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
41#include "internal.h"
42
43/*
44 * How many user pages to map in one call to iov_iter_extract_pages(). This
45 * determines the size of a structure in the slab cache
46 */
47#define DIO_PAGES 64
48
49/*
50 * Flags for dio_complete()
51 */
52#define DIO_COMPLETE_ASYNC 0x01 /* This is async IO */
53#define DIO_COMPLETE_INVALIDATE 0x02 /* Can invalidate pages */
54
55/*
56 * This code generally works in units of "dio_blocks". A dio_block is
57 * somewhere between the hard sector size and the filesystem block size. it
58 * is determined on a per-invocation basis. When talking to the filesystem
59 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
60 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
61 * to bio_block quantities by shifting left by blkfactor.
62 *
63 * If blkfactor is zero then the user's request was aligned to the filesystem's
64 * blocksize.
65 */
66
67/* dio_state only used in the submission path */
68
69struct dio_submit {
70 struct bio *bio; /* bio under assembly */
71 unsigned blkbits; /* doesn't change */
72 unsigned blkfactor; /* When we're using an alignment which
73 is finer than the filesystem's soft
74 blocksize, this specifies how much
75 finer. blkfactor=2 means 1/4-block
76 alignment. Does not change */
77 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
78 been performed at the start of a
79 write */
80 int pages_in_io; /* approximate total IO pages */
81 sector_t block_in_file; /* Current offset into the underlying
82 file in dio_block units. */
83 unsigned blocks_available; /* At block_in_file. changes */
84 int reap_counter; /* rate limit reaping */
85 sector_t final_block_in_request;/* doesn't change */
86 int boundary; /* prev block is at a boundary */
87 get_block_t *get_block; /* block mapping function */
88
89 loff_t logical_offset_in_bio; /* current first logical block in bio */
90 sector_t final_block_in_bio; /* current final block in bio + 1 */
91 sector_t next_block_for_io; /* next block to be put under IO,
92 in dio_blocks units */
93
94 /*
95 * Deferred addition of a page to the dio. These variables are
96 * private to dio_send_cur_page(), submit_page_section() and
97 * dio_bio_add_page().
98 */
99 struct page *cur_page; /* The page */
100 unsigned cur_page_offset; /* Offset into it, in bytes */
101 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
102 sector_t cur_page_block; /* Where it starts */
103 loff_t cur_page_fs_offset; /* Offset in file */
104
105 struct iov_iter *iter;
106 /*
107 * Page queue. These variables belong to dio_refill_pages() and
108 * dio_get_page().
109 */
110 unsigned head; /* next page to process */
111 unsigned tail; /* last valid page + 1 */
112 size_t from, to;
113};
114
115/* dio_state communicated between submission path and end_io */
116struct dio {
117 int flags; /* doesn't change */
118 blk_opf_t opf; /* request operation type and flags */
119 struct gendisk *bio_disk;
120 struct inode *inode;
121 loff_t i_size; /* i_size when submitted */
122 dio_iodone_t *end_io; /* IO completion function */
123 bool is_pinned; /* T if we have pins on the pages */
124
125 void *private; /* copy from map_bh.b_private */
126
127 /* BIO completion state */
128 spinlock_t bio_lock; /* protects BIO fields below */
129 int page_errors; /* err from iov_iter_extract_pages() */
130 int is_async; /* is IO async ? */
131 bool defer_completion; /* defer AIO completion to workqueue? */
132 bool should_dirty; /* if pages should be dirtied */
133 int io_error; /* IO error in completion path */
134 unsigned long refcount; /* direct_io_worker() and bios */
135 struct bio *bio_list; /* singly linked via bi_private */
136 struct task_struct *waiter; /* waiting task (NULL if none) */
137
138 /* AIO related stuff */
139 struct kiocb *iocb; /* kiocb */
140 ssize_t result; /* IO result */
141
142 /*
143 * pages[] (and any fields placed after it) are not zeroed out at
144 * allocation time. Don't add new fields after pages[] unless you
145 * wish that they not be zeroed.
146 */
147 union {
148 struct page *pages[DIO_PAGES]; /* page buffer */
149 struct work_struct complete_work;/* deferred AIO completion */
150 };
151} ____cacheline_aligned_in_smp;
152
153static struct kmem_cache *dio_cache __ro_after_init;
154
155/*
156 * How many pages are in the queue?
157 */
158static inline unsigned dio_pages_present(struct dio_submit *sdio)
159{
160 return sdio->tail - sdio->head;
161}
162
163/*
164 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
165 */
166static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
167{
168 struct page **pages = dio->pages;
169 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
170 ssize_t ret;
171
172 ret = iov_iter_extract_pages(i: sdio->iter, pages: &pages, LONG_MAX,
173 DIO_PAGES, extraction_flags: 0, offset0: &sdio->from);
174
175 if (ret < 0 && sdio->blocks_available && dio_op == REQ_OP_WRITE) {
176 /*
177 * A memory fault, but the filesystem has some outstanding
178 * mapped blocks. We need to use those blocks up to avoid
179 * leaking stale data in the file.
180 */
181 if (dio->page_errors == 0)
182 dio->page_errors = ret;
183 dio->pages[0] = ZERO_PAGE(0);
184 sdio->head = 0;
185 sdio->tail = 1;
186 sdio->from = 0;
187 sdio->to = PAGE_SIZE;
188 return 0;
189 }
190
191 if (ret >= 0) {
192 ret += sdio->from;
193 sdio->head = 0;
194 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
195 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
196 return 0;
197 }
198 return ret;
199}
200
201/*
202 * Get another userspace page. Returns an ERR_PTR on error. Pages are
203 * buffered inside the dio so that we can call iov_iter_extract_pages()
204 * against a decent number of pages, less frequently. To provide nicer use of
205 * the L1 cache.
