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);
69 mempool_destroy(client->pool);
71 return ERR_PTR(-ENOMEM);
73 EXPORT_SYMBOL(dm_io_client_create);
75 void dm_io_client_destroy(struct dm_io_client *client)
77 mempool_destroy(client->pool);
78 bioset_free(client->bios);
81 EXPORT_SYMBOL(dm_io_client_destroy);
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
89 *---------------------------------------------------------------*/
90 static void store_io_and_region_in_bio(struct bio *bio, struct io *io,
93 if (unlikely(!IS_ALIGNED((unsigned long)io, DM_IO_MAX_REGIONS))) {
94 DMCRIT("Unaligned struct io pointer %p", io);
98 bio->bi_private = (void *)((unsigned long)io | region);
101 static void retrieve_io_and_region_from_bio(struct bio *bio, struct io **io,
104 unsigned long val = (unsigned long)bio->bi_private;
106 *io = (void *)(val & -(unsigned long)DM_IO_MAX_REGIONS);
107 *region = val & (DM_IO_MAX_REGIONS - 1);
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 *---------------------------------------------------------------*/
114 static void complete_io(struct io *io)
116 unsigned long error_bits = io->error_bits;
117 io_notify_fn fn = io->callback;
118 void *context = io->context;
120 if (io->vma_invalidate_size)
121 invalidate_kernel_vmap_range(io->vma_invalidate_address,
122 io->vma_invalidate_size);
124 mempool_free(io, io->client->pool);
125 fn(error_bits, context);
128 static void dec_count(struct io *io, unsigned int region, int error)
131 set_bit(region, &io->error_bits);
133 if (atomic_dec_and_test(&io->count))
137 static void endio(struct bio *bio, int error)
142 if (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);
152 dec_count(io, region, error);
155 /*-----------------------------------------------------------------
156 * These little objects provide an abstraction for getting a new
157 * destination page for io.
158 *---------------------------------------------------------------*/
160 void (*get_page)(struct dpages *dp,
161 struct page **p, unsigned long *len, unsigned *offset);
162 void (*next_page)(struct dpages *dp);
167 void *vma_invalidate_address;
168 unsigned long vma_invalidate_size;
172 * Functions for getting the pages from a list.
174 static void list_get_page(struct dpages *dp,
175 struct page **p, unsigned long *len, unsigned *offset)
177 unsigned o = dp->context_u;
178 struct page_list *pl = (struct page_list *) dp->context_ptr;
181 *len = PAGE_SIZE - o;
185 static void list_next_page(struct dpages *dp)
187 struct page_list *pl = (struct page_list *) dp->context_ptr;
188 dp->context_ptr = pl->next;
192 static void list_dp_init(struct dpages *dp, struct page_list *pl, unsigned offset)
194 dp->get_page = list_get_page;
195 dp->next_page = list_next_page;
196 dp->context_u = offset;
197 dp->context_ptr = pl;
201 * Functions for getting the pages from a bvec.
203 static void bio_get_page(struct dpages *dp, struct page **p,
204 unsigned long *len, unsigned *offset)
206 struct bio_vec *bvec = dp->context_ptr;
208 *len = bvec->bv_len - dp->context_u;
209 *offset = bvec->bv_offset + dp->context_u;
212 static void bio_next_page(struct dpages *dp)
214 struct bio_vec *bvec = dp->context_ptr;
215 dp->context_ptr = bvec + 1;
219 static void bio_dp_init(struct dpages *dp, struct bio *bio)
221 dp->get_page = bio_get_page;
222 dp->next_page = bio_next_page;
223 dp->context_ptr = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
224 dp->context_u = bio->bi_iter.bi_bvec_done;
228 * Functions for getting the pages from a VMA.
