1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * fs/direct-io.c
4 *
5 * Copyright (C) 2002, Linus Torvalds.
6 *
7 * O_DIRECT
8 *
9 * 04Jul2002 Andrew Morton
10 * Initial version
11 * 11Sep2002 janetinc@us.ibm.com
12 * added readv/writev support.
13 * 29Oct2002 Andrew Morton
14 * rewrote bio_add_page() support.
15 * 30Oct2002 pbadari@us.ibm.com
16 * added support for non-aligned IO.
17 * 06Nov2002 pbadari@us.ibm.com
18 * added asynchronous IO support.
19 * 21Jul2003 nathans@sgi.com
20 * added IO completion notifier.
21 */
22
23 #include <linux/kernel.h>
24 #include <linux/module.h>
25 #include <linux/types.h>
26 #include <linux/fs.h>
27 #include <linux/mm.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/pagemap.h>
31 #include <linux/task_io_accounting_ops.h>
32 #include <linux/bio.h>
33 #include <linux/wait.h>
34 #include <linux/err.h>
35 #include <linux/blkdev.h>
36 #include <linux/buffer_head.h>
37 #include <linux/rwsem.h>
38 #include <linux/uio.h>
39 #include <linux/atomic.h>
40 #include <linux/prefetch.h>
41
42 #include "internal.h"
43
44 /*
45 * How many user pages to map in one call to iov_iter_extract_pages(). This
46 * determines the size of a structure in the slab cache
47 */
48 #define DIO_PAGES 64
49
50 /*
51 * Flags for dio_complete()
52 */
53 #define DIO_COMPLETE_ASYNC 0x01 /* This is async IO */
54 #define DIO_COMPLETE_INVALIDATE 0x02 /* Can invalidate pages */
55
56 /*
57 * This code generally works in units of "dio_blocks". A dio_block is
58 * somewhere between the hard sector size and the filesystem block size. it
59 * is determined on a per-invocation basis. When talking to the filesystem
60 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
61 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
62 * to bio_block quantities by shifting left by blkfactor.
63 *
64 * If blkfactor is zero then the user's request was aligned to the filesystem's
65 * blocksize.
66 */
67
68 /* dio_state only used in the submission path */
69
70 struct dio_submit {
71 struct bio *bio; /* bio under assembly */
72 unsigned blkbits; /* doesn't change */
73 unsigned blkfactor; /* When we're using an alignment which
74 is finer than the filesystem's soft
75 blocksize, this specifies how much
76 finer. blkfactor=2 means 1/4-block
77 alignment. Does not change */
78 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
79 been performed at the start of a
80 write */
81 int pages_in_io; /* approximate total IO pages */
82 sector_t block_in_file; /* Current offset into the underlying
83 file in dio_block units. */
84 unsigned blocks_available; /* At block_in_file. changes */
85 int reap_counter; /* rate limit reaping */
86 sector_t final_block_in_request;/* doesn't change */
87 int boundary; /* prev block is at a boundary */
88 get_block_t *get_block; /* block mapping function */
89
90 loff_t logical_offset_in_bio; /* current first logical block in bio */
91 sector_t final_block_in_bio; /* current final block in bio + 1 */
92 sector_t next_block_for_io; /* next block to be put under IO,
93 in dio_blocks units */
94
95 /*
96 * Deferred addition of a page to the dio. These variables are
97 * private to dio_send_cur_page(), submit_page_section() and
98 * dio_bio_add_page().
99 */
100 struct page *cur_page; /* The page */
101 unsigned cur_page_offset; /* Offset into it, in bytes */
102 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
103 sector_t cur_page_block; /* Where it starts */
104 loff_t cur_page_fs_offset; /* Offset in file */
105
106 struct iov_iter *iter;
107 /*
108 * Page queue. These variables belong to dio_refill_pages() and
109 * dio_get_page().
110 */
111 unsigned head; /* next page to process */
112 unsigned tail; /* last valid page + 1 */
113 size_t from, to;
114 };
115
116 /* dio_state communicated between submission path and end_io */
117 struct dio {
118 int flags; /* doesn't change */
119 blk_opf_t opf; /* request operation type and flags */
120 struct gendisk *bio_disk;
121 struct inode *inode;
122 loff_t i_size; /* i_size when submitted */
123 dio_iodone_t *end_io; /* IO completion function */
124 bool is_pinned; /* T if we have pins on the pages */
125
126 void *private; /* copy from map_bh.b_private */
127
128 /* BIO completion state */
129 spinlock_t bio_lock; /* protects BIO fields below */
130 int page_errors; /* err from iov_iter_extract_pages() */
131 int is_async; /* is IO async ? */
132 bool defer_completion; /* defer AIO completion to workqueue? */
133 bool should_dirty; /* if pages should be dirtied */
134 int io_error; /* IO error in completion path */
135 unsigned long refcount; /* direct_io_worker() and bios */
136 struct bio *bio_list; /* singly linked via bi_private */
137 struct task_struct *waiter; /* waiting task (NULL if none) */
138
139 /* AIO related stuff */
140 struct kiocb *iocb; /* kiocb */
141 ssize_t result; /* IO result */
142
143 /*
144 * pages[] (and any fields placed after it) are not zeroed out at
145 * allocation time. Don't add new fields after pages[] unless you
146 * wish that they not be zeroed.
147 */
148 union {
149 struct page *pages[DIO_PAGES]; /* page buffer */
150 struct work_struct complete_work;/* deferred AIO completion */
151 };
152 } ____cacheline_aligned_in_smp;
153
154 static struct kmem_cache *dio_cache __ro_after_init;
155
156 /*
157 * How many pages are in the queue?
