Line data Source code
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 get_user_pages(). This determines
46 : * 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 : dio_submit_t *submit_io; /* IO submition function */
90 :
91 : loff_t logical_offset_in_bio; /* current first logical block in bio */
92 : sector_t final_block_in_bio; /* current final block in bio + 1 */
93 : sector_t next_block_for_io; /* next block to be put under IO,
94 : in dio_blocks units */
95 :
96 : /*
97 : * Deferred addition of a page to the dio. These variables are
98 : * private to dio_send_cur_page(), submit_page_section() and
99 : * dio_bio_add_page().
100 : */
101 : struct page *cur_page; /* The page */
102 : unsigned cur_page_offset; /* Offset into it, in bytes */
103 : unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
104 : sector_t cur_page_block; /* Where it starts */
105 : loff_t cur_page_fs_offset; /* Offset in file */
106 :
107 : struct iov_iter *iter;
108 : /*
109 : * Page queue. These variables belong to dio_refill_pages() and
110 : * dio_get_page().
111 : */
112 : unsigned head; /* next page to process */
113 : unsigned tail; /* last valid page + 1 */
114 : size_t from, to;
115 : };
116 :
117 : /* dio_state communicated between submission path and end_io */
118 : struct dio {
119 : int flags; /* doesn't change */
120 : int op;
121 : int op_flags;
122 : struct gendisk *bio_disk;
123 : struct inode *inode;
124 : loff_t i_size; /* i_size when submitted */
125 : dio_iodone_t *end_io; /* IO completion function */
126 :
127 : void *private; /* copy from map_bh.b_private */
128 :
129 : /* BIO completion state */
130 : spinlock_t bio_lock; /* protects BIO fields below */
131 : int page_errors; /* errno from get_user_pages() */
132 : int is_async; /* is IO async ? */
133 : bool defer_completion; /* defer AIO completion to workqueue? */
134 : bool should_dirty; /* if pages should be dirtied */
135 : int io_error; /* IO error in completion path */
136 : unsigned long refcount; /* direct_io_worker() and bios */
137 : struct bio *bio_list; /* singly linked via bi_private */
138 : struct task_struct *waiter; /* waiting task (NULL if none) */
139 :
140 : /* AIO related stuff */
141 : struct kiocb *iocb; /* kiocb */
142 : ssize_t result; /* IO result */
143 :
144 : /*
145 : * pages[] (and any fields placed after it) are not zeroed out at
146 : * allocation time. Don't add new fields after pages[] unless you
147 : * wish that they not be zeroed.
148 : */
149 : union {
150 : struct page *pages[DIO_PAGES]; /* page buffer */
151 : struct work_struct complete_work;/* deferred AIO completion */
152 : };
153 : } ____cacheline_aligned_in_smp;
154 :
155 : static struct kmem_cache *dio_cache __read_mostly;
156 :
157 : /*
158 : * How many pages are in the queue?
159 : */
160 : static inline unsigned dio_pages_present(struct dio_submit *sdio)
161 : {
162 : return sdio->tail - sdio->head;
163 : }
164 :
165 : /*
166 : * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
167 : */
168 0 : static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
169 : {
170 : ssize_t ret;
171 :
172 0 : ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
173 : &sdio->from);
174 :
175 0 : if (ret < 0 && sdio->blocks_available && (dio->op == REQ_OP_WRITE)) {
176 0 : struct page *page = ZERO_PAGE(0);
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 0 : if (dio->page_errors == 0)
183 0 : dio->page_errors = ret;
184 0 : get_page(page);
185 0 : dio->pages[0] = page;
186 0 : sdio->head = 0;
187 0 : sdio->tail = 1;
188 0 : sdio->from = 0;
189 0 : sdio->to = PAGE_SIZE;
190 0 : return 0;
191 : }
192 :
193 0 : if (ret >= 0) {
194 0 : iov_iter_advance(sdio->iter, ret);
195 0 : ret += sdio->from;
196 0 : sdio->head = 0;
197 0 : sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
198 0 : sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
199 0 : return 0;
200 : }
201 0 : return ret;
202 : }
203 :
204 : /*
205 : * Get another userspace page. Returns an ERR_PTR on error. Pages are
206 : * buffered inside the dio so that we can call get_user_pages() against a
207 : * decent number of pages, less frequently. To provide nicer use of the
208 : * L1 cache.
209 : */
210 0 : static inline struct page *dio_get_page(struct dio *dio,
211 : struct dio_submit *sdio)
212 : {
213 0 : if (dio_pages_present(sdio) == 0) {
214 : int ret;
215 :
216 0 : ret = dio_refill_pages(dio, sdio);
217 0 : if (ret)
218 0 : return ERR_PTR(ret);
219 0 : BUG_ON(dio_pages_present(sdio) == 0);
220 : }
221 0 : return dio->pages[sdio->head];
222 : }
223 :
224 : /*
225 : * dio_complete() - called when all DIO BIO I/O has been completed
226 : *
227 : * This drops i_dio_count, lets interested parties know that a DIO operation
228 : * has completed, and calculates the resulting return code for the operation.
229 : *
230 : * It lets the filesystem know if it registered an interest earlier via
231 : * get_block. Pass the private field of the map buffer_head so that
232 : * filesystems can use it to hold additional state between get_block calls and
233 : * dio_complete.
234 : */
235 0 : static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags)
236 : {
237 0 : loff_t offset = dio->iocb->ki_pos;
238 0 : ssize_t transferred = 0;
239 : int err;
240 :
241 : /*
242 : * AIO submission can race with bio completion to get here while
243 : * expecting to have the last io completed by bio completion.
244 : * In that case -EIOCBQUEUED is in fact not an error we want
245 : * to preserve through this call.
