Linux Audio

Check our new training course

Loading...
v3.1
  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/mempool.h>
 14#include <linux/module.h>
 15#include <linux/sched.h>
 16#include <linux/slab.h>
 17#include <linux/dm-io.h>
 18
 19#define DM_MSG_PREFIX "io"
 20
 21#define DM_IO_MAX_REGIONS	BITS_PER_LONG
 22#define MIN_IOS		16
 23#define MIN_BIOS	16
 24
 25struct dm_io_client {
 26	mempool_t *pool;
 27	struct bio_set *bios;
 28};
 29
 30/*
 31 * Aligning 'struct io' reduces the number of bits required to store
 32 * its address.  Refer to store_io_and_region_in_bio() below.
 33 */
 34struct io {
 35	unsigned long error_bits;
 36	atomic_t count;
 37	struct task_struct *sleeper;
 38	struct dm_io_client *client;
 39	io_notify_fn callback;
 40	void *context;
 41	void *vma_invalidate_address;
 42	unsigned long vma_invalidate_size;
 43} __attribute__((aligned(DM_IO_MAX_REGIONS)));
 44
 45static struct kmem_cache *_dm_io_cache;
 46
 47/*
 48 * Create a client with mempool and bioset.
 49 */
 50struct dm_io_client *dm_io_client_create(void)
 51{
 52	struct dm_io_client *client;
 
 53
 54	client = kmalloc(sizeof(*client), GFP_KERNEL);
 55	if (!client)
 56		return ERR_PTR(-ENOMEM);
 57
 58	client->pool = mempool_create_slab_pool(MIN_IOS, _dm_io_cache);
 59	if (!client->pool)
 60		goto bad;
 61
 62	client->bios = bioset_create(MIN_BIOS, 0);
 63	if (!client->bios)
 64		goto bad;
 65
 66	return client;
 67
 68   bad:
 69	if (client->pool)
 70		mempool_destroy(client->pool);
 71	kfree(client);
 72	return ERR_PTR(-ENOMEM);
 73}
 74EXPORT_SYMBOL(dm_io_client_create);
 75
 76void dm_io_client_destroy(struct dm_io_client *client)
 77{
 78	mempool_destroy(client->pool);
 79	bioset_free(client->bios);
 80	kfree(client);
 81}
 82EXPORT_SYMBOL(dm_io_client_destroy);
 83
 84/*-----------------------------------------------------------------
 85 * We need to keep track of which region a bio is doing io for.
 86 * To avoid a memory allocation to store just 5 or 6 bits, we
 87 * ensure the 'struct io' pointer is aligned so enough low bits are
 88 * always zero and then combine it with the region number directly in
 89 * bi_private.
 90 *---------------------------------------------------------------*/
 91static void store_io_and_region_in_bio(struct bio *bio, struct io *io,
 92				       unsigned region)
 93{
 94	if (unlikely(!IS_ALIGNED((unsigned long)io, DM_IO_MAX_REGIONS))) {
 95		DMCRIT("Unaligned struct io pointer %p", io);
 96		BUG();
 97	}
 98
 99	bio->bi_private = (void *)((unsigned long)io | region);
100}
101
102static void retrieve_io_and_region_from_bio(struct bio *bio, struct io **io,
103				       unsigned *region)
104{
105	unsigned long val = (unsigned long)bio->bi_private;
106
107	*io = (void *)(val & -(unsigned long)DM_IO_MAX_REGIONS);
108	*region = val & (DM_IO_MAX_REGIONS - 1);
109}
110
111/*-----------------------------------------------------------------
112 * We need an io object to keep track of the number of bios that
113 * have been dispatched for a particular io.
114 *---------------------------------------------------------------*/
115static void dec_count(struct io *io, unsigned int region, int error)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
116{
117	if (error)
118		set_bit(region, &io->error_bits);
119
120	if (atomic_dec_and_test(&io->count)) {
121		if (io->vma_invalidate_size)
122			invalidate_kernel_vmap_range(io->vma_invalidate_address,
123						     io->vma_invalidate_size);
124
125		if (io->sleeper)
126			wake_up_process(io->sleeper);
127
128		else {
129			unsigned long r = io->error_bits;
130			io_notify_fn fn = io->callback;
131			void *context = io->context;
132
133			mempool_free(io, io->client->pool);
134			fn(r, context);
135		}
136	}
137}
138
139static void endio(struct bio *bio, int error)
140{
141	struct io *io;
142	unsigned region;
 