158 */
159 static inline unsigned dio_pages_present(struct dio_submit *sdio)
160 {
161 return sdio->tail - sdio->head;
162 }
163
164 /*
165 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
166 */
167 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
168 {
169 struct page **pages = dio->pages;
170 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
171 ssize_t ret;
172
173 ret = iov_iter_extract_pages(sdio->iter, &pages, LONG_MAX,
174 DIO_PAGES, 0, &sdio->from);
175
176 if (ret < 0 && sdio->blocks_available && dio_op == REQ_OP_WRITE) {
177 /*
178 * A memory fault, but the filesystem has some outstanding
179 * mapped blocks. We need to use those blocks up to avoid
180 * leaking stale data in the file.
181 */
182 if (dio->page_errors == 0)
183 dio->page_errors = ret;
184 dio->pages[0] = ZERO_PAGE(0);
185 sdio->head = 0;
186 sdio->tail = 1;
187 sdio->from = 0;
188 sdio->to = PAGE_SIZE;
189 return 0;
190 }
191
192 if (ret >= 0) {
193 ret += sdio->from;
194 sdio->head = 0;
195 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
196 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
197 return 0;
198 }
199 return ret;
200 }
201
202 /*
203 * Get another userspace page. Returns an ERR_PTR on error. Pages are
204 * buffered inside the dio so that we can call iov_iter_extract_pages()
205 * against a decent number of pages, less frequently. To provide nicer use of
206 * the L1 cache.
207 */
208 static inline struct page *dio_get_page(struct dio *dio,
209 struct dio_submit *sdio)
210 {
211 if (dio_pages_present(sdio) == 0) {
212 int ret;
213
214 ret = dio_refill_pages(dio, sdio);
215 if (ret)
216 return ERR_PTR(ret);
217 BUG_ON(dio_pages_present(sdio) == 0);
218 }
219 return dio->pages[sdio->head];
220 }
221
222 static void dio_pin_page(struct dio *dio, struct page *page)
223 {
224 if (dio->is_pinned)
225 folio_add_pin(page_folio(page));
226 }
227
228 static void dio_unpin_page(struct dio *dio, struct page *page)
229 {
230 if (dio->is_pinned)
231 unpin_user_page(page);
232 }
233
234 /*
235 * dio_complete() - called when all DIO BIO I/O has been completed
236 *
237 * This drops i_dio_count, lets interested parties know that a DIO operation
238 * has completed, and calculates the resulting return code for the operation.
239 *
240 * It lets the filesystem know if it registered an interest earlier via
241 * get_block. Pass the private field of the map buffer_head so that
242 * filesystems can use it to hold additional state between get_block calls and
243 * dio_complete.
244 */
245 static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags)
246 {
247 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
248 loff_t offset = dio->iocb->ki_pos;
249 ssize_t transferred = 0;
250 int err;
251
252 /*
253 * AIO submission can race with bio completion to get here while
254 * expecting to have the last io completed by bio completion.
255 * In that case -EIOCBQUEUED is in fact not an error we want
256 * to preserve through this call.
257 */
258 if (ret == -EIOCBQUEUED)
259 ret = 0;
260
261 if (dio->result) {
262 transferred = dio->result;
263
264 /* Check for short read case */
265 if (dio_op == REQ_OP_READ &&
266 ((offset + transferred) > dio->i_size))
267 transferred = dio->i_size - offset;
268 /* ignore EFAULT if some IO has been done */
269 if (unlikely(ret == -EFAULT) && transferred)
270 ret = 0;
271 }
272
273 if (ret == 0)
274 ret = dio->page_errors;
275 if (ret == 0)
276 ret = dio->io_error;
277 if (ret == 0)
278 ret = transferred;
279
280 if (dio->end_io) {
281 // XXX: ki_pos??
282 err = dio->end_io(dio->iocb, offset, ret, dio->private);
283 if (err)
284 ret = err;
285 }
286
287 /*
288 * Try again to invalidate clean pages which might have been cached by
289 * non-direct readahead, or faulted in by get_user_pages() if the source
290 * of the write was an mmap'ed region of the file we're writing. Either
291 * one is a pretty crazy thing to do, so we don't support it 100%. If
292 * this invalidation fails, tough, the write still worked...
293 *
294 * And this page cache invalidation has to be after dio->end_io(), as
295 * some filesystems convert unwritten extents to real allocations in
296 * end_io() when necessary, otherwise a racing buffer read would cache
297 * zeros from unwritten extents.
298 */
299 if (flags & DIO_COMPLETE_INVALIDATE &&
300 ret > 0 && dio_op == REQ_OP_WRITE)
301 kiocb_invalidate_post_direct_write(dio->iocb, ret);
302
303 inode_dio_end(dio->inode);
304
305 if (flags & DIO_COMPLETE_ASYNC) {
306 /*
307 * generic_write_sync expects ki_pos to have been updated
308 * already, but the submission path only does this for
309 * synchronous I/O.
310 */
311 dio->iocb->ki_pos += transferred;
312
313 if (ret > 0 && dio_op == REQ_OP_WRITE)
314 ret = generic_write_sync(dio->iocb, ret);
315 dio->iocb->ki_complete(dio->iocb, ret);
316 }
317
318 kmem_cache_free(dio_cache, dio);
319 return ret;
320 }
321
322 static void dio_aio_complete_work(struct work_struct *work)
323 {
324 struct dio *dio = container_of(work, struct dio, complete_work);
325
326 dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
327 }
328
329 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
330
331 /*
332 * Asynchronous IO callback.
