xref: /openbmc/linux/drivers/md/dm-io.c (revision 9470114d)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2003 Sistina Software
4  * Copyright (C) 2006 Red Hat GmbH
5  *
6  * This file is released under the GPL.
7  */
8 
9 #include "dm-core.h"
10 
11 #include <linux/device-mapper.h>
12 
13 #include <linux/bio.h>
14 #include <linux/completion.h>
15 #include <linux/mempool.h>
16 #include <linux/module.h>
17 #include <linux/sched.h>
18 #include <linux/slab.h>
19 #include <linux/dm-io.h>
20 
21 #define DM_MSG_PREFIX "io"
22 
23 #define DM_IO_MAX_REGIONS	BITS_PER_LONG
24 
25 struct 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  */
34 struct io {
35 	unsigned long error_bits;
36 	atomic_t count;
37 	struct dm_io_client *client;
38 	io_notify_fn callback;
39 	void *context;
40 	void *vma_invalidate_address;
41 	unsigned long vma_invalidate_size;
42 } __aligned(DM_IO_MAX_REGIONS);
43 
44 static struct kmem_cache *_dm_io_cache;
45 
46 /*
47  * Create a client with mempool and bioset.
48  */
49 struct dm_io_client *dm_io_client_create(void)
50 {
51 	struct dm_io_client *client;
52 	unsigned int min_ios = dm_get_reserved_bio_based_ios();
53 	int ret;
54 
55 	client = kzalloc(sizeof(*client), GFP_KERNEL);
56 	if (!client)
57 		return ERR_PTR(-ENOMEM);
58 
59 	ret = mempool_init_slab_pool(&client->pool, min_ios, _dm_io_cache);
60 	if (ret)
61 		goto bad;
62 
63 	ret = bioset_init(&client->bios, min_ios, 0, BIOSET_NEED_BVECS);
64 	if (ret)
65 		goto bad;
66 
67 	return client;
68 
69 bad:
70 	mempool_exit(&client->pool);
71 	kfree(client);
72 	return ERR_PTR(ret);
73 }
74 EXPORT_SYMBOL(dm_io_client_create);
75 
76 void dm_io_client_destroy(struct dm_io_client *client)
77 {
78 	mempool_exit(&client->pool);
79 	bioset_exit(&client->bios);
80 	kfree(client);
81 }
82 EXPORT_SYMBOL(dm_io_client_destroy);
83 
84 /*
85  *-------------------------------------------------------------------
86  * We need to keep track of which region a bio is doing io for.
87  * To avoid a memory allocation to store just 5 or 6 bits, we
88  * ensure the 'struct io' pointer is aligned so enough low bits are
89  * always zero and then combine it with the region number directly in
90  * bi_private.
91  *-------------------------------------------------------------------
92  */
93 static void store_io_and_region_in_bio(struct bio *bio, struct io *io,
94 				       unsigned int region)
95 {
96 	if (unlikely(!IS_ALIGNED((unsigned long)io, DM_IO_MAX_REGIONS))) {
97 		DMCRIT("Unaligned struct io pointer %p", io);
98 		BUG();
99 	}
100 
101 	bio->bi_private = (void *)((unsigned long)io | region);
102 }
103 
104 static void retrieve_io_and_region_from_bio(struct bio *bio, struct io **io,
105 				       unsigned int *region)
106 {
107 	unsigned long val = (unsigned long)bio->bi_private;
108 
109 	*io = (void *)(val & -(unsigned long)DM_IO_MAX_REGIONS);
110 	*region = val & (DM_IO_MAX_REGIONS - 1);
111 }
112 
113 /*
114  *--------------------------------------------------------------
115  * We need an io object to keep track of the number of bios that
116  * have been dispatched for a particular io.
117  *--------------------------------------------------------------
118  */
119 static void complete_io(struct io *io)
120 {
121 	unsigned long error_bits = io->error_bits;
122 	io_notify_fn fn = io->callback;
123 	void *context = io->context;
124 
125 	if (io->vma_invalidate_size)
126 		invalidate_kernel_vmap_range(io->vma_invalidate_address,
127 					     io->vma_invalidate_size);
128 
129 	mempool_free(io, &io->client->pool);
130 	fn(error_bits, context);
131 }
132 
133 static void dec_count(struct io *io, unsigned int region, blk_status_t error)
134 {
135 	if (error)
136 		set_bit(region, &io->error_bits);
137 
138 	if (atomic_dec_and_test(&io->count))
139 		complete_io(io);
140 }
141 
142 static void endio(struct bio *bio)
143 {
144 	struct io *io;
145 	unsigned int region;
146 	blk_status_t error;
147 
148 	if (bio->bi_status && bio_data_dir(bio) == READ)
149 		zero_fill_bio(bio);
150 
151 	/*
152 	 * The bio destructor in bio_put() may use the io object.
