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
70struct 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 */
117struct 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
154static struct kmem_cache *dio_cache __ro_after_init;
155
156/*
157 * How many pages are in the queue?
158 */
159static 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 */
167static 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(i: sdio->iter, pages: &pages, LONG_MAX,
174 DIO_PAGES, extraction_flags: 0, offset0: &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 */
208static 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(error: ret);
217 BUG_ON(dio_pages_present(sdio) == 0);
218 }
219 return dio->pages[sdio->head];
220}
221
222static 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
228static 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 */
245static 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(iocb: dio->iocb, count: ret);
302
303 inode_dio_end(inode: 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(iocb: dio->iocb, count: ret);
315 dio->iocb->ki_complete(dio->iocb, ret);
316 }
317
318 kmem_cache_free(s: dio_cache, objp: dio);
319 return ret;
320}
321
322static 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, ret: 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
327}
328
329static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
330
331/*
332 * Asynchronous IO callback.
333 */
334static 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(tsk: dio->waiter);
349 spin_unlock_irqrestore(lock: &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(wq: dio->inode->i_sb->s_dio_done_wq,
367 work: &dio->complete_work);
368 } else {
369 dio_complete(dio, ret: 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 */
381static 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(tsk: dio->waiter);
391 spin_unlock_irqrestore(lock: &dio->bio_lock, flags);
392}
393
394static inline void
395dio_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, opf: 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, bit: BIO_PAGE_PINNED);
413 sdio->bio = bio;
414 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
415}
416
417/*
418 * In the AIO read case we speculatively dirty the pages before starting IO.
419 * During IO completion, any of these pages which happen to have been written
420 * back will be redirtied by bio_check_pages_dirty().
421 *
422 * bios hold a dio reference between submit_bio and ->end_io.
423 */
424static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
425{
426 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
427 struct bio *bio = sdio->bio;
428 unsigned long flags;
429
430 bio->bi_private = dio;
431
432 spin_lock_irqsave(&dio->bio_lock, flags);
433 dio->refcount++;
434 spin_unlock_irqrestore(lock: &dio->bio_lock, flags);
435
436 if (dio->is_async && dio_op == REQ_OP_READ && dio->should_dirty)
437 bio_set_pages_dirty(bio);
438
439 dio->bio_disk = bio->bi_bdev->bd_disk;
440
441 submit_bio(bio);
442
443 sdio->bio = NULL;
444 sdio->boundary = 0;
445 sdio->logical_offset_in_bio = 0;
446}
447
448/*
449 * Release any resources in case of a failure
450 */
451static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
452{
453 if (dio->is_pinned)
454 unpin_user_pages(pages: dio->pages + sdio->head,
455 npages: sdio->tail - sdio->head);
456 sdio->head = sdio->tail;
457}
458
459/*
460 * Wait for the next BIO to complete. Remove it and return it. NULL is
461 * returned once all BIOs have been completed. This must only be called once
462 * all bios have been issued so that dio->refcount can only decrease. This
463 * requires that the caller hold a reference on the dio.
464 */
465static struct bio *dio_await_one(struct dio *dio)
466{
467 unsigned long flags;
468 struct bio *bio = NULL;
469
470 spin_lock_irqsave(&dio->bio_lock, flags);
471
472 /*
473 * Wait as long as the list is empty and there are bios in flight. bio
474 * completion drops the count, maybe adds to the list, and wakes while
475 * holding the bio_lock so we don't need set_current_state()'s barrier
476 * and can call it after testing our condition.
477 */
478 while (dio->refcount > 1 && dio->bio_list == NULL) {
479 __set_current_state(TASK_UNINTERRUPTIBLE);
480 dio->waiter = current;
481 spin_unlock_irqrestore(lock: &dio->bio_lock, flags);
482 blk_io_schedule();
483 /* wake up sets us TASK_RUNNING */
484 spin_lock_irqsave(&dio->bio_lock, flags);
485 dio->waiter = NULL;
486 }
487 if (dio->bio_list) {
488 bio = dio->bio_list;
489 dio->bio_list = bio->bi_private;
490 }
491 spin_unlock_irqrestore(lock: &dio->bio_lock, flags);
492 return bio;
493}
494
495/*
496 * Process one completed BIO. No locks are held.