246 : */
247 0 : if (ret == -EIOCBQUEUED)
248 0 : ret = 0;
249 :
250 0 : if (dio->result) {
251 0 : transferred = dio->result;
252 :
253 : /* Check for short read case */
254 0 : if ((dio->op == REQ_OP_READ) &&
255 0 : ((offset + transferred) > dio->i_size))
256 0 : transferred = dio->i_size - offset;
257 : /* ignore EFAULT if some IO has been done */
258 0 : if (unlikely(ret == -EFAULT) && transferred)
259 0 : ret = 0;
260 : }
261 :
262 0 : if (ret == 0)
263 0 : ret = dio->page_errors;
264 0 : if (ret == 0)
265 0 : ret = dio->io_error;
266 0 : if (ret == 0)
267 0 : ret = transferred;
268 :
269 0 : if (dio->end_io) {
270 : // XXX: ki_pos??
271 0 : err = dio->end_io(dio->iocb, offset, ret, dio->private);
272 0 : if (err)
273 0 : ret = err;
274 : }
275 :
276 : /*
277 : * Try again to invalidate clean pages which might have been cached by
278 : * non-direct readahead, or faulted in by get_user_pages() if the source
279 : * of the write was an mmap'ed region of the file we're writing. Either
280 : * one is a pretty crazy thing to do, so we don't support it 100%. If
281 : * this invalidation fails, tough, the write still worked...
282 : *
283 : * And this page cache invalidation has to be after dio->end_io(), as
284 : * some filesystems convert unwritten extents to real allocations in
285 : * end_io() when necessary, otherwise a racing buffer read would cache
286 : * zeros from unwritten extents.
287 : */
288 0 : if (flags & DIO_COMPLETE_INVALIDATE &&
289 0 : ret > 0 && dio->op == REQ_OP_WRITE &&
290 0 : dio->inode->i_mapping->nrpages) {
291 0 : err = invalidate_inode_pages2_range(dio->inode->i_mapping,
292 0 : offset >> PAGE_SHIFT,
293 0 : (offset + ret - 1) >> PAGE_SHIFT);
294 0 : if (err)
295 0 : dio_warn_stale_pagecache(dio->iocb->ki_filp);
296 : }
297 :
298 0 : inode_dio_end(dio->inode);
299 :
300 0 : if (flags & DIO_COMPLETE_ASYNC) {
301 : /*
302 : * generic_write_sync expects ki_pos to have been updated
303 : * already, but the submission path only does this for
304 : * synchronous I/O.
305 : */
306 0 : dio->iocb->ki_pos += transferred;
307 :
308 0 : if (ret > 0 && dio->op == REQ_OP_WRITE)
309 0 : ret = generic_write_sync(dio->iocb, ret);
310 0 : dio->iocb->ki_complete(dio->iocb, ret);
311 : }
312 :
313 0 : kmem_cache_free(dio_cache, dio);
314 0 : return ret;
315 : }
316 :
317 0 : static void dio_aio_complete_work(struct work_struct *work)
318 : {
319 0 : struct dio *dio = container_of(work, struct dio, complete_work);
320 :
321 0 : dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
322 0 : }
323 :
324 : static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
325 :
326 : /*
327 : * Asynchronous IO callback.
328 : */
329 0 : static void dio_bio_end_aio(struct bio *bio)
330 : {
331 0 : struct dio *dio = bio->bi_private;
332 : unsigned long remaining;
333 : unsigned long flags;
334 0 : bool defer_completion = false;
335 :
336 : /* cleanup the bio */
337 0 : dio_bio_complete(dio, bio);
338 :
339 0 : spin_lock_irqsave(&dio->bio_lock, flags);
340 0 : remaining = --dio->refcount;
341 0 : if (remaining == 1 && dio->waiter)
342 0 : wake_up_process(dio->waiter);
343 0 : spin_unlock_irqrestore(&dio->bio_lock, flags);
344 :
345 0 : if (remaining == 0) {
346 : /*
347 : * Defer completion when defer_completion is set or
348 : * when the inode has pages mapped and this is AIO write.
349 : * We need to invalidate those pages because there is a
350 : * chance they contain stale data in the case buffered IO
351 : * went in between AIO submission and completion into the
352 : * same region.
353 : */
354 0 : if (dio->result)
355 0 : defer_completion = dio->defer_completion ||
356 0 : (dio->op == REQ_OP_WRITE &&
357 0 : dio->inode->i_mapping->nrpages);
358 0 : if (defer_completion) {
359 0 : INIT_WORK(&dio->complete_work, dio_aio_complete_work);
360 0 : queue_work(dio->inode->i_sb->s_dio_done_wq,
361 : &dio->complete_work);
362 : } else {
363 0 : dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
364 : }
365 : }
366 0 : }
367 :
368 : /*
369 : * The BIO completion handler simply queues the BIO up for the process-context
370 : * handler.
371 : *
372 : * During I/O bi_private points at the dio. After I/O, bi_private is used to
373 : * implement a singly-linked list of completed BIOs, at dio->bio_list.
374 : */
375 0 : static void dio_bio_end_io(struct bio *bio)
376 : {
377 0 : struct dio *dio = bio->bi_private;
378 : unsigned long flags;
379 :
380 0 : spin_lock_irqsave(&dio->bio_lock, flags);
381 0 : bio->bi_private = dio->bio_list;
382 0 : dio->bio_list = bio;
383 0 : if (--dio->refcount == 1 && dio->waiter)
384 0 : wake_up_process(dio->waiter);
385 0 : spin_unlock_irqrestore(&dio->bio_lock, flags);
386 0 : }
387 :
388 : static inline void
389 0 : dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
390 : struct block_device *bdev,
391 : sector_t first_sector, int nr_vecs)
392 : {
393 : struct bio *bio;
394 :
395 : /*
396 : * bio_alloc() is guaranteed to return a bio when allowed to sleep and
397 : * we request a valid number of vectors.
398 : */
399 0 : bio = bio_alloc(bdev, nr_vecs, dio->op | dio->op_flags, GFP_KERNEL);
400 0 : bio->bi_iter.bi_sector = first_sector;
401 0 : if (dio->is_async)
402 0 : bio->bi_end_io = dio_bio_end_aio;
403 : else
404 0 : bio->bi_end_io = dio_bio_end_io;
405 0 : sdio->bio = bio;
406 0 : sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
407 0 : }
408 :
409 : /*
410 : * In the AIO read case we speculatively dirty the pages before starting IO.
