path: root/fs/direct-io.c (plain)
blob: fc90f0c33cbe41537e437b1ec64a85dee80f95d3

1	/*
2	* fs/direct-io.c
3	*
4	* Copyright (C) 2002, Linus Torvalds.
5	*
6	* O_DIRECT
7	*
8	* 04Jul2002 Andrew Morton
9	* Initial version
10	* 11Sep2002 janetinc@us.ibm.com
11	* added readv/writev support.
12	* 29Oct2002 Andrew Morton
13	* rewrote bio_add_page() support.
14	* 30Oct2002 pbadari@us.ibm.com
15	* added support for non-aligned IO.
16	* 06Nov2002 pbadari@us.ibm.com
17	* added asynchronous IO support.
18	* 21Jul2003 nathans@sgi.com
19	* added IO completion notifier.
20	*/
21
22	#include <linux/kernel.h>
23	#include <linux/module.h>
24	#include <linux/types.h>
25	#include <linux/fs.h>
26	#include <linux/mm.h>
27	#include <linux/slab.h>
28	#include <linux/highmem.h>
29	#include <linux/pagemap.h>
30	#include <linux/task_io_accounting_ops.h>
31	#include <linux/bio.h>
32	#include <linux/wait.h>
33	#include <linux/err.h>
34	#include <linux/blkdev.h>
35	#include <linux/buffer_head.h>
36	#include <linux/rwsem.h>
37	#include <linux/uio.h>
38	#include <linux/atomic.h>
39	#include <linux/prefetch.h>
40
41	/*
42	* How many user pages to map in one call to get_user_pages(). This determines
43	* the size of a structure in the slab cache
44	*/
45	#define DIO_PAGES 64
46
47	/*
48	* This code generally works in units of "dio_blocks". A dio_block is
49	* somewhere between the hard sector size and the filesystem block size. it
50	* is determined on a per-invocation basis. When talking to the filesystem
51	* we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
52	* down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
53	* to bio_block quantities by shifting left by blkfactor.
54	*
55	* If blkfactor is zero then the user's request was aligned to the filesystem's
56	* blocksize.
57	*/
58
59	/* dio_state only used in the submission path */
60
61	struct dio_submit {
62	struct bio *bio; /* bio under assembly */
63	unsigned blkbits; /* doesn't change */
64	unsigned blkfactor; /* When we're using an alignment which
65	is finer than the filesystem's soft
66	blocksize, this specifies how much
67	finer. blkfactor=2 means 1/4-block
68	alignment. Does not change */
69	unsigned start_zero_done; /* flag: sub-blocksize zeroing has
70	been performed at the start of a
71	write */
72	int pages_in_io; /* approximate total IO pages */
73	sector_t block_in_file; /* Current offset into the underlying
74	file in dio_block units. */
75	unsigned blocks_available; /* At block_in_file. changes */
76	int reap_counter; /* rate limit reaping */
77	sector_t final_block_in_request;/* doesn't change */
78	int boundary; /* prev block is at a boundary */
79	get_block_t *get_block; /* block mapping function */
80	dio_submit_t *submit_io; /* IO submition function */
81
82	loff_t logical_offset_in_bio; /* current first logical block in bio */
83	sector_t final_block_in_bio; /* current final block in bio + 1 */
84	sector_t next_block_for_io; /* next block to be put under IO,
85	in dio_blocks units */
86
87	/*
88	* Deferred addition of a page to the dio. These variables are
89	* private to dio_send_cur_page(), submit_page_section() and
90	* dio_bio_add_page().
91	*/
92	struct page *cur_page; /* The page */
93	unsigned cur_page_offset; /* Offset into it, in bytes */
94	unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
95	sector_t cur_page_block; /* Where it starts */
96	loff_t cur_page_fs_offset; /* Offset in file */
97
98	struct iov_iter *iter;
99	/*
100	* Page queue. These variables belong to dio_refill_pages() and
101	* dio_get_page().
102	*/
103	unsigned head; /* next page to process */
104	unsigned tail; /* last valid page + 1 */
105	size_t from, to;
106	};
107
108	/* dio_state communicated between submission path and end_io */
109	struct dio {
110	int flags; /* doesn't change */
111	int op;
112	int op_flags;
113	blk_qc_t bio_cookie;
114	struct block_device *bio_bdev;
115	struct inode *inode;
116	loff_t i_size; /* i_size when submitted */
117	dio_iodone_t *end_io; /* IO completion function */
118
119	void *private; /* copy from map_bh.b_private */
120
121	/* BIO completion state */
122	spinlock_t bio_lock; /* protects BIO fields below */
123	int page_errors; /* errno from get_user_pages() */
124	int is_async; /* is IO async ? */
125	bool defer_completion; /* defer AIO completion to workqueue? */
126	bool should_dirty; /* if pages should be dirtied */
127	int io_error; /* IO error in completion path */
128	unsigned long refcount; /* direct_io_worker() and bios */
129	struct bio *bio_list; /* singly linked via bi_private */
130	struct task_struct *waiter; /* waiting task (NULL if none) */
131
132	/* AIO related stuff */
133	struct kiocb *iocb; /* kiocb */
134	ssize_t result; /* IO result */
135
136	/*
137	* pages[] (and any fields placed after it) are not zeroed out at
138	* allocation time. Don't add new fields after pages[] unless you
139	* wish that they not be zeroed.
140	*/
141	union {
142	struct page *pages[DIO_PAGES]; /* page buffer */
143	struct work_struct complete_work;/* deferred AIO completion */
144	};
145	} ____cacheline_aligned_in_smp;
146
147	static struct kmem_cache *dio_cache __read_mostly;
148
149	/*
150	* How many pages are in the queue?
151	*/
152	static inline unsigned dio_pages_present(struct dio_submit *sdio)
153	{
154	return sdio->tail - sdio->head;
155	}
156
157	/*
158	* Go grab and pin some userspace pages. Typically we'll get 64 at a time.