497 */
498static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
499{
500 blk_status_t err = bio->bi_status;
501 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
502 bool should_dirty = dio_op == REQ_OP_READ && dio->should_dirty;
503
504 if (err) {
505 if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
506 dio->io_error = -EAGAIN;
507 else
508 dio->io_error = -EIO;
509 }
510
511 if (dio->is_async && should_dirty) {
512 bio_check_pages_dirty(bio); /* transfers ownership */
513 } else {
514 bio_release_pages(bio, mark_dirty: should_dirty);
515 bio_put(bio);
516 }
517 return err;
518}
519
520/*
521 * Wait on and process all in-flight BIOs. This must only be called once
522 * all bios have been issued so that the refcount can only decrease.
523 * This just waits for all bios to make it through dio_bio_complete. IO
524 * errors are propagated through dio->io_error and should be propagated via
525 * dio_complete().
526 */
527static void dio_await_completion(struct dio *dio)
528{
529 struct bio *bio;
530 do {
531 bio = dio_await_one(dio);
532 if (bio)
533 dio_bio_complete(dio, bio);
534 } while (bio);
535}
536
537/*
538 * A really large O_DIRECT read or write can generate a lot of BIOs. So
539 * to keep the memory consumption sane we periodically reap any completed BIOs
540 * during the BIO generation phase.
541 *
542 * This also helps to limit the peak amount of pinned userspace memory.
543 */
544static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
545{
546 int ret = 0;
547
548 if (sdio->reap_counter++ >= 64) {
549 while (dio->bio_list) {
550 unsigned long flags;
551 struct bio *bio;
552 int ret2;
553
554 spin_lock_irqsave(&dio->bio_lock, flags);
555 bio = dio->bio_list;
556 dio->bio_list = bio->bi_private;
557 spin_unlock_irqrestore(lock: &dio->bio_lock, flags);
558 ret2 = blk_status_to_errno(status: dio_bio_complete(dio, bio));
559 if (ret == 0)
560 ret = ret2;
561 }
562 sdio->reap_counter = 0;
563 }
564 return ret;
565}
566
567static int dio_set_defer_completion(struct dio *dio)
568{
569 struct super_block *sb = dio->inode->i_sb;
570
571 if (dio->defer_completion)
572 return 0;
573 dio->defer_completion = true;
574 if (!sb->s_dio_done_wq)
575 return sb_init_dio_done_wq(sb);
576 return 0;
577}
578
579/*
580 * Call into the fs to map some more disk blocks. We record the current number
581 * of available blocks at sdio->blocks_available. These are in units of the
582 * fs blocksize, i_blocksize(inode).
583 *
584 * The fs is allowed to map lots of blocks at once. If it wants to do that,
585 * it uses the passed inode-relative block number as the file offset, as usual.
586 *
587 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
588 * has remaining to do. The fs should not map more than this number of blocks.
589 *
590 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
591 * indicate how much contiguous disk space has been made available at
592 * bh->b_blocknr.
593 *
594 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
595 * This isn't very efficient...
596 *
597 * In the case of filesystem holes: the fs may return an arbitrarily-large
598 * hole by returning an appropriate value in b_size and by clearing
599 * buffer_mapped(). However the direct-io code will only process holes one
600 * block at a time - it will repeatedly call get_block() as it walks the hole.
601 */
602static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
603 struct buffer_head *map_bh)
604{
605 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
606 int ret;
607 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
608 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
609 unsigned long fs_count; /* Number of filesystem-sized blocks */
610 int create;
611 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
612 loff_t i_size;
613
614 /*
615 * If there was a memory error and we've overwritten all the
616 * mapped blocks then we can now return that memory error
617 */
618 ret = dio->page_errors;
619 if (ret == 0) {
620 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
621 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
622 fs_endblk = (sdio->final_block_in_request - 1) >>
623 sdio->blkfactor;
624 fs_count = fs_endblk - fs_startblk + 1;
625
626 map_bh->b_state = 0;
627 map_bh->b_size = fs_count << i_blkbits;
628
629 /*
630 * For writes that could fill holes inside i_size on a
631 * DIO_SKIP_HOLES filesystem we forbid block creations: only
632 * overwrites are permitted. We will return early to the caller
633 * once we see an unmapped buffer head returned, and the caller
634 * will fall back to buffered I/O.
