blob: 3aafb3343a65c76fb66211228564d73ce9097485 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * fs/direct-io.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 *
6 * O_DIRECT
7 *
Francois Camie1f8e872008-10-15 22:01:59 -07008 * 04Jul2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 * Initial version
10 * 11Sep2002 janetinc@us.ibm.com
11 * added readv/writev support.
Francois Camie1f8e872008-10-15 22:01:59 -070012 * 29Oct2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -070013 * 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>
Andrew Morton98c4d572006-12-10 02:19:47 -080030#include <linux/task_io_accounting_ops.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070031#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>
Arun Sharma600634972011-07-26 16:09:06 -070038#include <linux/atomic.h>
Andi Kleen65dd2aa2012-01-12 17:20:35 -080039#include <linux/prefetch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070040
41/*
42 * How many user pages to map in one call to get_user_pages(). This determines
Andi Kleencde1ecb2011-08-01 21:38:04 -070043 * the size of a structure in the slab cache
Linus Torvalds1da177e2005-04-16 15:20:36 -070044 */
45#define DIO_PAGES 64
46
47/*
Lukas Czernerffe51f02017-10-17 08:43:09 -060048 * Flags for dio_complete()
49 */
50#define DIO_COMPLETE_ASYNC 0x01 /* This is async IO */
51#define DIO_COMPLETE_INVALIDATE 0x02 /* Can invalidate pages */
52
53/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * This code generally works in units of "dio_blocks". A dio_block is
55 * somewhere between the hard sector size and the filesystem block size. it
56 * is determined on a per-invocation basis. When talking to the filesystem
57 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
58 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
59 * to bio_block quantities by shifting left by blkfactor.
60 *
61 * If blkfactor is zero then the user's request was aligned to the filesystem's
62 * blocksize.
Linus Torvalds1da177e2005-04-16 15:20:36 -070063 */
64
Andi Kleeneb28be22011-08-01 21:38:03 -070065/* dio_state only used in the submission path */
66
67struct dio_submit {
Linus Torvalds1da177e2005-04-16 15:20:36 -070068 struct bio *bio; /* bio under assembly */
Linus Torvalds1da177e2005-04-16 15:20:36 -070069 unsigned blkbits; /* doesn't change */
70 unsigned blkfactor; /* When we're using an alignment which
71 is finer than the filesystem's soft
72 blocksize, this specifies how much
73 finer. blkfactor=2 means 1/4-block
74 alignment. Does not change */
75 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
76 been performed at the start of a
77 write */
78 int pages_in_io; /* approximate total IO pages */
Linus Torvalds1da177e2005-04-16 15:20:36 -070079 sector_t block_in_file; /* Current offset into the underlying
80 file in dio_block units. */
81 unsigned blocks_available; /* At block_in_file. changes */
Andi Kleen0dc2bc42011-08-01 21:38:05 -070082 int reap_counter; /* rate limit reaping */
Linus Torvalds1da177e2005-04-16 15:20:36 -070083 sector_t final_block_in_request;/* doesn't change */
Linus Torvalds1da177e2005-04-16 15:20:36 -070084 int boundary; /* prev block is at a boundary */
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -080085 get_block_t *get_block; /* block mapping function */
Josef Bacikfacd07b2010-05-23 11:00:55 -040086 dio_submit_t *submit_io; /* IO submition function */
Andi Kleeneb28be22011-08-01 21:38:03 -070087
Josef Bacikfacd07b2010-05-23 11:00:55 -040088 loff_t logical_offset_in_bio; /* current first logical block in bio */
Linus Torvalds1da177e2005-04-16 15:20:36 -070089 sector_t final_block_in_bio; /* current final block in bio + 1 */
90 sector_t next_block_for_io; /* next block to be put under IO,
91 in dio_blocks units */
Linus Torvalds1da177e2005-04-16 15:20:36 -070092
93 /*
94 * Deferred addition of a page to the dio. These variables are
95 * private to dio_send_cur_page(), submit_page_section() and
96 * dio_bio_add_page().
97 */
98 struct page *cur_page; /* The page */
99 unsigned cur_page_offset; /* Offset into it, in bytes */
100 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
101 sector_t cur_page_block; /* Where it starts */
Josef Bacikfacd07b2010-05-23 11:00:55 -0400102 loff_t cur_page_fs_offset; /* Offset in file */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103
Al Viro7b2c99d2014-03-15 04:05:57 -0400104 struct iov_iter *iter;
Jeff Moyer23aee092009-12-15 16:47:49 -0800105 /*
106 * Page queue. These variables belong to dio_refill_pages() and
107 * dio_get_page().