159	*/
160	static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
161	{
162	ssize_t ret;
163
164	ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
165	&sdio->from);
166
167	if (ret < 0 && sdio->blocks_available && (dio->op == REQ_OP_WRITE)) {
168	struct page *page = ZERO_PAGE(0);
169	/*
170	* A memory fault, but the filesystem has some outstanding
171	* mapped blocks. We need to use those blocks up to avoid
172	* leaking stale data in the file.
173	*/
174	if (dio->page_errors == 0)
175	dio->page_errors = ret;
176	get_page(page);
177	dio->pages[0] = page;
178	sdio->head = 0;
179	sdio->tail = 1;
180	sdio->from = 0;
181	sdio->to = PAGE_SIZE;
182	return 0;
183	}
184
185	if (ret >= 0) {
186	iov_iter_advance(sdio->iter, ret);
187	ret += sdio->from;
188	sdio->head = 0;
189	sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
190	sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
191	return 0;
192	}
193	return ret;
194	}
195
196	/*
197	* Get another userspace page. Returns an ERR_PTR on error. Pages are
198	* buffered inside the dio so that we can call get_user_pages() against a
199	* decent number of pages, less frequently. To provide nicer use of the
200	* L1 cache.
201	*/
202	static inline struct page *dio_get_page(struct dio *dio,
203	struct dio_submit *sdio)
204	{
205	if (dio_pages_present(sdio) == 0) {
206	int ret;
207
208	ret = dio_refill_pages(dio, sdio);
209	if (ret)
210	return ERR_PTR(ret);
211	BUG_ON(dio_pages_present(sdio) == 0);
212	}
213	return dio->pages[sdio->head];
214	}
215
216	/**
217	* dio_complete() - called when all DIO BIO I/O has been completed
218	* @offset: the byte offset in the file of the completed operation
219	*
220	* This drops i_dio_count, lets interested parties know that a DIO operation
221	* has completed, and calculates the resulting return code for the operation.
222	*
223	* It lets the filesystem know if it registered an interest earlier via
224	* get_block. Pass the private field of the map buffer_head so that
225	* filesystems can use it to hold additional state between get_block calls and
226	* dio_complete.
227	*/
228	static ssize_t dio_complete(struct dio *dio, ssize_t ret, bool is_async)
229	{
230	loff_t offset = dio->iocb->ki_pos;
231	ssize_t transferred = 0;
232
233	/*
234	* AIO submission can race with bio completion to get here while
235	* expecting to have the last io completed by bio completion.
236	* In that case -EIOCBQUEUED is in fact not an error we want
237	* to preserve through this call.
238	*/
239	if (ret == -EIOCBQUEUED)
240	ret = 0;
241
242	if (dio->result) {
243	transferred = dio->result;
244
245	/* Check for short read case */
246	if ((dio->op == REQ_OP_READ) &&
247	((offset + transferred) > dio->i_size))
248	transferred = dio->i_size - offset;
249	/* ignore EFAULT if some IO has been done */
250	if (unlikely(ret == -EFAULT) && transferred)
251	ret = 0;
252	}
253
254	if (ret == 0)
255	ret = dio->page_errors;
256	if (ret == 0)
257	ret = dio->io_error;
258	if (ret == 0)
259	ret = transferred;
260
261	if (dio->end_io) {
262	int err;
263
264	// XXX: ki_pos??
265	err = dio->end_io(dio->iocb, offset, ret, dio->private);
266	if (err)
267	ret = err;
268	}
269
270	if (!(dio->flags & DIO_SKIP_DIO_COUNT))
271	inode_dio_end(dio->inode);
272
273	if (is_async) {
274	/*
275	* generic_write_sync expects ki_pos to have been updated
276	* already, but the submission path only does this for
277	* synchronous I/O.
278	*/
279	dio->iocb->ki_pos += transferred;
280
281	if (ret > 0 && dio->op == REQ_OP_WRITE)
282	ret = generic_write_sync(dio->iocb, ret);
283	dio->iocb->ki_complete(dio->iocb, ret, 0);
284	}
285
286	kmem_cache_free(dio_cache, dio);
287	return ret;
288	}
289
290	static void dio_aio_complete_work(struct work_struct *work)
291	{
292	struct dio *dio = container_of(work, struct dio, complete_work);
293
294	dio_complete(dio, 0, true);
295	}
296
297	static int dio_bio_complete(struct dio *dio, struct bio *bio);
298
299	/*
300	* Asynchronous IO callback.
301	*/
302	static void dio_bio_end_aio(struct bio *bio)
303	{
304	struct dio *dio = bio->bi_private;
305	unsigned long remaining;
306	unsigned long flags;
307
308	/* cleanup the bio */
309	dio_bio_complete(dio, bio);
310
311	spin_lock_irqsave(&dio->bio_lock, flags);
312	remaining = --dio->refcount;
313	if (remaining == 1 && dio->waiter)
314	wake_up_process(dio->waiter);
315	spin_unlock_irqrestore(&dio->bio_lock, flags);
316
317	if (remaining == 0) {
318	if (dio->result && dio->defer_completion) {
319	INIT_WORK(&dio->complete_work, dio_aio_complete_work);
320	queue_work(dio->inode->i_sb->s_dio_done_wq,
321	&dio->complete_work);
322	} else {
323	dio_complete(dio, 0, true);
324	}
325	}
326	}
327
328	/*
329	* The BIO completion handler simply queues the BIO up for the process-context
330	* handler.
331	*
332	* During I/O bi_private points at the dio. After I/O, bi_private is used to
333	* implement a singly-linked list of completed BIOs, at dio->bio_list.