635 *
636 * Otherwise the decision is left to the get_blocks method,
637 * which may decide to handle it or also return an unmapped
638 * buffer head.
639 */
640 create = dio_op == REQ_OP_WRITE;
641 if (dio->flags & DIO_SKIP_HOLES) {
642 i_size = i_size_read(inode: dio->inode);
643 if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits)
644 create = 0;
645 }
646
647 ret = (*sdio->get_block)(dio->inode, fs_startblk,
648 map_bh, create);
649
650 /* Store for completion */
651 dio->private = map_bh->b_private;
652
653 if (ret == 0 && buffer_defer_completion(bh: map_bh))
654 ret = dio_set_defer_completion(dio);
655 }
656 return ret;
657}
658
659/*
660 * There is no bio. Make one now.
661 */
662static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
663 sector_t start_sector, struct buffer_head *map_bh)
664{
665 sector_t sector;
666 int ret, nr_pages;
667
668 ret = dio_bio_reap(dio, sdio);
669 if (ret)
670 goto out;
671 sector = start_sector << (sdio->blkbits - 9);
672 nr_pages = bio_max_segs(nr_segs: sdio->pages_in_io);
673 BUG_ON(nr_pages <= 0);
674 dio_bio_alloc(dio, sdio, bdev: map_bh->b_bdev, first_sector: sector, nr_vecs: nr_pages);
675 sdio->boundary = 0;
676out:
677 return ret;
678}
679
680/*
681 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
682 * that was successful then update final_block_in_bio and take a ref against
683 * the just-added page.
684 *
685 * Return zero on success. Non-zero means the caller needs to start a new BIO.
686 */
687static inline int dio_bio_add_page(struct dio *dio, struct dio_submit *sdio)
688{
689 int ret;
690
691 ret = bio_add_page(bio: sdio->bio, page: sdio->cur_page,
692 len: sdio->cur_page_len, off: sdio->cur_page_offset);
693 if (ret == sdio->cur_page_len) {
694 /*
695 * Decrement count only, if we are done with this page
696 */
697 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
698 sdio->pages_in_io--;
699 dio_pin_page(dio, page: sdio->cur_page);
700 sdio->final_block_in_bio = sdio->cur_page_block +
701 (sdio->cur_page_len >> sdio->blkbits);
702 ret = 0;
703 } else {
704 ret = 1;
705 }
706 return ret;
707}
708
709/*
710 * Put cur_page under IO. The section of cur_page which is described by
711 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
712 * starts on-disk at cur_page_block.
713 *
714 * We take a ref against the page here (on behalf of its presence in the bio).
715 *
716 * The caller of this function is responsible for removing cur_page from the
717 * dio, and for dropping the refcount which came from that presence.
718 */
719static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
720 struct buffer_head *map_bh)
721{
722 int ret = 0;
723
724 if (sdio->bio) {
725 loff_t cur_offset = sdio->cur_page_fs_offset;
726 loff_t bio_next_offset = sdio->logical_offset_in_bio +
727 sdio->bio->bi_iter.bi_size;
728
729 /*
730 * See whether this new request is contiguous with the old.
731 *
732 * Btrfs cannot handle having logically non-contiguous requests
733 * submitted. For example if you have
734 *
735 * Logical: [0-4095][HOLE][8192-12287]
736 * Physical: [0-4095] [4096-8191]
737 *
738 * We cannot submit those pages together as one BIO. So if our
739 * current logical offset in the file does not equal what would
740 * be the next logical offset in the bio, submit the bio we
741 * have.