206 */
207static inline struct page *dio_get_page(struct dio *dio,
208 struct dio_submit *sdio)
209{
210 if (dio_pages_present(sdio) == 0) {
211 int ret;
212
213 ret = dio_refill_pages(dio, sdio);
214 if (ret)
215 return ERR_PTR(error: ret);
216 BUG_ON(dio_pages_present(sdio) == 0);
217 }
218 return dio->pages[sdio->head];
219}
220
221static void dio_pin_page(struct dio *dio, struct page *page)
222{
223 if (dio->is_pinned)
224 folio_add_pin(page_folio(page));
225}
226
227static void dio_unpin_page(struct dio *dio, struct page *page)
228{
229 if (dio->is_pinned)
230 unpin_user_page(page);
231}
232
233/*
234 * dio_complete() - called when all DIO BIO I/O has been completed
235 *
236 * This drops i_dio_count, lets interested parties know that a DIO operation
237 * has completed, and calculates the resulting return code for the operation.
238 *
239 * It lets the filesystem know if it registered an interest earlier via
240 * get_block. Pass the private field of the map buffer_head so that
241 * filesystems can use it to hold additional state between get_block calls and
242 * dio_complete.
243 */
244static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags)
245{
246 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
247 loff_t offset = dio->iocb->ki_pos;
248 ssize_t transferred = 0;
249 int err;
250
251 /*
252 * AIO submission can race with bio completion to get here while
253 * expecting to have the last io completed by bio completion.
254 * In that case -EIOCBQUEUED is in fact not an error we want
255 * to preserve through this call.
256 */
257 if (ret == -EIOCBQUEUED)
258 ret = 0;
259
260 if (dio->result) {
261 transferred = dio->result;
262
263 /* Check for short read case */
264 if (dio_op == REQ_OP_READ &&
265 ((offset + transferred) > dio->i_size))
266 transferred = dio->i_size - offset;
267 /* ignore EFAULT if some IO has been done */
268 if (unlikely(ret == -EFAULT) && transferred)
269 ret = 0;
270 }
271
272 if (ret == 0)
273 ret = dio->page_errors;
274 if (ret == 0)
275 ret = dio->io_error;
276 if (ret == 0)
277 ret = transferred;
278
279 if (dio->end_io) {
280 // XXX: ki_pos??
281 err = dio->end_io(dio->iocb, offset, ret, dio->private);
282 if (err)
283 ret = err;
284 }
285
286 /*
287 * Try again to invalidate clean pages which might have been cached by
288 * non-direct readahead, or faulted in by get_user_pages() if the source
289 * of the write was an mmap'ed region of the file we're writing. Either
290 * one is a pretty crazy thing to do, so we don't support it 100%. If
291 * this invalidation fails, tough, the write still worked...
292 *
293 * And this page cache invalidation has to be after dio->end_io(), as
294 * some filesystems convert unwritten extents to real allocations in
295 * end_io() when necessary, otherwise a racing buffer read would cache
296 * zeros from unwritten extents.
297 */
298 if (flags & DIO_COMPLETE_INVALIDATE &&
299 ret > 0 && dio_op == REQ_OP_WRITE)
300 kiocb_invalidate_post_direct_write(iocb: dio->iocb, count: ret);
301
302 inode_dio_end(inode: dio->inode);
303
304 if (flags & DIO_COMPLETE_ASYNC) {
305 /*
306 * generic_write_sync expects ki_pos to have been updated
307 * already, but the submission path only does this for
308 * synchronous I/O.
309 */
310 dio->iocb->ki_pos += transferred;
311
312 if (ret > 0 && dio_op == REQ_OP_WRITE)
313 ret = generic_write_sync(iocb: dio->iocb, count: ret);
314 dio->iocb->ki_complete(dio->iocb, ret);
315 }
316
317 kmem_cache_free(s: dio_cache, objp: dio);
318 return ret;
319}
320
321static void dio_aio_complete_work(struct work_struct *work)
322{
323 struct dio *dio = container_of(work, struct dio, complete_work);
324
325 dio_complete(dio, ret: 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
326}
327
328static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
329
330/*
331 * Asynchronous IO callback.
332 */
333static void dio_bio_end_aio(struct bio *bio)
334{
335 struct dio *dio = bio->bi_private;
336 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
337 unsigned long remaining;
338 unsigned long flags;
339 bool defer_completion = false;
340
341 /* cleanup the bio */
342 dio_bio_complete(dio, bio);
343
344 spin_lock_irqsave(&dio->bio_lock, flags);
345 remaining = --dio->refcount;
346 if (remaining == 1 && dio->waiter)
347 wake_up_process(tsk: dio->waiter);
348 spin_unlock_irqrestore(lock: &dio->bio_lock, flags);
349
350 if (remaining == 0) {
351 /*
352 * Defer completion when defer_completion is set or
353 * when the inode has pages mapped and this is AIO write.
354 * We need to invalidate those pages because there is a
355 * chance they contain stale data in the case buffered IO
356 * went in between AIO submission and completion into the
357 * same region.
