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/mempool.h>
14 #include <linux/module.h>
15 #include <linux/sched.h>
16 #include <linux/slab.h>
17 #include <linux/dm-io.h>
19 #define DM_MSG_PREFIX "io"
21 #define DM_IO_MAX_REGIONS BITS_PER_LONG
29 * Aligning 'struct io' reduces the number of bits required to store
30 * its address. Refer to store_io_and_region_in_bio() below.
33 unsigned long error_bits;
35 struct task_struct *sleeper;
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 dec_count(struct io *io, unsigned int region, int error)
117 set_bit(region, &io->error_bits);
119 if (atomic_dec_and_test(&io->count)) {
120 if (io->vma_invalidate_size)
121 invalidate_kernel_vmap_range(io->vma_invalidate_address,
122 io->vma_invalidate_size);
125 wake_up_process(io->sleeper);
128 unsigned long r = io->error_bits;
129 io_notify_fn fn = io->callback;
130 void *context = io->context;
132 mempool_free(io, io->client->pool);
138 static void endio(struct bio *bio, int error)
143 if (error && bio_data_dir(bio) == READ)
147 * The bio destructor in bio_put() may use the io object.
149 retrieve_io_and_region_from_bio(bio, &io, ®ion);
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 = ((unsigned long) data) & (PAGE_SIZE - 1);
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 = ((unsigned long) data) & (PAGE_SIZE - 1);
275 dp->context_ptr = data;
278 /*-----------------------------------------------------------------
279 * IO routines that accept a list of pages.
280 *---------------------------------------------------------------*/
281 static void do_region(int rw, unsigned region, struct dm_io_region *where,
282 struct dpages *dp, struct io *io)
289 sector_t remaining = where->count;
290 struct request_queue *q = bdev_get_queue(where->bdev);
291 unsigned short logical_block_size = queue_logical_block_size(q);
292 sector_t num_sectors;
295 * where->count may be zero if rw holds a flush and we need to
296 * send a zero-sized flush.
300 * Allocate a suitably sized-bio.
302 if ((rw & REQ_DISCARD) || (rw & REQ_WRITE_SAME))
305 num_bvecs = min_t(int, bio_get_nr_vecs(where->bdev),
306 dm_sector_div_up(remaining, (PAGE_SIZE >> SECTOR_SHIFT)));
308 bio = bio_alloc_bioset(GFP_NOIO, num_bvecs, io->client->bios);
309 bio->bi_iter.bi_sector = where->sector + (where->count - remaining);
310 bio->bi_bdev = where->bdev;
311 bio->bi_end_io = endio;
312 store_io_and_region_in_bio(bio, io, region);
314 if (rw & REQ_DISCARD) {
315 num_sectors = min_t(sector_t, q->limits.max_discard_sectors, remaining);
316 bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
317 remaining -= num_sectors;
318 } else if (rw & REQ_WRITE_SAME) {
320 * WRITE SAME only uses a single page.
322 dp->get_page(dp, &page, &len, &offset);
323 bio_add_page(bio, page, logical_block_size, offset);
324 num_sectors = min_t(sector_t, q->limits.max_write_same_sectors, remaining);
325 bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
328 remaining -= num_sectors;
330 } else while (remaining) {
332 * Try and add as many pages as possible.
334 dp->get_page(dp, &page, &len, &offset);
335 len = min(len, to_bytes(remaining));
336 if (!bio_add_page(bio, page, len, offset))
340 remaining -= to_sector(len);
344 atomic_inc(&io->count);
349 static void dispatch_io(int rw, unsigned int num_regions,
350 struct dm_io_region *where, struct dpages *dp,
351 struct io *io, int sync)
354 struct dpages old_pages = *dp;
356 BUG_ON(num_regions > DM_IO_MAX_REGIONS);
362 * For multiple regions we need to be careful to rewind
363 * the dp object for each call to do_region.
365 for (i = 0; i < num_regions; i++) {
367 if (where[i].count || (rw & REQ_FLUSH))
368 do_region(rw, i, where + i, dp, io);
372 * Drop the extra reference that we were holding to avoid
373 * the io being completed too early.