108 */
109 unsigned head; /* next page to process */
110 unsigned tail; /* last valid page + 1 */
Al Viro7b2c99d2014-03-15 04:05:57 -0400111 size_t from, to;
Andi Kleeneb28be22011-08-01 21:38:03 -0700112};
113
114/* dio_state communicated between submission path and end_io */
115struct dio {
116 int flags; /* doesn't change */
Mike Christie8a4c1e42016-06-05 14:31:50 -0500117 int op;
118 int op_flags;
Jens Axboe15c4f632015-10-27 14:09:51 +0900119 blk_qc_t bio_cookie;
Christoph Hellwig74d46992017-08-23 19:10:32 +0200120 struct gendisk *bio_disk;
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700121 struct inode *inode;
Andi Kleeneb28be22011-08-01 21:38:03 -0700122 loff_t i_size; /* i_size when submitted */
123 dio_iodone_t *end_io; /* IO completion function */
Andi Kleeneb28be22011-08-01 21:38:03 -0700124
Andi Kleen18772642011-08-01 21:38:07 -0700125 void *private; /* copy from map_bh.b_private */
Andi Kleeneb28be22011-08-01 21:38:03 -0700126
127 /* BIO completion state */
128 spinlock_t bio_lock; /* protects BIO fields below */
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700129 int page_errors; /* errno from get_user_pages() */
130 int is_async; /* is IO async ? */
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200131 bool defer_completion; /* defer AIO completion to workqueue? */
Ming Lei53cbf3b2015-08-17 10:31:46 +0800132 bool should_dirty; /* if pages should be dirtied */
Andi Kleen0dc2bc42011-08-01 21:38:05 -0700133 int io_error; /* IO error in completion path */
Andi Kleeneb28be22011-08-01 21:38:03 -0700134 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 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700140 ssize_t result; /* IO result */
141
Jeff Moyer23aee092009-12-15 16:47:49 -0800142 /*
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 */
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200147 union {
148 struct page *pages[DIO_PAGES]; /* page buffer */
149 struct work_struct complete_work;/* deferred AIO completion */
150 };
Andi Kleen6e8267f2011-08-01 21:38:06 -0700151} ____cacheline_aligned_in_smp;
152
153static struct kmem_cache *dio_cache __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154
155/*
156 * How many pages are in the queue?
157 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700158static inline unsigned dio_pages_present(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159{
Andi Kleeneb28be22011-08-01 21:38:03 -0700160 return sdio->tail - sdio->head;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161}
162
163/*
164 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
165 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700166static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167{
Al Viro7b2c99d2014-03-15 04:05:57 -0400168 ssize_t ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169
Miklos Szeredi2c809292014-09-24 17:09:11 +0200170 ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
Al Viro7b2c99d2014-03-15 04:05:57 -0400171 &sdio->from);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172
Mike Christie8a4c1e42016-06-05 14:31:50 -0500173 if (ret < 0 && sdio->blocks_available && (dio->op == REQ_OP_WRITE)) {
Nick Piggin557ed1f2007-10-16 01:24:40 -0700174 struct page *page = ZERO_PAGE(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175 /*
176 * A memory fault, but the filesystem has some outstanding
177 * mapped blocks. We need to use those blocks up to avoid
178 * leaking stale data in the file.
179 */
180 if (dio->page_errors == 0)
181 dio->page_errors = ret;
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300182 get_page(page);
Nick Pigginb5810032005-10-29 18:16:12 -0700183 dio->pages[0] = page;
Andi Kleeneb28be22011-08-01 21:38:03 -0700184 sdio->head = 0;
185 sdio->tail = 1;
Al Viro7b2c99d2014-03-15 04:05:57 -0400186 sdio->from = 0;
187 sdio->to = PAGE_SIZE;
188 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189 }
190
191 if (ret >= 0) {
Al Viro7b2c99d2014-03-15 04:05:57 -0400192 iov_iter_advance(sdio->iter, ret);
193 ret += sdio->from;
Andi Kleeneb28be22011-08-01 21:38:03 -0700194 sdio->head = 0;
Al Viro7b2c99d2014-03-15 04:05:57 -0400195 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
196 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
197 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199 return ret;
200}
201
202/*
203 * Get another userspace page. Returns an ERR_PTR on error. Pages are
204 * buffered inside the dio so that we can call get_user_pages() against a
205 * decent number of pages, less frequently. To provide nicer use of the
206 * L1 cache.