143
144	if (error && bio_data_dir(bio) == READ)
145		zero_fill_bio(bio);
146
147	/*
148	 * The bio destructor in bio_put() may use the io object.
149	 */
150	retrieve_io_and_region_from_bio(bio, &io, &region);
151
 
152	bio_put(bio);
153
154	dec_count(io, region, error);
155}
156
157/*-----------------------------------------------------------------
158 * These little objects provide an abstraction for getting a new
159 * destination page for io.
160 *---------------------------------------------------------------*/
161struct dpages {
162	void (*get_page)(struct dpages *dp,
163			 struct page **p, unsigned long *len, unsigned *offset);
164	void (*next_page)(struct dpages *dp);
165
166	unsigned context_u;
 
 
 
167	void *context_ptr;
168
169	void *vma_invalidate_address;
170	unsigned long vma_invalidate_size;
171};
172
173/*
174 * Functions for getting the pages from a list.
175 */
176static void list_get_page(struct dpages *dp,
177		  struct page **p, unsigned long *len, unsigned *offset)
178{
179	unsigned o = dp->context_u;
180	struct page_list *pl = (struct page_list *) dp->context_ptr;
181
182	*p = pl->page;
183	*len = PAGE_SIZE - o;
184	*offset = o;
185}
186
187static void list_next_page(struct dpages *dp)
188{
189	struct page_list *pl = (struct page_list *) dp->context_ptr;
190	dp->context_ptr = pl->next;
191	dp->context_u = 0;
192}
193
194static void list_dp_init(struct dpages *dp, struct page_list *pl, unsigned offset)
195{
196	dp->get_page = list_get_page;
197	dp->next_page = list_next_page;
198	dp->context_u = offset;
199	dp->context_ptr = pl;
200}
201
202/*
203 * Functions for getting the pages from a bvec.
204 */
205static void bvec_get_page(struct dpages *dp,
206		  struct page **p, unsigned long *len, unsigned *offset)
207{
208	struct bio_vec *bvec = (struct bio_vec *) dp->context_ptr;
209	*p = bvec->bv_page;
210	*len = bvec->bv_len;
211	*offset = bvec->bv_offset;
 
 
 
 
 
212}
213
214static void bvec_next_page(struct dpages *dp)
215{
216	struct bio_vec *bvec = (struct bio_vec *) dp->context_ptr;
217	dp->context_ptr = bvec + 1;
 
 
218}
219
220static void bvec_dp_init(struct dpages *dp, struct bio_vec *bvec)
221{
222	dp->get_page = bvec_get_page;
223	dp->next_page = bvec_next_page;
224	dp->context_ptr = bvec;
 
 
 
 
 
 
225}
226
227/*
228 * Functions for getting the pages from a VMA.
229 */
230static void vm_get_page(struct dpages *dp,
231		 struct page **p, unsigned long *len, unsigned *offset)
232{
233	*p = vmalloc_to_page(dp->context_ptr);
234	*offset = dp->context_u;
235	*len = PAGE_SIZE - dp->context_u;
236}
237
238static void vm_next_page(struct dpages *dp)
239{
240	dp->context_ptr += PAGE_SIZE - dp->context_u;
241	dp->context_u = 0;
242}
243
244static void vm_dp_init(struct dpages *dp, void *data)
245{
246	dp->get_page = vm_get_page;
247	dp->next_page = vm_next_page;
248	dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1);
249	dp->context_ptr = data;
250}
251
252static void dm_bio_destructor(struct bio *bio)
253{
254	unsigned region;
255	struct io *io;
256
257	retrieve_io_and_region_from_bio(bio, &io, &region);
258
259	bio_free(bio, io->client->bios);
260}
261
262/*
263 * Functions for getting the pages from kernel memory.
264 */
265static void km_get_page(struct dpages *dp, struct page **p, unsigned long *len,
266			unsigned *offset)
267{
268	*p = virt_to_page(dp->context_ptr);
269	*offset = dp->context_u;
270	*len = PAGE_SIZE - dp->context_u;
271}
272
273static void km_next_page(struct dpages *dp)
274{
275	dp->context_ptr += PAGE_SIZE - dp->context_u;
276	dp->context_u = 0;
277}
278
279static void km_dp_init(struct dpages *dp, void *data)
280{
281	dp->get_page = km_get_page;
282	dp->next_page = km_next_page;
283	dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1);
284	dp->context_ptr = data;
285}
286
287/*-----------------------------------------------------------------
288 * IO routines that accept a list of pages.
289 *---------------------------------------------------------------*/
290static void do_region(int rw, unsigned region, struct dm_io_region *where,
291		      struct dpages *dp, struct io *io)
 