333 */
334 static void dio_bio_end_aio(struct bio *bio)
335 {
336 struct dio *dio = bio->bi_private;
337 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
338 unsigned long remaining;
339 unsigned long flags;
340 bool defer_completion = false;
341
342 /* cleanup the bio */
343 dio_bio_complete(dio, bio);
344
345 spin_lock_irqsave(&dio->bio_lock, flags);
346 remaining = --dio->refcount;
347 if (remaining == 1 && dio->waiter)
348 wake_up_process(dio->waiter);
349 spin_unlock_irqrestore(&dio->bio_lock, flags);
350
351 if (remaining == 0) {
352 /*
353 * Defer completion when defer_completion is set or
354 * when the inode has pages mapped and this is AIO write.
355 * We need to invalidate those pages because there is a
356 * chance they contain stale data in the case buffered IO
357 * went in between AIO submission and completion into the
358 * same region.
359 */
360 if (dio->result)
361 defer_completion = dio->defer_completion ||
362 (dio_op == REQ_OP_WRITE &&
363 dio->inode->i_mapping->nrpages);
364 if (defer_completion) {
365 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
366 queue_work(dio->inode->i_sb->s_dio_done_wq,
367 &dio->complete_work);
368 } else {
369 dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
370 }
371 }
372 }
373
374 /*
375 * The BIO completion handler simply queues the BIO up for the process-context
376 * handler.
377 *
378 * During I/O bi_private points at the dio. After I/O, bi_private is used to
379 * implement a singly-linked list of completed BIOs, at dio->bio_list.
380 */
381 static void dio_bio_end_io(struct bio *bio)
382 {
383 struct dio *dio = bio->bi_private;
384 unsigned long flags;
385
386 spin_lock_irqsave(&dio->bio_lock, flags);
387 bio->bi_private = dio->bio_list;
388 dio->bio_list = bio;
389 if (--dio->refcount == 1 && dio->waiter)
390 wake_up_process(dio->waiter);
391 spin_unlock_irqrestore(&dio->bio_lock, flags);
392 }
393
394 static inline void
395 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
396 struct block_device *bdev,
397 sector_t first_sector, int nr_vecs)
398 {
399 struct bio *bio;
400
401 /*
402 * bio_alloc() is guaranteed to return a bio when allowed to sleep and
403 * we request a valid number of vectors.
404 */
405 bio = bio_alloc(bdev, nr_vecs, dio->opf, GFP_KERNEL);
406 bio->bi_iter.bi_sector = first_sector;
407 if (dio->is_async)
408 bio->bi_end_io = dio_bio_end_aio;
409 else
410 bio->bi_end_io = dio_bio_end_io;
411 if (dio->is_pinned)
412 bio_set_flag(bio, BIO_PAGE_PINNED);
413 bio->bi_write_hint = file_inode(dio->iocb->ki_filp)->i_write_hint;
414
415 sdio->bio = bio;
416 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
417 }
418
419 /*
420 * In the AIO read case we speculatively dirty the pages before starting IO.
421 * During IO completion, any of these pages which happen to have been written
422 * back will be redirtied by bio_check_pages_dirty().
423 *
424 * bios hold a dio reference between submit_bio and ->end_io.
425 */
426 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
427 {
428 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
429 struct bio *bio = sdio->bio;
430 unsigned long flags;
431
432 bio->bi_private = dio;
433
434 spin_lock_irqsave(&dio->bio_lock, flags);
435 dio->refcount++;
436 spin_unlock_irqrestore(&dio->bio_lock, flags);
437
438 if (dio->is_async && dio_op == REQ_OP_READ && dio->should_dirty)
439 bio_set_pages_dirty(bio);
440
441 dio->bio_disk = bio->bi_bdev->bd_disk;
442
443 submit_bio(bio);
444
445 sdio->bio = NULL;
446 sdio->boundary = 0;
447 sdio->logical_offset_in_bio = 0;
448 }
449
450 /*
451 * Release any resources in case of a failure
452 */
453 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
454 {
455 if (dio->is_pinned)
456 unpin_user_pages(dio->pages + sdio->head,
457 sdio->tail - sdio->head);
458 sdio->head = sdio->tail;
459 }
460
461 /*
462 * Wait for the next BIO to complete. Remove it and return it. NULL is
463 * returned once all BIOs have been completed. This must only be called once
464 * all bios have been issued so that dio->refcount can only decrease. This
465 * requires that the caller hold a reference on the dio.
466 */
467 static struct bio *dio_await_one(struct dio *dio)
468 {
469 unsigned long flags;
470 struct bio *bio = NULL;
471
472 spin_lock_irqsave(&dio->bio_lock, flags);
473
474 /*
475 * Wait as long as the list is empty and there are bios in flight. bio
476 * completion drops the count, maybe adds to the list, and wakes while
477 * holding the bio_lock so we don't need set_current_state()'s barrier
478 * and can call it after testing our condition.
479 */
480 while (dio->refcount > 1 && dio->bio_list == NULL) {
481 __set_current_state(TASK_UNINTERRUPTIBLE);
482 dio->waiter = current;
483 spin_unlock_irqrestore(&dio->bio_lock, flags);
484 blk_io_schedule();
485 /* wake up sets us TASK_RUNNING */
486 spin_lock_irqsave(&dio->bio_lock, flags);
487 dio->waiter = NULL;
488 }
489 if (dio->bio_list) {
490 bio = dio->bio_list;
491 dio->bio_list = bio->bi_private;
492 }
493 spin_unlock_irqrestore(&dio->bio_lock, flags);
494 return bio;
495 }
496
497 /*
498 * Process one completed BIO. No locks are held.