334	*/
335	static void dio_bio_end_io(struct bio *bio)
336	{
337	struct dio *dio = bio->bi_private;
338	unsigned long flags;
339
340	spin_lock_irqsave(&dio->bio_lock, flags);
341	bio->bi_private = dio->bio_list;
342	dio->bio_list = bio;
343	if (--dio->refcount == 1 && dio->waiter)
344	wake_up_process(dio->waiter);
345	spin_unlock_irqrestore(&dio->bio_lock, flags);
346	}
347
348	/**
349	* dio_end_io - handle the end io action for the given bio
350	* @bio: The direct io bio thats being completed
351	* @error: Error if there was one
352	*
353	* This is meant to be called by any filesystem that uses their own dio_submit_t
354	* so that the DIO specific endio actions are dealt with after the filesystem
355	* has done it's completion work.
356	*/
357	void dio_end_io(struct bio *bio, int error)
358	{
359	struct dio *dio = bio->bi_private;
360
361	if (dio->is_async)
362	dio_bio_end_aio(bio);
363	else
364	dio_bio_end_io(bio);
365	}
366	EXPORT_SYMBOL_GPL(dio_end_io);
367
368	static inline void
369	dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
370	struct block_device *bdev,
371	sector_t first_sector, int nr_vecs)
372	{
373	struct bio *bio;
374
375	/*
376	* bio_alloc() is guaranteed to return a bio when called with
377	* __GFP_RECLAIM and we request a valid number of vectors.
378	*/
379	bio = bio_alloc(GFP_KERNEL, nr_vecs);
380
381	bio->bi_bdev = bdev;
382	bio->bi_iter.bi_sector = first_sector;
383	bio_set_op_attrs(bio, dio->op, dio->op_flags);
384	if (dio->is_async)
385	bio->bi_end_io = dio_bio_end_aio;
386	else
387	bio->bi_end_io = dio_bio_end_io;
388
389	sdio->bio = bio;
390	sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
391	}
392
393	/*
394	* In the AIO read case we speculatively dirty the pages before starting IO.
395	* During IO completion, any of these pages which happen to have been written
396	* back will be redirtied by bio_check_pages_dirty().
397	*
398	* bios hold a dio reference between submit_bio and ->end_io.
399	*/
400	static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
401	{
402	struct bio *bio = sdio->bio;
403	unsigned long flags;
404
405	bio->bi_private = dio;
406
407	spin_lock_irqsave(&dio->bio_lock, flags);
408	dio->refcount++;
409	spin_unlock_irqrestore(&dio->bio_lock, flags);
410
411	if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty)
412	bio_set_pages_dirty(bio);
413
414	dio->bio_bdev = bio->bi_bdev;
415
416	if (sdio->submit_io) {
417	sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio);
418	dio->bio_cookie = BLK_QC_T_NONE;
419	} else
420	dio->bio_cookie = submit_bio(bio);
421
422	sdio->bio = NULL;
423	sdio->boundary = 0;
424	sdio->logical_offset_in_bio = 0;
425	}
426
427	/*
428	* Release any resources in case of a failure
429	*/
430	static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
431	{
432	while (sdio->head < sdio->tail)
433	put_page(dio->pages[sdio->head++]);
434	}
435
436	/*
437	* Wait for the next BIO to complete. Remove it and return it. NULL is
438	* returned once all BIOs have been completed. This must only be called once
439	* all bios have been issued so that dio->refcount can only decrease. This
440	* requires that that the caller hold a reference on the dio.
441	*/
442	static struct bio *dio_await_one(struct dio *dio)
443	{
444	unsigned long flags;
445	struct bio *bio = NULL;
446
447	spin_lock_irqsave(&dio->bio_lock, flags);
448
449	/*
450	* Wait as long as the list is empty and there are bios in flight. bio
451	* completion drops the count, maybe adds to the list, and wakes while
452	* holding the bio_lock so we don't need set_current_state()'s barrier
453	* and can call it after testing our condition.
454	*/
455	while (dio->refcount > 1 && dio->bio_list == NULL) {
456	__set_current_state(TASK_UNINTERRUPTIBLE);
457	dio->waiter = current;
458	spin_unlock_irqrestore(&dio->bio_lock, flags);
459	if (!(dio->iocb->ki_flags & IOCB_HIPRI) ||
460	!blk_poll(bdev_get_queue(dio->bio_bdev), dio->bio_cookie))
461	io_schedule();
462	/* wake up sets us TASK_RUNNING */
463	spin_lock_irqsave(&dio->bio_lock, flags);
464	dio->waiter = NULL;
465	}
466	if (dio->bio_list) {
467	bio = dio->bio_list;
468	dio->bio_list = bio->bi_private;
469	}
470	spin_unlock_irqrestore(&dio->bio_lock, flags);
471	return bio;
472	}
473
474	/*
475	* Process one completed BIO. No locks are held.
476	*/
477	static int dio_bio_complete(struct dio *dio, struct bio *bio)
478	{
479	struct bio_vec *bvec;
480	unsigned i;
481	int err;
482
483	if (bio->bi_error)
484	dio->io_error = -EIO;
485
486	if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty) {
487	err = bio->bi_error;
488	bio_check_pages_dirty(bio); /* transfers ownership */
489	} else {
490	bio_for_each_segment_all(bvec, bio, i) {
491	struct page *page = bvec->bv_page;
492
493	if (dio->op == REQ_OP_READ && !PageCompound(page) &&
494	dio->should_dirty)
495	set_page_dirty_lock(page);
496	put_page(page);
497	}
498	err = bio->bi_error;
499	bio_put(bio);
500	}
501	return err;
502	}
503
504	/*
505	* Wait on and process all in-flight BIOs. This must only be called once
506	* all bios have been issued so that the refcount can only decrease.
507	* This just waits for all bios to make it through dio_bio_complete. IO
508	* errors are propagated through dio->io_error and should be propagated via
509	* dio_complete().
