2 * Copyright (C) 2003 Sistina Software
3 * Copyright (C) 2006 Red Hat GmbH
5 * This file is released under the GPL.
10 #include <linux/device-mapper.h>
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>
20 #define DM_MSG_PREFIX "io"
22 #define DM_IO_MAX_REGIONS BITS_PER_LONG
30 * Aligning 'struct io' reduces the number of bits required to store
31 * its address. Refer to store_io_and_region_in_bio() below.
34 unsigned long error_bits;
36 struct dm_io_client *client;
37 io_notify_fn callback;
39 void *vma_invalidate_address;
40 unsigned long vma_invalidate_size;
41 } __attribute__((aligned(DM_IO_MAX_REGIONS)));
43 static struct kmem_cache *_dm_io_cache;
46 * Create a client with mempool and bioset.
48 struct dm_io_client *dm_io_client_create(void)
50 struct dm_io_client *client;
51 unsigned min_ios = dm_get_reserved_bio_based_ios();
53 client = kmalloc(sizeof(*client), GFP_KERNEL);
55 return ERR_PTR(-ENOMEM);
57 client->pool = mempool_create_slab_pool(min_ios, _dm_io_cache);
61 client->bios = bioset_create(min_ios, 0);
68 mempool_destroy(client->pool);
70 return ERR_PTR(-ENOMEM);
72 EXPORT_SYMBOL(dm_io_client_create);
74 void dm_io_client_destroy(struct dm_io_client *client)
76 mempool_destroy(client->pool);
77 bioset_free(client->bios);
80 EXPORT_SYMBOL(dm_io_client_destroy);
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
88 *---------------------------------------------------------------*/
89 static void store_io_and_region_in_bio(struct bio *bio, struct io *io,
92 if (unlikely(!IS_ALIGNED((unsigned long)io, DM_IO_MAX_REGIONS))) {
93 DMCRIT("Unaligned struct io pointer %p", io);
97 bio->bi_private = (void *)((unsigned long)io | region);
100 static void retrieve_io_and_region_from_bio(struct bio *bio, struct io **io,
103 unsigned long val = (unsigned long)bio->bi_private;
105 *io = (void *)(val & -(unsigned long)DM_IO_MAX_REGIONS);
106 *region = val & (DM_IO_MAX_REGIONS - 1);
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 *---------------------------------------------------------------*/
113 static void complete_io(struct io *io)
115 unsigned long error_bits = io->error_bits;
116 io_notify_fn fn = io->callback;
117 void *context = io->context;
119 if (io->vma_invalidate_size)
120 invalidate_kernel_vmap_range(io->vma_invalidate_address,
121 io->vma_invalidate_size);
123 mempool_free(io, io->client->pool);
124 fn(error_bits, context);
127 static void dec_count(struct io *io, unsigned int region, int error)
130 set_bit(region, &io->error_bits);
132 if (atomic_dec_and_test(&io->count))
136 static void endio(struct bio *bio)
142 if (bio->bi_error && bio_data_dir(bio) == READ)
146 * The bio destructor in bio_put() may use the io object.
148 retrieve_io_and_region_from_bio(bio, &io, ®ion);
150 error = bio->bi_error;
153 dec_count(io, region, error);
156 /*-----------------------------------------------------------------
157 * These little objects provide an abstraction for getting a new
158 * destination page for io.
159 *---------------------------------------------------------------*/
161 void (*get_page)(struct dpages *dp,
162 struct page **p, unsigned long *len, unsigned *offset);
163 void (*next_page)(struct dpages *dp);
168 void *vma_invalidate_address;
169 unsigned long vma_invalidate_size;
173 * Functions for getting the pages from a list.
175 static void list_get_page(struct dpages *dp,
176 struct page **p, unsigned long *len, unsigned *offset)
178 unsigned o = dp->context_u;
179 struct page_list *pl = (struct page_list *) dp->context_ptr;
182 *len = PAGE_SIZE - o;
186 static void list_next_page(struct dpages *dp)
188 struct page_list *pl = (struct page_list *) dp->context_ptr;
189 dp->context_ptr = pl->next;
193 static void list_dp_init(struct dpages *dp, struct page_list *pl, unsigned offset)
195 dp->get_page = list_get_page;
196 dp->next_page = list_next_page;
197 dp->context_u = offset;
198 dp->context_ptr = pl;
202 * Functions for getting the pages from a bvec.