411 : * During IO completion, any of these pages which happen to have been written
412 : * back will be redirtied by bio_check_pages_dirty().
413 : *
414 : * bios hold a dio reference between submit_bio and ->end_io.
415 : */
416 0 : static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
417 : {
418 0 : struct bio *bio = sdio->bio;
419 : unsigned long flags;
420 :
421 0 : bio->bi_private = dio;
422 : /* don't account direct I/O as memory stall */
423 0 : bio_clear_flag(bio, BIO_WORKINGSET);
424 :
425 0 : spin_lock_irqsave(&dio->bio_lock, flags);
426 0 : dio->refcount++;
427 0 : spin_unlock_irqrestore(&dio->bio_lock, flags);
428 :
429 0 : if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty)
430 0 : bio_set_pages_dirty(bio);
431 :
432 0 : dio->bio_disk = bio->bi_bdev->bd_disk;
433 :
434 0 : if (sdio->submit_io)
435 0 : sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio);
436 : else
437 0 : submit_bio(bio);
438 :
439 0 : sdio->bio = NULL;
440 0 : sdio->boundary = 0;
441 0 : sdio->logical_offset_in_bio = 0;
442 0 : }
443 :
444 : /*
445 : * Release any resources in case of a failure
446 : */
447 : static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
448 : {
449 0 : while (sdio->head < sdio->tail)
450 0 : put_page(dio->pages[sdio->head++]);
451 : }
452 :
453 : /*
454 : * Wait for the next BIO to complete. Remove it and return it. NULL is
455 : * returned once all BIOs have been completed. This must only be called once
456 : * all bios have been issued so that dio->refcount can only decrease. This
457 : * requires that the caller hold a reference on the dio.
458 : */
459 0 : static struct bio *dio_await_one(struct dio *dio)
460 : {
461 : unsigned long flags;
462 0 : struct bio *bio = NULL;
463 :
464 0 : spin_lock_irqsave(&dio->bio_lock, flags);
465 :
466 : /*
467 : * Wait as long as the list is empty and there are bios in flight. bio
468 : * completion drops the count, maybe adds to the list, and wakes while
469 : * holding the bio_lock so we don't need set_current_state()'s barrier
470 : * and can call it after testing our condition.
471 : */
472 0 : while (dio->refcount > 1 && dio->bio_list == NULL) {
473 0 : __set_current_state(TASK_UNINTERRUPTIBLE);
474 0 : dio->waiter = current;
475 0 : spin_unlock_irqrestore(&dio->bio_lock, flags);
476 0 : blk_io_schedule();
477 : /* wake up sets us TASK_RUNNING */
478 0 : spin_lock_irqsave(&dio->bio_lock, flags);
479 0 : dio->waiter = NULL;
480 : }
481 0 : if (dio->bio_list) {
482 0 : bio = dio->bio_list;
483 0 : dio->bio_list = bio->bi_private;
484 : }
485 0 : spin_unlock_irqrestore(&dio->bio_lock, flags);
486 0 : return bio;
487 : }
488 :
489 : /*
490 : * Process one completed BIO. No locks are held.
491 : */
492 0 : static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
493 : {
494 0 : blk_status_t err = bio->bi_status;
495 0 : bool should_dirty = dio->op == REQ_OP_READ && dio->should_dirty;
496 :
497 0 : if (err) {
498 0 : if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
499 0 : dio->io_error = -EAGAIN;
500 : else
501 0 : dio->io_error = -EIO;
502 : }
503 :
504 0 : if (dio->is_async && should_dirty) {
505 0 : bio_check_pages_dirty(bio); /* transfers ownership */
506 : } else {
507 0 : bio_release_pages(bio, should_dirty);
508 0 : bio_put(bio);
509 : }
510 0 : return err;
511 : }
512 :
513 : /*
514 : * Wait on and process all in-flight BIOs. This must only be called once
515 : * all bios have been issued so that the refcount can only decrease.
516 : * This just waits for all bios to make it through dio_bio_complete. IO
517 : * errors are propagated through dio->io_error and should be propagated via
518 : * dio_complete().
519 : */
520 0 : static void dio_await_completion(struct dio *dio)
521 : {
522 : struct bio *bio;
523 : do {
524 0 : bio = dio_await_one(dio);
525 0 : if (bio)
526 0 : dio_bio_complete(dio, bio);
527 0 : } while (bio);
528 0 : }
529 :
530 : /*
531 : * A really large O_DIRECT read or write can generate a lot of BIOs. So
532 : * to keep the memory consumption sane we periodically reap any completed BIOs
533 : * during the BIO generation phase.
534 : *
535 : * This also helps to limit the peak amount of pinned userspace memory.
536 : */
537 0 : static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
538 : {
539 0 : int ret = 0;
540 :
541 0 : if (sdio->reap_counter++ >= 64) {
542 0 : while (dio->bio_list) {
543 : unsigned long flags;
544 : struct bio *bio;
545 : int ret2;
546 :
547 0 : spin_lock_irqsave(&dio->bio_lock, flags);
548 0 : bio = dio->bio_list;
549 0 : dio->bio_list = bio->bi_private;
550 0 : spin_unlock_irqrestore(&dio->bio_lock, flags);
551 0 : ret2 = blk_status_to_errno(dio_bio_complete(dio, bio));
552 0 : if (ret == 0)
553 0 : ret = ret2;
554 : }
555 0 : sdio->reap_counter = 0;
556 : }
557 0 : return ret;
558 : }
559 :
560 : /*
561 : * Create workqueue for deferred direct IO completions. We allocate the
562 : * workqueue when it's first needed. This avoids creating workqueue for
563 : * filesystems that don't need it and also allows us to create the workqueue
564 : * late enough so the we can include s_id in the name of the workqueue.