742 */
743 if (sdio->final_block_in_bio != sdio->cur_page_block ||
744 cur_offset != bio_next_offset)
745 dio_bio_submit(dio, sdio);
746 }
747
748 if (sdio->bio == NULL) {
749 ret = dio_new_bio(dio, sdio, start_sector: sdio->cur_page_block, map_bh);
750 if (ret)
751 goto out;
752 }
753
754 if (dio_bio_add_page(dio, sdio) != 0) {
755 dio_bio_submit(dio, sdio);
756 ret = dio_new_bio(dio, sdio, start_sector: sdio->cur_page_block, map_bh);
757 if (ret == 0) {
758 ret = dio_bio_add_page(dio, sdio);
759 BUG_ON(ret != 0);
760 }
761 }
762out:
763 return ret;
764}
765
766/*
767 * An autonomous function to put a chunk of a page under deferred IO.
768 *
769 * The caller doesn't actually know (or care) whether this piece of page is in
770 * a BIO, or is under IO or whatever. We just take care of all possible
771 * situations here. The separation between the logic of do_direct_IO() and
772 * that of submit_page_section() is important for clarity. Please don't break.
773 *
774 * The chunk of page starts on-disk at blocknr.
775 *
776 * We perform deferred IO, by recording the last-submitted page inside our
777 * private part of the dio structure. If possible, we just expand the IO
778 * across that page here.
779 *
780 * If that doesn't work out then we put the old page into the bio and add this
781 * page to the dio instead.
782 */
783static inline int
784submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
785 unsigned offset, unsigned len, sector_t blocknr,
786 struct buffer_head *map_bh)
787{
788 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
789 int ret = 0;
790 int boundary = sdio->boundary; /* dio_send_cur_page may clear it */
791
792 if (dio_op == REQ_OP_WRITE) {
793 /*
794 * Read accounting is performed in submit_bio()
795 */
796 task_io_account_write(bytes: len);
797 }
798
799 /*
800 * Can we just grow the current page's presence in the dio?
801 */
802 if (sdio->cur_page == page &&
803 sdio->cur_page_offset + sdio->cur_page_len == offset &&
804 sdio->cur_page_block +
805 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
806 sdio->cur_page_len += len;
807 goto out;
808 }
809
810 /*
811 * If there's a deferred page already there then send it.
812 */
813 if (sdio->cur_page) {
814 ret = dio_send_cur_page(dio, sdio, map_bh);
815 dio_unpin_page(dio, page: sdio->cur_page);
816 sdio->cur_page = NULL;
817 if (ret)
818 return ret;
819 }
820
821 dio_pin_page(dio, page); /* It is in dio */
822 sdio->cur_page = page;
823 sdio->cur_page_offset = offset;
824 sdio->cur_page_len = len;
825 sdio->cur_page_block = blocknr;
826 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
827out:
828 /*
829 * If boundary then we want to schedule the IO now to
830 * avoid metadata seeks.
831 */
832 if (boundary) {
833 ret = dio_send_cur_page(dio, sdio, map_bh);
834 if (sdio->bio)
835 dio_bio_submit(dio, sdio);
836 dio_unpin_page(dio, page: sdio->cur_page);
837 sdio->cur_page = NULL;
838 }
839 return ret;
840}
841
842/*
843 * If we are not writing the entire block and get_block() allocated
844 * the block for us, we need to fill-in the unused portion of the
845 * block with zeros. This happens only if user-buffer, fileoffset or
846 * io length is not filesystem block-size multiple.
847 *
848 * `end' is zero if we're doing the start of the IO, 1 at the end of the
849 * IO.
850 */
851static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
852 int end, struct buffer_head *map_bh)
853{
854 unsigned dio_blocks_per_fs_block;
855 unsigned this_chunk_blocks; /* In dio_blocks */
856 unsigned this_chunk_bytes;
857 struct page *page;
858
859 sdio->start_zero_done = 1;
860 if (!sdio->blkfactor || !buffer_new(bh: map_bh))
861 return;
862
863 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
864 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
865
866 if (!this_chunk_blocks)
867 return;
868
869 /*
870 * We need to zero out part of an fs block. It is either at the
871 * beginning or the end of the fs block.