358 */
359 if (dio->result)
360 defer_completion = dio->defer_completion ||
361 (dio_op == REQ_OP_WRITE &&
362 dio->inode->i_mapping->nrpages);
363 if (defer_completion) {
364 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
365 queue_work(wq: dio->inode->i_sb->s_dio_done_wq,
366 work: &dio->complete_work);
367 } else {
368 dio_complete(dio, ret: 0, DIO_COMPLETE_ASYNC);
369 }
370 }
371}
372
373/*
374 * The BIO completion handler simply queues the BIO up for the process-context
375 * handler.
376 *
377 * During I/O bi_private points at the dio. After I/O, bi_private is used to
378 * implement a singly-linked list of completed BIOs, at dio->bio_list.
379 */
380static void dio_bio_end_io(struct bio *bio)
381{
382 struct dio *dio = bio->bi_private;
383 unsigned long flags;
384
385 spin_lock_irqsave(&dio->bio_lock, flags);
386 bio->bi_private = dio->bio_list;
387 dio->bio_list = bio;
388 if (--dio->refcount == 1 && dio->waiter)
389 wake_up_process(tsk: dio->waiter);
390 spin_unlock_irqrestore(lock: &dio->bio_lock, flags);
391}
392
393static inline void
394dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
395 struct block_device *bdev,
396 sector_t first_sector, int nr_vecs)
397{
398 struct bio *bio;
399
400 /*
401 * bio_alloc() is guaranteed to return a bio when allowed to sleep and
402 * we request a valid number of vectors.
403 */
404 bio = bio_alloc(bdev, nr_vecs, opf: dio->opf, GFP_KERNEL);
405 bio->bi_iter.bi_sector = first_sector;
406 if (dio->is_async)
407 bio->bi_end_io = dio_bio_end_aio;
408 else
409 bio->bi_end_io = dio_bio_end_io;
410 if (dio->is_pinned)
411 bio_set_flag(bio, bit: BIO_PAGE_PINNED);
412 bio->bi_write_hint = file_inode(f: dio->iocb->ki_filp)->i_write_hint;
413
414 sdio->bio = bio;
415 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
416}
417
418/*
419 * In the AIO read case we speculatively dirty the pages before starting IO.
420 * During IO completion, any of these pages which happen to have been written
421 * back will be redirtied by bio_check_pages_dirty().
422 *
423 * bios hold a dio reference between submit_bio and ->end_io.
424 */
425static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
426{
427 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
428 struct bio *bio = sdio->bio;
429 unsigned long flags;
430
431 bio->bi_private = dio;
432
433 spin_lock_irqsave(&dio->bio_lock, flags);
434 dio->refcount++;
435 spin_unlock_irqrestore(lock: &dio->bio_lock, flags);
436
437 if (dio->is_async && dio_op == REQ_OP_READ && dio->should_dirty)
438 bio_set_pages_dirty(bio);
439
440 dio->bio_disk = bio->bi_bdev->bd_disk;
441
442 submit_bio(bio);
443
444 sdio->bio = NULL;
445 sdio->boundary = 0;
446 sdio->logical_offset_in_bio = 0;
447}
448
449/*
450 * Release any resources in case of a failure
451 */
452static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
453{
454 if (dio->is_pinned)
455 unpin_user_pages(pages: dio->pages + sdio->head,
456 npages: sdio->tail - sdio->head);
457 sdio->head = sdio->tail;
458}
459
460/*
461 * Wait for the next BIO to complete. Remove it and return it. NULL is
462 * returned once all BIOs have been completed. This must only be called once
463 * all bios have been issued so that dio->refcount can only decrease. This
464 * requires that the caller hold a reference on the dio.
465 */
466static struct bio *dio_await_one(struct dio *dio)
467{
468 unsigned long flags;
469 struct bio *bio = NULL;
470
471 spin_lock_irqsave(&dio->bio_lock, flags);
472
473 /*
474 * Wait as long as the list is empty and there are bios in flight. bio
475 * completion drops the count, maybe adds to the list, and wakes while
476 * holding the bio_lock so we don't need set_current_state()'s barrier
477 * and can call it after testing our condition.
478 */
479 while (dio->refcount > 1 && dio->bio_list == NULL) {
480 __set_current_state(TASK_UNINTERRUPTIBLE);
481 dio->waiter = current;
482 spin_unlock_irqrestore(lock: &dio->bio_lock, flags);
483 blk_io_schedule();
484 /* wake up sets us TASK_RUNNING */
485 spin_lock_irqsave(&dio->bio_lock, flags);
486 dio->waiter = NULL;
487 }
488 if (dio->bio_list) {
489 bio = dio->bio_list;
490 dio->bio_list = bio->bi_private;
491 }
492 spin_unlock_irqrestore(lock: &dio->bio_lock, flags);
493 return bio;
494}
495
496/*
497 * Process one completed BIO. No locks are held.
498 */
499static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
500{
501 blk_status_t err = bio->bi_status;
502 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
503 bool should_dirty = dio_op == REQ_OP_READ && dio->should_dirty;
504
505 if (err) {
506 if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
507 dio->io_error = -EAGAIN;
508 else
509 dio->io_error = -EIO;
510 }
511
512 if (dio->is_async && should_dirty) {
513 bio_check_pages_dirty(bio); /* transfers ownership */
514 } else {
515 bio_release_pages(bio, mark_dirty: should_dirty);
516 bio_put(bio);
517 }
518 return err;
519}
520
521/*
522 * Wait on and process all in-flight BIOs. This must only be called once
523 * all bios have been issued so that the refcount can only decrease.
524 * This just waits for all bios to make it through dio_bio_complete. IO
525 * errors are propagated through dio->io_error and should be propagated via
526 * dio_complete().