204 static void bio_get_page(struct dpages *dp, struct page **p,
205 unsigned long *len, unsigned *offset)
207 struct bio_vec *bvec = dp->context_ptr;
209 *len = bvec->bv_len - dp->context_u;
210 *offset = bvec->bv_offset + dp->context_u;
213 static void bio_next_page(struct dpages *dp)
215 struct bio_vec *bvec = dp->context_ptr;
216 dp->context_ptr = bvec + 1;
220 static void bio_dp_init(struct dpages *dp, struct bio *bio)
222 dp->get_page = bio_get_page;
223 dp->next_page = bio_next_page;
224 dp->context_ptr = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
225 dp->context_u = bio->bi_iter.bi_bvec_done;
229 * Functions for getting the pages from a VMA.
231 static void vm_get_page(struct dpages *dp,
232 struct page **p, unsigned long *len, unsigned *offset)
234 *p = vmalloc_to_page(dp->context_ptr);
235 *offset = dp->context_u;
236 *len = PAGE_SIZE - dp->context_u;
239 static void vm_next_page(struct dpages *dp)
241 dp->context_ptr += PAGE_SIZE - dp->context_u;
245 static void vm_dp_init(struct dpages *dp, void *data)
247 dp->get_page = vm_get_page;
248 dp->next_page = vm_next_page;
249 dp->context_u = offset_in_page(data);
250 dp->context_ptr = data;
254 * Functions for getting the pages from kernel memory.
256 static void km_get_page(struct dpages *dp, struct page **p, unsigned long *len,
259 *p = virt_to_page(dp->context_ptr);
260 *offset = dp->context_u;
261 *len = PAGE_SIZE - dp->context_u;
264 static void km_next_page(struct dpages *dp)
266 dp->context_ptr += PAGE_SIZE - dp->context_u;
270 static void km_dp_init(struct dpages *dp, void *data)
272 dp->get_page = km_get_page;
273 dp->next_page = km_next_page;
274 dp->context_u = offset_in_page(data);
275 dp->context_ptr = data;
278 /*-----------------------------------------------------------------
279 * IO routines that accept a list of pages.
280 *---------------------------------------------------------------*/
281 static void do_region(int op, int op_flags, unsigned region,
282 struct dm_io_region *where, struct dpages *dp,
290 sector_t remaining = where->count;
291 struct request_queue *q = bdev_get_queue(where->bdev);
292 unsigned short logical_block_size = queue_logical_block_size(q);
293 sector_t num_sectors;
294 unsigned int uninitialized_var(special_cmd_max_sectors);
297 * Reject unsupported discard and write same requests.
299 if (op == REQ_OP_DISCARD)
300 special_cmd_max_sectors = q->limits.max_discard_sectors;
301 else if (op == REQ_OP_WRITE_SAME)
302 special_cmd_max_sectors = q->limits.max_write_same_sectors;
303 if ((op == REQ_OP_DISCARD || op == REQ_OP_WRITE_SAME) &&
304 special_cmd_max_sectors == 0) {
305 dec_count(io, region, -EOPNOTSUPP);
310 * where->count may be zero if op holds a flush and we need to
311 * send a zero-sized flush.
315 * Allocate a suitably sized-bio.
317 if ((op == REQ_OP_DISCARD) || (op == REQ_OP_WRITE_SAME))
320 num_bvecs = min_t(int, BIO_MAX_PAGES,
321 dm_sector_div_up(remaining, (PAGE_SIZE >> SECTOR_SHIFT)));
323 bio = bio_alloc_bioset(GFP_NOIO, num_bvecs, io->client->bios);
324 bio->bi_iter.bi_sector = where->sector + (where->count - remaining);
325 bio->bi_bdev = where->bdev;
326 bio->bi_end_io = endio;
327 bio_set_op_attrs(bio, op, op_flags);
328 store_io_and_region_in_bio(bio, io, region);
330 if (op == REQ_OP_DISCARD) {
331 num_sectors = min_t(sector_t, special_cmd_max_sectors, remaining);
332 bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
333 remaining -= num_sectors;
334 } else if (op == REQ_OP_WRITE_SAME) {
336 * WRITE SAME only uses a single page.
338 dp->get_page(dp, &page, &len, &offset);
339 bio_add_page(bio, page, logical_block_size, offset);
340 num_sectors = min_t(sector_t, special_cmd_max_sectors, remaining);
341 bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
344 remaining -= num_sectors;
346 } else while (remaining) {
348 * Try and add as many pages as possible.