499 */
500 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
501 {
502 blk_status_t err = bio->bi_status;
503 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
504 bool should_dirty = dio_op == REQ_OP_READ && dio->should_dirty;
505
506 if (err) {
507 if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
508 dio->io_error = -EAGAIN;
509 else
510 dio->io_error = -EIO;
511 }
512
513 if (dio->is_async && should_dirty) {
514 bio_check_pages_dirty(bio); /* transfers ownership */
515 } else {
516 bio_release_pages(bio, should_dirty);
517 bio_put(bio);
518 }
519 return err;
520 }
521
522 /*
523 * Wait on and process all in-flight BIOs. This must only be called once
524 * all bios have been issued so that the refcount can only decrease.
525 * This just waits for all bios to make it through dio_bio_complete. IO
526 * errors are propagated through dio->io_error and should be propagated via
527 * dio_complete().
528 */
529 static void dio_await_completion(struct dio *dio)
530 {
531 struct bio *bio;
532 do {
533 bio = dio_await_one(dio);
534 if (bio)
535 dio_bio_complete(dio, bio);
536 } while (bio);
537 }
538
539 /*
540 * A really large O_DIRECT read or write can generate a lot of BIOs. So
541 * to keep the memory consumption sane we periodically reap any completed BIOs
542 * during the BIO generation phase.
543 *
544 * This also helps to limit the peak amount of pinned userspace memory.
545 */
546 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
547 {
548 int ret = 0;
549
550 if (sdio->reap_counter++ >= 64) {
551 while (dio->bio_list) {
552 unsigned long flags;
553 struct bio *bio;
554 int ret2;
555
556 spin_lock_irqsave(&dio->bio_lock, flags);
557 bio = dio->bio_list;
558 dio->bio_list = bio->bi_private;
559 spin_unlock_irqrestore(&dio->bio_lock, flags);
560 ret2 = blk_status_to_errno(dio_bio_complete(dio, bio));
561 if (ret == 0)
562 ret = ret2;
563 }
564 sdio->reap_counter = 0;
565 }
566 return ret;
567 }
568
569 static int dio_set_defer_completion(struct dio *dio)
570 {
571 struct super_block *sb = dio->inode->i_sb;
572
573 if (dio->defer_completion)
574 return 0;
575 dio->defer_completion = true;
576 if (!sb->s_dio_done_wq)
577 return sb_init_dio_done_wq(sb);
578 return 0;
579 }
580
581 /*
582 * Call into the fs to map some more disk blocks. We record the current number
583 * of available blocks at sdio->blocks_available. These are in units of the
584 * fs blocksize, i_blocksize(inode).
585 *
586 * The fs is allowed to map lots of blocks at once. If it wants to do that,
587 * it uses the passed inode-relative block number as the file offset, as usual.
588 *
589 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
590 * has remaining to do. The fs should not map more than this number of blocks.
591 *
592 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
593 * indicate how much contiguous disk space has been made available at
594 * bh->b_blocknr.
595 *
596 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
597 * This isn't very efficient...
598 *
599 * In the case of filesystem holes: the fs may return an arbitrarily-large
600 * hole by returning an appropriate value in b_size and by clearing
601 * buffer_mapped(). However the direct-io code will only process holes one
602 * block at a time - it will repeatedly call get_block() as it walks the hole.
603 */
604 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
605 struct buffer_head *map_bh)
606 {
607 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
608 int ret;
609 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
610 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
611 unsigned long fs_count; /* Number of filesystem-sized blocks */
612 int create;
613 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
614 loff_t i_size;
615
616 /*
617 * If there was a memory error and we've overwritten all the
618 * mapped blocks then we can now return that memory error
619 */
620 ret = dio->page_errors;
621 if (ret == 0) {
622 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
623 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
624 fs_endblk = (sdio->final_block_in_request - 1) >>
625 sdio->blkfactor;
626 fs_count = fs_endblk - fs_startblk + 1;
627
628 map_bh->b_state = 0;
629 map_bh->b_size = fs_count << i_blkbits;
630
631 /*
632 * For writes that could fill holes inside i_size on a
633 * DIO_SKIP_HOLES filesystem we forbid block creations: only
634 * overwrites are permitted. We will return early to the caller
635 * once we see an unmapped buffer head returned, and the caller
636 * will fall back to buffered I/O.
637 *
638 * Otherwise the decision is left to the get_blocks method,
639 * which may decide to handle it or also return an unmapped
640 * buffer head.
641 */
642 create = dio_op == REQ_OP_WRITE;
643 if (dio->flags & DIO_SKIP_HOLES) {
644 i_size = i_size_read(dio->inode);
645 if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits)
646 create = 0;
647 }
648
649 ret = (*sdio->get_block)(dio->inode, fs_startblk,
650 map_bh, create);
651
652 /* Store for completion */
653 dio->private = map_bh->b_private;
654
655 if (ret == 0 && buffer_defer_completion(map_bh))
656 ret = dio_set_defer_completion(dio);
657 }
658 return ret;
659 }
660
661 /*
662 * There is no bio. Make one now.