565 : */
566 0 : int sb_init_dio_done_wq(struct super_block *sb)
567 : {
568 : struct workqueue_struct *old;
569 0 : struct workqueue_struct *wq = alloc_workqueue("dio/%s",
570 : WQ_MEM_RECLAIM, 0,
571 0 : sb->s_id);
572 0 : if (!wq)
573 : return -ENOMEM;
574 : /*
575 : * This has to be atomic as more DIOs can race to create the workqueue
576 : */
577 0 : old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
578 : /* Someone created workqueue before us? Free ours... */
579 0 : if (old)
580 0 : destroy_workqueue(wq);
581 : return 0;
582 : }
583 :
584 : static int dio_set_defer_completion(struct dio *dio)
585 : {
586 0 : struct super_block *sb = dio->inode->i_sb;
587 :
588 0 : if (dio->defer_completion)
589 : return 0;
590 0 : dio->defer_completion = true;
591 0 : if (!sb->s_dio_done_wq)
592 0 : return sb_init_dio_done_wq(sb);
593 : return 0;
594 : }
595 :
596 : /*
597 : * Call into the fs to map some more disk blocks. We record the current number
598 : * of available blocks at sdio->blocks_available. These are in units of the
599 : * fs blocksize, i_blocksize(inode).
600 : *
601 : * The fs is allowed to map lots of blocks at once. If it wants to do that,
602 : * it uses the passed inode-relative block number as the file offset, as usual.
603 : *
604 : * get_block() is passed the number of i_blkbits-sized blocks which direct_io
605 : * has remaining to do. The fs should not map more than this number of blocks.
606 : *
607 : * If the fs has mapped a lot of blocks, it should populate bh->b_size to
608 : * indicate how much contiguous disk space has been made available at
609 : * bh->b_blocknr.
610 : *
611 : * If *any* of the mapped blocks are new, then the fs must set buffer_new().
612 : * This isn't very efficient...
613 : *
614 : * In the case of filesystem holes: the fs may return an arbitrarily-large
615 : * hole by returning an appropriate value in b_size and by clearing
616 : * buffer_mapped(). However the direct-io code will only process holes one
617 : * block at a time - it will repeatedly call get_block() as it walks the hole.
618 : */
619 0 : static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
620 : struct buffer_head *map_bh)
621 : {
622 : int ret;
623 : sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
624 : sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
625 : unsigned long fs_count; /* Number of filesystem-sized blocks */
626 : int create;
627 0 : unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
628 : loff_t i_size;
629 :
630 : /*
631 : * If there was a memory error and we've overwritten all the
632 : * mapped blocks then we can now return that memory error
633 : */
634 0 : ret = dio->page_errors;
635 0 : if (ret == 0) {
636 0 : BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
637 0 : fs_startblk = sdio->block_in_file >> sdio->blkfactor;
638 0 : fs_endblk = (sdio->final_block_in_request - 1) >>
639 : sdio->blkfactor;
640 0 : fs_count = fs_endblk - fs_startblk + 1;
641 :
642 0 : map_bh->b_state = 0;
643 0 : map_bh->b_size = fs_count << i_blkbits;
644 :
645 : /*
646 : * For writes that could fill holes inside i_size on a
647 : * DIO_SKIP_HOLES filesystem we forbid block creations: only
648 : * overwrites are permitted. We will return early to the caller
649 : * once we see an unmapped buffer head returned, and the caller
650 : * will fall back to buffered I/O.
651 : *
652 : * Otherwise the decision is left to the get_blocks method,
653 : * which may decide to handle it or also return an unmapped
654 : * buffer head.
655 : */
656 0 : create = dio->op == REQ_OP_WRITE;
657 0 : if (dio->flags & DIO_SKIP_HOLES) {
658 0 : i_size = i_size_read(dio->inode);
659 0 : if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits)
660 0 : create = 0;
661 : }
662 :
663 0 : ret = (*sdio->get_block)(dio->inode, fs_startblk,
664 : map_bh, create);
665 :
666 : /* Store for completion */
667 0 : dio->private = map_bh->b_private;
668 :
669 0 : if (ret == 0 && buffer_defer_completion(map_bh))
670 0 : ret = dio_set_defer_completion(dio);
671 : }
672 0 : return ret;
673 : }
674 :
675 : /*
676 : * There is no bio. Make one now.
677 : */
678 0 : static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
679 : sector_t start_sector, struct buffer_head *map_bh)
680 : {
681 : sector_t sector;
682 : int ret, nr_pages;
683 :
684 0 : ret = dio_bio_reap(dio, sdio);
685 0 : if (ret)
686 : goto out;
687 0 : sector = start_sector << (sdio->blkbits - 9);
688 0 : nr_pages = bio_max_segs(sdio->pages_in_io);
689 0 : BUG_ON(nr_pages <= 0);
690 0 : dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
691 0 : sdio->boundary = 0;
692 : out:
693 0 : return ret;
694 : }
695 :
696 : /*
697 : * Attempt to put the current chunk of 'cur_page' into the current BIO. If
698 : * that was successful then update final_block_in_bio and take a ref against
699 : * the just-added page.
700 : *
701 : * Return zero on success. Non-zero means the caller needs to start a new BIO.
702 : */
703 0 : static inline int dio_bio_add_page(struct dio_submit *sdio)
704 : {
705 : int ret;
706 :
707 0 : ret = bio_add_page(sdio->bio, sdio->cur_page,
708 : sdio->cur_page_len, sdio->cur_page_offset);
709 0 : if (ret == sdio->cur_page_len) {
710 : /*
711 : * Decrement count only, if we are done with this page
712 : */
713 0 : if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
714 0 : sdio->pages_in_io--;
715 0 : get_page(sdio->cur_page);
716 0 : sdio->final_block_in_bio = sdio->cur_page_block +
717 0 : (sdio->cur_page_len >> sdio->blkbits);
718 0 : ret = 0;
719 : } else {
720 : ret = 1;
721 : }
722 0 : return ret;
723 : }
724 :
725 : /*
726 : * Put cur_page under IO. The section of cur_page which is described by
727 : * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
728 : * starts on-disk at cur_page_block.
729 : *
730 : * We take a ref against the page here (on behalf of its presence in the bio).
731 : *
732 : * The caller of this function is responsible for removing cur_page from the
733 : * dio, and for dropping the refcount which came from that presence.