207 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700208static inline struct page *dio_get_page(struct dio *dio,
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300209 struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210{
Andi Kleeneb28be22011-08-01 21:38:03 -0700211 if (dio_pages_present(sdio) == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212 int ret;
213
Andi Kleeneb28be22011-08-01 21:38:03 -0700214 ret = dio_refill_pages(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215 if (ret)
216 return ERR_PTR(ret);
Andi Kleeneb28be22011-08-01 21:38:03 -0700217 BUG_ON(dio_pages_present(sdio) == 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218 }
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300219 return dio->pages[sdio->head];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220}
221
Zach Brown6d544bb2006-12-10 02:20:54 -0800222/**
223 * dio_complete() - called when all DIO BIO I/O has been completed
224 * @offset: the byte offset in the file of the completed operation
225 *
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200226 * This drops i_dio_count, lets interested parties know that a DIO operation
227 * has completed, and calculates the resulting return code for the operation.
Zach Brown6d544bb2006-12-10 02:20:54 -0800228 *
229 * It lets the filesystem know if it registered an interest earlier via
230 * get_block. Pass the private field of the map buffer_head so that
231 * filesystems can use it to hold additional state between get_block calls and
232 * dio_complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233 */
Lukas Czernerffe51f02017-10-17 08:43:09 -0600234static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700235{
Christoph Hellwig716b9bc2016-04-07 08:51:59 -0700236 loff_t offset = dio->iocb->ki_pos;
Zach Brown6d544bb2006-12-10 02:20:54 -0800237 ssize_t transferred = 0;
Lukas Czerner332391a2017-09-21 08:16:29 -0600238 int err;
Zach Brown6d544bb2006-12-10 02:20:54 -0800239
Zach Brown8459d862006-12-10 02:21:05 -0800240 /*
241 * AIO submission can race with bio completion to get here while
242 * expecting to have the last io completed by bio completion.
243 * In that case -EIOCBQUEUED is in fact not an error we want
244 * to preserve through this call.
245 */
246 if (ret == -EIOCBQUEUED)
247 ret = 0;
248
Zach Brown6d544bb2006-12-10 02:20:54 -0800249 if (dio->result) {
250 transferred = dio->result;
251
252 /* Check for short read case */
Mike Christie8a4c1e42016-06-05 14:31:50 -0500253 if ((dio->op == REQ_OP_READ) &&
254 ((offset + transferred) > dio->i_size))
Zach Brown6d544bb2006-12-10 02:20:54 -0800255 transferred = dio->i_size - offset;
Al Viro4038acd2016-10-03 20:38:55 -0400256 /* ignore EFAULT if some IO has been done */
257 if (unlikely(ret == -EFAULT) && transferred)
258 ret = 0;
Zach Brown6d544bb2006-12-10 02:20:54 -0800259 }
260
Zach Brown6d544bb2006-12-10 02:20:54 -0800261 if (ret == 0)
262 ret = dio->page_errors;
263 if (ret == 0)
264 ret = dio->io_error;
265 if (ret == 0)
266 ret = transferred;
267
Eryu Guan5e25c262017-10-13 09:47:46 -0700268 if (dio->end_io) {
269 // XXX: ki_pos??
270 err = dio->end_io(dio->iocb, offset, ret, dio->private);
271 if (err)
272 ret = err;
273 }
274
Lukas Czerner332391a2017-09-21 08:16:29 -0600275 /*
276 * Try again to invalidate clean pages which might have been cached by
277 * non-direct readahead, or faulted in by get_user_pages() if the source
278 * of the write was an mmap'ed region of the file we're writing. Either
279 * one is a pretty crazy thing to do, so we don't support it 100%. If
280 * this invalidation fails, tough, the write still worked...
Eryu Guan5e25c262017-10-13 09:47:46 -0700281 *
282 * And this page cache invalidation has to be after dio->end_io(), as
283 * some filesystems convert unwritten extents to real allocations in
284 * end_io() when necessary, otherwise a racing buffer read would cache
285 * zeros from unwritten extents.
Lukas Czerner332391a2017-09-21 08:16:29 -0600286 */
Lukas Czernerffe51f02017-10-17 08:43:09 -0600287 if (flags & DIO_COMPLETE_INVALIDATE &&
288 ret > 0 && dio->op == REQ_OP_WRITE &&
Lukas Czerner332391a2017-09-21 08:16:29 -0600289 dio->inode->i_mapping->nrpages) {
290 err = invalidate_inode_pages2_range(dio->inode->i_mapping,
291 offset >> PAGE_SHIFT,
292 (offset + ret - 1) >> PAGE_SHIFT);
293 WARN_ON_ONCE(err);
294 }
295
Jens Axboefe0f07d2015-04-15 17:05:48 -0600296 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
297 inode_dio_end(dio->inode);
298
Lukas Czernerffe51f02017-10-17 08:43:09 -0600299 if (flags & DIO_COMPLETE_ASYNC) {
Christoph Hellwige2592212016-04-07 08:52:01 -0700300 /*
301 * generic_write_sync expects ki_pos to have been updated
302 * already, but the submission path only does this for
303 * synchronous I/O.
