119			io_notify_fn fn = io->callback;
120			void *context = io->context;
121
122			mempool_free(io, io->client->pool);
123			fn(r, context);
124		}
125	}
126}
127
128static void endio(struct bio *bio, int error)
129{
130	struct io *io;
131	unsigned region;
132
133	if (error && bio_data_dir(bio) == READ)
134		zero_fill_bio(bio);
135
136	/*
137	 * The bio destructor in bio_put() may use the io object.
138	 */
139	io = bio->bi_private;
140	region = bio_get_region(bio);
141
142	bio->bi_max_vecs++;
143	bio_put(bio);
144
145	dec_count(io, region, error);
146}
147
148/*-----------------------------------------------------------------
149 * These little objects provide an abstraction for getting a new
150 * destination page for io.
151 *---------------------------------------------------------------*/
152struct dpages {
153	void (*get_page)(struct dpages *dp,
154			 struct page **p, unsigned long *len, unsigned *offset);
155	void (*next_page)(struct dpages *dp);
156
157	unsigned context_u;
158	void *context_ptr;
159};
160
161/*
162 * Functions for getting the pages from a list.
163 */
164static void list_get_page(struct dpages *dp,
165		  struct page **p, unsigned long *len, unsigned *offset)
166{
167	unsigned o = dp->context_u;
168	struct page_list *pl = (struct page_list *) dp->context_ptr;
169
170	*p = pl->page;
171	*len = PAGE_SIZE - o;
172	*offset = o;
173}
174
175static void list_next_page(struct dpages *dp)
176{
177	struct page_list *pl = (struct page_list *) dp->context_ptr;
178	dp->context_ptr = pl->next;
179	dp->context_u = 0;
180}
181
182static void list_dp_init(struct dpages *dp, struct page_list *pl, unsigned offset)
183{
184	dp->get_page = list_get_page;
185	dp->next_page = list_next_page;
186	dp->context_u = offset;
187	dp->context_ptr = pl;
188}
189
190/*
191 * Functions for getting the pages from a bvec.
192 */
193static void bvec_get_page(struct dpages *dp,
194		  struct page **p, unsigned long *len, unsigned *offset)
195{
196	struct bio_vec *bvec = (struct bio_vec *) dp->context_ptr;
197	*p = bvec->bv_page;
198	*len = bvec->bv_len;
199	*offset = bvec->bv_offset;
200}
201
202static void bvec_next_page(struct dpages *dp)
203{
204	struct bio_vec *bvec = (struct bio_vec *) dp->context_ptr;
205	dp->context_ptr = bvec + 1;
206}
207
208static void bvec_dp_init(struct dpages *dp, struct bio_vec *bvec)
209{
210	dp->get_page = bvec_get_page;
211	dp->next_page = bvec_next_page;
212	dp->context_ptr = bvec;
213}
214
215/*
216 * Functions for getting the pages from a VMA.
217 */
218static void vm_get_page(struct dpages *dp,
219		 struct page **p, unsigned long *len, unsigned *offset)
220{
221	*p = vmalloc_to_page(dp->context_ptr);
222	*offset = dp->context_u;
223	*len = PAGE_SIZE - dp->context_u;
224}
225
226static void vm_next_page(struct dpages *dp)
227{
228	dp->context_ptr += PAGE_SIZE - dp->context_u;
229	dp->context_u = 0;
230}
231
232static void vm_dp_init(struct dpages *dp, void *data)
233{
234	dp->get_page = vm_get_page;
235	dp->next_page = vm_next_page;
236	dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1);
237	dp->context_ptr = data;
238}
239
240static void dm_bio_destructor(struct bio *bio)
241{
242	struct io *io = bio->bi_private;
243
244	bio_free(bio, io->client->bios);
245}
246
247/*
248 * Functions for getting the pages from kernel memory.
249 */
250static void km_get_page(struct dpages *dp, struct page **p, unsigned long *len,
251			unsigned *offset)
252{
253	*p = virt_to_page(dp->context_ptr);
254	*offset = dp->context_u;
255	*len = PAGE_SIZE - dp->context_u;
256}
257
258static void km_next_page(struct dpages *dp)
259{
260	dp->context_ptr += PAGE_SIZE - dp->context_u;
261	dp->context_u = 0;
262}
263
264static void km_dp_init(struct dpages *dp, void *data)
265{
266	dp->get_page = km_get_page;
267	dp->next_page = km_next_page;
268	dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1);
269	dp->context_ptr = data;
270}
271
272/*-----------------------------------------------------------------
273 * IO routines that accept a list of pages.
274 *---------------------------------------------------------------*/
275static void do_region(int rw, unsigned region, struct dm_io_region *where,
276		      struct dpages *dp, struct io *io)
277{
278	struct bio *bio;
279	struct page *page;
280	unsigned long len;
281	unsigned offset;
282	unsigned num_bvecs;
283	sector_t remaining = where->count;
284
285	while (remaining) {
286		/*
287		 * Allocate a suitably sized-bio: we add an extra
288		 * bvec for bio_get/set_region() and decrement bi_max_vecs
289		 * to hide it from bio_add_page().
290		 */
291		num_bvecs = dm_sector_div_up(remaining,
292					     (PAGE_SIZE >> SECTOR_SHIFT));
293		num_bvecs = 1 + min_t(int, bio_get_nr_vecs(where->bdev),
294				      num_bvecs);
295		if (unlikely(num_bvecs > BIO_MAX_PAGES))
296			num_bvecs = BIO_MAX_PAGES;
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 dm_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_SYNCIO) | (1 << BIO_RW_UNPLUG);
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 dm_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 & RW_MASK) != WRITE) {
359		WARN_ON(1);
360		return -EIO;
361	}
362
363	io.error_bits = 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_bits;
385
386	return io.error_bits ? -EIO : 0;
387}
388
389static int async_io(struct dm_io_client *client, unsigned int num_regions,
390		    struct dm_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 & RW_MASK) != 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_bits = 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 *
443 * If the IO is asynchronous (i.e. it has notify.fn), you must either unplug
444 * the queue with blk_unplug() some time later or set the BIO_RW_SYNC bit in
445 * io_req->bi_rw. If you fail to do one of these, the IO will be submitted to
446 * the disk after q->unplug_delay, which defaults to 3ms in blk-settings.c.
447 */
448int dm_io(struct dm_io_request *io_req, unsigned num_regions,
449	  struct dm_io_region *where, unsigned long *sync_error_bits)
450{
451	int r;
452	struct dpages dp;
453
454	r = dp_init(io_req, &dp);
455	if (r)
456		return r;
457
458	if (!io_req->notify.fn)
459		return sync_io(io_req->client, num_regions, where,
460			       io_req->bi_rw, &dp, sync_error_bits);
461
462	return async_io(io_req->client, num_regions, where, io_req->bi_rw,
463			&dp, io_req->notify.fn, io_req->notify.context);
464}
465EXPORT_SYMBOL(dm_io);
466