153 	 */
154 	retrieve_io_and_region_from_bio(bio, &io, &region);
155 
156 	error = bio->bi_status;
157 	bio_put(bio);
158 
159 	dec_count(io, region, error);
160 }
161 
162 /*
163  *--------------------------------------------------------------
164  * These little objects provide an abstraction for getting a new
165  * destination page for io.
166  *--------------------------------------------------------------
167  */
168 struct dpages {
169 	void (*get_page)(struct dpages *dp,
170 			 struct page **p, unsigned long *len, unsigned int *offset);
171 	void (*next_page)(struct dpages *dp);
172 
173 	union {
174 		unsigned int context_u;
175 		struct bvec_iter context_bi;
176 	};
177 	void *context_ptr;
178 
179 	void *vma_invalidate_address;
180 	unsigned long vma_invalidate_size;
181 };
182 
183 /*
184  * Functions for getting the pages from a list.
185  */
186 static void list_get_page(struct dpages *dp,
187 		  struct page **p, unsigned long *len, unsigned int *offset)
188 {
189 	unsigned int o = dp->context_u;
190 	struct page_list *pl = dp->context_ptr;
191 
192 	*p = pl->page;
193 	*len = PAGE_SIZE - o;
194 	*offset = o;
195 }
196 
197 static void list_next_page(struct dpages *dp)
198 {
199 	struct page_list *pl = dp->context_ptr;
200 
201 	dp->context_ptr = pl->next;
202 	dp->context_u = 0;
203 }
204 
205 static void list_dp_init(struct dpages *dp, struct page_list *pl, unsigned int offset)
206 {
207 	dp->get_page = list_get_page;
208 	dp->next_page = list_next_page;
209 	dp->context_u = offset;
210 	dp->context_ptr = pl;
211 }
212 
213 /*
214  * Functions for getting the pages from a bvec.
215  */
216 static void bio_get_page(struct dpages *dp, struct page **p,
217 			 unsigned long *len, unsigned int *offset)
218 {
219 	struct bio_vec bvec = bvec_iter_bvec((struct bio_vec *)dp->context_ptr,
220 					     dp->context_bi);
221 
222 	*p = bvec.bv_page;
223 	*len = bvec.bv_len;
224 	*offset = bvec.bv_offset;
225 
226 	/* avoid figuring it out again in bio_next_page() */
227 	dp->context_bi.bi_sector = (sector_t)bvec.bv_len;
228 }
229 
230 static void bio_next_page(struct dpages *dp)
231 {
232 	unsigned int len = (unsigned int)dp->context_bi.bi_sector;
233 
234 	bvec_iter_advance((struct bio_vec *)dp->context_ptr,
235 			  &dp->context_bi, len);
236 }
237 
238 static void bio_dp_init(struct dpages *dp, struct bio *bio)
239 {
240 	dp->get_page = bio_get_page;
241 	dp->next_page = bio_next_page;
242 
243 	/*
244 	 * We just use bvec iterator to retrieve pages, so it is ok to
245 	 * access the bvec table directly here
246 	 */
247 	dp->context_ptr = bio->bi_io_vec;
248 	dp->context_bi = bio->bi_iter;
249 }
250 
251 /*
252  * Functions for getting the pages from a VMA.
253  */
254 static void vm_get_page(struct dpages *dp,
255 		 struct page **p, unsigned long *len, unsigned int *offset)
256 {
257 	*p = vmalloc_to_page(dp->context_ptr);
258 	*offset = dp->context_u;
259 	*len = PAGE_SIZE - dp->context_u;
260 }
261 
262 static void vm_next_page(struct dpages *dp)
263 {
264 	dp->context_ptr += PAGE_SIZE - dp->context_u;
265 	dp->context_u = 0;
266 }
267 
268 static void vm_dp_init(struct dpages *dp, void *data)
269 {
270 	dp->get_page = vm_get_page;
271 	dp->next_page = vm_next_page;
272 	dp->context_u = offset_in_page(data);
273 	dp->context_ptr = data;
274 }
275 
276 /*
277  * Functions for getting the pages from kernel memory.
278  */
279 static void km_get_page(struct dpages *dp, struct page **p, unsigned long *len,
280 			unsigned int *offset)
281 {
282 	*p = virt_to_page(dp->context_ptr);
283 	*offset = dp->context_u;
284 	*len = PAGE_SIZE - dp->context_u;
285 }
286 
287 static void km_next_page(struct dpages *dp)
288 {
289 	dp->context_ptr += PAGE_SIZE - dp->context_u;
290 	dp->context_u = 0;
291 }
292 
293 static void km_dp_init(struct dpages *dp, void *data)
294 {
295 	dp->get_page = km_get_page;
296 	dp->next_page = km_next_page;
297 	dp->context_u = offset_in_page(data);
298 	dp->context_ptr = data;
299 }
300 
301 /*
302  *---------------------------------------------------------------
303  * IO routines that accept a list of pages.