304 */
305 dio->iocb->ki_pos += transferred;
Christoph Hellwig02afc272013-09-04 15:04:40 +0200306
Mike Christie8a4c1e42016-06-05 14:31:50 -0500307 if (dio->op == REQ_OP_WRITE)
Christoph Hellwige2592212016-04-07 08:52:01 -0700308 ret = generic_write_sync(dio->iocb, transferred);
Christoph Hellwig04b2fa92015-02-02 14:49:06 +0100309 dio->iocb->ki_complete(dio->iocb, ret, 0);
Christoph Hellwig02afc272013-09-04 15:04:40 +0200310 }
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200311
312 kmem_cache_free(dio_cache, dio);
Zach Brown6d544bb2006-12-10 02:20:54 -0800313 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314}
315
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200316static void dio_aio_complete_work(struct work_struct *work)
317{
318 struct dio *dio = container_of(work, struct dio, complete_work);
319
Lukas Czernerffe51f02017-10-17 08:43:09 -0600320 dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200321}
322
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +0200323static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200324
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325/*
326 * Asynchronous IO callback.
327 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200328static void dio_bio_end_aio(struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329{
330 struct dio *dio = bio->bi_private;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800331 unsigned long remaining;
332 unsigned long flags;
Lukas Czerner332391a2017-09-21 08:16:29 -0600333 bool defer_completion = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335 /* cleanup the bio */
336 dio_bio_complete(dio, bio);
Zach Brown02732012006-12-10 02:20:59 -0800337
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800338 spin_lock_irqsave(&dio->bio_lock, flags);
339 remaining = --dio->refcount;
340 if (remaining == 1 && dio->waiter)
Zach Brown20258b2b2006-12-10 02:21:01 -0800341 wake_up_process(dio->waiter);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800342 spin_unlock_irqrestore(&dio->bio_lock, flags);
Zach Brown20258b2b2006-12-10 02:21:01 -0800343
Zach Brown8459d862006-12-10 02:21:05 -0800344 if (remaining == 0) {
Lukas Czerner332391a2017-09-21 08:16:29 -0600345 /*
346 * Defer completion when defer_completion is set or
347 * when the inode has pages mapped and this is AIO write.
348 * We need to invalidate those pages because there is a
349 * chance they contain stale data in the case buffered IO
350 * went in between AIO submission and completion into the
351 * same region.
352 */
353 if (dio->result)
354 defer_completion = dio->defer_completion ||
355 (dio->op == REQ_OP_WRITE &&
356 dio->inode->i_mapping->nrpages);
357 if (defer_completion) {
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200358 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
359 queue_work(dio->inode->i_sb->s_dio_done_wq,
360 &dio->complete_work);
361 } else {
Lukas Czernerffe51f02017-10-17 08:43:09 -0600362 dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200363 }
Zach Brown8459d862006-12-10 02:21:05 -0800364 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365}
366
367/*
368 * The BIO completion handler simply queues the BIO up for the process-context
369 * handler.
370 *
371 * During I/O bi_private points at the dio. After I/O, bi_private is used to
372 * implement a singly-linked list of completed BIOs, at dio->bio_list.
373 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200374static void dio_bio_end_io(struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375{
376 struct dio *dio = bio->bi_private;
377 unsigned long flags;
378
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379 spin_lock_irqsave(&dio->bio_lock, flags);
380 bio->bi_private = dio->bio_list;
381 dio->bio_list = bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800382 if (--dio->refcount == 1 && dio->waiter)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383 wake_up_process(dio->waiter);
384 spin_unlock_irqrestore(&dio->bio_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385}
386
Josef Bacikfacd07b2010-05-23 11:00:55 -0400387/**
388 * dio_end_io - handle the end io action for the given bio
389 * @bio: The direct io bio thats being completed
Josef Bacikfacd07b2010-05-23 11:00:55 -0400390 *
391 * This is meant to be called by any filesystem that uses their own dio_submit_t
392 * so that the DIO specific endio actions are dealt with after the filesystem
393 * has done it's completion work.
394 */
Christoph Hellwig40553512017-06-03 09:37:58 +0200395void dio_end_io(struct bio *bio)
Josef Bacikfacd07b2010-05-23 11:00:55 -0400396{
397 struct dio *dio = bio->bi_private;
398
399 if (dio->is_async)
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200400 dio_bio_end_aio(bio);
Josef Bacikfacd07b2010-05-23 11:00:55 -0400401 else
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200402 dio_bio_end_io(bio);
Josef Bacikfacd07b2010-05-23 11:00:55 -0400403}
404EXPORT_SYMBOL_GPL(dio_end_io);
405
Andi Kleenba253fb2011-08-01 21:38:08 -0700406static inline void
Andi Kleeneb28be22011-08-01 21:38:03 -0700407dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
408 struct block_device *bdev,
409 sector_t first_sector, int nr_vecs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410{
411 struct bio *bio;
412
David Dillow20d96002011-01-20 14:44:22 -0800413 /*
414 * bio_alloc() is guaranteed to return a bio when called with
Mel Gorman71baba42015-11-06 16:28:28 -0800415 * __GFP_RECLAIM and we request a valid number of vectors.