350 dp->get_page(dp, &page, &len, &offset);
351 len = min(len, to_bytes(remaining));
352 if (!bio_add_page(bio, page, len, offset))
356 remaining -= to_sector(len);
360 atomic_inc(&io->count);
365 static void dispatch_io(int op, int op_flags, unsigned int num_regions,
366 struct dm_io_region *where, struct dpages *dp,
367 struct io *io, int sync)
370 struct dpages old_pages = *dp;
372 BUG_ON(num_regions > DM_IO_MAX_REGIONS);
375 op_flags |= REQ_SYNC;
378 * For multiple regions we need to be careful to rewind
379 * the dp object for each call to do_region.
381 for (i = 0; i < num_regions; i++) {
383 if (where[i].count || (op_flags & REQ_PREFLUSH))
384 do_region(op, op_flags, i, where + i, dp, io);
388 * Drop the extra reference that we were holding to avoid
389 * the io being completed too early.
395 unsigned long error_bits;
396 struct completion wait;
399 static void sync_io_complete(unsigned long error, void *context)
401 struct sync_io *sio = context;
403 sio->error_bits = error;
404 complete(&sio->wait);
407 static int sync_io(struct dm_io_client *client, unsigned int num_regions,
408 struct dm_io_region *where, int op, int op_flags,
409 struct dpages *dp, unsigned long *error_bits)
414 if (num_regions > 1 && !op_is_write(op)) {
419 init_completion(&sio.wait);
421 io = mempool_alloc(client->pool, GFP_NOIO);
423 atomic_set(&io->count, 1); /* see dispatch_io() */
425 io->callback = sync_io_complete;
428 io->vma_invalidate_address = dp->vma_invalidate_address;
429 io->vma_invalidate_size = dp->vma_invalidate_size;
431 dispatch_io(op, op_flags, num_regions, where, dp, io, 1);
433 wait_for_completion_io(&sio.wait);
436 *error_bits = sio.error_bits;
438 return sio.error_bits ? -EIO : 0;
441 static int async_io(struct dm_io_client *client, unsigned int num_regions,
442 struct dm_io_region *where, int op, int op_flags,
443 struct dpages *dp, io_notify_fn fn, void *context)
447 if (num_regions > 1 && !op_is_write(op)) {
453 io = mempool_alloc(client->pool, GFP_NOIO);
455 atomic_set(&io->count, 1); /* see dispatch_io() */
458 io->context = context;
460 io->vma_invalidate_address = dp->vma_invalidate_address;
461 io->vma_invalidate_size = dp->vma_invalidate_size;
463 dispatch_io(op, op_flags, num_regions, where, dp, io, 0);
467 static int dp_init(struct dm_io_request *io_req, struct dpages *dp,
470 /* Set up dpages based on memory type */
472 dp->vma_invalidate_address = NULL;
473 dp->vma_invalidate_size = 0;
475 switch (io_req->mem.type) {
476 case DM_IO_PAGE_LIST:
477 list_dp_init(dp, io_req->mem.ptr.pl, io_req->mem.offset);
481 bio_dp_init(dp, io_req->mem.ptr.bio);
485 flush_kernel_vmap_range(io_req->mem.ptr.vma, size);
486 if (io_req->bi_op == REQ_OP_READ) {
487 dp->vma_invalidate_address = io_req->mem.ptr.vma;
488 dp->vma_invalidate_size = size;
490 vm_dp_init(dp, io_req->mem.ptr.vma);
494 km_dp_init(dp, io_req->mem.ptr.addr);
505 * New collapsed (a)synchronous interface.
507 * If the IO is asynchronous (i.e. it has notify.fn), you must either unplug
508 * the queue with blk_unplug() some time later or set REQ_SYNC in
509 * io_req->bi_opf. If you fail to do one of these, the IO will be submitted to
510 * the disk after q->unplug_delay, which defaults to 3ms in blk-settings.c.
512 int dm_io(struct dm_io_request *io_req, unsigned num_regions,
513 struct dm_io_region *where, unsigned long *sync_error_bits)
518 r = dp_init(io_req, &dp, (unsigned long)where->count << SECTOR_SHIFT);
522 if (!io_req->notify.fn)
523 return sync_io(io_req->client, num_regions, where,
524 io_req->bi_op, io_req->bi_op_flags, &dp,
527 return async_io(io_req->client, num_regions, where, io_req->bi_op,
528 io_req->bi_op_flags, &dp, io_req->notify.fn,
529 io_req->notify.context);
531 EXPORT_SYMBOL(dm_io);
533 int __init dm_io_init(void)
535 _dm_io_cache = KMEM_CACHE(io, 0);
542 void dm_io_exit(void)
544 kmem_cache_destroy(_dm_io_cache);