734 : */
735 0 : static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
736 : struct buffer_head *map_bh)
737 : {
738 0 : int ret = 0;
739 :
740 0 : if (sdio->bio) {
741 0 : loff_t cur_offset = sdio->cur_page_fs_offset;
742 0 : loff_t bio_next_offset = sdio->logical_offset_in_bio +
743 0 : sdio->bio->bi_iter.bi_size;
744 :
745 : /*
746 : * See whether this new request is contiguous with the old.
747 : *
748 : * Btrfs cannot handle having logically non-contiguous requests
749 : * submitted. For example if you have
750 : *
751 : * Logical: [0-4095][HOLE][8192-12287]
752 : * Physical: [0-4095] [4096-8191]
753 : *
754 : * We cannot submit those pages together as one BIO. So if our
755 : * current logical offset in the file does not equal what would
756 : * be the next logical offset in the bio, submit the bio we
757 : * have.
758 : */
759 0 : if (sdio->final_block_in_bio != sdio->cur_page_block ||
760 : cur_offset != bio_next_offset)
761 0 : dio_bio_submit(dio, sdio);
762 : }
763 :
764 0 : if (sdio->bio == NULL) {
765 0 : ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
766 0 : if (ret)
767 : goto out;
768 : }
769 :
770 0 : if (dio_bio_add_page(sdio) != 0) {
771 0 : dio_bio_submit(dio, sdio);
772 0 : ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
773 0 : if (ret == 0) {
774 0 : ret = dio_bio_add_page(sdio);
775 0 : BUG_ON(ret != 0);
776 : }
777 : }
778 : out:
779 0 : return ret;
780 : }
781 :
782 : /*
783 : * An autonomous function to put a chunk of a page under deferred IO.
784 : *
785 : * The caller doesn't actually know (or care) whether this piece of page is in
786 : * a BIO, or is under IO or whatever. We just take care of all possible
787 : * situations here. The separation between the logic of do_direct_IO() and
788 : * that of submit_page_section() is important for clarity. Please don't break.
789 : *
790 : * The chunk of page starts on-disk at blocknr.
791 : *
792 : * We perform deferred IO, by recording the last-submitted page inside our
793 : * private part of the dio structure. If possible, we just expand the IO
794 : * across that page here.
795 : *
796 : * If that doesn't work out then we put the old page into the bio and add this
797 : * page to the dio instead.
798 : */
799 : static inline int
800 0 : submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
801 : unsigned offset, unsigned len, sector_t blocknr,
802 : struct buffer_head *map_bh)
803 : {
804 0 : int ret = 0;
805 0 : int boundary = sdio->boundary; /* dio_send_cur_page may clear it */
806 :
807 : if (dio->op == REQ_OP_WRITE) {
808 : /*
809 : * Read accounting is performed in submit_bio()
810 : */
811 : task_io_account_write(len);
812 : }
813 :
814 : /*
815 : * Can we just grow the current page's presence in the dio?
816 : */
817 0 : if (sdio->cur_page == page &&
818 0 : sdio->cur_page_offset + sdio->cur_page_len == offset &&
819 0 : sdio->cur_page_block +
820 0 : (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
821 0 : sdio->cur_page_len += len;
822 0 : goto out;
823 : }
824 :
825 : /*
826 : * If there's a deferred page already there then send it.
827 : */
828 0 : if (sdio->cur_page) {
829 0 : ret = dio_send_cur_page(dio, sdio, map_bh);
830 0 : put_page(sdio->cur_page);
831 0 : sdio->cur_page = NULL;
832 0 : if (ret)
833 : return ret;
834 : }
835 :
836 0 : get_page(page); /* It is in dio */
837 0 : sdio->cur_page = page;
838 0 : sdio->cur_page_offset = offset;
839 0 : sdio->cur_page_len = len;
840 0 : sdio->cur_page_block = blocknr;
841 0 : sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
842 : out:
843 : /*
844 : * If boundary then we want to schedule the IO now to
845 : * avoid metadata seeks.
846 : */
847 0 : if (boundary) {
848 0 : ret = dio_send_cur_page(dio, sdio, map_bh);
849 0 : if (sdio->bio)
850 0 : dio_bio_submit(dio, sdio);
851 0 : put_page(sdio->cur_page);
852 0 : sdio->cur_page = NULL;
853 : }
854 : return ret;
855 : }
856 :
857 : /*
858 : * If we are not writing the entire block and get_block() allocated
859 : * the block for us, we need to fill-in the unused portion of the
860 : * block with zeros. This happens only if user-buffer, fileoffset or
861 : * io length is not filesystem block-size multiple.
862 : *
863 : * `end' is zero if we're doing the start of the IO, 1 at the end of the
864 : * IO.
865 : */
866 0 : static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
867 : int end, struct buffer_head *map_bh)
868 : {
869 : unsigned dio_blocks_per_fs_block;
870 : unsigned this_chunk_blocks; /* In dio_blocks */
871 : unsigned this_chunk_bytes;
872 : struct page *page;
873 :
874 0 : sdio->start_zero_done = 1;
875 0 : if (!sdio->blkfactor || !buffer_new(map_bh))
876 : return;
877 :
878 0 : dio_blocks_per_fs_block = 1 << sdio->blkfactor;
879 0 : this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
880 :
881 0 : if (!this_chunk_blocks)
882 : return;
883 :
884 : /*
885 : * We need to zero out part of an fs block. It is either at the
886 : * beginning or the end of the fs block.
887 : */
888 0 : if (end)
889 0 : this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
890 :
891 0 : this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
892 :
893 0 : page = ZERO_PAGE(0);
894 0 : if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
895 : sdio->next_block_for_io, map_bh))
896 : return;
897 :
898 0 : sdio->next_block_for_io += this_chunk_blocks;
899 : }
900 :
901 : /*
902 : * Walk the user pages, and the file, mapping blocks to disk and generating
903 : * a sequence of (page,offset,len,block) mappings. These mappings are injected
904 : * into submit_page_section(), which takes care of the next stage of submission
905 : *
906 : * Direct IO against a blockdev is different from a file. Because we can
907 : * happily perform page-sized but 512-byte aligned IOs. It is important that
908 : * blockdev IO be able to have fine alignment and large sizes.
