Loading...
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, ®ion);
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, ®ion);
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 struct request_queue *q = bdev_get_queue(where->bdev);
300 sector_t discard_sectors;
301
302 /*
303 * where->count may be zero if rw holds a flush and we need to
304 * send a zero-sized flush.
305 */
306 do {
307 /*
308 * Allocate a suitably sized-bio.
309 */
310 if (rw & REQ_DISCARD)
311 num_bvecs = 1;
312 else
313 num_bvecs = min_t(int, bio_get_nr_vecs(where->bdev),
314 dm_sector_div_up(remaining, (PAGE_SIZE >> SECTOR_SHIFT)));
315
316 bio = bio_alloc_bioset(GFP_NOIO, num_bvecs, io->client->bios);
317 bio->bi_sector = where->sector + (where->count - remaining);
318 bio->bi_bdev = where->bdev;
319 bio->bi_end_io = endio;
320 bio->bi_destructor = dm_bio_destructor;
321 store_io_and_region_in_bio(bio, io, region);
322
323 if (rw & REQ_DISCARD) {
324 discard_sectors = min_t(sector_t, q->limits.max_discard_sectors, remaining);
325 bio->bi_size = discard_sectors << SECTOR_SHIFT;
326 remaining -= discard_sectors;
327 } else while (remaining) {
328 /*
329 * Try and add as many pages as possible.
330 */
331 dp->get_page(dp, &page, &len, &offset);
332 len = min(len, to_bytes(remaining));
333 if (!bio_add_page(bio, page, len, offset))
334 break;
335
336 offset = 0;
337 remaining -= to_sector(len);
338 dp->next_page(dp);
339 }
340
341 atomic_inc(&io->count);
342 submit_bio(rw, bio);
343 } while (remaining);
344}
345
346static void dispatch_io(int rw, unsigned int num_regions,
347 struct dm_io_region *where, struct dpages *dp,
348 struct io *io, int sync)
349{
350 int i;
351 struct dpages old_pages = *dp;
352
353 BUG_ON(num_regions > DM_IO_MAX_REGIONS);
354
355 if (sync)
356 rw |= REQ_SYNC;
357
358 /*
359 * For multiple regions we need to be careful to rewind
360 * the dp object for each call to do_region.
361 */
362 for (i = 0; i < num_regions; i++) {
363 *dp = old_pages;
364 if (where[i].count || (rw & REQ_FLUSH))
365 do_region(rw, i, where + i, dp, io);
366 }
367
368 /*
369 * Drop the extra reference that we were holding to avoid
370 * the io being completed too early.
371 */
372 dec_count(io, 0, 0);
373}
374
375static int sync_io(struct dm_io_client *client, unsigned int num_regions,
376 struct dm_io_region *where, int rw, struct dpages *dp,
377 unsigned long *error_bits)
378{
379 /*
380 * gcc <= 4.3 can't do the alignment for stack variables, so we must
381 * align it on our own.
382 * volatile prevents the optimizer from removing or reusing
383 * "io_" field from the stack frame (allowed in ANSI C).