527 */
528static void dio_await_completion(struct dio *dio)
529{
530 struct bio *bio;
531 do {
532 bio = dio_await_one(dio);
533 if (bio)
534 dio_bio_complete(dio, bio);
535 } while (bio);
536}
537
538/*
539 * A really large O_DIRECT read or write can generate a lot of BIOs. So
540 * to keep the memory consumption sane we periodically reap any completed BIOs
541 * during the BIO generation phase.
542 *
543 * This also helps to limit the peak amount of pinned userspace memory.
544 */
545static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
546{
547 int ret = 0;
548
549 if (sdio->reap_counter++ >= 64) {
550 while (dio->bio_list) {
551 unsigned long flags;
552 struct bio *bio;
553 int ret2;
554
555 spin_lock_irqsave(&dio->bio_lock, flags);
556 bio = dio->bio_list;
557 dio->bio_list = bio->bi_private;
558 spin_unlock_irqrestore(lock: &dio->bio_lock, flags);
559 ret2 = blk_status_to_errno(status: dio_bio_complete(dio, bio));
560 if (ret == 0)
561 ret = ret2;
562 }
563 sdio->reap_counter = 0;
564 }
565 return ret;
566}
567
568static int dio_set_defer_completion(struct dio *dio)
569{
570 struct super_block *sb = dio->inode->i_sb;
571
572 if (dio->defer_completion)
573 return 0;
574 dio->defer_completion = true;
575 if (!sb->s_dio_done_wq)
576 return sb_init_dio_done_wq(sb);
577 return 0;
578}
579
580/*
581 * Call into the fs to map some more disk blocks. We record the current number
582 * of available blocks at sdio->blocks_available. These are in units of the
583 * fs blocksize, i_blocksize(inode).
584 *
585 * The fs is allowed to map lots of blocks at once. If it wants to do that,
586 * it uses the passed inode-relative block number as the file offset, as usual.
587 *
588 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
589 * has remaining to do. The fs should not map more than this number of blocks.
590 *
591 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
592 * indicate how much contiguous disk space has been made available at
593 * bh->b_blocknr.
594 *
595 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
596 * This isn't very efficient...
597 *
598 * In the case of filesystem holes: the fs may return an arbitrarily-large
599 * hole by returning an appropriate value in b_size and by clearing
600 * buffer_mapped(). However the direct-io code will only process holes one
601 * block at a time - it will repeatedly call get_block() as it walks the hole.
602 */
603static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
604 struct buffer_head *map_bh)
605{
606 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
607 int ret;
608 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
609 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
610 unsigned long fs_count; /* Number of filesystem-sized blocks */
611 int create;
612 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
613 loff_t i_size;
614
615 /*
616 * If there was a memory error and we've overwritten all the
617 * mapped blocks then we can now return that memory error
618 */
619 ret = dio->page_errors;
620 if (ret == 0) {
621 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
622 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
623 fs_endblk = (sdio->final_block_in_request - 1) >>
624 sdio->blkfactor;
625 fs_count = fs_endblk - fs_startblk + 1;
626
627 map_bh->b_state = 0;
628 map_bh->b_size = fs_count << i_blkbits;
629
630 /*
631 * For writes that could fill holes inside i_size on a
632 * DIO_SKIP_HOLES filesystem we forbid block creations: only
633 * overwrites are permitted. We will return early to the caller
634 * once we see an unmapped buffer head returned, and the caller
635 * will fall back to buffered I/O.
636 *
637 * Otherwise the decision is left to the get_blocks method,
638 * which may decide to handle it or also return an unmapped
639 * buffer head.
640 */
641 create = dio_op == REQ_OP_WRITE;
642 if (dio->flags & DIO_SKIP_HOLES) {
643 i_size = i_size_read(inode: dio->inode);
644 if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits)
645 create = 0;
646 }
647
648 ret = (*sdio->get_block)(dio->inode, fs_startblk,
649 map_bh, create);
650
651 /* Store for completion */
652 dio->private = map_bh->b_private;
653
654 if (ret == 0 && buffer_defer_completion(bh: map_bh))
655 ret = dio_set_defer_completion(dio);
656 }
657 return ret;
658}
659
660/*
661 * There is no bio. Make one now.
662 */
663static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
664 sector_t start_sector, struct buffer_head *map_bh)
665{
666 sector_t sector;
667 int ret, nr_pages;
668
669 ret = dio_bio_reap(dio, sdio);
670 if (ret)
671 goto out;
672 sector = start_sector << (sdio->blkbits - 9);
673 nr_pages = bio_max_segs(nr_segs: sdio->pages_in_io);
674 BUG_ON(nr_pages <= 0);
675 dio_bio_alloc(dio, sdio, bdev: map_bh->b_bdev, first_sector: sector, nr_vecs: nr_pages);
676 sdio->boundary = 0;
677out:
678 return ret;
679}
680
681/*
682 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
683 * that was successful then update final_block_in_bio and take a ref against
684 * the just-added page.
685 *
686 * Return zero on success. Non-zero means the caller needs to start a new BIO.
687 */
688static inline int dio_bio_add_page(struct dio *dio, struct dio_submit *sdio)
689{
690 int ret;
691
692 ret = bio_add_page(bio: sdio->bio, page: sdio->cur_page,
693 len: sdio->cur_page_len, off: sdio->cur_page_offset);
694 if (ret == sdio->cur_page_len) {
695 /*
696 * Decrement count only, if we are done with this page
697 */
698 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
699 sdio->pages_in_io--;
700 dio_pin_page(dio, page: sdio->cur_page);
701 sdio->final_block_in_bio = sdio->cur_page_block +
702 (sdio->cur_page_len >> sdio->blkbits);
703 ret = 0;
704 } else {
705 ret = 1;
706 }
707 return ret;
708}
709
710/*
711 * Put cur_page under IO. The section of cur_page which is described by
712 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
713 * starts on-disk at cur_page_block.