909 : *
910 : * So what we do is to permit the ->get_block function to populate bh.b_size
911 : * with the size of IO which is permitted at this offset and this i_blkbits.
912 : *
913 : * For best results, the blockdev should be set up with 512-byte i_blkbits and
914 : * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
915 : * fine alignment but still allows this function to work in PAGE_SIZE units.
916 : */
917 0 : static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
918 : struct buffer_head *map_bh)
919 : {
920 0 : const unsigned blkbits = sdio->blkbits;
921 0 : const unsigned i_blkbits = blkbits + sdio->blkfactor;
922 0 : int ret = 0;
923 :
924 0 : while (sdio->block_in_file < sdio->final_block_in_request) {
925 : struct page *page;
926 : size_t from, to;
927 :
928 0 : page = dio_get_page(dio, sdio);
929 0 : if (IS_ERR(page)) {
930 0 : ret = PTR_ERR(page);
931 0 : goto out;
932 : }
933 0 : from = sdio->head ? 0 : sdio->from;
934 0 : to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
935 0 : sdio->head++;
936 :
937 0 : while (from < to) {
938 : unsigned this_chunk_bytes; /* # of bytes mapped */
939 : unsigned this_chunk_blocks; /* # of blocks */
940 : unsigned u;
941 :
942 0 : if (sdio->blocks_available == 0) {
943 : /*
944 : * Need to go and map some more disk
945 : */
946 : unsigned long blkmask;
947 : unsigned long dio_remainder;
948 :
949 0 : ret = get_more_blocks(dio, sdio, map_bh);
950 0 : if (ret) {
951 0 : put_page(page);
952 0 : goto out;
953 : }
954 0 : if (!buffer_mapped(map_bh))
955 : goto do_holes;
956 :
957 0 : sdio->blocks_available =
958 0 : map_bh->b_size >> blkbits;
959 0 : sdio->next_block_for_io =
960 0 : map_bh->b_blocknr << sdio->blkfactor;
961 0 : if (buffer_new(map_bh)) {
962 0 : clean_bdev_aliases(
963 : map_bh->b_bdev,
964 : map_bh->b_blocknr,
965 0 : map_bh->b_size >> i_blkbits);
966 : }
967 :
968 0 : if (!sdio->blkfactor)
969 : goto do_holes;
970 :
971 0 : blkmask = (1 << sdio->blkfactor) - 1;
972 0 : dio_remainder = (sdio->block_in_file & blkmask);
973 :
974 : /*
975 : * If we are at the start of IO and that IO
976 : * starts partway into a fs-block,
977 : * dio_remainder will be non-zero. If the IO
978 : * is a read then we can simply advance the IO
979 : * cursor to the first block which is to be
980 : * read. But if the IO is a write and the
981 : * block was newly allocated we cannot do that;
982 : * the start of the fs block must be zeroed out
983 : * on-disk
984 : */
985 0 : if (!buffer_new(map_bh))
986 0 : sdio->next_block_for_io += dio_remainder;
987 0 : sdio->blocks_available -= dio_remainder;
988 : }
989 : do_holes:
990 : /* Handle holes */
991 0 : if (!buffer_mapped(map_bh)) {
992 : loff_t i_size_aligned;
993 :
994 : /* AKPM: eargh, -ENOTBLK is a hack */
995 0 : if (dio->op == REQ_OP_WRITE) {
996 0 : put_page(page);
997 0 : return -ENOTBLK;
998 : }
999 :
1000 : /*
1001 : * Be sure to account for a partial block as the
1002 : * last block in the file
1003 : */
1004 0 : i_size_aligned = ALIGN(i_size_read(dio->inode),
1005 : 1 << blkbits);
1006 0 : if (sdio->block_in_file >=
1007 0 : i_size_aligned >> blkbits) {
1008 : /* We hit eof */
1009 0 : put_page(page);
1010 0 : goto out;
1011 : }
1012 0 : zero_user(page, from, 1 << blkbits);
1013 0 : sdio->block_in_file++;
1014 0 : from += 1 << blkbits;
1015 0 : dio->result += 1 << blkbits;
1016 0 : goto next_block;
1017 : }
1018 :
1019 : /*
1020 : * If we're performing IO which has an alignment which
1021 : * is finer than the underlying fs, go check to see if
1022 : * we must zero out the start of this block.
1023 : */
1024 0 : if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1025 0 : dio_zero_block(dio, sdio, 0, map_bh);
1026 :
1027 : /*
1028 : * Work out, in this_chunk_blocks, how much disk we
1029 : * can add to this page
1030 : */
1031 0 : this_chunk_blocks = sdio->blocks_available;
1032 0 : u = (to - from) >> blkbits;
1033 0 : if (this_chunk_blocks > u)
1034 0 : this_chunk_blocks = u;
1035 0 : u = sdio->final_block_in_request - sdio->block_in_file;
1036 0 : if (this_chunk_blocks > u)
1037 0 : this_chunk_blocks = u;
1038 0 : this_chunk_bytes = this_chunk_blocks << blkbits;
1039 0 : BUG_ON(this_chunk_bytes == 0);
1040 :
1041 0 : if (this_chunk_blocks == sdio->blocks_available)
1042 0 : sdio->boundary = buffer_boundary(map_bh);
1043 0 : ret = submit_page_section(dio, sdio, page,
1044 : from,
1045 : this_chunk_bytes,
1046 : sdio->next_block_for_io,
1047 : map_bh);
1048 0 : if (ret) {
1049 0 : put_page(page);
1050 0 : goto out;
1051 : }
1052 0 : sdio->next_block_for_io += this_chunk_blocks;
1053 :
1054 0 : sdio->block_in_file += this_chunk_blocks;
1055 0 : from += this_chunk_bytes;
1056 0 : dio->result += this_chunk_bytes;
1057 0 : sdio->blocks_available -= this_chunk_blocks;
1058 : next_block:
1059 0 : BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1060 0 : if (sdio->block_in_file == sdio->final_block_in_request)
1061 : break;
1062 : }
1063 :
1064 : /* Drop the ref which was taken in get_user_pages() */
1065 0 : put_page(page);
1066 : }
1067 : out:
1068 : return ret;
1069 : }
1070 :
1071 : static inline int drop_refcount(struct dio *dio)
1072 : {
1073 : int ret2;
1074 : unsigned long flags;
1075 :
1076 : /*
1077 : * Sync will always be dropping the final ref and completing the
1078 : * operation. AIO can if it was a broken operation described above or
1079 : * in fact if all the bios race to complete before we get here. In
1080 : * that case dio_complete() translates the EIOCBQUEUED into the proper
1081 : * return code that the caller will hand to ->complete().