David Dillow20d96002011-01-20 14:44:22 -0800416 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417 bio = bio_alloc(GFP_KERNEL, nr_vecs);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418
Christoph Hellwig74d46992017-08-23 19:10:32 +0200419 bio_set_dev(bio, bdev);
Kent Overstreet4f024f32013-10-11 15:44:27 -0700420 bio->bi_iter.bi_sector = first_sector;
Mike Christie8a4c1e42016-06-05 14:31:50 -0500421 bio_set_op_attrs(bio, dio->op, dio->op_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 if (dio->is_async)
423 bio->bi_end_io = dio_bio_end_aio;
424 else
425 bio->bi_end_io = dio_bio_end_io;
426
Jens Axboe45d06cf2017-06-27 11:01:22 -0600427 bio->bi_write_hint = dio->iocb->ki_hint;
428
Andi Kleeneb28be22011-08-01 21:38:03 -0700429 sdio->bio = bio;
430 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431}
432
433/*
434 * In the AIO read case we speculatively dirty the pages before starting IO.
435 * During IO completion, any of these pages which happen to have been written
436 * back will be redirtied by bio_check_pages_dirty().
Zach Brown02732012006-12-10 02:20:59 -0800437 *
438 * bios hold a dio reference between submit_bio and ->end_io.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700440static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441{
Andi Kleeneb28be22011-08-01 21:38:03 -0700442 struct bio *bio = sdio->bio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800443 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444
445 bio->bi_private = dio;
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800446
447 spin_lock_irqsave(&dio->bio_lock, flags);
448 dio->refcount++;
449 spin_unlock_irqrestore(&dio->bio_lock, flags);
450
Mike Christie8a4c1e42016-06-05 14:31:50 -0500451 if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452 bio_set_pages_dirty(bio);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800453
Christoph Hellwig74d46992017-08-23 19:10:32 +0200454 dio->bio_disk = bio->bi_disk;
Jens Axboec1c53462015-11-10 10:14:38 -0700455
Jens Axboe15c4f632015-10-27 14:09:51 +0900456 if (sdio->submit_io) {
Mike Christie8a4c1e42016-06-05 14:31:50 -0500457 sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio);
Jens Axboe15c4f632015-10-27 14:09:51 +0900458 dio->bio_cookie = BLK_QC_T_NONE;
Jens Axboec1c53462015-11-10 10:14:38 -0700459 } else
Mike Christie4e49ea42016-06-05 14:31:41 -0500460 dio->bio_cookie = submit_bio(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461
Andi Kleeneb28be22011-08-01 21:38:03 -0700462 sdio->bio = NULL;
463 sdio->boundary = 0;
464 sdio->logical_offset_in_bio = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465}
466
467/*
468 * Release any resources in case of a failure
469 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700470static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471{
Al Viro7b2c99d2014-03-15 04:05:57 -0400472 while (sdio->head < sdio->tail)
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300473 put_page(dio->pages[sdio->head++]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474}
475
476/*
Zach Brown02732012006-12-10 02:20:59 -0800477 * Wait for the next BIO to complete. Remove it and return it. NULL is
478 * returned once all BIOs have been completed. This must only be called once
479 * all bios have been issued so that dio->refcount can only decrease. This
480 * requires that that the caller hold a reference on the dio.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 */
482static struct bio *dio_await_one(struct dio *dio)
483{
484 unsigned long flags;
Zach Brown02732012006-12-10 02:20:59 -0800485 struct bio *bio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486
487 spin_lock_irqsave(&dio->bio_lock, flags);
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800488
489 /*
490 * Wait as long as the list is empty and there are bios in flight. bio
491 * completion drops the count, maybe adds to the list, and wakes while
492 * holding the bio_lock so we don't need set_current_state()'s barrier
493 * and can call it after testing our condition.