714 *
715 * We take a ref against the page here (on behalf of its presence in the bio).
716 *
717 * The caller of this function is responsible for removing cur_page from the
718 * dio, and for dropping the refcount which came from that presence.
719 */
720static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
721 struct buffer_head *map_bh)
722{
723 int ret = 0;
724
725 if (sdio->bio) {
726 loff_t cur_offset = sdio->cur_page_fs_offset;
727 loff_t bio_next_offset = sdio->logical_offset_in_bio +
728 sdio->bio->bi_iter.bi_size;
729
730 /*
731 * See whether this new request is contiguous with the old.
732 *
733 * Btrfs cannot handle having logically non-contiguous requests
734 * submitted. For example if you have
735 *
736 * Logical: [0-4095][HOLE][8192-12287]
737 * Physical: [0-4095] [4096-8191]
738 *
739 * We cannot submit those pages together as one BIO. So if our
740 * current logical offset in the file does not equal what would
741 * be the next logical offset in the bio, submit the bio we
742 * have.
743 */
744 if (sdio->final_block_in_bio != sdio->cur_page_block ||
745 cur_offset != bio_next_offset)
746 dio_bio_submit(dio, sdio);
747 }
748
749 if (sdio->bio == NULL) {
750 ret = dio_new_bio(dio, sdio, start_sector: sdio->cur_page_block, map_bh);
751 if (ret)
752 goto out;
753 }
754
755 if (dio_bio_add_page(dio, sdio) != 0) {
756 dio_bio_submit(dio, sdio);
757 ret = dio_new_bio(dio, sdio, start_sector: sdio->cur_page_block, map_bh);
758 if (ret == 0) {
759 ret = dio_bio_add_page(dio, sdio);
760 BUG_ON(ret != 0);
761 }
762 }
763out:
764 return ret;
765}
766
767/*
768 * An autonomous function to put a chunk of a page under deferred IO.
769 *
770 * The caller doesn't actually know (or care) whether this piece of page is in
771 * a BIO, or is under IO or whatever. We just take care of all possible
772 * situations here. The separation between the logic of do_direct_IO() and
773 * that of submit_page_section() is important for clarity. Please don't break.
774 *
775 * The chunk of page starts on-disk at blocknr.
776 *
777 * We perform deferred IO, by recording the last-submitted page inside our
778 * private part of the dio structure. If possible, we just expand the IO
779 * across that page here.
780 *
781 * If that doesn't work out then we put the old page into the bio and add this
782 * page to the dio instead.
783 */
784static inline int
785submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
786 unsigned offset, unsigned len, sector_t blocknr,
787 struct buffer_head *map_bh)
788{
789 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
790 int ret = 0;
791 int boundary = sdio->boundary; /* dio_send_cur_page may clear it */
792
793 if (dio_op == REQ_OP_WRITE) {
794 /*
795 * Read accounting is performed in submit_bio()
796 */
797 task_io_account_write(bytes: len);
798 }
799
800 /*
801 * Can we just grow the current page's presence in the dio?
802 */
803 if (sdio->cur_page == page &&
804 sdio->cur_page_offset + sdio->cur_page_len == offset &&
805 sdio->cur_page_block +
806 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
807 sdio->cur_page_len += len;
808 goto out;
809 }
810
811 /*
812 * If there's a deferred page already there then send it.
813 */
814 if (sdio->cur_page) {
815 ret = dio_send_cur_page(dio, sdio, map_bh);
816 dio_unpin_page(dio, page: sdio->cur_page);
817 sdio->cur_page = NULL;
818 if (ret)
819 return ret;
820 }
821
822 dio_pin_page(dio, page); /* It is in dio */
823 sdio->cur_page = page;
824 sdio->cur_page_offset = offset;
825 sdio->cur_page_len = len;
826 sdio->cur_page_block = blocknr;
827 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
828out:
829 /*
830 * If boundary then we want to schedule the IO now to
831 * avoid metadata seeks.
832 */
833 if (boundary) {
834 ret = dio_send_cur_page(dio, sdio, map_bh);
835 if (sdio->bio)
836 dio_bio_submit(dio, sdio);
837 dio_unpin_page(dio, page: sdio->cur_page);
838 sdio->cur_page = NULL;
839 }
840 return ret;
841}
842
843/*
844 * If we are not writing the entire block and get_block() allocated
845 * the block for us, we need to fill-in the unused portion of the
846 * block with zeros. This happens only if user-buffer, fileoffset or
847 * io length is not filesystem block-size multiple.
848 *
849 * `end' is zero if we're doing the start of the IO, 1 at the end of the
850 * IO.
851 */
852static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
853 int end, struct buffer_head *map_bh)
854{
855 unsigned dio_blocks_per_fs_block;
856 unsigned this_chunk_blocks; /* In dio_blocks */
857 unsigned this_chunk_bytes;
858 struct page *page;
859
860 sdio->start_zero_done = 1;
861 if (!sdio->blkfactor || !buffer_new(bh: map_bh))
862 return;
863
864 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
865 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
866
867 if (!this_chunk_blocks)
868 return;
869
870 /*
871 * We need to zero out part of an fs block. It is either at the
872 * beginning or the end of the fs block.