230 static void vm_get_page(struct dpages *dp,
231 struct page **p, unsigned long *len, unsigned *offset)
233 *p = vmalloc_to_page(dp->context_ptr);
234 *offset = dp->context_u;
235 *len = PAGE_SIZE - dp->context_u;
238 static void vm_next_page(struct dpages *dp)
240 dp->context_ptr += PAGE_SIZE - dp->context_u;
244 static void vm_dp_init(struct dpages *dp, void *data)
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;
253 * Functions for getting the pages from kernel memory.
255 static void km_get_page(struct dpages *dp, struct page **p, unsigned long *len,
258 *p = virt_to_page(dp->context_ptr);
259 *offset = dp->context_u;
260 *len = PAGE_SIZE - dp->context_u;
263 static void km_next_page(struct dpages *dp)
265 dp->context_ptr += PAGE_SIZE - dp->context_u;
269 static void km_dp_init(struct dpages *dp, void *data)
271 dp->get_page = km_get_page;
272 dp->next_page = km_next_page;
273 dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1);
274 dp->context_ptr = data;
277 /*-----------------------------------------------------------------
278 * IO routines that accept a list of pages.
279 *---------------------------------------------------------------*/
280 static void do_region(int rw, unsigned region, struct dm_io_region *where,
281 struct dpages *dp, struct io *io)
288 sector_t remaining = where->count;
289 struct request_queue *q = bdev_get_queue(where->bdev);
290 unsigned short logical_block_size = queue_logical_block_size(q);
291 sector_t num_sectors;
292 unsigned int uninitialized_var(special_cmd_max_sectors);
295 * Reject unsupported discard and write same requests.
297 if (rw & REQ_DISCARD)
298 special_cmd_max_sectors = q->limits.max_discard_sectors;
299 else if (rw & REQ_WRITE_SAME)
300 special_cmd_max_sectors = q->limits.max_write_same_sectors;
301 if ((rw & (REQ_DISCARD | REQ_WRITE_SAME)) && special_cmd_max_sectors == 0) {
302 dec_count(io, region, -EOPNOTSUPP);
307 * where->count may be zero if rw holds a flush and we need to
308 * send a zero-sized flush.
312 * Allocate a suitably sized-bio.
314 if ((rw & REQ_DISCARD) || (rw & REQ_WRITE_SAME))
317 num_bvecs = min_t(int, bio_get_nr_vecs(where->bdev),
318 dm_sector_div_up(remaining, (PAGE_SIZE >> SECTOR_SHIFT)));
320 bio = bio_alloc_bioset(GFP_NOIO, num_bvecs, io->client->bios);
321 bio->bi_iter.bi_sector = where->sector + (where->count - remaining);
322 bio->bi_bdev = where->bdev;
323 bio->bi_end_io = endio;
324 store_io_and_region_in_bio(bio, io, region);
326 if (rw & REQ_DISCARD) {
327 num_sectors = min_t(sector_t, special_cmd_max_sectors, remaining);
328 bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
329 remaining -= num_sectors;
330 } else if (rw & REQ_WRITE_SAME) {
332 * WRITE SAME only uses a single page.
334 dp->get_page(dp, &page, &len, &offset);
335 bio_add_page(bio, page, logical_block_size, offset);
336 num_sectors = min_t(sector_t, special_cmd_max_sectors, remaining);
337 bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
340 remaining -= num_sectors;
342 } else while (remaining) {
344 * Try and add as many pages as possible.
346 dp->get_page(dp, &page, &len, &offset);
347 len = min(len, to_bytes(remaining));
348 if (!bio_add_page(bio, page, len, offset))
352 remaining -= to_sector(len);
356 atomic_inc(&io->count);
361 static void dispatch_io(int rw, unsigned int num_regions,
362 struct dm_io_region *where, struct dpages *dp,
363 struct io *io, int sync)
366 struct dpages old_pages = *dp;
368 BUG_ON(num_regions > DM_IO_MAX_REGIONS);
374 * For multiple regions we need to be careful to rewind
375 * the dp object for each call to do_region.