494 */
495 while (dio->refcount > 1 && dio->bio_list == NULL) {
496 __set_current_state(TASK_UNINTERRUPTIBLE);
497 dio->waiter = current;
498 spin_unlock_irqrestore(&dio->bio_lock, flags);
Christoph Hellwigc43c83a2016-03-03 16:04:02 +0100499 if (!(dio->iocb->ki_flags & IOCB_HIPRI) ||
Christoph Hellwigea435e12017-11-02 21:29:54 +0300500 !blk_poll(dio->bio_disk->queue, dio->bio_cookie))
Jens Axboe15c4f632015-10-27 14:09:51 +0900501 io_schedule();
Zach Brown5eb6c7a2006-12-10 02:21:07 -0800502 /* wake up sets us TASK_RUNNING */
503 spin_lock_irqsave(&dio->bio_lock, flags);
504 dio->waiter = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505 }
Zach Brown02732012006-12-10 02:20:59 -0800506 if (dio->bio_list) {
507 bio = dio->bio_list;
508 dio->bio_list = bio->bi_private;
509 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 spin_unlock_irqrestore(&dio->bio_lock, flags);
511 return bio;
512}
513
514/*
515 * Process one completed BIO. No locks are held.
516 */
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +0200517static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518{
Kent Overstreetcb34e052012-09-05 15:22:02 -0700519 struct bio_vec *bvec;
520 unsigned i;
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +0200521 blk_status_t err = bio->bi_status;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522
Goldwyn Rodrigues03a07c92017-06-20 07:05:46 -0500523 if (err) {
524 if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
525 dio->io_error = -EAGAIN;
526 else
527 dio->io_error = -EIO;
528 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529
Mike Christie8a4c1e42016-06-05 14:31:50 -0500530 if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty) {
Mike Krinkin7ddc9712016-01-30 19:09:59 +0300531 bio_check_pages_dirty(bio); /* transfers ownership */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532 } else {
Kent Overstreetcb34e052012-09-05 15:22:02 -0700533 bio_for_each_segment_all(bvec, bio, i) {
534 struct page *page = bvec->bv_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535
Mike Christie8a4c1e42016-06-05 14:31:50 -0500536 if (dio->op == REQ_OP_READ && !PageCompound(page) &&
Ming Lei53cbf3b2015-08-17 10:31:46 +0800537 dio->should_dirty)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538 set_page_dirty_lock(page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300539 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540 }
541 bio_put(bio);
542 }
Sasha Levin9b81c842015-08-10 19:05:18 -0400543 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544}
545
546/*
Zach Brown02732012006-12-10 02:20:59 -0800547 * Wait on and process all in-flight BIOs. This must only be called once
548 * all bios have been issued so that the refcount can only decrease.
549 * This just waits for all bios to make it through dio_bio_complete. IO
Robert P. J. Daybeb7dd82007-05-09 07:14:03 +0200550 * errors are propagated through dio->io_error and should be propagated via
Zach Brown02732012006-12-10 02:20:59 -0800551 * dio_complete().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800553static void dio_await_completion(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554{
Zach Brown02732012006-12-10 02:20:59 -0800555 struct bio *bio;
556 do {
557 bio = dio_await_one(dio);
558 if (bio)
559 dio_bio_complete(dio, bio);
560 } while (bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561}
562
563/*
564 * A really large O_DIRECT read or write can generate a lot of BIOs. So
565 * to keep the memory consumption sane we periodically reap any completed BIOs
566 * during the BIO generation phase.
567 *
568 * This also helps to limit the peak amount of pinned userspace memory.
569 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700570static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571{
572 int ret = 0;
573
Andi Kleeneb28be22011-08-01 21:38:03 -0700574 if (sdio->reap_counter++ >= 64) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575 while (dio->bio_list) {
576 unsigned long flags;
577 struct bio *bio;
578 int ret2;
579
580 spin_lock_irqsave(&dio->bio_lock, flags);
581 bio = dio->bio_list;
582 dio->bio_list = bio->bi_private;
583 spin_unlock_irqrestore(&dio->bio_lock, flags);
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +0200584 ret2 = blk_status_to_errno(dio_bio_complete(dio, bio));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 if (ret == 0)
586 ret = ret2;
587 }
Andi Kleeneb28be22011-08-01 21:38:03 -0700588 sdio->reap_counter = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589 }
590 return ret;
591}
592
593/*
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200594 * Create workqueue for deferred direct IO completions. We allocate the
595 * workqueue when it's first needed. This avoids creating workqueue for
596 * filesystems that don't need it and also allows us to create the workqueue
597 * late enough so the we can include s_id in the name of the workqueue.