292{
293	struct bio *bio;
294	struct page *page;
295	unsigned long len;
296	unsigned offset;
297	unsigned num_bvecs;
298	sector_t remaining = where->count;
 
 
 
 
299
300	/*
301	 * where->count may be zero if rw holds a flush and we need to
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
302	 * send a zero-sized flush.
303	 */
304	do {
305		/*
306		 * Allocate a suitably sized-bio.
307		 */
308		num_bvecs = dm_sector_div_up(remaining,
309					     (PAGE_SIZE >> SECTOR_SHIFT));
310		num_bvecs = min_t(int, bio_get_nr_vecs(where->bdev), num_bvecs);
 
 
 
 
 
 
 
 
 
 
311		bio = bio_alloc_bioset(GFP_NOIO, num_bvecs, io->client->bios);
312		bio->bi_sector = where->sector + (where->count - remaining);
313		bio->bi_bdev = where->bdev;
314		bio->bi_end_io = endio;
315		bio->bi_destructor = dm_bio_destructor;
316		store_io_and_region_in_bio(bio, io, region);
317
318		/*
319		 * Try and add as many pages as possible.
320		 */
321		while (remaining) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
322			dp->get_page(dp, &page, &len, &offset);
323			len = min(len, to_bytes(remaining));
324			if (!bio_add_page(bio, page, len, offset))
325				break;
326
327			offset = 0;
328			remaining -= to_sector(len);
329			dp->next_page(dp);
330		}
331
332		atomic_inc(&io->count);
333		submit_bio(rw, bio);
334	} while (remaining);
335}
336
337static void dispatch_io(int rw, unsigned int num_regions,
338			struct dm_io_region *where, struct dpages *dp,
339			struct io *io, int sync)
340{
341	int i;
342	struct dpages old_pages = *dp;
343
344	BUG_ON(num_regions > DM_IO_MAX_REGIONS);
345
346	if (sync)
347		rw |= REQ_SYNC;
348
349	/*
350	 * For multiple regions we need to be careful to rewind
351	 * the dp object for each call to do_region.
352	 */
353	for (i = 0; i < num_regions; i++) {
354		*dp = old_pages;
355		if (where[i].count || (rw & REQ_FLUSH))
356			do_region(rw, i, where + i, dp, io);
357	}
358
359	/*
360	 * Drop the extra reference that we were holding to avoid
361	 * the io being completed too early.
362	 */
363	dec_count(io, 0, 0);
364}
365
 
 
 
 
 
 
 
 
 
 
 
 
 
366static int sync_io(struct dm_io_client *client, unsigned int num_regions,
367		   struct dm_io_region *where, int rw, struct dpages *dp,
368		   unsigned long *error_bits)
369{
370	/*
371	 * gcc <= 4.3 can't do the alignment for stack variables, so we must
372	 * align it on our own.
373	 * volatile prevents the optimizer from removing or reusing
374	 * "io_" field from the stack frame (allowed in ANSI C).
375	 */
376	volatile char io_[sizeof(struct io) + __alignof__(struct io) - 1];
377	struct io *io = (struct io *)PTR_ALIGN(&io_, __alignof__(struct io));
378
379	if (num_regions > 1 && (rw & RW_MASK) != WRITE) {
380		WARN_ON(1);
381		return -EIO;
382	}
383
 
 
 
384	io->error_bits = 0;
385	atomic_set(&io->count, 1); /* see dispatch_io() */
386	io->sleeper = current;
387	io->client = client;
 