873 */
874 if (end)
875 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
876
877 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
878
879 page = ZERO_PAGE(0);
880 if (submit_page_section(dio, sdio, page, offset: 0, len: this_chunk_bytes,
881 blocknr: sdio->next_block_for_io, map_bh))
882 return;
883
884 sdio->next_block_for_io += this_chunk_blocks;
885}
886
887/*
888 * Walk the user pages, and the file, mapping blocks to disk and generating
889 * a sequence of (page,offset,len,block) mappings. These mappings are injected
890 * into submit_page_section(), which takes care of the next stage of submission
891 *
892 * Direct IO against a blockdev is different from a file. Because we can
893 * happily perform page-sized but 512-byte aligned IOs. It is important that
894 * blockdev IO be able to have fine alignment and large sizes.
895 *
896 * So what we do is to permit the ->get_block function to populate bh.b_size
897 * with the size of IO which is permitted at this offset and this i_blkbits.
898 *
899 * For best results, the blockdev should be set up with 512-byte i_blkbits and
900 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
901 * fine alignment but still allows this function to work in PAGE_SIZE units.
902 */
903static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
904 struct buffer_head *map_bh)
905{
906 const enum req_op dio_op = dio->opf & REQ_OP_MASK;
907 const unsigned blkbits = sdio->blkbits;
908 const unsigned i_blkbits = blkbits + sdio->blkfactor;
909 int ret = 0;
910
911 while (sdio->block_in_file < sdio->final_block_in_request) {
912 struct page *page;
913 size_t from, to;
914
915 page = dio_get_page(dio, sdio);
916 if (IS_ERR(ptr: page)) {
917 ret = PTR_ERR(ptr: page);
918 goto out;
919 }
920 from = sdio->head ? 0 : sdio->from;
921 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
922 sdio->head++;
923
924 while (from < to) {
925 unsigned this_chunk_bytes; /* # of bytes mapped */
926 unsigned this_chunk_blocks; /* # of blocks */
927 unsigned u;
928
929 if (sdio->blocks_available == 0) {
930 /*
931 * Need to go and map some more disk
932 */
933 unsigned long blkmask;
934 unsigned long dio_remainder;
935
936 ret = get_more_blocks(dio, sdio, map_bh);
937 if (ret) {
938 dio_unpin_page(dio, page);
939 goto out;
940 }
941 if (!buffer_mapped(bh: map_bh))
942 goto do_holes;
943
944 sdio->blocks_available =
945 map_bh->b_size >> blkbits;
946 sdio->next_block_for_io =
947 map_bh->b_blocknr << sdio->blkfactor;
948 if (buffer_new(bh: map_bh)) {
949 clean_bdev_aliases(
950 bdev: map_bh->b_bdev,
951 block: map_bh->b_blocknr,
952 len: map_bh->b_size >> i_blkbits);
953 }
954
955 if (!sdio->blkfactor)
956 goto do_holes;
957
958 blkmask = (1 << sdio->blkfactor) - 1;
959 dio_remainder = (sdio->block_in_file & blkmask);
960
961 /*
962 * If we are at the start of IO and that IO
963 * starts partway into a fs-block,
964 * dio_remainder will be non-zero. If the IO
965 * is a read then we can simply advance the IO
966 * cursor to the first block which is to be
967 * read. But if the IO is a write and the
968 * block was newly allocated we cannot do that;
969 * the start of the fs block must be zeroed out
970 * on-disk
971 */
972 if (!buffer_new(bh: map_bh))
973 sdio->next_block_for_io += dio_remainder;
974 sdio->blocks_available -= dio_remainder;
975 }
976do_holes:
977 /* Handle holes */
978 if (!buffer_mapped(bh: map_bh)) {
979 loff_t i_size_aligned;
980
981 /* AKPM: eargh, -ENOTBLK is a hack */
982 if (dio_op == REQ_OP_WRITE) {
983 dio_unpin_page(dio, page);
984 return -ENOTBLK;
985 }
986
987 /*
988 * Be sure to account for a partial block as the
989 * last block in the file
990 */
991 i_size_aligned = ALIGN(i_size_read(dio->inode),
992 1 << blkbits);
993 if (sdio->block_in_file >=
994 i_size_aligned >> blkbits) {
995 /* We hit eof */
996 dio_unpin_page(dio, page);
997 goto out;
998 }
999 zero_user(page, start: from, size: 1 << blkbits);
1000 sdio->block_in_file++;
1001 from += 1 << blkbits;
1002 dio->result += 1 << blkbits;
1003 goto next_block;
1004 }
1005
1006 /*
1007 * If we're performing IO which has an alignment which
1008 * is finer than the underlying fs, go check to see if
1009 * we must zero out the start of this block.