1082 : *
1083 : * This is managed by the bio_lock instead of being an atomic_t so that
1084 : * completion paths can drop their ref and use the remaining count to
1085 : * decide to wake the submission path atomically.
1086 : */
1087 0 : spin_lock_irqsave(&dio->bio_lock, flags);
1088 0 : ret2 = --dio->refcount;
1089 0 : spin_unlock_irqrestore(&dio->bio_lock, flags);
1090 : return ret2;
1091 : }
1092 :
1093 : /*
1094 : * This is a library function for use by filesystem drivers.
1095 : *
1096 : * The locking rules are governed by the flags parameter:
1097 : * - if the flags value contains DIO_LOCKING we use a fancy locking
1098 : * scheme for dumb filesystems.
1099 : * For writes this function is called under i_mutex and returns with
1100 : * i_mutex held, for reads, i_mutex is not held on entry, but it is
1101 : * taken and dropped again before returning.
1102 : * - if the flags value does NOT contain DIO_LOCKING we don't use any
1103 : * internal locking but rather rely on the filesystem to synchronize
1104 : * direct I/O reads/writes versus each other and truncate.
1105 : *
1106 : * To help with locking against truncate we incremented the i_dio_count
1107 : * counter before starting direct I/O, and decrement it once we are done.
1108 : * Truncate can wait for it to reach zero to provide exclusion. It is
1109 : * expected that filesystem provide exclusion between new direct I/O
1110 : * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1111 : * but other filesystems need to take care of this on their own.
1112 : *
1113 : * NOTE: if you pass "sdio" to anything by pointer make sure that function
1114 : * is always inlined. Otherwise gcc is unable to split the structure into
1115 : * individual fields and will generate much worse code. This is important
1116 : * for the whole file.
1117 : */
1118 : static inline ssize_t
1119 0 : do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1120 : struct block_device *bdev, struct iov_iter *iter,
1121 : get_block_t get_block, dio_iodone_t end_io,
1122 : dio_submit_t submit_io, int flags)
1123 : {
1124 0 : unsigned i_blkbits = READ_ONCE(inode->i_blkbits);
1125 0 : unsigned blkbits = i_blkbits;
1126 0 : unsigned blocksize_mask = (1 << blkbits) - 1;
1127 0 : ssize_t retval = -EINVAL;
1128 0 : const size_t count = iov_iter_count(iter);
1129 0 : loff_t offset = iocb->ki_pos;
1130 0 : const loff_t end = offset + count;
1131 : struct dio *dio;
1132 0 : struct dio_submit sdio = { 0, };
1133 0 : struct buffer_head map_bh = { 0, };
1134 : struct blk_plug plug;
1135 0 : unsigned long align = offset | iov_iter_alignment(iter);
1136 :
1137 : /*
1138 : * Avoid references to bdev if not absolutely needed to give
1139 : * the early prefetch in the caller enough time.
1140 : */
1141 :
1142 : /* watch out for a 0 len io from a tricksy fs */
1143 0 : if (iov_iter_rw(iter) == READ && !count)
1144 : return 0;
1145 :
1146 0 : dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1147 0 : if (!dio)
1148 : return -ENOMEM;
1149 : /*
1150 : * Believe it or not, zeroing out the page array caused a .5%
1151 : * performance regression in a database benchmark. So, we take
1152 : * care to only zero out what's needed.
1153 : */
1154 0 : memset(dio, 0, offsetof(struct dio, pages));
1155 :
1156 0 : dio->flags = flags;
1157 0 : if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) {
1158 : /* will be released by direct_io_worker */
1159 : inode_lock(inode);
1160 : }
1161 :
1162 : /* Once we sampled i_size check for reads beyond EOF */
1163 0 : dio->i_size = i_size_read(inode);
1164 0 : if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1165 : retval = 0;
1166 : goto fail_dio;
1167 : }
1168 :
1169 0 : if (align & blocksize_mask) {
1170 0 : if (bdev)
1171 0 : blkbits = blksize_bits(bdev_logical_block_size(bdev));
1172 0 : blocksize_mask = (1 << blkbits) - 1;
1173 0 : if (align & blocksize_mask)
1174 : goto fail_dio;
1175 : }
1176 :
1177 0 : if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ) {
1178 0 : struct address_space *mapping = iocb->ki_filp->f_mapping;
1179 :
1180 0 : retval = filemap_write_and_wait_range(mapping, offset, end - 1);
1181 0 : if (retval)
1182 : goto fail_dio;
1183 : }
1184 :
1185 : /*
1186 : * For file extending writes updating i_size before data writeouts
1187 : * complete can expose uninitialized blocks in dumb filesystems.
1188 : * In that case we need to wait for I/O completion even if asked
1189 : * for an asynchronous write.
1190 : */
1191 0 : if (is_sync_kiocb(iocb))
1192 0 : dio->is_async = false;
1193 0 : else if (iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
1194 0 : dio->is_async = false;
1195 : else
1196 0 : dio->is_async = true;
1197 :
1198 0 : dio->inode = inode;
1199 0 : if (iov_iter_rw(iter) == WRITE) {
1200 0 : dio->op = REQ_OP_WRITE;
1201 0 : dio->op_flags = REQ_SYNC | REQ_IDLE;
1202 0 : if (iocb->ki_flags & IOCB_NOWAIT)
1203 0 : dio->op_flags |= REQ_NOWAIT;
1204 : } else {
1205 0 : dio->op = REQ_OP_READ;
1206 : }
1207 :
1208 : /*
1209 : * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1210 : * so that we can call ->fsync.
