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 int ret;
53
54 client = kzalloc(sizeof(*client), GFP_KERNEL);
55 if (!client)
56 return ERR_PTR(-ENOMEM);
57
58 ret = mempool_init_slab_pool(&client->pool, min_ios, _dm_io_cache);
59 if (ret)
60 goto bad;
61
62 ret = bioset_init(&client->bios, min_ios, 0, BIOSET_NEED_BVECS);
63 if (ret)
64 goto bad;
65
66 return client;
67
68 bad:
69 mempool_exit(&client->pool);
70 kfree(client);
71 return ERR_PTR(ret);
72}
73EXPORT_SYMBOL(dm_io_client_create);
74
75void dm_io_client_destroy(struct dm_io_client *client)
76{
77 mempool_exit(&client->pool);
78 bioset_exit(&client->bios);
79 kfree(client);
80}
81EXPORT_SYMBOL(dm_io_client_destroy);
82
83/*-----------------------------------------------------------------
84 * We need to keep track of which region a bio is doing io for.
85 * To avoid a memory allocation to store just 5 or 6 bits, we
86 * ensure the 'struct io' pointer is aligned so enough low bits are
87 * always zero and then combine it with the region number directly in
88 * bi_private.
89 *---------------------------------------------------------------*/
90static void store_io_and_region_in_bio(struct bio *bio, struct io *io,
91 unsigned region)
92{
93 if (unlikely(!IS_ALIGNED((unsigned long)io, DM_IO_MAX_REGIONS))) {
94 DMCRIT("Unaligned struct io pointer %p", io);
95 BUG();
96 }
97
98 bio->bi_private = (void *)((unsigned long)io | region);
99}
100
101static void retrieve_io_and_region_from_bio(struct bio *bio, struct io **io,
102 unsigned *region)
103{
104 unsigned long val = (unsigned long)bio->bi_private;
105
106 *io = (void *)(val & -(unsigned long)DM_IO_MAX_REGIONS);
107 *region = val & (DM_IO_MAX_REGIONS - 1);
108}
109
110/*-----------------------------------------------------------------
111 * We need an io object to keep track of the number of bios that
112 * have been dispatched for a particular io.
113 *---------------------------------------------------------------*/
114static void complete_io(struct io *io)
115{
116 unsigned long error_bits = io->error_bits;
117 io_notify_fn fn = io->callback;
118 void *context = io->context;
119
120 if (io->vma_invalidate_size)
121 invalidate_kernel_vmap_range(io->vma_invalidate_address,
122 io->vma_invalidate_size);
123
124 mempool_free(io, &io->client->pool);
125 fn(error_bits, context);
126}
127
128static void dec_count(struct io *io, unsigned int region, blk_status_t error)
129{
130 if (error)
131 set_bit(region, &io->error_bits);
132
133 if (atomic_dec_and_test(&io->count))
134 complete_io(io);
135}
136
137static void endio(struct bio *bio)
138{
139 struct io *io;
140 unsigned region;
141 blk_status_t error;
142
143 if (bio->bi_status && bio_data_dir(bio) == READ)
144 zero_fill_bio(bio);
145
146 /*
147 * The bio destructor in bio_put() may use the io object.
148 */
149 retrieve_io_and_region_from_bio(bio, &io, &region);
150
151 error = bio->bi_status;
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 union {
167 unsigned context_u;
168 struct bvec_iter context_bi;
169 };
170 void *context_ptr;
171
172 void *vma_invalidate_address;
173 unsigned long vma_invalidate_size;
174};
175
176/*
177 * Functions for getting the pages from a list.
178 */
179static void list_get_page(struct dpages *dp,
180 struct page **p, unsigned long *len, unsigned *offset)
181{
182 unsigned o = dp->context_u;
183 struct page_list *pl = (struct page_list *) dp->context_ptr;
184
185 *p = pl->page;
186 *len = PAGE_SIZE - o;
187 *offset = o;
188}
189
190static void list_next_page(struct dpages *dp)
191{
192 struct page_list *pl = (struct page_list *) dp->context_ptr;
193 dp->context_ptr = pl->next;
194 dp->context_u = 0;
195}
196
197static void list_dp_init(struct dpages *dp, struct page_list *pl, unsigned offset)
198{
199 dp->get_page = list_get_page;
200 dp->next_page = list_next_page;
201 dp->context_u = offset;
202 dp->context_ptr = pl;
203}
204
205/*
206 * Functions for getting the pages from a bvec.