 
388
389	io->vma_invalidate_address = dp->vma_invalidate_address;
390	io->vma_invalidate_size = dp->vma_invalidate_size;
391
392	dispatch_io(rw, num_regions, where, dp, io, 1);
393
394	while (1) {
395		set_current_state(TASK_UNINTERRUPTIBLE);
396
397		if (!atomic_read(&io->count))
398			break;
399
400		io_schedule();
401	}
402	set_current_state(TASK_RUNNING);
403
404	if (error_bits)
405		*error_bits = io->error_bits;
406
407	return io->error_bits ? -EIO : 0;
408}
409
410static int async_io(struct dm_io_client *client, unsigned int num_regions,
411		    struct dm_io_region *where, int rw, struct dpages *dp,
412		    io_notify_fn fn, void *context)
413{
414	struct io *io;
415
416	if (num_regions > 1 && (rw & RW_MASK) != WRITE) {
417		WARN_ON(1);
418		fn(1, context);
419		return -EIO;
420	}
421
422	io = mempool_alloc(client->pool, GFP_NOIO);
423	io->error_bits = 0;
424	atomic_set(&io->count, 1); /* see dispatch_io() */
425	io->sleeper = NULL;
426	io->client = client;
427	io->callback = fn;
428	io->context = context;
429
430	io->vma_invalidate_address = dp->vma_invalidate_address;
431	io->vma_invalidate_size = dp->vma_invalidate_size;
432
433	dispatch_io(rw, num_regions, where, dp, io, 0);
434	return 0;
435}
436
437static int dp_init(struct dm_io_request *io_req, struct dpages *dp,
438		   unsigned long size)
439{
440	/* Set up dpages based on memory type */
441
442	dp->vma_invalidate_address = NULL;
443	dp->vma_invalidate_size = 0;
444
445	switch (io_req->mem.type) {
446	case DM_IO_PAGE_LIST:
447		list_dp_init(dp, io_req->mem.ptr.pl, io_req->mem.offset);
448		break;
449
450	case DM_IO_BVEC:
451		bvec_dp_init(dp, io_req->mem.ptr.bvec);
452		break;
453
454	case DM_IO_VMA:
455		flush_kernel_vmap_range(io_req->mem.ptr.vma, size);
456		if ((io_req->bi_rw & RW_MASK) == READ) {
457			dp->vma_invalidate_address = io_req->mem.ptr.vma;
458			dp->vma_invalidate_size = size;
459		}
460		vm_dp_init(dp, io_req->mem.ptr.vma);
461		break;
462
463	case DM_IO_KMEM:
464		km_dp_init(dp, io_req->mem.ptr.addr);
465		break;
466
467	default:
468		return -EINVAL;
469	}
470
471	return 0;
472}
473
474/*
475 * New collapsed (a)synchronous interface.
476 *
477 * If the IO is asynchronous (i.e. it has notify.fn), you must either unplug
478 * the queue with blk_unplug() some time later or set REQ_SYNC in
479io_req->bi_rw. If you fail to do one of these, the IO will be submitted to
480 * the disk after q->unplug_delay, which defaults to 3ms in blk-settings.c.
481 */
482int dm_io(struct dm_io_request *io_req, unsigned num_regions,
483	  struct dm_io_region *where, unsigned long *sync_error_bits)
484{
485	int r;
486	struct dpages dp;
487
488	r = dp_init(io_req, &dp, (unsigned long)where->count << SECTOR_SHIFT);
489	if (r)
490		return r;
491
492	if (!io_req->notify.fn)
493		return sync_io(io_req->client, num_regions, where,
494			       io_req->bi_rw, &dp, sync_error_bits);
 
495
496	return async_io(io_req->client, num_regions, where, io_req->bi_rw,
497			&dp, io_req->notify.fn, io_req->notify.context);
 