510	*/
511	static void dio_await_completion(struct dio *dio)
512	{
513	struct bio *bio;
514	do {
515	bio = dio_await_one(dio);
516	if (bio)
517	dio_bio_complete(dio, bio);
518	} while (bio);
519	}
520
521	/*
522	* A really large O_DIRECT read or write can generate a lot of BIOs. So
523	* to keep the memory consumption sane we periodically reap any completed BIOs
524	* during the BIO generation phase.
525	*
526	* This also helps to limit the peak amount of pinned userspace memory.
527	*/
528	static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
529	{
530	int ret = 0;
531
532	if (sdio->reap_counter++ >= 64) {
533	while (dio->bio_list) {
534	unsigned long flags;
535	struct bio *bio;
536	int ret2;
537
538	spin_lock_irqsave(&dio->bio_lock, flags);
539	bio = dio->bio_list;
540	dio->bio_list = bio->bi_private;
541	spin_unlock_irqrestore(&dio->bio_lock, flags);
542	ret2 = dio_bio_complete(dio, bio);
543	if (ret == 0)
544	ret = ret2;
545	}
546	sdio->reap_counter = 0;
547	}
548	return ret;
549	}
550
551	/*
552	* Create workqueue for deferred direct IO completions. We allocate the
553	* workqueue when it's first needed. This avoids creating workqueue for
554	* filesystems that don't need it and also allows us to create the workqueue
555	* late enough so the we can include s_id in the name of the workqueue.
556	*/
557	static int sb_init_dio_done_wq(struct super_block *sb)
558	{
559	struct workqueue_struct *old;
560	struct workqueue_struct *wq = alloc_workqueue("dio/%s",
561	WQ_MEM_RECLAIM, 0,
562	sb->s_id);
563	if (!wq)
564	return -ENOMEM;
565	/*
566	* This has to be atomic as more DIOs can race to create the workqueue
567	*/
568	old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
569	/* Someone created workqueue before us? Free ours... */
570	if (old)
571	destroy_workqueue(wq);
572	return 0;
573	}
574
575	static int dio_set_defer_completion(struct dio *dio)
576	{
577	struct super_block *sb = dio->inode->i_sb;
578
579	if (dio->defer_completion)
580	return 0;
581	dio->defer_completion = true;
582	if (!sb->s_dio_done_wq)
583	return sb_init_dio_done_wq(sb);
584	return 0;
585	}
586
587	/*
588	* Call into the fs to map some more disk blocks. We record the current number
589	* of available blocks at sdio->blocks_available. These are in units of the
590	* fs blocksize, i_blocksize(inode).
591	*
592	* The fs is allowed to map lots of blocks at once. If it wants to do that,
593	* it uses the passed inode-relative block number as the file offset, as usual.
594	*
595	* get_block() is passed the number of i_blkbits-sized blocks which direct_io
596	* has remaining to do. The fs should not map more than this number of blocks.
597	*
598	* If the fs has mapped a lot of blocks, it should populate bh->b_size to
599	* indicate how much contiguous disk space has been made available at
600	* bh->b_blocknr.
601	*
602	* If *any* of the mapped blocks are new, then the fs must set buffer_new().
603	* This isn't very efficient...
604	*
605	* In the case of filesystem holes: the fs may return an arbitrarily-large
606	* hole by returning an appropriate value in b_size and by clearing
607	* buffer_mapped(). However the direct-io code will only process holes one
608	* block at a time - it will repeatedly call get_block() as it walks the hole.
609	*/
610	static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
611	struct buffer_head *map_bh)
612	{
613	int ret;
614	sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
615	sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
616	unsigned long fs_count; /* Number of filesystem-sized blocks */
617	int create;
618	unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
619	loff_t i_size;
620
621	/*
622	* If there was a memory error and we've overwritten all the
623	* mapped blocks then we can now return that memory error
624	*/
625	ret = dio->page_errors;
626	if (ret == 0) {
627	BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
628	fs_startblk = sdio->block_in_file >> sdio->blkfactor;
629	fs_endblk = (sdio->final_block_in_request - 1) >>
630	sdio->blkfactor;
631	fs_count = fs_endblk - fs_startblk + 1;
632
633	map_bh->b_state = 0;
634	map_bh->b_size = fs_count << i_blkbits;
635
636	/*
637	* For writes that could fill holes inside i_size on a
638	* DIO_SKIP_HOLES filesystem we forbid block creations: only
639	* overwrites are permitted. We will return early to the caller
640	* once we see an unmapped buffer head returned, and the caller
641	* will fall back to buffered I/O.
642	*
643	* Otherwise the decision is left to the get_blocks method,
644	* which may decide to handle it or also return an unmapped
645	* buffer head.
646	*/
647	create = dio->op == REQ_OP_WRITE;
648	if (dio->flags & DIO_SKIP_HOLES) {
649	i_size = i_size_read(dio->inode);
650	if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits)
651	create = 0;
652	}
653
654	ret = (*sdio->get_block)(dio->inode, fs_startblk,
655	map_bh, create);
656
657	/* Store for completion */
658	dio->private = map_bh->b_private;
659
660	if (ret == 0 && buffer_defer_completion(map_bh))
661	ret = dio_set_defer_completion(dio);
662	}
663	return ret;
664	}
665
666	/*
667	* There is no bio. Make one now.
668	*/
669	static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
670	sector_t start_sector, struct buffer_head *map_bh)
671	{
672	sector_t sector;
673	int ret, nr_pages;
674
675	ret = dio_bio_reap(dio, sdio);
676	if (ret)
677	goto out;
678	sector = start_sector << (sdio->blkbits - 9);
679	nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES);
680	BUG_ON(nr_pages <= 0);
681	dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
682	sdio->boundary = 0;
683	out:
684	return ret;
685	}
686
687	/*
688	* Attempt to put the current chunk of 'cur_page' into the current BIO. If
689	* that was successful then update final_block_in_bio and take a ref against
690	* the just-added page.