378 static int sync_io(struct dm_io_client *client, unsigned int num_regions,
379 struct dm_io_region *where, int rw, struct dpages *dp,
380 unsigned long *error_bits)
383 * gcc <= 4.3 can't do the alignment for stack variables, so we must
384 * align it on our own.
385 * volatile prevents the optimizer from removing or reusing
386 * "io_" field from the stack frame (allowed in ANSI C).
388 volatile char io_[sizeof(struct io) + __alignof__(struct io) - 1];
389 struct io *io = (struct io *)PTR_ALIGN(&io_, __alignof__(struct io));
391 if (num_regions > 1 && (rw & RW_MASK) != WRITE) {
397 atomic_set(&io->count, 1); /* see dispatch_io() */
398 io->sleeper = current;
401 io->vma_invalidate_address = dp->vma_invalidate_address;
402 io->vma_invalidate_size = dp->vma_invalidate_size;
404 dispatch_io(rw, num_regions, where, dp, io, 1);
407 set_current_state(TASK_UNINTERRUPTIBLE);
409 if (!atomic_read(&io->count))
414 set_current_state(TASK_RUNNING);
417 *error_bits = io->error_bits;
419 return io->error_bits ? -EIO : 0;
422 static int async_io(struct dm_io_client *client, unsigned int num_regions,
423 struct dm_io_region *where, int rw, struct dpages *dp,
424 io_notify_fn fn, void *context)
428 if (num_regions > 1 && (rw & RW_MASK) != WRITE) {
434 io = mempool_alloc(client->pool, GFP_NOIO);
436 atomic_set(&io->count, 1); /* see dispatch_io() */
440 io->context = context;
442 io->vma_invalidate_address = dp->vma_invalidate_address;
443 io->vma_invalidate_size = dp->vma_invalidate_size;
445 dispatch_io(rw, num_regions, where, dp, io, 0);
449 static int dp_init(struct dm_io_request *io_req, struct dpages *dp,
452 /* Set up dpages based on memory type */
454 dp->vma_invalidate_address = NULL;
455 dp->vma_invalidate_size = 0;
457 switch (io_req->mem.type) {
458 case DM_IO_PAGE_LIST:
459 list_dp_init(dp, io_req->mem.ptr.pl, io_req->mem.offset);
463 bio_dp_init(dp, io_req->mem.ptr.bio);
467 flush_kernel_vmap_range(io_req->mem.ptr.vma, size);
468 if ((io_req->bi_rw & RW_MASK) == READ) {
469 dp->vma_invalidate_address = io_req->mem.ptr.vma;
470 dp->vma_invalidate_size = size;
472 vm_dp_init(dp, io_req->mem.ptr.vma);
476 km_dp_init(dp, io_req->mem.ptr.addr);
487 * New collapsed (a)synchronous interface.
489 * If the IO is asynchronous (i.e. it has notify.fn), you must either unplug
490 * the queue with blk_unplug() some time later or set REQ_SYNC in
491 io_req->bi_rw. If you fail to do one of these, the IO will be submitted to
492 * the disk after q->unplug_delay, which defaults to 3ms in blk-settings.c.
494 int dm_io(struct dm_io_request *io_req, unsigned num_regions,
495 struct dm_io_region *where, unsigned long *sync_error_bits)
500 r = dp_init(io_req, &dp, (unsigned long)where->count << SECTOR_SHIFT);
504 if (!io_req->notify.fn)
505 return sync_io(io_req->client, num_regions, where,
506 io_req->bi_rw, &dp, sync_error_bits);
508 return async_io(io_req->client, num_regions, where, io_req->bi_rw,
509 &dp, io_req->notify.fn, io_req->notify.context);
511 EXPORT_SYMBOL(dm_io);
513 int __init dm_io_init(void)
515 _dm_io_cache = KMEM_CACHE(io, 0);
522 void dm_io_exit(void)
524 kmem_cache_destroy(_dm_io_cache);