872 */
873 if (end)
874 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
875
876 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
877
878 page = ZERO_PAGE(0);
879 if (submit_page_section(dio, sdio, page, offset: 0, len: this_chunk_bytes,
880 blocknr: sdio->next_block_for_io, map_bh))
881 return;
882
883 sdio->next_block_for_io += this_chunk_blocks;
884}
885
886/*
887 * Walk the user pages, and the file, mapping blocks to disk and generating
888 * a sequence of (page,offset,len,block) mappings. These mappings are injected
889 * into submit_page_section(), which takes care of the next stage of submission
890 *
891 * Direct IO against a blockdev is different from a file. Because we can
892 * happily perform page-sized but 512-byte aligned IOs. It is important that
893 * blockdev IO be able to have fine alignment and large sizes.
894 *
895 * So what we do is to permit the ->get_block function to populate bh.b_size
896 * with the size of IO which is permitted at this offset and this i_blkbits.
897 *
898 * For best results, the blockdev should be set up with 512-byte i_blkbits and
899 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
900 * fine alignment but still allows this function to work in PAGE_SIZE units.
901 */
902static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
903 struct buffer_head *map_bh)
904{
905 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
906 const unsigned blkbits = sdio->blkbits;
907 const unsigned i_blkbits = blkbits + sdio->blkfactor;
908 int ret = 0;
909
910 while (sdio->block_in_file < sdio->final_block_in_request) {
911 struct page *page;
912 size_t from, to;
913
914 page = dio_get_page(dio, sdio);
915 if (IS_ERR(ptr: page)) {
916 ret = PTR_ERR(ptr: page);
917 goto out;
918 }
919 from = sdio->head ? 0 : sdio->from;
920 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
921 sdio->head++;
922
923 while (from < to) {
924 unsigned this_chunk_bytes; /* # of bytes mapped */
925 unsigned this_chunk_blocks; /* # of blocks */
926 unsigned u;
927
928 if (sdio->blocks_available == 0) {
929 /*
930 * Need to go and map some more disk
931 */
932 unsigned long blkmask;
933 unsigned long dio_remainder;
934
935 ret = get_more_blocks(dio, sdio, map_bh);
936 if (ret) {
937 dio_unpin_page(dio, page);
938 goto out;
939 }
940 if (!buffer_mapped(bh: map_bh))
941 goto do_holes;
942
943 sdio->blocks_available =
944 map_bh->b_size >> blkbits;
945 sdio->next_block_for_io =
946 map_bh->b_blocknr << sdio->blkfactor;
947 if (buffer_new(bh: map_bh)) {
948 clean_bdev_aliases(
949 bdev: map_bh->b_bdev,
950 block: map_bh->b_blocknr,
951 len: map_bh->b_size >> i_blkbits);
952 }
953
954 if (!sdio->blkfactor)
955 goto do_holes;
956
957 blkmask = (1 << sdio->blkfactor) - 1;
958 dio_remainder = (sdio->block_in_file & blkmask);
959
960 /*
961 * If we are at the start of IO and that IO
962 * starts partway into a fs-block,
963 * dio_remainder will be non-zero. If the IO
964 * is a read then we can simply advance the IO
965 * cursor to the first block which is to be
966 * read. But if the IO is a write and the
967 * block was newly allocated we cannot do that;
968 * the start of the fs block must be zeroed out
969 * on-disk
970 */
971 if (!buffer_new(bh: map_bh))
972 sdio->next_block_for_io += dio_remainder;
973 sdio->blocks_available -= dio_remainder;
974 }
975do_holes:
976 /* Handle holes */
977 if (!buffer_mapped(bh: map_bh)) {
978 loff_t i_size_aligned;
979
980 /* AKPM: eargh, -ENOTBLK is a hack */
981 if (dio_op == REQ_OP_WRITE) {
982 dio_unpin_page(dio, page);
983 return -ENOTBLK;
984 }
985
986 /*
987 * Be sure to account for a partial block as the
988 * last block in the file
989 */
990 i_size_aligned = ALIGN(i_size_read(dio->inode),
991 1 << blkbits);
992 if (sdio->block_in_file >=
993 i_size_aligned >> blkbits) {
994 /* We hit eof */
995 dio_unpin_page(dio, page);
996 goto out;
997 }
998 zero_user(page, start: from, size: 1 << blkbits);
999 sdio->block_in_file++;
1000 from += 1 << blkbits;
1001 dio->result += 1 << blkbits;
1002 goto next_block;
1003 }
1004
1005 /*
1006 * If we're performing IO which has an alignment which
1007 * is finer than the underlying fs, go check to see if
1008 * we must zero out the start of this block.