663 */
664 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
665 sector_t start_sector, struct buffer_head *map_bh)
666 {
667 sector_t sector;
668 int ret, nr_pages;
669
670 ret = dio_bio_reap(dio, sdio);
671 if (ret)
672 goto out;
673 sector = start_sector << (sdio->blkbits - 9);
674 nr_pages = bio_max_segs(sdio->pages_in_io);
675 BUG_ON(nr_pages <= 0);
676 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
677 sdio->boundary = 0;
678 out:
679 return ret;
680 }
681
682 /*
683 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
684 * that was successful then update final_block_in_bio and take a ref against
685 * the just-added page.
686 *
687 * Return zero on success. Non-zero means the caller needs to start a new BIO.
688 */
689 static inline int dio_bio_add_page(struct dio *dio, struct dio_submit *sdio)
690 {
691 int ret;
692
693 ret = bio_add_page(sdio->bio, sdio->cur_page,
694 sdio->cur_page_len, sdio->cur_page_offset);
695 if (ret == sdio->cur_page_len) {
696 /*
697 * Decrement count only, if we are done with this page
698 */
699 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
700 sdio->pages_in_io--;
701 dio_pin_page(dio, sdio->cur_page);
702 sdio->final_block_in_bio = sdio->cur_page_block +
703 (sdio->cur_page_len >> sdio->blkbits);
704 ret = 0;
705 } else {
706 ret = 1;
707 }
708 return ret;
709 }
710
711 /*
712 * Put cur_page under IO. The section of cur_page which is described by
713 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
714 * starts on-disk at cur_page_block.
715 *
716 * We take a ref against the page here (on behalf of its presence in the bio).
717 *
718 * The caller of this function is responsible for removing cur_page from the
719 * dio, and for dropping the refcount which came from that presence.
720 */
721 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
722 struct buffer_head *map_bh)
723 {
724 int ret = 0;
725
726 if (sdio->bio) {
727 loff_t cur_offset = sdio->cur_page_fs_offset;
728 loff_t bio_next_offset = sdio->logical_offset_in_bio +
729 sdio->bio->bi_iter.bi_size;
730
731 /*
732 * See whether this new request is contiguous with the old.
733 *
734 * Btrfs cannot handle having logically non-contiguous requests
735 * submitted. For example if you have
736 *
737 * Logical: [0-4095][HOLE][8192-12287]
738 * Physical: [0-4095] [4096-8191]
739 *
740 * We cannot submit those pages together as one BIO. So if our
741 * current logical offset in the file does not equal what would
742 * be the next logical offset in the bio, submit the bio we
743 * have.
744 */
745 if (sdio->final_block_in_bio != sdio->cur_page_block ||
746 cur_offset != bio_next_offset)
747 dio_bio_submit(dio, sdio);
748 }
749
750 if (sdio->bio == NULL) {
751 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
752 if (ret)
753 goto out;
754 }
755
756 if (dio_bio_add_page(dio, sdio) != 0) {
757 dio_bio_submit(dio, sdio);
758 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
759 if (ret == 0) {
760 ret = dio_bio_add_page(dio, sdio);
761 BUG_ON(ret != 0);
762 }
763 }
764 out:
765 return ret;
766 }
767
768 /*
769 * An autonomous function to put a chunk of a page under deferred IO.
770 *
771 * The caller doesn't actually know (or care) whether this piece of page is in
772 * a BIO, or is under IO or whatever. We just take care of all possible
773 * situations here. The separation between the logic of do_direct_IO() and
774 * that of submit_page_section() is important for clarity. Please don't break.
775 *
776 * The chunk of page starts on-disk at blocknr.
777 *
778 * We perform deferred IO, by recording the last-submitted page inside our
779 * private part of the dio structure. If possible, we just expand the IO
780 * across that page here.
781 *
782 * If that doesn't work out then we put the old page into the bio and add this
783 * page to the dio instead.
784 */
785 static inline int
786 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
787 unsigned offset, unsigned len, sector_t blocknr,
788 struct buffer_head *map_bh)
789 {
790 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
791 int ret = 0;
792 int boundary = sdio->boundary; /* dio_send_cur_page may clear it */
793
794 if (dio_op == REQ_OP_WRITE) {
795 /*
796 * Read accounting is performed in submit_bio()
797 */
798 task_io_account_write(len);
799 }
800
801 /*
802 * Can we just grow the current page's presence in the dio?
803 */
804 if (sdio->cur_page == page &&
805 sdio->cur_page_offset + sdio->cur_page_len == offset &&
806 sdio->cur_page_block +
807 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
808 sdio->cur_page_len += len;
809 goto out;
810 }
811
812 /*
813 * If there's a deferred page already there then send it.
814 */
815 if (sdio->cur_page) {
816 ret = dio_send_cur_page(dio, sdio, map_bh);
817 dio_unpin_page(dio, sdio->cur_page);
818 sdio->cur_page = NULL;
819 if (ret)
820 return ret;
821 }
822
823 dio_pin_page(dio, page); /* It is in dio */
824 sdio->cur_page = page;
825 sdio->cur_page_offset = offset;
826 sdio->cur_page_len = len;
827 sdio->cur_page_block = blocknr;
828 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
829 out:
830 /*
831 * If boundary then we want to schedule the IO now to
832 * avoid metadata seeks.
833 */
834 if (boundary) {
835 ret = dio_send_cur_page(dio, sdio, map_bh);
836 if (sdio->bio)
837 dio_bio_submit(dio, sdio);
838 dio_unpin_page(dio, sdio->cur_page);
839 sdio->cur_page = NULL;
840 }
841 return ret;
842 }
843
844 /*
845 * If we are not writing the entire block and get_block() allocated
846 * the block for us, we need to fill-in the unused portion of the
847 * block with zeros. This happens only if user-buffer, fileoffset or
848 * io length is not filesystem block-size multiple.