598 */
Christoph Hellwigec1b8262016-11-30 14:33:53 +1100599int sb_init_dio_done_wq(struct super_block *sb)
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200600{
Olof Johansson45150c42013-09-09 10:34:23 -0700601 struct workqueue_struct *old;
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200602 struct workqueue_struct *wq = alloc_workqueue("dio/%s",
603 WQ_MEM_RECLAIM, 0,
604 sb->s_id);
605 if (!wq)
606 return -ENOMEM;
607 /*
608 * This has to be atomic as more DIOs can race to create the workqueue
609 */
Olof Johansson45150c42013-09-09 10:34:23 -0700610 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200611 /* Someone created workqueue before us? Free ours... */
Olof Johansson45150c42013-09-09 10:34:23 -0700612 if (old)
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200613 destroy_workqueue(wq);
614 return 0;
615}
616
617static int dio_set_defer_completion(struct dio *dio)
618{
619 struct super_block *sb = dio->inode->i_sb;
620
621 if (dio->defer_completion)
622 return 0;
623 dio->defer_completion = true;
624 if (!sb->s_dio_done_wq)
625 return sb_init_dio_done_wq(sb);
626 return 0;
627}
628
629/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630 * Call into the fs to map some more disk blocks. We record the current number
Andi Kleeneb28be22011-08-01 21:38:03 -0700631 * of available blocks at sdio->blocks_available. These are in units of the
Fabian Frederick93407472017-02-27 14:28:32 -0800632 * fs blocksize, i_blocksize(inode).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700633 *
634 * The fs is allowed to map lots of blocks at once. If it wants to do that,
635 * it uses the passed inode-relative block number as the file offset, as usual.
636 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800637 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638 * has remaining to do. The fs should not map more than this number of blocks.
639 *
640 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
641 * indicate how much contiguous disk space has been made available at
642 * bh->b_blocknr.
643 *
644 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
645 * This isn't very efficient...
646 *
647 * In the case of filesystem holes: the fs may return an arbitrarily-large
648 * hole by returning an appropriate value in b_size and by clearing
649 * buffer_mapped(). However the direct-io code will only process holes one
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800650 * block at a time - it will repeatedly call get_block() as it walks the hole.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 */
Andi Kleen18772642011-08-01 21:38:07 -0700652static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
653 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654{
655 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
Tao Maae55e1a2012-01-12 17:20:33 -0800657 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 unsigned long fs_count; /* Number of filesystem-sized blocks */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659 int create;
Linus Torvaldsab738572012-11-29 12:27:00 -0800660 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661
662 /*
663 * If there was a memory error and we've overwritten all the
664 * mapped blocks then we can now return that memory error
665 */
666 ret = dio->page_errors;
667 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700668 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
669 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
Tao Maae55e1a2012-01-12 17:20:33 -0800670 fs_endblk = (sdio->final_block_in_request - 1) >>
671 sdio->blkfactor;
672 fs_count = fs_endblk - fs_startblk + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673
Nathan Scott3c674e72006-03-29 09:26:15 +1000674 map_bh->b_state = 0;
Linus Torvaldsab738572012-11-29 12:27:00 -0800675 map_bh->b_size = fs_count << i_blkbits;
Nathan Scott3c674e72006-03-29 09:26:15 +1000676
Christoph Hellwig5fe878ae2009-12-15 16:47:50 -0800677 /*
Eryu Guan9ecd10b2016-05-27 14:27:18 -0700678 * For writes that could fill holes inside i_size on a
679 * DIO_SKIP_HOLES filesystem we forbid block creations: only
680 * overwrites are permitted. We will return early to the caller
681 * once we see an unmapped buffer head returned, and the caller
682 * will fall back to buffered I/O.
Christoph Hellwig5fe878ae2009-12-15 16:47:50 -0800683 *
684 * Otherwise the decision is left to the get_blocks method,
685 * which may decide to handle it or also return an unmapped
686 * buffer head.
687 */
Mike Christie8a4c1e42016-06-05 14:31:50 -0500688 create = dio->op == REQ_OP_WRITE;
Christoph Hellwig5fe878ae2009-12-15 16:47:50 -0800689 if (dio->flags & DIO_SKIP_HOLES) {
Eryu Guan9ecd10b2016-05-27 14:27:18 -0700690 if (fs_startblk <= ((i_size_read(dio->inode) - 1) >>
691 i_blkbits))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692 create = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693 }
Nathan Scott3c674e72006-03-29 09:26:15 +1000694
Andi Kleeneb28be22011-08-01 21:38:03 -0700695 ret = (*sdio->get_block)(dio->inode, fs_startblk,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696 map_bh, create);
Andi Kleen18772642011-08-01 21:38:07 -0700697
698 /* Store for completion */
699 dio->private = map_bh->b_private;
Christoph Hellwig7b7a8662013-09-04 15:04:39 +0200700
701 if (ret == 0 && buffer_defer_completion(map_bh))
702 ret = dio_set_defer_completion(dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703 }
704 return ret;
705}
706
707/*
708 * There is no bio. Make one now.