691	*
692	* Return zero on success. Non-zero means the caller needs to start a new BIO.
693	*/
694	static inline int dio_bio_add_page(struct dio_submit *sdio)
695	{
696	int ret;
697
698	ret = bio_add_page(sdio->bio, sdio->cur_page,
699	sdio->cur_page_len, sdio->cur_page_offset);
700	if (ret == sdio->cur_page_len) {
701	/*
702	* Decrement count only, if we are done with this page
703	*/
704	if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
705	sdio->pages_in_io--;
706	get_page(sdio->cur_page);
707	sdio->final_block_in_bio = sdio->cur_page_block +
708	(sdio->cur_page_len >> sdio->blkbits);
709	ret = 0;
710	} else {
711	ret = 1;
712	}
713	return ret;
714	}
715
716	/*
717	* Put cur_page under IO. The section of cur_page which is described by
718	* cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
719	* starts on-disk at cur_page_block.
720	*
721	* We take a ref against the page here (on behalf of its presence in the bio).
722	*
723	* The caller of this function is responsible for removing cur_page from the
724	* dio, and for dropping the refcount which came from that presence.
725	*/
726	static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
727	struct buffer_head *map_bh)
728	{
729	int ret = 0;
730
731	if (sdio->bio) {
732	loff_t cur_offset = sdio->cur_page_fs_offset;
733	loff_t bio_next_offset = sdio->logical_offset_in_bio +
734	sdio->bio->bi_iter.bi_size;
735
736	/*
737	* See whether this new request is contiguous with the old.
738	*
739	* Btrfs cannot handle having logically non-contiguous requests
740	* submitted. For example if you have
741	*
742	* Logical: [0-4095][HOLE][8192-12287]
743	* Physical: [0-4095] [4096-8191]
744	*
745	* We cannot submit those pages together as one BIO. So if our
746	* current logical offset in the file does not equal what would
747	* be the next logical offset in the bio, submit the bio we
748	* have.
749	*/
750	if (sdio->final_block_in_bio != sdio->cur_page_block ||
751	cur_offset != bio_next_offset)
752	dio_bio_submit(dio, sdio);
753	}
754
755	if (sdio->bio == NULL) {
756	ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
757	if (ret)
758	goto out;
759	}
760
761	if (dio_bio_add_page(sdio) != 0) {
762	dio_bio_submit(dio, sdio);
763	ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
764	if (ret == 0) {
765	ret = dio_bio_add_page(sdio);
766	BUG_ON(ret != 0);
767	}
768	}
769	out:
770	return ret;
771	}
772
773	/*
774	* An autonomous function to put a chunk of a page under deferred IO.
775	*
776	* The caller doesn't actually know (or care) whether this piece of page is in
777	* a BIO, or is under IO or whatever. We just take care of all possible
778	* situations here. The separation between the logic of do_direct_IO() and
779	* that of submit_page_section() is important for clarity. Please don't break.
780	*
781	* The chunk of page starts on-disk at blocknr.
782	*
783	* We perform deferred IO, by recording the last-submitted page inside our
784	* private part of the dio structure. If possible, we just expand the IO
785	* across that page here.
786	*
787	* If that doesn't work out then we put the old page into the bio and add this
788	* page to the dio instead.
789	*/
790	static inline int
791	submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
792	unsigned offset, unsigned len, sector_t blocknr,
793	struct buffer_head *map_bh)
794	{
795	int ret = 0;
796
797	if (dio->op == REQ_OP_WRITE) {
798	/*
799	* Read accounting is performed in submit_bio()
800	*/
801	task_io_account_write(len);
802	}
803
804	/*
805	* Can we just grow the current page's presence in the dio?
806	*/
807	if (sdio->cur_page == page &&
808	sdio->cur_page_offset + sdio->cur_page_len == offset &&
809	sdio->cur_page_block +
810	(sdio->cur_page_len >> sdio->blkbits) == blocknr) {
811	sdio->cur_page_len += len;
812	goto out;
813	}
814
815	/*
816	* If there's a deferred page already there then send it.
817	*/
818	if (sdio->cur_page) {
819	ret = dio_send_cur_page(dio, sdio, map_bh);
820	put_page(sdio->cur_page);
821	sdio->cur_page = NULL;
822	if (ret)
823	return ret;
824	}
825
826	get_page(page); /* It is in dio */
827	sdio->cur_page = page;
828	sdio->cur_page_offset = offset;
829	sdio->cur_page_len = len;
830	sdio->cur_page_block = blocknr;
831	sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
832	out:
833	/*
834	* If sdio->boundary then we want to schedule the IO now to
835	* avoid metadata seeks.
836	*/
837	if (sdio->boundary) {
838	ret = dio_send_cur_page(dio, sdio, map_bh);
839	if (sdio->bio)
840	dio_bio_submit(dio, sdio);
841	put_page(sdio->cur_page);
842	sdio->cur_page = NULL;
843	}
844	return ret;
845	}
846
847	/*
848	* Clean any dirty buffers in the blockdev mapping which alias newly-created
849	* file blocks. Only called for S_ISREG files - blockdevs do not set
850	* buffer_new
851	*/
852	static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
853	{
854	unsigned i;
855	unsigned nblocks;
856
857	nblocks = map_bh->b_size >> dio->inode->i_blkbits;
858
859	for (i = 0; i < nblocks; i++) {
860	unmap_underlying_metadata(map_bh->b_bdev,
861	map_bh->b_blocknr + i);
862	}
863	}
864
865	/*
866	* If we are not writing the entire block and get_block() allocated
867	* the block for us, we need to fill-in the unused portion of the
868	* block with zeros. This happens only if user-buffer, fileoffset or
869	* io length is not filesystem block-size multiple.