1211 : */
1212 0 : if (dio->is_async && iov_iter_rw(iter) == WRITE) {
1213 0 : retval = 0;
1214 0 : if (iocb->ki_flags & IOCB_DSYNC)
1215 0 : retval = dio_set_defer_completion(dio);
1216 0 : else if (!dio->inode->i_sb->s_dio_done_wq) {
1217 : /*
1218 : * In case of AIO write racing with buffered read we
1219 : * need to defer completion. We can't decide this now,
1220 : * however the workqueue needs to be initialized here.
1221 : */
1222 0 : retval = sb_init_dio_done_wq(dio->inode->i_sb);
1223 : }
1224 0 : if (retval)
1225 : goto fail_dio;
1226 : }
1227 :
1228 : /*
1229 : * Will be decremented at I/O completion time.
1230 : */
1231 0 : inode_dio_begin(inode);
1232 :
1233 0 : retval = 0;
1234 0 : sdio.blkbits = blkbits;
1235 0 : sdio.blkfactor = i_blkbits - blkbits;
1236 0 : sdio.block_in_file = offset >> blkbits;
1237 :
1238 0 : sdio.get_block = get_block;
1239 0 : dio->end_io = end_io;
1240 0 : sdio.submit_io = submit_io;
1241 0 : sdio.final_block_in_bio = -1;
1242 0 : sdio.next_block_for_io = -1;
1243 :
1244 0 : dio->iocb = iocb;
1245 :
1246 0 : spin_lock_init(&dio->bio_lock);
1247 0 : dio->refcount = 1;
1248 :
1249 0 : dio->should_dirty = iter_is_iovec(iter) && iov_iter_rw(iter) == READ;
1250 0 : sdio.iter = iter;
1251 0 : sdio.final_block_in_request = end >> blkbits;
1252 :
1253 : /*
1254 : * In case of non-aligned buffers, we may need 2 more
1255 : * pages since we need to zero out first and last block.
1256 : */
1257 0 : if (unlikely(sdio.blkfactor))
1258 0 : sdio.pages_in_io = 2;
1259 :
1260 0 : sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
1261 :
1262 0 : blk_start_plug(&plug);
1263 :
1264 0 : retval = do_direct_IO(dio, &sdio, &map_bh);
1265 0 : if (retval)
1266 : dio_cleanup(dio, &sdio);
1267 :
1268 0 : if (retval == -ENOTBLK) {
1269 : /*
1270 : * The remaining part of the request will be
1271 : * handled by buffered I/O when we return
1272 : */
1273 0 : retval = 0;
1274 : }
1275 : /*
1276 : * There may be some unwritten disk at the end of a part-written
1277 : * fs-block-sized block. Go zero that now.
1278 : */
1279 0 : dio_zero_block(dio, &sdio, 1, &map_bh);
1280 :
1281 0 : if (sdio.cur_page) {
1282 : ssize_t ret2;
1283 :
1284 0 : ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1285 0 : if (retval == 0)
1286 0 : retval = ret2;
1287 0 : put_page(sdio.cur_page);
1288 0 : sdio.cur_page = NULL;
1289 : }
1290 0 : if (sdio.bio)
1291 0 : dio_bio_submit(dio, &sdio);
1292 :
1293 0 : blk_finish_plug(&plug);
1294 :
1295 : /*
1296 : * It is possible that, we return short IO due to end of file.
1297 : * In that case, we need to release all the pages we got hold on.
1298 : */
1299 0 : dio_cleanup(dio, &sdio);
1300 :
1301 : /*
1302 : * All block lookups have been performed. For READ requests
1303 : * we can let i_mutex go now that its achieved its purpose
1304 : * of protecting us from looking up uninitialized blocks.
1305 : */
1306 0 : if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
1307 0 : inode_unlock(dio->inode);
1308 :
1309 : /*
1310 : * The only time we want to leave bios in flight is when a successful
1311 : * partial aio read or full aio write have been setup. In that case
1312 : * bio completion will call aio_complete. The only time it's safe to
1313 : * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1314 : * This had *better* be the only place that raises -EIOCBQUEUED.
1315 : */
1316 0 : BUG_ON(retval == -EIOCBQUEUED);
1317 0 : if (dio->is_async && retval == 0 && dio->result &&
1318 0 : (iov_iter_rw(iter) == READ || dio->result == count))
1319 : retval = -EIOCBQUEUED;
1320 : else
1321 0 : dio_await_completion(dio);
1322 :
1323 0 : if (drop_refcount(dio) == 0) {
1324 0 : retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE);
1325 : } else
1326 0 : BUG_ON(retval != -EIOCBQUEUED);
1327 :
1328 : return retval;
1329 :
1330 : fail_dio:
1331 0 : if (dio->flags & DIO_LOCKING && iov_iter_rw(iter) == READ)
1332 : inode_unlock(inode);
1333 :
1334 0 : kmem_cache_free(dio_cache, dio);
1335 0 : return retval;
1336 : }
1337 :
1338 0 : ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1339 : struct block_device *bdev, struct iov_iter *iter,
1340 : get_block_t get_block,
1341 : dio_iodone_t end_io, dio_submit_t submit_io,
1342 : int flags)
1343 : {
1344 : /*
1345 : * The block device state is needed in the end to finally
1346 : * submit everything. Since it's likely to be cache cold
1347 : * prefetch it here as first thing to hide some of the
1348 : * latency.
1349 : *
1350 : * Attempt to prefetch the pieces we likely need later.
1351 : */
1352 0 : prefetch(&bdev->bd_disk->part_tbl);
1353 0 : prefetch(bdev->bd_disk->queue);
1354 0 : prefetch((char *)bdev->bd_disk->queue + SMP_CACHE_BYTES);
1355 :
1356 0 : return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block,
1357 : end_io, submit_io, flags);
1358 : }
1359 :
1360 : EXPORT_SYMBOL(__blockdev_direct_IO);
1361 :
1362 1 : static __init int dio_init(void)
1363 : {
1364 1 : dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1365 1 : return 0;
1366 : }
1367 : module_init(dio_init)
|