384 */
385 volatile char io_[sizeof(struct io) + __alignof__(struct io) - 1];
386 struct io *io = (struct io *)PTR_ALIGN(&io_, __alignof__(struct io));
387
388 if (num_regions > 1 && (rw & RW_MASK) != WRITE) {
389 WARN_ON(1);
390 return -EIO;
391 }
392
393 io->error_bits = 0;
394 atomic_set(&io->count, 1); /* see dispatch_io() */
395 io->sleeper = current;
396 io->client = client;
397
398 io->vma_invalidate_address = dp->vma_invalidate_address;
399 io->vma_invalidate_size = dp->vma_invalidate_size;
400
401 dispatch_io(rw, num_regions, where, dp, io, 1);
402
403 while (1) {
404 set_current_state(TASK_UNINTERRUPTIBLE);
405
406 if (!atomic_read(&io->count))
407 break;
408
409 io_schedule();
410 }
411 set_current_state(TASK_RUNNING);
412
413 if (error_bits)
414 *error_bits = io->error_bits;
415
416 return io->error_bits ? -EIO : 0;
417}
418
419static int async_io(struct dm_io_client *client, unsigned int num_regions,
420 struct dm_io_region *where, int rw, struct dpages *dp,
421 io_notify_fn fn, void *context)
422{
423 struct io *io;
424
425 if (num_regions > 1 && (rw & RW_MASK) != WRITE) {
426 WARN_ON(1);
427 fn(1, context);
428 return -EIO;
429 }
430
431 io = mempool_alloc(client->pool, GFP_NOIO);
432 io->error_bits = 0;
433 atomic_set(&io->count, 1); /* see dispatch_io() */
434 io->sleeper = NULL;
435 io->client = client;
436 io->callback = fn;
437 io->context = context;
438
439 io->vma_invalidate_address = dp->vma_invalidate_address;
440 io->vma_invalidate_size = dp->vma_invalidate_size;
441
442 dispatch_io(rw, num_regions, where, dp, io, 0);
443 return 0;
444}
445
446static int dp_init(struct dm_io_request *io_req, struct dpages *dp,
447 unsigned long size)
448{
449 /* Set up dpages based on memory type */
450
451 dp->vma_invalidate_address = NULL;
452 dp->vma_invalidate_size = 0;
453
454 switch (io_req->mem.type) {
455 case DM_IO_PAGE_LIST:
456 list_dp_init(dp, io_req->mem.ptr.pl, io_req->mem.offset);
457 break;
458
459 case DM_IO_BVEC:
460 bvec_dp_init(dp, io_req->mem.ptr.bvec);
461 break;
462
463 case DM_IO_VMA:
464 flush_kernel_vmap_range(io_req->mem.ptr.vma, size);
465 if ((io_req->bi_rw & RW_MASK) == READ) {
466 dp->vma_invalidate_address = io_req->mem.ptr.vma;
467 dp->vma_invalidate_size = size;
468 }
469 vm_dp_init(dp, io_req->mem.ptr.vma);
470 break;
471
472 case DM_IO_KMEM:
473 km_dp_init(dp, io_req->mem.ptr.addr);
474 break;
475
476 default:
477 return -EINVAL;
478 }
479
480 return 0;
481}
482
483/*
484 * New collapsed (a)synchronous interface.
485 *
486 * If the IO is asynchronous (i.e. it has notify.fn), you must either unplug
487 * the queue with blk_unplug() some time later or set REQ_SYNC in
488io_req->bi_rw. If you fail to do one of these, the IO will be submitted to
489 * the disk after q->unplug_delay, which defaults to 3ms in blk-settings.c.
490 */
491int dm_io(struct dm_io_request *io_req, unsigned num_regions,
492 struct dm_io_region *where, unsigned long *sync_error_bits)
493{
494 int r;
495 struct dpages dp;
496
497 r = dp_init(io_req, &dp, (unsigned long)where->count << SECTOR_SHIFT);
498 if (r)
499 return r;
500
501 if (!io_req->notify.fn)
502 return sync_io(io_req->client, num_regions, where,
503 io_req->bi_rw, &dp, sync_error_bits);
504
505 return async_io(io_req->client, num_regions, where, io_req->bi_rw,
506 &dp, io_req->notify.fn, io_req->notify.context);
507}
508EXPORT_SYMBOL(dm_io);
509
510int __init dm_io_init(void)
511{
512 _dm_io_cache = KMEM_CACHE(io, 0);
513 if (!_dm_io_cache)
514 return -ENOMEM;
515
516 return 0;
517}
518
519void dm_io_exit(void)
520{
521 kmem_cache_destroy(_dm_io_cache);
522 _dm_io_cache = NULL;
523}