207 */
208static void bio_get_page(struct dpages *dp, struct page **p,
209 unsigned long *len, unsigned *offset)
210{
211 struct bio_vec bvec = bvec_iter_bvec((struct bio_vec *)dp->context_ptr,
212 dp->context_bi);
213
214 *p = bvec.bv_page;
215 *len = bvec.bv_len;
216 *offset = bvec.bv_offset;
217
218 /* avoid figuring it out again in bio_next_page() */
219 dp->context_bi.bi_sector = (sector_t)bvec.bv_len;
220}
221
222static void bio_next_page(struct dpages *dp)
223{
224 unsigned int len = (unsigned int)dp->context_bi.bi_sector;
225
226 bvec_iter_advance((struct bio_vec *)dp->context_ptr,
227 &dp->context_bi, len);
228}
229
230static void bio_dp_init(struct dpages *dp, struct bio *bio)
231{
232 dp->get_page = bio_get_page;
233 dp->next_page = bio_next_page;
234
235 /*
236 * We just use bvec iterator to retrieve pages, so it is ok to
237 * access the bvec table directly here
238 */
239 dp->context_ptr = bio->bi_io_vec;
240 dp->context_bi = bio->bi_iter;
241}
242
243/*
244 * Functions for getting the pages from a VMA.
245 */
246static void vm_get_page(struct dpages *dp,
247 struct page **p, unsigned long *len, unsigned *offset)
248{
249 *p = vmalloc_to_page(dp->context_ptr);
250 *offset = dp->context_u;
251 *len = PAGE_SIZE - dp->context_u;
252}
253
254static void vm_next_page(struct dpages *dp)
255{
256 dp->context_ptr += PAGE_SIZE - dp->context_u;
257 dp->context_u = 0;
258}
259
260static void vm_dp_init(struct dpages *dp, void *data)
261{
262 dp->get_page = vm_get_page;
263 dp->next_page = vm_next_page;
264 dp->context_u = offset_in_page(data);
265 dp->context_ptr = data;
266}
267
268/*
269 * Functions for getting the pages from kernel memory.
270 */
271static void km_get_page(struct dpages *dp, struct page **p, unsigned long *len,
272 unsigned *offset)
273{
274 *p = virt_to_page(dp->context_ptr);
275 *offset = dp->context_u;
276 *len = PAGE_SIZE - dp->context_u;
277}
278
279static void km_next_page(struct dpages *dp)
280{
281 dp->context_ptr += PAGE_SIZE - dp->context_u;
282 dp->context_u = 0;
283}
284
285static void km_dp_init(struct dpages *dp, void *data)
286{
287 dp->get_page = km_get_page;
288 dp->next_page = km_next_page;
289 dp->context_u = offset_in_page(data);
290 dp->context_ptr = data;
291}
292
293/*-----------------------------------------------------------------
294 * IO routines that accept a list of pages.
295 *---------------------------------------------------------------*/
296static void do_region(int op, int op_flags, unsigned region,
297 struct dm_io_region *where, struct dpages *dp,
298 struct io *io)
299{
300 struct bio *bio;
301 struct page *page;
302 unsigned long len;
303 unsigned offset;
304 unsigned num_bvecs;
305 sector_t remaining = where->count;
306 struct request_queue *q = bdev_get_queue(where->bdev);
307 unsigned short logical_block_size = queue_logical_block_size(q);
308 sector_t num_sectors;
309 unsigned int uninitialized_var(special_cmd_max_sectors);
310
311 /*
312 * Reject unsupported discard and write same requests.
313 */
314 if (op == REQ_OP_DISCARD)
315 special_cmd_max_sectors = q->limits.max_discard_sectors;
316 else if (op == REQ_OP_WRITE_ZEROES)
317 special_cmd_max_sectors = q->limits.max_write_zeroes_sectors;
318 else if (op == REQ_OP_WRITE_SAME)
319 special_cmd_max_sectors = q->limits.max_write_same_sectors;
320 if ((op == REQ_OP_DISCARD || op == REQ_OP_WRITE_ZEROES ||
321 op == REQ_OP_WRITE_SAME) && special_cmd_max_sectors == 0) {
322 atomic_inc(&io->count);
323 dec_count(io, region, BLK_STS_NOTSUPP);
324 return;
325 }
326
327 /*
328 * where->count may be zero if op holds a flush and we need to
329 * send a zero-sized flush.