377 for (i = 0; i < num_regions; i++) {
379 if (where[i].count || (rw & REQ_FLUSH))
380 do_region(rw, i, where + i, dp, io);
384 * Drop the extra reference that we were holding to avoid
385 * the io being completed too early.
391 unsigned long error_bits;
392 struct completion wait;
395 static void sync_io_complete(unsigned long error, void *context)
397 struct sync_io *sio = context;
399 sio->error_bits = error;
400 complete(&sio->wait);
403 static int sync_io(struct dm_io_client *client, unsigned int num_regions,
404 struct dm_io_region *where, int rw, struct dpages *dp,
405 unsigned long *error_bits)
410 if (num_regions > 1 && (rw & RW_MASK) != WRITE) {
415 init_completion(&sio.wait);
417 io = mempool_alloc(client->pool, GFP_NOIO);
419 atomic_set(&io->count, 1); /* see dispatch_io() */
421 io->callback = sync_io_complete;
424 io->vma_invalidate_address = dp->vma_invalidate_address;
425 io->vma_invalidate_size = dp->vma_invalidate_size;
427 dispatch_io(rw, num_regions, where, dp, io, 1);
429 wait_for_completion_io(&sio.wait);
432 *error_bits = sio.error_bits;
434 return sio.error_bits ? -EIO : 0;
437 static int async_io(struct dm_io_client *client, unsigned int num_regions,
438 struct dm_io_region *where, int rw, struct dpages *dp,
439 io_notify_fn fn, void *context)
443 if (num_regions > 1 && (rw & RW_MASK) != WRITE) {
449 io = mempool_alloc(client->pool, GFP_NOIO);
451 atomic_set(&io->count, 1); /* see dispatch_io() */
454 io->context = context;
456 io->vma_invalidate_address = dp->vma_invalidate_address;
457 io->vma_invalidate_size = dp->vma_invalidate_size;
459 dispatch_io(rw, num_regions, where, dp, io, 0);
463 static int dp_init(struct dm_io_request *io_req, struct dpages *dp,
466 /* Set up dpages based on memory type */
468 dp->vma_invalidate_address = NULL;
469 dp->vma_invalidate_size = 0;
471 switch (io_req->mem.type) {
472 case DM_IO_PAGE_LIST:
473 list_dp_init(dp, io_req->mem.ptr.pl, io_req->mem.offset);
477 bio_dp_init(dp, io_req->mem.ptr.bio);
481 flush_kernel_vmap_range(io_req->mem.ptr.vma, size);
482 if ((io_req->bi_rw & RW_MASK) == READ) {
483 dp->vma_invalidate_address = io_req->mem.ptr.vma;
484 dp->vma_invalidate_size = size;
486 vm_dp_init(dp, io_req->mem.ptr.vma);
490 km_dp_init(dp, io_req->mem.ptr.addr);
501 * New collapsed (a)synchronous interface.
503 * If the IO is asynchronous (i.e. it has notify.fn), you must either unplug
504 * the queue with blk_unplug() some time later or set REQ_SYNC in io_req->bi_rw.
505 * If you fail to do one of these, the IO will be submitted to the disk after
506 * q->unplug_delay, which defaults to 3ms in blk-settings.c.
508 int dm_io(struct dm_io_request *io_req, unsigned num_regions,
509 struct dm_io_region *where, unsigned long *sync_error_bits)
514 r = dp_init(io_req, &dp, (unsigned long)where->count << SECTOR_SHIFT);
518 if (!io_req->notify.fn)
519 return sync_io(io_req->client, num_regions, where,
520 io_req->bi_rw, &dp, sync_error_bits);
522 return async_io(io_req->client, num_regions, where, io_req->bi_rw,
523 &dp, io_req->notify.fn, io_req->notify.context);
525 EXPORT_SYMBOL(dm_io);
527 int __init dm_io_init(void)
529 _dm_io_cache = KMEM_CACHE(io, 0);
536 void dm_io_exit(void)
538 kmem_cache_destroy(_dm_io_cache);