1010 */
1011 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1012 dio_zero_block(dio, sdio, end: 0, map_bh);
1013
1014 /*
1015 * Work out, in this_chunk_blocks, how much disk we
1016 * can add to this page
1017 */
1018 this_chunk_blocks = sdio->blocks_available;
1019 u = (to - from) >> blkbits;
1020 if (this_chunk_blocks > u)
1021 this_chunk_blocks = u;
1022 u = sdio->final_block_in_request - sdio->block_in_file;
1023 if (this_chunk_blocks > u)
1024 this_chunk_blocks = u;
1025 this_chunk_bytes = this_chunk_blocks << blkbits;
1026 BUG_ON(this_chunk_bytes == 0);
1027
1028 if (this_chunk_blocks == sdio->blocks_available)
1029 sdio->boundary = buffer_boundary(bh: map_bh);
1030 ret = submit_page_section(dio, sdio, page,
1031 offset: from,
1032 len: this_chunk_bytes,
1033 blocknr: sdio->next_block_for_io,
1034 map_bh);
1035 if (ret) {
1036 dio_unpin_page(dio, page);
1037 goto out;
1038 }
1039 sdio->next_block_for_io += this_chunk_blocks;
1040
1041 sdio->block_in_file += this_chunk_blocks;
1042 from += this_chunk_bytes;
1043 dio->result += this_chunk_bytes;
1044 sdio->blocks_available -= this_chunk_blocks;
1045next_block:
1046 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1047 if (sdio->block_in_file == sdio->final_block_in_request)
1048 break;
1049 }
1050
1051 /* Drop the pin which was taken in get_user_pages() */
1052 dio_unpin_page(dio, page);
1053 }
1054out:
1055 return ret;
1056}
1057
1058static inline int drop_refcount(struct dio *dio)
1059{
1060 int ret2;
1061 unsigned long flags;
1062
1063 /*
1064 * Sync will always be dropping the final ref and completing the
1065 * operation. AIO can if it was a broken operation described above or
1066 * in fact if all the bios race to complete before we get here. In
1067 * that case dio_complete() translates the EIOCBQUEUED into the proper
1068 * return code that the caller will hand to ->complete().
1069 *
1070 * This is managed by the bio_lock instead of being an atomic_t so that
1071 * completion paths can drop their ref and use the remaining count to
1072 * decide to wake the submission path atomically.
1073 */
1074 spin_lock_irqsave(&dio->bio_lock, flags);
1075 ret2 = --dio->refcount;
1076 spin_unlock_irqrestore(lock: &dio->bio_lock, flags);
1077 return ret2;
1078}
1079
1080/*
1081 * This is a library function for use by filesystem drivers.
1082 *
1083 * The locking rules are governed by the flags parameter:
1084 * - if the flags value contains DIO_LOCKING we use a fancy locking
1085 * scheme for dumb filesystems.
1086 * For writes this function is called under i_mutex and returns with
1087 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1088 * taken and dropped again before returning.
1089 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1090 * internal locking but rather rely on the filesystem to synchronize
1091 * direct I/O reads/writes versus each other and truncate.
1092 *
1093 * To help with locking against truncate we incremented the i_dio_count
1094 * counter before starting direct I/O, and decrement it once we are done.
1095 * Truncate can wait for it to reach zero to provide exclusion. It is
1096 * expected that filesystem provide exclusion between new direct I/O
1097 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1098 * but other filesystems need to take care of this on their own.
1099 *
1100 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1101 * is always inlined. Otherwise gcc is unable to split the structure into
1102 * individual fields and will generate much worse code. This is important
1103 * for the whole file.
1104 */
1105ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1106 struct block_device *bdev, struct iov_iter *iter,
1107 get_block_t get_block, dio_iodone_t end_io,
1108 int flags)
1109{
1110 unsigned i_blkbits = READ_ONCE(inode->i_blkbits);
1111 unsigned blkbits = i_blkbits;
1112 unsigned blocksize_mask = (1 << blkbits) - 1;
1113 ssize_t retval = -EINVAL;
1114 const size_t count = iov_iter_count(i: iter);
1115 loff_t offset = iocb->ki_pos;
1116 const loff_t end = offset + count;
1117 struct dio *dio;
1118 struct dio_submit sdio = { NULL, };
1119 struct buffer_head map_bh = { 0, };
1120 struct blk_plug plug;
1121 unsigned long align = offset | iov_iter_alignment(i: iter);
1122
1123 /* watch out for a 0 len io from a tricksy fs */
1124 if (iov_iter_rw(i: iter) == READ && !count)
1125 return 0;
1126
1127 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1128 if (!dio)
1129 return -ENOMEM;
1130 /*
1131 * Believe it or not, zeroing out the page array caused a .5%
1132 * performance regression in a database benchmark. So, we take
1133 * care to only zero out what's needed.
1134 */
1135 memset(dio, 0, offsetof(struct dio, pages));
1136
1137 dio->flags = flags;
1138 if (dio->flags & DIO_LOCKING && iov_iter_rw(i: iter) == READ) {
1139 /* will be released by direct_io_worker */
1140 inode_lock(inode);
1141 }
1142 dio->is_pinned = iov_iter_extract_will_pin(iter);
1143
1144 /* Once we sampled i_size check for reads beyond EOF */
1145 dio->i_size = i_size_read(inode);
1146 if (iov_iter_rw(i: iter) == READ && offset >= dio->i_size) {
1147 retval = 0;
1148 goto fail_dio;
1149 }
1150
1151 if (align & blocksize_mask) {
1152 if (bdev)
1153 blkbits = blksize_bits(size: bdev_logical_block_size(bdev));
1154 blocksize_mask = (1 << blkbits) - 1;
1155 if (align & blocksize_mask)
1156 goto fail_dio;
1157 }
1158
1159 if (dio->flags & DIO_LOCKING && iov_iter_rw(i: iter) == READ) {
1160 struct address_space *mapping = iocb->ki_filp->f_mapping;
1161
1162 retval = filemap_write_and_wait_range(mapping, lstart: offset, lend: end - 1);
1163 if (retval)
1164 goto fail_dio;
1165 }
1166
1167 /*
1168 * For file extending writes updating i_size before data writeouts
1169 * complete can expose uninitialized blocks in dumb filesystems.
1170 * In that case we need to wait for I/O completion even if asked
1171 * for an asynchronous write.