498}
499EXPORT_SYMBOL(dm_io);
500
501int __init dm_io_init(void)
502{
503	_dm_io_cache = KMEM_CACHE(io, 0);
504	if (!_dm_io_cache)
505		return -ENOMEM;
506
507	return 0;
508}
509
510void dm_io_exit(void)
511{
512	kmem_cache_destroy(_dm_io_cache);
513	_dm_io_cache = NULL;
514}
v4.17
  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-core.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, BIOSET_NEED_BVECS);
 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, blk_status_t 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	blk_status_t error;
141
142	if (bio->bi_status && 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_status;
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	union {
166		unsigned context_u;
167		struct bvec_iter context_bi;
168	};
169	void *context_ptr;
170
171	void *vma_invalidate_address;
172	unsigned long vma_invalidate_size;
173};
174
175/*
176 * Functions for getting the pages from a list.
177 */
178static void list_get_page(struct dpages *dp,
179		  struct page **p, unsigned long *len, unsigned *offset)
180{
181	unsigned o = dp->context_u;
182	struct page_list *pl = (struct page_list *) dp->context_ptr;
183
184	*p = pl->page;
185	*len = PAGE_SIZE - o;
186	*offset = o;
187}
188
189static void list_next_page(struct dpages *dp)
190{
191	struct page_list *pl = (struct page_list *) dp->context_ptr;
192	dp->context_ptr = pl->next;
193	dp->context_u = 0;
194}
195
196static void list_dp_init(struct dpages *dp, struct page_list *pl, unsigned offset)
197{
198	dp->get_page = list_get_page;
199	dp->next_page = list_next_page;
200	dp->context_u = offset;
201	dp->context_ptr = pl;
202}
203
204/*
205 * Functions for getting the pages from a bvec.
206 */
207static void bio_get_page(struct dpages *dp, struct page **p,
208			 unsigned long *len, unsigned *offset)
209{
210	struct bio_vec bvec = bvec_iter_bvec((struct bio_vec *)dp->context_ptr,
211					     dp->context_bi);
212
213	*p = bvec.bv_page;
214	*len = bvec.bv_len;
215	*offset = bvec.bv_offset;
216
217	/* avoid figuring it out again in bio_next_page() */
218	dp->context_bi.bi_sector = (sector_t)bvec.bv_len;
219}
220
221static void bio_next_page(struct dpages *dp)
222{
223	unsigned int len = (unsigned int)dp->context_bi.bi_sector;
224
225	bvec_iter_advance((struct bio_vec *)dp->context_ptr,
226			  &dp->context_bi, len);
227}
228
229static void bio_dp_init(struct dpages *dp, struct bio *bio)
230{
231	dp->get_page = bio_get_page;
232	dp->next_page = bio_next_page;
233
234	/*
235	 * We just use bvec iterator to retrieve pages, so it is ok to
236	 * access the bvec table directly here
237	 */
238	dp->context_ptr = bio->bi_io_vec;
239	dp->context_bi = bio->bi_iter;
240}
241
242/*
243 * Functions for getting the pages from a VMA.
244 */
245static void vm_get_page(struct dpages *dp,
246		 struct page **p, unsigned long *len, unsigned *offset)
247{
248	*p = vmalloc_to_page(dp->context_ptr);
249	*offset = dp->context_u;
250	*len = PAGE_SIZE - dp->context_u;
251}
252
253static void vm_next_page(struct dpages *dp)
254{
255	dp->context_ptr += PAGE_SIZE - dp->context_u;
256	dp->context_u = 0;
257}
258
259static void vm_dp_init(struct dpages *dp, void *data)
260{
261	dp->get_page = vm_get_page;
262	dp->next_page = vm_next_page;
263	dp->context_u = offset_in_page(data);
264	dp->context_ptr = data;
265}
266
 
 
 
 
 
 
 
 
 