304  *---------------------------------------------------------------
305  */
306 static void do_region(const blk_opf_t opf, unsigned int region,
307 		      struct dm_io_region *where, struct dpages *dp,
308 		      struct io *io, unsigned short ioprio)
309 {
310 	struct bio *bio;
311 	struct page *page;
312 	unsigned long len;
313 	unsigned int offset;
314 	unsigned int num_bvecs;
315 	sector_t remaining = where->count;
316 	struct request_queue *q = bdev_get_queue(where->bdev);
317 	sector_t num_sectors;
318 	unsigned int special_cmd_max_sectors;
319 	const enum req_op op = opf & REQ_OP_MASK;
320 
321 	/*
322 	 * Reject unsupported discard and write same requests.
323 	 */
324 	if (op == REQ_OP_DISCARD)
325 		special_cmd_max_sectors = bdev_max_discard_sectors(where->bdev);
326 	else if (op == REQ_OP_WRITE_ZEROES)
327 		special_cmd_max_sectors = q->limits.max_write_zeroes_sectors;
328 	if ((op == REQ_OP_DISCARD || op == REQ_OP_WRITE_ZEROES) &&
329 	    special_cmd_max_sectors == 0) {
330 		atomic_inc(&io->count);
331 		dec_count(io, region, BLK_STS_NOTSUPP);
332 		return;
333 	}
334 
335 	/*
336 	 * where->count may be zero if op holds a flush and we need to
337 	 * send a zero-sized flush.
338 	 */
339 	do {
340 		/*
341 		 * Allocate a suitably sized-bio.
342 		 */
343 		switch (op) {
344 		case REQ_OP_DISCARD:
345 		case REQ_OP_WRITE_ZEROES:
346 			num_bvecs = 0;
347 			break;
348 		default:
349 			num_bvecs = bio_max_segs(dm_sector_div_up(remaining,
350 						(PAGE_SIZE >> SECTOR_SHIFT)));
351 		}
352 
353 		bio = bio_alloc_bioset(where->bdev, num_bvecs, opf, GFP_NOIO,
354 				       &io->client->bios);
355 		bio->bi_iter.bi_sector = where->sector + (where->count - remaining);
356 		bio->bi_end_io = endio;
357 		bio->bi_ioprio = ioprio;
358 		store_io_and_region_in_bio(bio, io, region);
359 
360 		if (op == REQ_OP_DISCARD || op == REQ_OP_WRITE_ZEROES) {
361 			num_sectors = min_t(sector_t, special_cmd_max_sectors, remaining);
362 			bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
363 			remaining -= num_sectors;
364 		} else {
365 			while (remaining) {
366 				/*
367 				 * Try and add as many pages as possible.
368 				 */
369 				dp->get_page(dp, &page, &len, &offset);
370 				len = min(len, to_bytes(remaining));
371 				if (!bio_add_page(bio, page, len, offset))
372 					break;
373 
374 				offset = 0;
375 				remaining -= to_sector(len);
376 				dp->next_page(dp);
377 			}
378 		}
379 
380 		atomic_inc(&io->count);
381 		submit_bio(bio);
382 	} while (remaining);
383 }
384 
385 static void dispatch_io(blk_opf_t opf, unsigned int num_regions,
386 			struct dm_io_region *where, struct dpages *dp,
387 			struct io *io, int sync, unsigned short ioprio)
388 {
389 	int i;
390 	struct dpages old_pages = *dp;
391 
392 	BUG_ON(num_regions > DM_IO_MAX_REGIONS);
393 
394 	if (sync)
395 		opf |= REQ_SYNC;
396 
397 	/*
398 	 * For multiple regions we need to be careful to rewind
399 	 * the dp object for each call to do_region.
400 	 */
401 	for (i = 0; i < num_regions; i++) {
402 		*dp = old_pages;
403 		if (where[i].count || (opf & REQ_PREFLUSH))
404 			do_region(opf, i, where + i, dp, io, ioprio);
405 	}
406 
407 	/*
408 	 * Drop the extra reference that we were holding to avoid
409 	 * the io being completed too early.