870	*
871	* `end' is zero if we're doing the start of the IO, 1 at the end of the
872	* IO.
873	*/
874	static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
875	int end, struct buffer_head *map_bh)
876	{
877	unsigned dio_blocks_per_fs_block;
878	unsigned this_chunk_blocks; /* In dio_blocks */
879	unsigned this_chunk_bytes;
880	struct page *page;
881
882	sdio->start_zero_done = 1;
883	if (!sdio->blkfactor || !buffer_new(map_bh))
884	return;
885
886	dio_blocks_per_fs_block = 1 << sdio->blkfactor;
887	this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
888
889	if (!this_chunk_blocks)
890	return;
891
892	/*
893	* We need to zero out part of an fs block. It is either at the
894	* beginning or the end of the fs block.
895	*/
896	if (end)
897	this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
898
899	this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
900
901	page = ZERO_PAGE(0);
902	if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
903	sdio->next_block_for_io, map_bh))
904	return;
905
906	sdio->next_block_for_io += this_chunk_blocks;
907	}
908
909	/*
910	* Walk the user pages, and the file, mapping blocks to disk and generating
911	* a sequence of (page,offset,len,block) mappings. These mappings are injected
912	* into submit_page_section(), which takes care of the next stage of submission
913	*
914	* Direct IO against a blockdev is different from a file. Because we can
915	* happily perform page-sized but 512-byte aligned IOs. It is important that
916	* blockdev IO be able to have fine alignment and large sizes.
917	*
918	* So what we do is to permit the ->get_block function to populate bh.b_size
919	* with the size of IO which is permitted at this offset and this i_blkbits.
920	*
921	* For best results, the blockdev should be set up with 512-byte i_blkbits and
922	* it should set b_size to PAGE_SIZE or more inside get_block(). This gives
923	* fine alignment but still allows this function to work in PAGE_SIZE units.
924	*/
925	static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
926	struct buffer_head *map_bh)
927	{
928	const unsigned blkbits = sdio->blkbits;
929	int ret = 0;
930
931	while (sdio->block_in_file < sdio->final_block_in_request) {
932	struct page *page;
933	size_t from, to;
934
935	page = dio_get_page(dio, sdio);
936	if (IS_ERR(page)) {
937	ret = PTR_ERR(page);
938	goto out;
939	}
940	from = sdio->head ? 0 : sdio->from;
941	to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
942	sdio->head++;
943
944	while (from < to) {
945	unsigned this_chunk_bytes; /* # of bytes mapped */
946	unsigned this_chunk_blocks; /* # of blocks */
947	unsigned u;
948
949	if (sdio->blocks_available == 0) {
950	/*
951	* Need to go and map some more disk
952	*/
953	unsigned long blkmask;
954	unsigned long dio_remainder;
955
956	ret = get_more_blocks(dio, sdio, map_bh);
957	if (ret) {
958	put_page(page);
959	goto out;
960	}
961	if (!buffer_mapped(map_bh))
962	goto do_holes;
963
964	sdio->blocks_available =
965	map_bh->b_size >> sdio->blkbits;
966	sdio->next_block_for_io =
967	map_bh->b_blocknr << sdio->blkfactor;
968	if (buffer_new(map_bh))
969	clean_blockdev_aliases(dio, map_bh);
970
971	if (!sdio->blkfactor)
972	goto do_holes;
973
974	blkmask = (1 << sdio->blkfactor) - 1;
975	dio_remainder = (sdio->block_in_file & blkmask);
976
977	/*
978	* If we are at the start of IO and that IO
979	* starts partway into a fs-block,
980	* dio_remainder will be non-zero. If the IO
981	* is a read then we can simply advance the IO
982	* cursor to the first block which is to be
983	* read. But if the IO is a write and the
984	* block was newly allocated we cannot do that;
985	* the start of the fs block must be zeroed out
986	* on-disk
987	*/
988	if (!buffer_new(map_bh))
989	sdio->next_block_for_io += dio_remainder;
990	sdio->blocks_available -= dio_remainder;
991	}
992	do_holes:
993	/* Handle holes */
994	if (!buffer_mapped(map_bh)) {
995	loff_t i_size_aligned;
996
997	/* AKPM: eargh, -ENOTBLK is a hack */
998	if (dio->op == REQ_OP_WRITE) {
999	put_page(page);
1000	return -ENOTBLK;
1001	}
1002
1003	/*
1004	* Be sure to account for a partial block as the
1005	* last block in the file
1006	*/
1007	i_size_aligned = ALIGN(i_size_read(dio->inode),
1008	1 << blkbits);
1009	if (sdio->block_in_file >=
1010	i_size_aligned >> blkbits) {
1011	/* We hit eof */
1012	put_page(page);
1013	goto out;
1014	}
1015	zero_user(page, from, 1 << blkbits);
1016	sdio->block_in_file++;
1017	from += 1 << blkbits;
1018	dio->result += 1 << blkbits;
1019	goto next_block;
1020	}
1021
1022	/*
1023	* If we're performing IO which has an alignment which
1024	* is finer than the underlying fs, go check to see if
1025	* we must zero out the start of this block.