849 *
850 * `end' is zero if we're doing the start of the IO, 1 at the end of the
851 * IO.
852 */
853 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
854 int end, struct buffer_head *map_bh)
855 {
856 unsigned dio_blocks_per_fs_block;
857 unsigned this_chunk_blocks; /* In dio_blocks */
858 unsigned this_chunk_bytes;
859 struct page *page;
860
861 sdio->start_zero_done = 1;
862 if (!sdio->blkfactor || !buffer_new(map_bh))
863 return;
864
865 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
866 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
867
868 if (!this_chunk_blocks)
869 return;
870
871 /*
872 * We need to zero out part of an fs block. It is either at the
873 * beginning or the end of the fs block.
874 */
875 if (end)
876 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
877
878 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
879
880 page = ZERO_PAGE(0);
881 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
882 sdio->next_block_for_io, map_bh))
883 return;
884
885 sdio->next_block_for_io += this_chunk_blocks;
886 }
887
888 /*
889 * Walk the user pages, and the file, mapping blocks to disk and generating
890 * a sequence of (page,offset,len,block) mappings. These mappings are injected
891 * into submit_page_section(), which takes care of the next stage of submission
892 *
893 * Direct IO against a blockdev is different from a file. Because we can
894 * happily perform page-sized but 512-byte aligned IOs. It is important that
895 * blockdev IO be able to have fine alignment and large sizes.
896 *
897 * So what we do is to permit the ->get_block function to populate bh.b_size
898 * with the size of IO which is permitted at this offset and this i_blkbits.
899 *
900 * For best results, the blockdev should be set up with 512-byte i_blkbits and
901 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
902 * fine alignment but still allows this function to work in PAGE_SIZE units.
903 */
904 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
905 struct buffer_head *map_bh)
906 {
907 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
908 const unsigned blkbits = sdio->blkbits;
909 const unsigned i_blkbits = blkbits + sdio->blkfactor;
910 int ret = 0;
911
912 while (sdio->block_in_file < sdio->final_block_in_request) {
913 struct page *page;
914 size_t from, to;
915
916 page = dio_get_page(dio, sdio);
917 if (IS_ERR(page)) {
918 ret = PTR_ERR(page);
919 goto out;
920 }
921 from = sdio->head ? 0 : sdio->from;
922 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
923 sdio->head++;
924
925 while (from < to) {
926 unsigned this_chunk_bytes; /* # of bytes mapped */
927 unsigned this_chunk_blocks; /* # of blocks */
928 unsigned u;
929
930 if (sdio->blocks_available == 0) {
931 /*
932 * Need to go and map some more disk
933 */
934 unsigned long blkmask;
935 unsigned long dio_remainder;
936
937 ret = get_more_blocks(dio, sdio, map_bh);
938 if (ret) {
939 dio_unpin_page(dio, page);
940 goto out;
941 }
942 if (!buffer_mapped(map_bh))
943 goto do_holes;
944
945 sdio->blocks_available =
946 map_bh->b_size >> blkbits;
947 sdio->next_block_for_io =
948 map_bh->b_blocknr << sdio->blkfactor;
949 if (buffer_new(map_bh)) {
950 clean_bdev_aliases(
951 map_bh->b_bdev,
952 map_bh->b_blocknr,
953 map_bh->b_size >> i_blkbits);
954 }
955
956 if (!sdio->blkfactor)
957 goto do_holes;
958
959 blkmask = (1 << sdio->blkfactor) - 1;
960 dio_remainder = (sdio->block_in_file & blkmask);
961
962 /*
963 * If we are at the start of IO and that IO
964 * starts partway into a fs-block,
965 * dio_remainder will be non-zero. If the IO
966 * is a read then we can simply advance the IO
967 * cursor to the first block which is to be
968 * read. But if the IO is a write and the
969 * block was newly allocated we cannot do that;
970 * the start of the fs block must be zeroed out
971 * on-disk
972 */
973 if (!buffer_new(map_bh))
974 sdio->next_block_for_io += dio_remainder;
975 sdio->blocks_available -= dio_remainder;
976 }
977 do_holes:
978 /* Handle holes */
979 if (!buffer_mapped(map_bh)) {
980 loff_t i_size_aligned;
981
982 /* AKPM: eargh, -ENOTBLK is a hack */
983 if (dio_op == REQ_OP_WRITE) {
984 dio_unpin_page(dio, page);
985 return -ENOTBLK;
986 }
987
988 /*
989 * Be sure to account for a partial block as the
990 * last block in the file
991 */
992 i_size_aligned = ALIGN(i_size_read(dio->inode),
993 1 << blkbits);
994 if (sdio->block_in_file >=
995 i_size_aligned >> blkbits) {
996 /* We hit eof */
997 dio_unpin_page(dio, page);
998 goto out;
999 }
1000 zero_user(page, from, 1 << blkbits);
1001 sdio->block_in_file++;
1002 from += 1 << blkbits;
1003 dio->result += 1 << blkbits;
1004 goto next_block;
1005 }
1006
1007 /*
1008 * If we're performing IO which has an alignment which
1009 * is finer than the underlying fs, go check to see if
1010 * we must zero out the start of this block.