1009 */
1010 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1011 dio_zero_block(dio, sdio, end: 0, map_bh);
1012
1013 /*
1014 * Work out, in this_chunk_blocks, how much disk we
1015 * can add to this page
1016 */
1017 this_chunk_blocks = sdio->blocks_available;
1018 u = (to - from) >> blkbits;
1019 if (this_chunk_blocks > u)
1020 this_chunk_blocks = u;
1021 u = sdio->final_block_in_request - sdio->block_in_file;
1022 if (this_chunk_blocks > u)
1023 this_chunk_blocks = u;
1024 this_chunk_bytes = this_chunk_blocks << blkbits;
1025 BUG_ON(this_chunk_bytes == 0);
1026
1027 if (this_chunk_blocks == sdio->blocks_available)
1028 sdio->boundary = buffer_boundary(bh: map_bh);
1029 ret = submit_page_section(dio, sdio, page,
1030 offset: from,
1031 len: this_chunk_bytes,
1032 blocknr: sdio->next_block_for_io,
1033 map_bh);
1034 if (ret) {
1035 dio_unpin_page(dio, page);
1036 goto out;
1037 }
1038 sdio->next_block_for_io += this_chunk_blocks;
1039
1040 sdio->block_in_file += this_chunk_blocks;
1041 from += this_chunk_bytes;
1042 dio->result += this_chunk_bytes;
1043 sdio->blocks_available -= this_chunk_blocks;
1044next_block:
1045 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1046 if (sdio->block_in_file == sdio->final_block_in_request)
1047 break;
1048 }
1049
1050 /* Drop the pin which was taken in get_user_pages() */
1051 dio_unpin_page(dio, page);
1052 }
1053out:
1054 return ret;
1055}
1056
1057static inline int drop_refcount(struct dio *dio)
1058{
1059 int ret2;
1060 unsigned long flags;
1061
1062 /*
1063 * Sync will always be dropping the final ref and completing the
1064 * operation. AIO can if it was a broken operation described above or
1065 * in fact if all the bios race to complete before we get here. In
1066 * that case dio_complete() translates the EIOCBQUEUED into the proper
1067 * return code that the caller will hand to ->complete().
1068 *
1069 * This is managed by the bio_lock instead of being an atomic_t so that
1070 * completion paths can drop their ref and use the remaining count to
1071 * decide to wake the submission path atomically.
1072 */
1073 spin_lock_irqsave(&dio->bio_lock, flags);
1074 ret2 = --dio->refcount;
1075 spin_unlock_irqrestore(lock: &dio->bio_lock, flags);
1076 return ret2;
1077}
1078
1079/*
1080 * This is a library function for use by filesystem drivers.
1081 *
1082 * The locking rules are governed by the flags parameter:
1083 * - if the flags value contains DIO_LOCKING we use a fancy locking
1084 * scheme for dumb filesystems.
1085 * For writes this function is called under i_mutex and returns with
1086 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1087 * taken and dropped again before returning.
1088 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1089 * internal locking but rather rely on the filesystem to synchronize
1090 * direct I/O reads/writes versus each other and truncate.
1091 *
1092 * To help with locking against truncate we incremented the i_dio_count
1093 * counter before starting direct I/O, and decrement it once we are done.
1094 * Truncate can wait for it to reach zero to provide exclusion. It is
1095 * expected that filesystem provide exclusion between new direct I/O
1096 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1097 * but other filesystems need to take care of this on their own.
1098 *
1099 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1100 * is always inlined. Otherwise gcc is unable to split the structure into
1101 * individual fields and will generate much worse code. This is important
1102 * for the whole file.