330 */
331 do {
332 /*
333 * Allocate a suitably sized-bio.
334 */
335 switch (op) {
336 case REQ_OP_DISCARD:
337 case REQ_OP_WRITE_ZEROES:
338 num_bvecs = 0;
339 break;
340 case REQ_OP_WRITE_SAME:
341 num_bvecs = 1;
342 break;
343 default:
344 num_bvecs = min_t(int, BIO_MAX_PAGES,
345 dm_sector_div_up(remaining, (PAGE_SIZE >> SECTOR_SHIFT)));
346 }
347
348 bio = bio_alloc_bioset(GFP_NOIO, num_bvecs, &io->client->bios);
349 bio->bi_iter.bi_sector = where->sector + (where->count - remaining);
350 bio_set_dev(bio, where->bdev);
351 bio->bi_end_io = endio;
352 bio_set_op_attrs(bio, op, op_flags);
353 store_io_and_region_in_bio(bio, io, region);
354
355 if (op == REQ_OP_DISCARD || op == REQ_OP_WRITE_ZEROES) {
356 num_sectors = min_t(sector_t, special_cmd_max_sectors, remaining);
357 bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
358 remaining -= num_sectors;
359 } else if (op == REQ_OP_WRITE_SAME) {
360 /*
361 * WRITE SAME only uses a single page.
362 */
363 dp->get_page(dp, &page, &len, &offset);
364 bio_add_page(bio, page, logical_block_size, offset);
365 num_sectors = min_t(sector_t, special_cmd_max_sectors, remaining);
366 bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
367
368 offset = 0;
369 remaining -= num_sectors;
370 dp->next_page(dp);
371 } else while (remaining) {
372 /*
373 * Try and add as many pages as possible.
374 */
375 dp->get_page(dp, &page, &len, &offset);
376 len = min(len, to_bytes(remaining));
377 if (!bio_add_page(bio, page, len, offset))
378 break;
379
380 offset = 0;
381 remaining -= to_sector(len);
382 dp->next_page(dp);
383 }
384
385 atomic_inc(&io->count);
386 submit_bio(bio);
387 } while (remaining);
388}
389
390static void dispatch_io(int op, int op_flags, unsigned int num_regions,
391 struct dm_io_region *where, struct dpages *dp,
392 struct io *io, int sync)
393{
394 int i;
395 struct dpages old_pages = *dp;
396
397 BUG_ON(num_regions > DM_IO_MAX_REGIONS);
398
399 if (sync)
400 op_flags |= REQ_SYNC;
401
402 /*
403 * For multiple regions we need to be careful to rewind
404 * the dp object for each call to do_region.
405 */
406 for (i = 0; i < num_regions; i++) {
407 *dp = old_pages;
408 if (where[i].count || (op_flags & REQ_PREFLUSH))
409 do_region(op, op_flags, i, where + i, dp, io);
410 }
411
412 /*
413 * Drop the extra reference that we were holding to avoid
414 * the io being completed too early.