410 	 */
411 	dec_count(io, 0, 0);
412 }
413 
414 struct sync_io {
415 	unsigned long error_bits;
416 	struct completion wait;
417 };
418 
419 static void sync_io_complete(unsigned long error, void *context)
420 {
421 	struct sync_io *sio = context;
422 
423 	sio->error_bits = error;
424 	complete(&sio->wait);
425 }
426 
427 static int sync_io(struct dm_io_client *client, unsigned int num_regions,
428 		   struct dm_io_region *where, blk_opf_t opf, struct dpages *dp,
429 		   unsigned long *error_bits, unsigned short ioprio)
430 {
431 	struct io *io;
432 	struct sync_io sio;
433 
434 	if (num_regions > 1 && !op_is_write(opf)) {
435 		WARN_ON(1);
436 		return -EIO;
437 	}
438 
439 	init_completion(&sio.wait);
440 
441 	io = mempool_alloc(&client->pool, GFP_NOIO);
442 	io->error_bits = 0;
443 	atomic_set(&io->count, 1); /* see dispatch_io() */
444 	io->client = client;
445 	io->callback = sync_io_complete;
446 	io->context = &sio;
447 
448 	io->vma_invalidate_address = dp->vma_invalidate_address;
449 	io->vma_invalidate_size = dp->vma_invalidate_size;
450 
451 	dispatch_io(opf, num_regions, where, dp, io, 1, ioprio);
452 
453 	wait_for_completion_io(&sio.wait);
454 
455 	if (error_bits)
456 		*error_bits = sio.error_bits;
457 
458 	return sio.error_bits ? -EIO : 0;
459 }
460 
461 static int async_io(struct dm_io_client *client, unsigned int num_regions,
462 		    struct dm_io_region *where, blk_opf_t opf,
463 		    struct dpages *dp, io_notify_fn fn, void *context,
464 		    unsigned short ioprio)
465 {
466 	struct io *io;
467 
468 	if (num_regions > 1 && !op_is_write(opf)) {
469 		WARN_ON(1);
470 		fn(1, context);
471 		return -EIO;
472 	}
473 
474 	io = mempool_alloc(&client->pool, GFP_NOIO);
475 	io->error_bits = 0;
476 	atomic_set(&io->count, 1); /* see dispatch_io() */
477 	io->client = client;
478 	io->callback = fn;
479 	io->context = context;
480 
481 	io->vma_invalidate_address = dp->vma_invalidate_address;
482 	io->vma_invalidate_size = dp->vma_invalidate_size;
483 
484 	dispatch_io(opf, num_regions, where, dp, io, 0, ioprio);
485 	return 0;
486 }
487 
488 static int dp_init(struct dm_io_request *io_req, struct dpages *dp,
489 		   unsigned long size)
490 {
491 	/* Set up dpages based on memory type */
492 
493 	dp->vma_invalidate_address = NULL;
494 	dp->vma_invalidate_size = 0;
495 
496 	switch (io_req->mem.type) {
497 	case DM_IO_PAGE_LIST:
498 		list_dp_init(dp, io_req->mem.ptr.pl, io_req->mem.offset);
499 		break;
500 
501 	case DM_IO_BIO:
502 		bio_dp_init(dp, io_req->mem.ptr.bio);
503 		break;
504 
505 	case DM_IO_VMA:
506 		flush_kernel_vmap_range(io_req->mem.ptr.vma, size);
507 		if ((io_req->bi_opf & REQ_OP_MASK) == REQ_OP_READ) {
508 			dp->vma_invalidate_address = io_req->mem.ptr.vma;
509 			dp->vma_invalidate_size = size;
510 		}
511 		vm_dp_init(dp, io_req->mem.ptr.vma);
512 		break;
513 
514 	case DM_IO_KMEM:
515 		km_dp_init(dp, io_req->mem.ptr.addr);
516 		break;
517 
518 	default:
519 		return -EINVAL;
520 	}
521 
522 	return 0;
523 }
524 
525 int dm_io(struct dm_io_request *io_req, unsigned int num_regions,
526 	  struct dm_io_region *where, unsigned long *sync_error_bits,
527 	  unsigned short ioprio)
528 {
529 	int r;
530 	struct dpages dp;
531 
532 	r = dp_init(io_req, &dp, (unsigned long)where->count << SECTOR_SHIFT);
533 	if (r)
534 		return r;
535 
536 	if (!io_req->notify.fn)
537 		return sync_io(io_req->client, num_regions, where,
538 			       io_req->bi_opf, &dp, sync_error_bits, ioprio);
539 
540 	return async_io(io_req->client, num_regions, where,
541 			io_req->bi_opf, &dp, io_req->notify.fn,
542 			io_req->notify.context, ioprio);
543 }
544 EXPORT_SYMBOL(dm_io);
545 
546 int __init dm_io_init(void)
547 {
548 	_dm_io_cache = KMEM_CACHE(io, 0);
549 	if (!_dm_io_cache)
550 		return -ENOMEM;
551 
552 	return 0;
553 }
554 
555 void dm_io_exit(void)
556 {
557 	kmem_cache_destroy(_dm_io_cache);
558 	_dm_io_cache = NULL;
559 }
560