1026	*/
1027	if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1028	dio_zero_block(dio, sdio, 0, map_bh);
1029
1030	/*
1031	* Work out, in this_chunk_blocks, how much disk we
1032	* can add to this page
1033	*/
1034	this_chunk_blocks = sdio->blocks_available;
1035	u = (to - from) >> blkbits;
1036	if (this_chunk_blocks > u)
1037	this_chunk_blocks = u;
1038	u = sdio->final_block_in_request - sdio->block_in_file;
1039	if (this_chunk_blocks > u)
1040	this_chunk_blocks = u;
1041	this_chunk_bytes = this_chunk_blocks << blkbits;
1042	BUG_ON(this_chunk_bytes == 0);
1043
1044	if (this_chunk_blocks == sdio->blocks_available)
1045	sdio->boundary = buffer_boundary(map_bh);
1046	ret = submit_page_section(dio, sdio, page,
1047	from,
1048	this_chunk_bytes,
1049	sdio->next_block_for_io,
1050	map_bh);
1051	if (ret) {
1052	put_page(page);
1053	goto out;
1054	}
1055	sdio->next_block_for_io += this_chunk_blocks;
1056
1057	sdio->block_in_file += this_chunk_blocks;
1058	from += this_chunk_bytes;
1059	dio->result += this_chunk_bytes;
1060	sdio->blocks_available -= this_chunk_blocks;
1061	next_block:
1062	BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1063	if (sdio->block_in_file == sdio->final_block_in_request)
1064	break;
1065	}
1066
1067	/* Drop the ref which was taken in get_user_pages() */
1068	put_page(page);
1069	}
1070	out:
1071	return ret;
1072	}
1073
1074	static inline int drop_refcount(struct dio *dio)
1075	{
1076	int ret2;
1077	unsigned long flags;
1078
1079	/*
1080	* Sync will always be dropping the final ref and completing the
1081	* operation. AIO can if it was a broken operation described above or
1082	* in fact if all the bios race to complete before we get here. In
1083	* that case dio_complete() translates the EIOCBQUEUED into the proper
1084	* return code that the caller will hand to ->complete().
1085	*
1086	* This is managed by the bio_lock instead of being an atomic_t so that
1087	* completion paths can drop their ref and use the remaining count to
1088	* decide to wake the submission path atomically.
1089	*/
1090	spin_lock_irqsave(&dio->bio_lock, flags);
1091	ret2 = --dio->refcount;
1092	spin_unlock_irqrestore(&dio->bio_lock, flags);
1093	return ret2;
1094	}
1095
1096	/*
1097	* This is a library function for use by filesystem drivers.
1098	*
1099	* The locking rules are governed by the flags parameter:
1100	* - if the flags value contains DIO_LOCKING we use a fancy locking
1101	* scheme for dumb filesystems.
1102	* For writes this function is called under i_mutex and returns with
1103	* i_mutex held, for reads, i_mutex is not held on entry, but it is
1104	* taken and dropped again before returning.
1105	* - if the flags value does NOT contain DIO_LOCKING we don't use any
1106	* internal locking but rather rely on the filesystem to synchronize
1107	* direct I/O reads/writes versus each other and truncate.
1108	*
1109	* To help with locking against truncate we incremented the i_dio_count
1110	* counter before starting direct I/O, and decrement it once we are done.
1111	* Truncate can wait for it to reach zero to provide exclusion. It is
1112	* expected that filesystem provide exclusion between new direct I/O
1113	* and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1114	* but other filesystems need to take care of this on their own.
1115	*
1116	* NOTE: if you pass "sdio" to anything by pointer make sure that function
1117	* is always inlined. Otherwise gcc is unable to split the structure into
1118	* individual fields and will generate much worse code. This is important
1119	* for the whole file.
1120	*/
1121	static inline ssize_t
1122	do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1123	struct block_device *bdev, struct iov_iter *iter,
1124	get_block_t get_block, dio_iodone_t end_io,
1125	dio_submit_t submit_io, int flags)
1126	{
1127	unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1128	unsigned blkbits = i_blkbits;
1129	unsigned blocksize_mask = (1 << blkbits) - 1;
1130	ssize_t retval = -EINVAL;
1131	size_t count = iov_iter_count(iter);
1132	loff_t offset = iocb->ki_pos;
1133	loff_t end = offset + count;
1134	struct dio *dio;
1135	struct dio_submit sdio = { 0, };
1136	struct buffer_head map_bh = { 0, };
1137	struct blk_plug plug;
1138	unsigned long align = offset | iov_iter_alignment(iter);
1139
1140	/*
1141	* Avoid references to bdev if not absolutely needed to give
1142	* the early prefetch in the caller enough time.
1143	*/
1144
1145	if (align & blocksize_mask) {
1146	if (bdev)
1147	blkbits = blksize_bits(bdev_logical_block_size(bdev));
1148	blocksize_mask = (1 << blkbits) - 1;
1149	if (align & blocksize_mask)
1150	goto out;
1151	}
1152
1153	/* watch out for a 0 len io from a tricksy fs */
1154	if (iov_iter_rw(iter) == READ && !iov_iter_count(iter))
1155	return 0;
1156
1157	dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1158	retval = -ENOMEM;
1159	if (!dio)
1160	goto out;
1161	/*
1162	* Believe it or not, zeroing out the page array caused a .5%
1163	* performance regression in a database benchmark. So, we take
1164	* care to only zero out what's needed.
1165	*/
1166	memset(dio, 0, offsetof(struct dio, pages));
1167
1168	dio->flags = flags;
1169	if (dio->flags & DIO_LOCKING) {
1170	if (iov_iter_rw(iter) == READ) {
1171	struct address_space *mapping =
1172	iocb->ki_filp->f_mapping;
1173
1174	/* will be released by direct_io_worker */
1175	inode_lock(inode);
1176
1177	retval = filemap_write_and_wait_range(mapping, offset,
1178	end - 1);
1179	if (retval) {
1180	inode_unlock(inode);
1181	kmem_cache_free(dio_cache, dio);
1182	goto out;
1183	}
1184	}
1185	}
1186
1187	/* Once we sampled i_size check for reads beyond EOF */
1188	dio->i_size = i_size_read(inode);
1189	if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1190	if (dio->flags & DIO_LOCKING)
1191	inode_unlock(inode);
1192	kmem_cache_free(dio_cache, dio);
1193	retval = 0;
1194	goto out;
1195	}
1196
1197	/*
1198	* For file extending writes updating i_size before data writeouts
1199	* complete can expose uninitialized blocks in dumb filesystems.