1011 */
1012 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1013 dio_zero_block(dio, sdio, 0, map_bh);
1014
1015 /*
1016 * Work out, in this_chunk_blocks, how much disk we
1017 * can add to this page
1018 */
1019 this_chunk_blocks = sdio->blocks_available;
1020 u = (to - from) >> blkbits;
1021 if (this_chunk_blocks > u)
1022 this_chunk_blocks = u;
1023 u = sdio->final_block_in_request - sdio->block_in_file;
1024 if (this_chunk_blocks > u)
1025 this_chunk_blocks = u;
1026 this_chunk_bytes = this_chunk_blocks << blkbits;
1027 BUG_ON(this_chunk_bytes == 0);
1028
1029 if (this_chunk_blocks == sdio->blocks_available)
1030 sdio->boundary = buffer_boundary(map_bh);
1031 ret = submit_page_section(dio, sdio, page,
1032 from,
1033 this_chunk_bytes,
1034 sdio->next_block_for_io,
1035 map_bh);
1036 if (ret) {
1037 dio_unpin_page(dio, page);
1038 goto out;
1039 }
1040 sdio->next_block_for_io += this_chunk_blocks;
1041
1042 sdio->block_in_file += this_chunk_blocks;
1043 from += this_chunk_bytes;
1044 dio->result += this_chunk_bytes;
1045 sdio->blocks_available -= this_chunk_blocks;
1046 next_block:
1047 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1048 if (sdio->block_in_file == sdio->final_block_in_request)
1049 break;
1050 }
1051
1052 /* Drop the pin which was taken in get_user_pages() */
1053 dio_unpin_page(dio, page);
1054 }
1055 out:
1056 return ret;
1057 }
1058
1059 static inline int drop_refcount(struct dio *dio)
1060 {
1061 int ret2;
1062 unsigned long flags;
1063
1064 /*
1065 * Sync will always be dropping the final ref and completing the
1066 * operation. AIO can if it was a broken operation described above or
1067 * in fact if all the bios race to complete before we get here. In
1068 * that case dio_complete() translates the EIOCBQUEUED into the proper
1069 * return code that the caller will hand to ->complete().
1070 *
1071 * This is managed by the bio_lock instead of being an atomic_t so that
1072 * completion paths can drop their ref and use the remaining count to
1073 * decide to wake the submission path atomically.
1074 */
1075 spin_lock_irqsave(&dio->bio_lock, flags);
1076 ret2 = --dio->refcount;
1077 spin_unlock_irqrestore(&dio->bio_lock, flags);
1078 return ret2;
1079 }
1080
1081 /*
1082 * This is a library function for use by filesystem drivers.
1083 *
1084 * The locking rules are governed by the flags parameter:
1085 * - if the flags value contains DIO_LOCKING we use a fancy locking
1086 * scheme for dumb filesystems.
1087 * For writes this function is called under i_mutex and returns with
1088 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1089 * taken and dropped again before returning.
1090 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1091 * internal locking but rather rely on the filesystem to synchronize
1092 * direct I/O reads/writes versus each other and truncate.
1093 *
1094 * To help with locking against truncate we incremented the i_dio_count
1095 * counter before starting direct I/O, and decrement it once we are done.
1096 * Truncate can wait for it to reach zero to provide exclusion. It is
1097 * expected that filesystem provide exclusion between new direct I/O
1098 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1099 * but other filesystems need to take care of this on their own.
1100 *
1101 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1102 * is always inlined. Otherwise gcc is unable to split the structure into
1103 * individual fields and will generate much worse code. This is important
1104 * for the whole file.
1105 */
1106 ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1107 struct block_device *bdev, struct iov_iter *iter,
1108 get_block_t get_block, dio_iodone_t end_io,
1109 int flags)
1110 {
1111 unsigned i_blkbits = READ_ONCE(inode->i_blkbits);
1112 unsigned blkbits = i_blkbits;
1113 unsigned blocksize_mask = (1 << blkbits) - 1;
1114 ssize_t retval = -EINVAL;
1115 const size_t count = iov_iter_count(iter);
1116 loff_t offset = iocb->ki_pos;
1117 const loff_t end = offset + count;
1118 struct dio *dio;
1119 struct dio_submit sdio = { NULL, };
1120 struct buffer_head map_bh = { 0, };
1121 struct blk_plug plug;
1122 unsigned long align = offset | iov_iter_alignment(iter);
1123
1124 /*
1125 * Avoid references to bdev if not absolutely needed to give
1126 * the early prefetch in the caller enough time.
1127 */
1128
1129 /* watch out for a 0 len io from a tricksy fs */
1130 if (iov_iter_rw(iter) == READ && !count)
1131 return 0;
1132
1133 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1134 if (!dio)
1135 return -ENOMEM;
1136 /*
1137 * Believe it or not, zeroing out the page array caused a .5%
1138 * performance regression in a database benchmark. So, we take
1139 * care to only zero out what's needed.
1140 */
1141 memset(dio, 0, offsetof(struct dio, pages));
1142
1143 dio->flags = flags;
1144 if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) {
1145 /* will be released by direct_io_worker */
1146 inode_lock(inode);
1147 }
1148 dio->is_pinned = iov_iter_extract_will_pin(iter);
1149
1150 /* Once we sampled i_size check for reads beyond EOF */
1151 dio->i_size = i_size_read(inode);
1152 if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1153 retval = 0;
1154 goto fail_dio;
1155 }
1156
1157 if (align & blocksize_mask) {
1158 if (bdev)
1159 blkbits = blksize_bits(bdev_logical_block_size(bdev));
1160 blocksize_mask = (1 << blkbits) - 1;
1161 if (align & blocksize_mask)
1162 goto fail_dio;
1163 }
1164
1165 if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) {
1166 struct address_space *mapping = iocb->ki_filp->f_mapping;
1167
1168 retval = filemap_write_and_wait_range(mapping, offset, end - 1);
1169 if (retval)
1170 goto fail_dio;
1171 }
1172
1173 /*
1174 * For file extending writes updating i_size before data writeouts
1175 * complete can expose uninitialized blocks in dumb filesystems.