1103 */
1104ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1105 struct block_device *bdev, struct iov_iter *iter,
1106 get_block_t get_block, dio_iodone_t end_io,
1107 int flags)
1108{
1109 unsigned i_blkbits = READ_ONCE(inode->i_blkbits);
1110 unsigned blkbits = i_blkbits;
1111 unsigned blocksize_mask = (1 << blkbits) - 1;
1112 ssize_t retval = -EINVAL;
1113 const size_t count = iov_iter_count(i: iter);
1114 loff_t offset = iocb->ki_pos;
1115 const loff_t end = offset + count;
1116 struct dio *dio;
1117 struct dio_submit sdio = { 0, };
1118 struct buffer_head map_bh = { 0, };
1119 struct blk_plug plug;
1120 unsigned long align = offset | iov_iter_alignment(i: iter);
1121
1122 /*
1123 * Avoid references to bdev if not absolutely needed to give
1124 * the early prefetch in the caller enough time.
1125 */
1126
1127 /* watch out for a 0 len io from a tricksy fs */
1128 if (iov_iter_rw(i: iter) == READ && !count)
1129 return 0;
1130
1131 dio = kmem_cache_alloc(cachep: dio_cache, GFP_KERNEL);
1132 if (!dio)
1133 return -ENOMEM;
1134 /*
1135 * Believe it or not, zeroing out the page array caused a .5%
1136 * performance regression in a database benchmark. So, we take
1137 * care to only zero out what's needed.
1138 */
1139 memset(dio, 0, offsetof(struct dio, pages));
1140
1141 dio->flags = flags;
1142 if (dio->flags & DIO_LOCKING && iov_iter_rw(i: iter) == READ) {
1143 /* will be released by direct_io_worker */
1144 inode_lock(inode);
1145 }
1146 dio->is_pinned = iov_iter_extract_will_pin(iter);
1147
1148 /* Once we sampled i_size check for reads beyond EOF */
1149 dio->i_size = i_size_read(inode);
1150 if (iov_iter_rw(i: iter) == READ && offset >= dio->i_size) {
1151 retval = 0;
1152 goto fail_dio;
1153 }
1154
1155 if (align & blocksize_mask) {
1156 if (bdev)
1157 blkbits = blksize_bits(size: bdev_logical_block_size(bdev));
1158 blocksize_mask = (1 << blkbits) - 1;
1159 if (align & blocksize_mask)
1160 goto fail_dio;
1161 }
1162
1163 if (dio->flags & DIO_LOCKING && iov_iter_rw(i: iter) == READ) {
1164 struct address_space *mapping = iocb->ki_filp->f_mapping;
1165
1166 retval = filemap_write_and_wait_range(mapping, lstart: offset, lend: end - 1);
1167 if (retval)
1168 goto fail_dio;
1169 }
1170
1171 /*
1172 * For file extending writes updating i_size before data writeouts
1173 * complete can expose uninitialized blocks in dumb filesystems.
1174 * In that case we need to wait for I/O completion even if asked
1175 * for an asynchronous write.
1176 */
1177 if (is_sync_kiocb(kiocb: iocb))
1178 dio->is_async = false;
1179 else if (iov_iter_rw(i: iter) == WRITE && end > i_size_read(inode))
1180 dio->is_async = false;
1181 else
1182 dio->is_async = true;
1183
1184 dio->inode = inode;
1185 if (iov_iter_rw(i: iter) == WRITE) {
1186 dio->opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
1187 if (iocb->ki_flags & IOCB_NOWAIT)
1188 dio->opf |= REQ_NOWAIT;
1189 } else {
1190 dio->opf = REQ_OP_READ;
1191 }
1192
1193 /*
1194 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1195 * so that we can call ->fsync.
1196 */
1197 if (dio->is_async && iov_iter_rw(i: iter) == WRITE) {
1198 retval = 0;
1199 if (iocb_is_dsync(iocb))
1200 retval = dio_set_defer_completion(dio);
1201 else if (!dio->inode->i_sb->s_dio_done_wq) {
1202 /*
1203 * In case of AIO write racing with buffered read we
1204 * need to defer completion. We can't decide this now,
1205 * however the workqueue needs to be initialized here.
1206 */
1207 retval = sb_init_dio_done_wq(sb: dio->inode->i_sb);
1208 }
1209 if (retval)
1210 goto fail_dio;
1211 }
1212
1213 /*
1214 * Will be decremented at I/O completion time.