415 */
416 dec_count(io, 0, 0);
417}
418
419struct sync_io {
420 unsigned long error_bits;
421 struct completion wait;
422};
423
424static void sync_io_complete(unsigned long error, void *context)
425{
426 struct sync_io *sio = context;
427
428 sio->error_bits = error;
429 complete(&sio->wait);
430}
431
432static int sync_io(struct dm_io_client *client, unsigned int num_regions,
433 struct dm_io_region *where, int op, int op_flags,
434 struct dpages *dp, unsigned long *error_bits)
435{
436 struct io *io;
437 struct sync_io sio;
438
439 if (num_regions > 1 && !op_is_write(op)) {
440 WARN_ON(1);
441 return -EIO;
442 }
443
444 init_completion(&sio.wait);
445
446 io = mempool_alloc(&client->pool, GFP_NOIO);
447 io->error_bits = 0;
448 atomic_set(&io->count, 1); /* see dispatch_io() */
449 io->client = client;
450 io->callback = sync_io_complete;
451 io->context = &sio;
452
453 io->vma_invalidate_address = dp->vma_invalidate_address;
454 io->vma_invalidate_size = dp->vma_invalidate_size;
455
456 dispatch_io(op, op_flags, num_regions, where, dp, io, 1);
457
458 wait_for_completion_io(&sio.wait);
459
460 if (error_bits)
461 *error_bits = sio.error_bits;
462
463 return sio.error_bits ? -EIO : 0;
464}
465
466static int async_io(struct dm_io_client *client, unsigned int num_regions,
467 struct dm_io_region *where, int op, int op_flags,
468 struct dpages *dp, io_notify_fn fn, void *context)
469{
470 struct io *io;
471
472 if (num_regions > 1 && !op_is_write(op)) {
473 WARN_ON(1);
474 fn(1, context);
475 return -EIO;
476 }
477
478 io = mempool_alloc(&client->pool, GFP_NOIO);
479 io->error_bits = 0;
480 atomic_set(&io->count, 1); /* see dispatch_io() */
481 io->client = client;
482 io->callback = fn;
483 io->context = context;
484
485 io->vma_invalidate_address = dp->vma_invalidate_address;
486 io->vma_invalidate_size = dp->vma_invalidate_size;
487
488 dispatch_io(op, op_flags, num_regions, where, dp, io, 0);
489 return 0;
490}
491
492static int dp_init(struct dm_io_request *io_req, struct dpages *dp,
493 unsigned long size)
494{
495 /* Set up dpages based on memory type */
496
497 dp->vma_invalidate_address = NULL;
498 dp->vma_invalidate_size = 0;
499
500 switch (io_req->mem.type) {
501 case DM_IO_PAGE_LIST:
502 list_dp_init(dp, io_req->mem.ptr.pl, io_req->mem.offset);
503 break;
504
505 case DM_IO_BIO:
506 bio_dp_init(dp, io_req->mem.ptr.bio);
507 break;
508
509 case DM_IO_VMA:
510 flush_kernel_vmap_range(io_req->mem.ptr.vma, size);
511 if (io_req->bi_op == REQ_OP_READ) {
512 dp->vma_invalidate_address = io_req->mem.ptr.vma;
513 dp->vma_invalidate_size = size;
514 }
515 vm_dp_init(dp, io_req->mem.ptr.vma);
516 break;
517
518 case DM_IO_KMEM:
519 km_dp_init(dp, io_req->mem.ptr.addr);
520 break;
521
522 default:
523 return -EINVAL;
524 }
525
526 return 0;
527}
528
529/*
530 * New collapsed (a)synchronous interface.
531 *
532 * If the IO is asynchronous (i.e. it has notify.fn), you must either unplug
533 * the queue with blk_unplug() some time later or set REQ_SYNC in
534 * io_req->bi_opf. If you fail to do one of these, the IO will be submitted to
535 * the disk after q->unplug_delay, which defaults to 3ms in blk-settings.c.
536 */
537int dm_io(struct dm_io_request *io_req, unsigned num_regions,
538 struct dm_io_region *where, unsigned long *sync_error_bits)
539{
540 int r;
541 struct dpages dp;
542
543 r = dp_init(io_req, &dp, (unsigned long)where->count << SECTOR_SHIFT);
544 if (r)
545 return r;
546
547 if (!io_req->notify.fn)
548 return sync_io(io_req->client, num_regions, where,
549 io_req->bi_op, io_req->bi_op_flags, &dp,
550 sync_error_bits);
551
552 return async_io(io_req->client, num_regions, where, io_req->bi_op,
553 io_req->bi_op_flags, &dp, io_req->notify.fn,
554 io_req->notify.context);
555}
556EXPORT_SYMBOL(dm_io);
557
558int __init dm_io_init(void)
559{
560 _dm_io_cache = KMEM_CACHE(io, 0);
561 if (!_dm_io_cache)
562 return -ENOMEM;
563
564 return 0;
565}
566
567void dm_io_exit(void)
568{
569 kmem_cache_destroy(_dm_io_cache);
570 _dm_io_cache = NULL;
571}
572