1172 */
1173 if (is_sync_kiocb(kiocb: iocb))
1174 dio->is_async = false;
1175 else if (iov_iter_rw(i: iter) == WRITE && end > i_size_read(inode))
1176 dio->is_async = false;
1177 else
1178 dio->is_async = true;
1179
1180 dio->inode = inode;
1181 if (iov_iter_rw(i: iter) == WRITE) {
1182 dio->opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
1183 if (iocb->ki_flags & IOCB_NOWAIT)
1184 dio->opf |= REQ_NOWAIT;
1185 } else {
1186 dio->opf = REQ_OP_READ;
1187 }
1188
1189 /*
1190 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1191 * so that we can call ->fsync.
1192 */
1193 if (dio->is_async && iov_iter_rw(i: iter) == WRITE) {
1194 retval = 0;
1195 if (iocb_is_dsync(iocb))
1196 retval = dio_set_defer_completion(dio);
1197 else if (!dio->inode->i_sb->s_dio_done_wq) {
1198 /*
1199 * In case of AIO write racing with buffered read we
1200 * need to defer completion. We can't decide this now,
1201 * however the workqueue needs to be initialized here.
1202 */
1203 retval = sb_init_dio_done_wq(sb: dio->inode->i_sb);
1204 }
1205 if (retval)
1206 goto fail_dio;
1207 }
1208
1209 /*
1210 * Will be decremented at I/O completion time.
1211 */
1212 inode_dio_begin(inode);
1213
1214 sdio.blkbits = blkbits;
1215 sdio.blkfactor = i_blkbits - blkbits;
1216 sdio.block_in_file = offset >> blkbits;
1217
1218 sdio.get_block = get_block;
1219 dio->end_io = end_io;
1220 sdio.final_block_in_bio = -1;
1221 sdio.next_block_for_io = -1;
1222
1223 dio->iocb = iocb;
1224
1225 spin_lock_init(&dio->bio_lock);
1226 dio->refcount = 1;
1227
1228 dio->should_dirty = user_backed_iter(i: iter) && iov_iter_rw(i: iter) == READ;
1229 sdio.iter = iter;
1230 sdio.final_block_in_request = end >> blkbits;
1231
1232 /*
1233 * In case of non-aligned buffers, we may need 2 more
1234 * pages since we need to zero out first and last block.
1235 */
1236 if (unlikely(sdio.blkfactor))
1237 sdio.pages_in_io = 2;
1238
1239 sdio.pages_in_io += iov_iter_npages(i: iter, INT_MAX);
1240
1241 blk_start_plug(&plug);
1242
1243 retval = do_direct_IO(dio, sdio: &sdio, map_bh: &map_bh);
1244 if (retval)
1245 dio_cleanup(dio, sdio: &sdio);
1246
1247 if (retval == -ENOTBLK) {
1248 /*
1249 * The remaining part of the request will be
1250 * handled by buffered I/O when we return
1251 */
1252 retval = 0;
1253 }
1254 /*
1255 * There may be some unwritten disk at the end of a part-written
1256 * fs-block-sized block. Go zero that now.
1257 */
1258 dio_zero_block(dio, sdio: &sdio, end: 1, map_bh: &map_bh);
1259
1260 if (sdio.cur_page) {
1261 ssize_t ret2;
1262
1263 ret2 = dio_send_cur_page(dio, sdio: &sdio, map_bh: &map_bh);
1264 if (retval == 0)
1265 retval = ret2;
1266 dio_unpin_page(dio, page: sdio.cur_page);
1267 sdio.cur_page = NULL;
1268 }
1269 if (sdio.bio)
1270 dio_bio_submit(dio, sdio: &sdio);
1271
1272 blk_finish_plug(&plug);
1273
1274 /*
1275 * It is possible that, we return short IO due to end of file.
1276 * In that case, we need to release all the pages we got hold on.
1277 */
1278 dio_cleanup(dio, sdio: &sdio);
1279
1280 /*
1281 * All block lookups have been performed. For READ requests
1282 * we can let i_mutex go now that its achieved its purpose
1283 * of protecting us from looking up uninitialized blocks.
1284 */
1285 if (iov_iter_rw(i: iter) == READ && (dio->flags & DIO_LOCKING))
1286 inode_unlock(inode: dio->inode);
1287
1288 /*
1289 * The only time we want to leave bios in flight is when a successful
1290 * partial aio read or full aio write have been setup. In that case
1291 * bio completion will call aio_complete. The only time it's safe to
1292 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1293 * This had *better* be the only place that raises -EIOCBQUEUED.
1294 */
1295 BUG_ON(retval == -EIOCBQUEUED);
1296 if (dio->is_async && retval == 0 && dio->result &&
1297 (iov_iter_rw(i: iter) == READ || dio->result == count))
1298 retval = -EIOCBQUEUED;
1299 else
1300 dio_await_completion(dio);
1301
1302 if (drop_refcount(dio) == 0) {
1303 retval = dio_complete(dio, ret: retval, DIO_COMPLETE_INVALIDATE);
1304 } else
1305 BUG_ON(retval != -EIOCBQUEUED);
1306
1307 return retval;
1308
1309fail_dio:
1310 if (dio->flags & DIO_LOCKING && iov_iter_rw(i: iter) == READ)
1311 inode_unlock(inode);
1312
1313 kmem_cache_free(s: dio_cache, objp: dio);
1314 return retval;
1315}
1316EXPORT_SYMBOL(__blockdev_direct_IO);
1317
1318static __init int dio_init(void)
1319{
1320 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1321 return 0;
1322}
1323module_init(dio_init)
1324

Provided by KDAB

Privacy Policy
Improve your Profiling and Debugging skills
Find out more

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