1176 * In that case we need to wait for I/O completion even if asked
1177 * for an asynchronous write.
1178 */
1179 if (is_sync_kiocb(iocb))
1180 dio->is_async = false;
1181 else if (iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
1182 dio->is_async = false;
1183 else
1184 dio->is_async = true;
1185
1186 dio->inode = inode;
1187 if (iov_iter_rw(iter) == WRITE) {
1188 dio->opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
1189 if (iocb->ki_flags & IOCB_NOWAIT)
1190 dio->opf |= REQ_NOWAIT;
1191 } else {
1192 dio->opf = REQ_OP_READ;
1193 }
1194
1195 /*
1196 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1197 * so that we can call ->fsync.
1198 */
1199 if (dio->is_async && iov_iter_rw(iter) == WRITE) {
1200 retval = 0;
1201 if (iocb_is_dsync(iocb))
1202 retval = dio_set_defer_completion(dio);
1203 else if (!dio->inode->i_sb->s_dio_done_wq) {
1204 /*
1205 * In case of AIO write racing with buffered read we
1206 * need to defer completion. We can't decide this now,
1207 * however the workqueue needs to be initialized here.
1208 */
1209 retval = sb_init_dio_done_wq(dio->inode->i_sb);
1210 }
1211 if (retval)
1212 goto fail_dio;
1213 }
1214
1215 /*
1216 * Will be decremented at I/O completion time.
1217 */
1218 inode_dio_begin(inode);
1219
1220 sdio.blkbits = blkbits;
1221 sdio.blkfactor = i_blkbits - blkbits;
1222 sdio.block_in_file = offset >> blkbits;
1223
1224 sdio.get_block = get_block;
1225 dio->end_io = end_io;
1226 sdio.final_block_in_bio = -1;
1227 sdio.next_block_for_io = -1;
1228
1229 dio->iocb = iocb;
1230
1231 spin_lock_init(&dio->bio_lock);
1232 dio->refcount = 1;
1233
1234 dio->should_dirty = user_backed_iter(iter) && iov_iter_rw(iter) == READ;
1235 sdio.iter = iter;
1236 sdio.final_block_in_request = end >> blkbits;
1237
1238 /*
1239 * In case of non-aligned buffers, we may need 2 more
1240 * pages since we need to zero out first and last block.
1241 */
1242 if (unlikely(sdio.blkfactor))
1243 sdio.pages_in_io = 2;
1244
1245 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
1246
1247 blk_start_plug(&plug);
1248
1249 retval = do_direct_IO(dio, &sdio, &map_bh);
1250 if (retval)
1251 dio_cleanup(dio, &sdio);
1252
1253 if (retval == -ENOTBLK) {
1254 /*
1255 * The remaining part of the request will be
1256 * handled by buffered I/O when we return
1257 */
1258 retval = 0;
1259 }
1260 /*
1261 * There may be some unwritten disk at the end of a part-written
1262 * fs-block-sized block. Go zero that now.
1263 */
1264 dio_zero_block(dio, &sdio, 1, &map_bh);
1265
1266 if (sdio.cur_page) {
1267 ssize_t ret2;
1268
1269 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1270 if (retval == 0)
1271 retval = ret2;
1272 dio_unpin_page(dio, sdio.cur_page);
1273 sdio.cur_page = NULL;
1274 }
1275 if (sdio.bio)
1276 dio_bio_submit(dio, &sdio);
1277
1278 blk_finish_plug(&plug);
1279
1280 /*
1281 * It is possible that, we return short IO due to end of file.
1282 * In that case, we need to release all the pages we got hold on.
1283 */
1284 dio_cleanup(dio, &sdio);
1285
1286 /*
1287 * All block lookups have been performed. For READ requests
1288 * we can let i_mutex go now that its achieved its purpose
1289 * of protecting us from looking up uninitialized blocks.
1290 */
1291 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
1292 inode_unlock(dio->inode);
1293
1294 /*
1295 * The only time we want to leave bios in flight is when a successful
1296 * partial aio read or full aio write have been setup. In that case
1297 * bio completion will call aio_complete. The only time it's safe to
1298 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1299 * This had *better* be the only place that raises -EIOCBQUEUED.
1300 */
1301 BUG_ON(retval == -EIOCBQUEUED);
1302 if (dio->is_async && retval == 0 && dio->result &&
1303 (iov_iter_rw(iter) == READ || dio->result == count))
1304 retval = -EIOCBQUEUED;
1305 else
1306 dio_await_completion(dio);
1307
1308 if (drop_refcount(dio) == 0) {
1309 retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE);
1310 } else
1311 BUG_ON(retval != -EIOCBQUEUED);
1312
1313 return retval;
1314
1315 fail_dio:
1316 if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ)
1317 inode_unlock(inode);
1318
1319 kmem_cache_free(dio_cache, dio);
1320 return retval;
1321 }
1322 EXPORT_SYMBOL(__blockdev_direct_IO);
1323
1324 static __init int dio_init(void)
1325 {
1326 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1327 return 0;
1328 }
1329 module_init(dio_init)
1330
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.