1215 */
1216 inode_dio_begin(inode);
1217
1218 retval = 0;
1219 sdio.blkbits = blkbits;
1220 sdio.blkfactor = i_blkbits - blkbits;
1221 sdio.block_in_file = offset >> blkbits;
1222
1223 sdio.get_block = get_block;
1224 dio->end_io = end_io;
1225 sdio.final_block_in_bio = -1;
1226 sdio.next_block_for_io = -1;
1227
1228 dio->iocb = iocb;
1229
1230 spin_lock_init(&dio->bio_lock);
1231 dio->refcount = 1;
1232
1233 dio->should_dirty = user_backed_iter(i: iter) && iov_iter_rw(i: iter) == READ;
1234 sdio.iter = iter;
1235 sdio.final_block_in_request = end >> blkbits;
1236
1237 /*
1238 * In case of non-aligned buffers, we may need 2 more
1239 * pages since we need to zero out first and last block.
1240 */
1241 if (unlikely(sdio.blkfactor))
1242 sdio.pages_in_io = 2;
1243
1244 sdio.pages_in_io += iov_iter_npages(i: iter, INT_MAX);
1245
1246 blk_start_plug(&plug);
1247
1248 retval = do_direct_IO(dio, sdio: &sdio, map_bh: &map_bh);
1249 if (retval)
1250 dio_cleanup(dio, sdio: &sdio);
1251
1252 if (retval == -ENOTBLK) {
1253 /*
1254 * The remaining part of the request will be
1255 * handled by buffered I/O when we return
1256 */
1257 retval = 0;
1258 }
1259 /*
1260 * There may be some unwritten disk at the end of a part-written
1261 * fs-block-sized block. Go zero that now.
1262 */
1263 dio_zero_block(dio, sdio: &sdio, end: 1, map_bh: &map_bh);
1264
1265 if (sdio.cur_page) {
1266 ssize_t ret2;
1267
1268 ret2 = dio_send_cur_page(dio, sdio: &sdio, map_bh: &map_bh);
1269 if (retval == 0)
1270 retval = ret2;
1271 dio_unpin_page(dio, page: sdio.cur_page);
1272 sdio.cur_page = NULL;
1273 }
1274 if (sdio.bio)
1275 dio_bio_submit(dio, sdio: &sdio);
1276
1277 blk_finish_plug(&plug);
1278
1279 /*
1280 * It is possible that, we return short IO due to end of file.
1281 * In that case, we need to release all the pages we got hold on.
1282 */
1283 dio_cleanup(dio, sdio: &sdio);
1284
1285 /*
1286 * All block lookups have been performed. For READ requests
1287 * we can let i_mutex go now that its achieved its purpose
1288 * of protecting us from looking up uninitialized blocks.
1289 */
1290 if (iov_iter_rw(i: iter) == READ && (dio->flags & DIO_LOCKING))
1291 inode_unlock(inode: dio->inode);
1292
1293 /*
1294 * The only time we want to leave bios in flight is when a successful
1295 * partial aio read or full aio write have been setup. In that case
1296 * bio completion will call aio_complete. The only time it's safe to
1297 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1298 * This had *better* be the only place that raises -EIOCBQUEUED.
1299 */
1300 BUG_ON(retval == -EIOCBQUEUED);
1301 if (dio->is_async && retval == 0 && dio->result &&
1302 (iov_iter_rw(i: iter) == READ || dio->result == count))
1303 retval = -EIOCBQUEUED;
1304 else
1305 dio_await_completion(dio);
1306
1307 if (drop_refcount(dio) == 0) {
1308 retval = dio_complete(dio, ret: retval, DIO_COMPLETE_INVALIDATE);
1309 } else
1310 BUG_ON(retval != -EIOCBQUEUED);
1311
1312 return retval;
1313
1314fail_dio:
1315 if (dio->flags & DIO_LOCKING && iov_iter_rw(i: iter) == READ)
1316 inode_unlock(inode);
1317
1318 kmem_cache_free(s: dio_cache, objp: dio);
1319 return retval;
1320}
1321EXPORT_SYMBOL(__blockdev_direct_IO);
1322
1323static __init int dio_init(void)
1324{
1325 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1326 return 0;
1327}
1328module_init(dio_init)
1329

source code of linux/fs/direct-io.c