1200	* In that case we need to wait for I/O completion even if asked
1201	* for an asynchronous write.
1202	*/
1203	if (is_sync_kiocb(iocb))
1204	dio->is_async = false;
1205	else if (!(dio->flags & DIO_ASYNC_EXTEND) &&
1206	iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
1207	dio->is_async = false;
1208	else
1209	dio->is_async = true;
1210
1211	dio->inode = inode;
1212	if (iov_iter_rw(iter) == WRITE) {
1213	dio->op = REQ_OP_WRITE;
1214	dio->op_flags = WRITE_ODIRECT;
1215	} else {
1216	dio->op = REQ_OP_READ;
1217	}
1218
1219	/*
1220	* For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1221	* so that we can call ->fsync.
1222	*/
1223	if (dio->is_async && iov_iter_rw(iter) == WRITE &&
1224	((iocb->ki_filp->f_flags & O_DSYNC) ||
1225	IS_SYNC(iocb->ki_filp->f_mapping->host))) {
1226	retval = dio_set_defer_completion(dio);
1227	if (retval) {
1228	/*
1229	* We grab i_mutex only for reads so we don't have
1230	* to release it here
1231	*/
1232	kmem_cache_free(dio_cache, dio);
1233	goto out;
1234	}
1235	}
1236
1237	/*
1238	* Will be decremented at I/O completion time.
1239	*/
1240	if (!(dio->flags & DIO_SKIP_DIO_COUNT))
1241	inode_dio_begin(inode);
1242
1243	retval = 0;
1244	sdio.blkbits = blkbits;
1245	sdio.blkfactor = i_blkbits - blkbits;
1246	sdio.block_in_file = offset >> blkbits;
1247
1248	sdio.get_block = get_block;
1249	dio->end_io = end_io;
1250	sdio.submit_io = submit_io;
1251	sdio.final_block_in_bio = -1;
1252	sdio.next_block_for_io = -1;
1253
1254	dio->iocb = iocb;
1255
1256	spin_lock_init(&dio->bio_lock);
1257	dio->refcount = 1;
1258
1259	dio->should_dirty = (iter->type == ITER_IOVEC);
1260	sdio.iter = iter;
1261	sdio.final_block_in_request =
1262	(offset + iov_iter_count(iter)) >> blkbits;
1263
1264	/*
1265	* In case of non-aligned buffers, we may need 2 more
1266	* pages since we need to zero out first and last block.
1267	*/
1268	if (unlikely(sdio.blkfactor))
1269	sdio.pages_in_io = 2;
1270
1271	sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
1272
1273	blk_start_plug(&plug);
1274
1275	retval = do_direct_IO(dio, &sdio, &map_bh);
1276	if (retval)
1277	dio_cleanup(dio, &sdio);
1278
1279	if (retval == -ENOTBLK) {
1280	/*
1281	* The remaining part of the request will be
1282	* be handled by buffered I/O when we return
1283	*/
1284	retval = 0;
1285	}
1286	/*
1287	* There may be some unwritten disk at the end of a part-written
1288	* fs-block-sized block. Go zero that now.
1289	*/
1290	dio_zero_block(dio, &sdio, 1, &map_bh);
1291
1292	if (sdio.cur_page) {
1293	ssize_t ret2;
1294
1295	ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1296	if (retval == 0)
1297	retval = ret2;
1298	put_page(sdio.cur_page);
1299	sdio.cur_page = NULL;
1300	}
1301	if (sdio.bio)
1302	dio_bio_submit(dio, &sdio);
1303
1304	blk_finish_plug(&plug);
1305
1306	/*
1307	* It is possible that, we return short IO due to end of file.
1308	* In that case, we need to release all the pages we got hold on.
1309	*/
1310	dio_cleanup(dio, &sdio);
1311
1312	/*
1313	* All block lookups have been performed. For READ requests
1314	* we can let i_mutex go now that its achieved its purpose
1315	* of protecting us from looking up uninitialized blocks.
1316	*/
1317	if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
1318	inode_unlock(dio->inode);
1319
1320	/*
1321	* The only time we want to leave bios in flight is when a successful
1322	* partial aio read or full aio write have been setup. In that case
1323	* bio completion will call aio_complete. The only time it's safe to
1324	* call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1325	* This had *better* be the only place that raises -EIOCBQUEUED.
1326	*/
1327	BUG_ON(retval == -EIOCBQUEUED);
1328	if (dio->is_async && retval == 0 && dio->result &&
1329	(iov_iter_rw(iter) == READ || dio->result == count))
1330	retval = -EIOCBQUEUED;
1331	else
1332	dio_await_completion(dio);
1333
1334	if (drop_refcount(dio) == 0) {
1335	retval = dio_complete(dio, retval, false);
1336	} else
1337	BUG_ON(retval != -EIOCBQUEUED);
1338
1339	out:
1340	return retval;
1341	}
1342
1343	ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1344	struct block_device *bdev, struct iov_iter *iter,
1345	get_block_t get_block,
1346	dio_iodone_t end_io, dio_submit_t submit_io,
1347	int flags)
1348	{
1349	/*
1350	* The block device state is needed in the end to finally
1351	* submit everything. Since it's likely to be cache cold
1352	* prefetch it here as first thing to hide some of the
1353	* latency.
1354	*
1355	* Attempt to prefetch the pieces we likely need later.
1356	*/
1357	prefetch(&bdev->bd_disk->part_tbl);
1358	prefetch(bdev->bd_queue);
1359	prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1360
1361	return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block,
1362	end_io, submit_io, flags);
1363	}
1364
1365	EXPORT_SYMBOL(__blockdev_direct_IO);
1366
1367	static __init int dio_init(void)
1368	{
1369	dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1370	return 0;
1371	}
1372	module_init(dio_init)
1373