 
267/*
268 * Functions for getting the pages from kernel memory.
269 */
270static void km_get_page(struct dpages *dp, struct page **p, unsigned long *len,
271			unsigned *offset)
272{
273	*p = virt_to_page(dp->context_ptr);
274	*offset = dp->context_u;
275	*len = PAGE_SIZE - dp->context_u;
276}
277
278static void km_next_page(struct dpages *dp)
279{
280	dp->context_ptr += PAGE_SIZE - dp->context_u;
281	dp->context_u = 0;
282}
283
284static void km_dp_init(struct dpages *dp, void *data)
285{
286	dp->get_page = km_get_page;
287	dp->next_page = km_next_page;
288	dp->context_u = offset_in_page(data);
289	dp->context_ptr = data;
290}
291
292/*-----------------------------------------------------------------
293 * IO routines that accept a list of pages.
294 *---------------------------------------------------------------*/
295static void do_region(int op, int op_flags, unsigned region,
296		      struct dm_io_region *where, struct dpages *dp,
297		      struct io *io)
298{
299	struct bio *bio;
300	struct page *page;
301	unsigned long len;
302	unsigned offset;
303	unsigned num_bvecs;
304	sector_t remaining = where->count;
305	struct request_queue *q = bdev_get_queue(where->bdev);
306	unsigned short logical_block_size = queue_logical_block_size(q);
307	sector_t num_sectors;
308	unsigned int uninitialized_var(special_cmd_max_sectors);
309
310	/*
311	 * Reject unsupported discard and write same requests.
312	 */
313	if (op == REQ_OP_DISCARD)
314		special_cmd_max_sectors = q->limits.max_discard_sectors;
315	else if (op == REQ_OP_WRITE_ZEROES)
316		special_cmd_max_sectors = q->limits.max_write_zeroes_sectors;
317	else if (op == REQ_OP_WRITE_SAME)
318		special_cmd_max_sectors = q->limits.max_write_same_sectors;
319	if ((op == REQ_OP_DISCARD || op == REQ_OP_WRITE_ZEROES ||
320	     op == REQ_OP_WRITE_SAME) && special_cmd_max_sectors == 0) {
321		atomic_inc(&io->count);
322		dec_count(io, region, BLK_STS_NOTSUPP);
323		return;
324	}
325
326	/*
327	 * where->count may be zero if op holds a flush and we need to
328	 * send a zero-sized flush.
329	 */
330	do {
331		/*
332		 * Allocate a suitably sized-bio.
333		 */
334		switch (op) {
335		case REQ_OP_DISCARD:
336		case REQ_OP_WRITE_ZEROES:
337			num_bvecs = 0;
338			break;
339		case REQ_OP_WRITE_SAME:
340			num_bvecs = 1;
341			break;
342		default:
343			num_bvecs = min_t(int, BIO_MAX_PAGES,
344					  dm_sector_div_up(remaining, (PAGE_SIZE >> SECTOR_SHIFT)));
345		}
346
347		bio = bio_alloc_bioset(GFP_NOIO, num_bvecs, io->client->bios);
348		bio->bi_iter.bi_sector = where->sector + (where->count - remaining);
349		bio_set_dev(bio, where->bdev);
350		bio->bi_end_io = endio;
351		bio_set_op_attrs(bio, op, op_flags);
352		store_io_and_region_in_bio(bio, io, region);
353
354		if (op == REQ_OP_DISCARD || op == REQ_OP_WRITE_ZEROES) {
355			num_sectors = min_t(sector_t, special_cmd_max_sectors, remaining);
356			bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
357			remaining -= num_sectors;
358		} else if (op == REQ_OP_WRITE_SAME) {
359			/*
360			 * WRITE SAME only uses a single page.
361			 */
362			dp->get_page(dp, &page, &len, &offset);
363			bio_add_page(bio, page, logical_block_size, offset);
364			num_sectors = min_t(sector_t, special_cmd_max_sectors, remaining);
365			bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
366
367			offset = 0;
368			remaining -= num_sectors;
369			dp->next_page(dp);
370		} else while (remaining) {
371			/*
372			 * Try and add as many pages as possible.
373			 */
374			dp->get_page(dp, &page, &len, &offset);
375			len = min(len, to_bytes(remaining));
376			if (!bio_add_page(bio, page, len, offset))
377				break;
378
379			offset = 0;
380			remaining -= to_sector(len);
381			dp->next_page(dp);
382		}
383
384		atomic_inc(&io->count);
385		submit_bio(bio);
386	} while (remaining);
387}
388
389static void dispatch_io(int op, int op_flags, unsigned int num_regions,
390			struct dm_io_region *where, struct dpages *dp,
391			struct io *io, int sync)
392{
393	int i;
394	struct dpages old_pages = *dp;
395
396	BUG_ON(num_regions > DM_IO_MAX_REGIONS);
397
398	if (sync)
399		op_flags |= REQ_SYNC;
400
401	/*
402	 * For multiple regions we need to be careful to rewind
403	 * the dp object for each call to do_region.
404	 */
405	for (i = 0; i < num_regions; i++) {
406		*dp = old_pages;
407		if (where[i].count || (op_flags & REQ_PREFLUSH))
408			do_region(op, op_flags, i, where + i, dp, io);
409	}
410
411	/*
412	 * Drop the extra reference that we were holding to avoid
413	 * the io being completed too early.
414	 */
415	dec_count(io, 0, 0);
416}
417
418struct sync_io {
419	unsigned long error_bits;
420	struct completion wait;
421};
422
423static void sync_io_complete(unsigned long error, void *context)
424{
425	struct sync_io *sio = context;
426
427	sio->error_bits = error;
428	complete(&sio->wait);
429}
430
431static int sync_io(struct dm_io_client *client, unsigned int num_regions,
432		   struct dm_io_region *where, int op, int op_flags,
433		   struct dpages *dp, unsigned long *error_bits)
434{
435	struct io *io;
436	struct sync_io sio;
 
 
 
 
 
 
437
438	if (num_regions > 1 && !op_is_write(op)) {
439		WARN_ON(1);
440		return -EIO;
441	}
442
443	init_completion(&sio.wait);
444
445	io = mempool_alloc(client->pool, GFP_NOIO);
446	io->error_bits = 0;
447	atomic_set(&io->count, 1); /* see dispatch_io() */
 
448	io->client = client;
449	io->callback = sync_io_complete;
450	io->context = &sio;
451
452	io->vma_invalidate_address = dp->vma_invalidate_address;
453	io->vma_invalidate_size = dp->vma_invalidate_size;
454
455	dispatch_io(op, op_flags, num_regions, where, dp, io, 1);
 
 
 
456
457	wait_for_completion_io(&sio.wait);
 
 
 
 
 
458
459	if (error_bits)
460		*error_bits = sio.error_bits;
461
462	return sio.error_bits ? -EIO : 0;
463}
464
465static int async_io(struct dm_io_client *client, unsigned int num_regions,
466		    struct dm_io_region *where, int op, int op_flags,
467		    struct dpages *dp, io_notify_fn fn, void *context)
468{
469	struct io *io;
470
471	if (num_regions > 1 && !op_is_write(op)) {
472		WARN_ON(1);
473		fn(1, context);
474		return -EIO;
475	}
476
477	io = mempool_alloc(client->pool, GFP_NOIO);
478	io->error_bits = 0;
479	atomic_set(&io->count, 1); /* see dispatch_io() */
 
480	io->client = client;
481	io->callback = fn;
482	io->context = context;
483
484	io->vma_invalidate_address = dp->vma_invalidate_address;
485	io->vma_invalidate_size = dp->vma_invalidate_size;
486
487	dispatch_io(op, op_flags, num_regions, where, dp, io, 0);
488	return 0;
489}
490
491static int dp_init(struct dm_io_request *io_req, struct dpages *dp,
492		   unsigned long size)
493{
494	/* Set up dpages based on memory type */
495
496	dp->vma_invalidate_address = NULL;
497	dp->vma_invalidate_size = 0;
498
499	switch (io_req->mem.type) {
500	case DM_IO_PAGE_LIST:
501		list_dp_init(dp, io_req->mem.ptr.pl, io_req->mem.offset);
502		break;
503
504	case DM_IO_BIO:
505		bio_dp_init(dp, io_req->mem.ptr.bio);
506		break;
507
508	case DM_IO_VMA:
509		flush_kernel_vmap_range(io_req->mem.ptr.vma, size);
510		if (io_req->bi_op == REQ_OP_READ) {
511			dp->vma_invalidate_address = io_req->mem.ptr.vma;
512			dp->vma_invalidate_size = size;
513		}
514		vm_dp_init(dp, io_req->mem.ptr.vma);
515		break;
516
517	case DM_IO_KMEM:
518		km_dp_init(dp, io_req->mem.ptr.addr);
519		break;
520
521	default:
522		return -EINVAL;
523	}
524
525	return 0;
526}
527
528/*
529 * New collapsed (a)synchronous interface.
530 *
531 * If the IO is asynchronous (i.e. it has notify.fn), you must either unplug
532 * the queue with blk_unplug() some time later or set REQ_SYNC in
533 * io_req->bi_opf. If you fail to do one of these, the IO will be submitted to
534 * the disk after q->unplug_delay, which defaults to 3ms in blk-settings.c.
535 */
536int dm_io(struct dm_io_request *io_req, unsigned num_regions,
537	  struct dm_io_region *where, unsigned long *sync_error_bits)
538{
539	int r;
540	struct dpages dp;
541
542	r = dp_init(io_req, &dp, (unsigned long)where->count << SECTOR_SHIFT);
543	if (r)
544		return r;
545
546	if (!io_req->notify.fn)
547		return sync_io(io_req->client, num_regions, where,
548			       io_req->bi_op, io_req->bi_op_flags, &dp,
549			       sync_error_bits);
550
551	return async_io(io_req->client, num_regions, where, io_req->bi_op,
552			io_req->bi_op_flags, &dp, io_req->notify.fn,
553			io_req->notify.context);
554}
555EXPORT_SYMBOL(dm_io);
556
557int __init dm_io_init(void)
558{
559	_dm_io_cache = KMEM_CACHE(io, 0);
560	if (!_dm_io_cache)
561		return -ENOMEM;
562
563	return 0;
564}
565
566void dm_io_exit(void)
567{
568	kmem_cache_destroy(_dm_io_cache);
569	_dm_io_cache = NULL;
570}