709 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700710static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
711 sector_t start_sector, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712{
713 sector_t sector;
714 int ret, nr_pages;
715
Andi Kleeneb28be22011-08-01 21:38:03 -0700716 ret = dio_bio_reap(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717 if (ret)
718 goto out;
Andi Kleeneb28be22011-08-01 21:38:03 -0700719 sector = start_sector << (sdio->blkbits - 9);
Kent Overstreetb54ffb72015-05-19 14:31:01 +0200720 nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721 BUG_ON(nr_pages <= 0);
Andi Kleen18772642011-08-01 21:38:07 -0700722 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
Andi Kleeneb28be22011-08-01 21:38:03 -0700723 sdio->boundary = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724out:
725 return ret;
726}
727
728/*
729 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
730 * that was successful then update final_block_in_bio and take a ref against
731 * the just-added page.
732 *
733 * Return zero on success. Non-zero means the caller needs to start a new BIO.
734 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700735static inline int dio_bio_add_page(struct dio_submit *sdio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736{
737 int ret;
738
Andi Kleeneb28be22011-08-01 21:38:03 -0700739 ret = bio_add_page(sdio->bio, sdio->cur_page,
740 sdio->cur_page_len, sdio->cur_page_offset);
741 if (ret == sdio->cur_page_len) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700742 /*
743 * Decrement count only, if we are done with this page
744 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700745 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
746 sdio->pages_in_io--;
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300747 get_page(sdio->cur_page);
Andi Kleeneb28be22011-08-01 21:38:03 -0700748 sdio->final_block_in_bio = sdio->cur_page_block +
749 (sdio->cur_page_len >> sdio->blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750 ret = 0;
751 } else {
752 ret = 1;
753 }
754 return ret;
755}
756
757/*
758 * Put cur_page under IO. The section of cur_page which is described by
759 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
760 * starts on-disk at cur_page_block.
761 *
762 * We take a ref against the page here (on behalf of its presence in the bio).
763 *
764 * The caller of this function is responsible for removing cur_page from the
765 * dio, and for dropping the refcount which came from that presence.
766 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700767static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
768 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769{
770 int ret = 0;
771
Andi Kleeneb28be22011-08-01 21:38:03 -0700772 if (sdio->bio) {
773 loff_t cur_offset = sdio->cur_page_fs_offset;
774 loff_t bio_next_offset = sdio->logical_offset_in_bio +
Kent Overstreet4f024f32013-10-11 15:44:27 -0700775 sdio->bio->bi_iter.bi_size;
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400776
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777 /*
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400778 * See whether this new request is contiguous with the old.
779 *
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900780 * Btrfs cannot handle having logically non-contiguous requests
781 * submitted. For example if you have
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400782 *
783 * Logical: [0-4095][HOLE][8192-12287]
Namhyung Kimf0940ce2011-01-11 21:15:03 +0900784 * Physical: [0-4095] [4096-8191]
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400785 *
786 * We cannot submit those pages together as one BIO. So if our
787 * current logical offset in the file does not equal what would
788 * be the next logical offset in the bio, submit the bio we
789 * have.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700791 if (sdio->final_block_in_bio != sdio->cur_page_block ||
Josef Bacikc2c6ca42010-05-23 11:00:55 -0400792 cur_offset != bio_next_offset)
Andi Kleeneb28be22011-08-01 21:38:03 -0700793 dio_bio_submit(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 }
795
Andi Kleeneb28be22011-08-01 21:38:03 -0700796 if (sdio->bio == NULL) {
Andi Kleen18772642011-08-01 21:38:07 -0700797 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798 if (ret)
799 goto out;
800 }
801
Andi Kleeneb28be22011-08-01 21:38:03 -0700802 if (dio_bio_add_page(sdio) != 0) {
803 dio_bio_submit(dio, sdio);
Andi Kleen18772642011-08-01 21:38:07 -0700804 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805 if (ret == 0) {
Andi Kleeneb28be22011-08-01 21:38:03 -0700806 ret = dio_bio_add_page(sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807 BUG_ON(ret != 0);
808 }
809 }
810out:
811 return ret;
812}
813
814/*
815 * An autonomous function to put a chunk of a page under deferred IO.
816 *
817 * The caller doesn't actually know (or care) whether this piece of page is in
818 * a BIO, or is under IO or whatever. We just take care of all possible
819 * situations here. The separation between the logic of do_direct_IO() and
820 * that of submit_page_section() is important for clarity. Please don't break.
821 *
822 * The chunk of page starts on-disk at blocknr.
823 *
824 * We perform deferred IO, by recording the last-submitted page inside our
825 * private part of the dio structure. If possible, we just expand the IO
826 * across that page here.
827 *
828 * If that doesn't work out then we put the old page into the bio and add this
829 * page to the dio instead.
830 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700831static inline int
Andi Kleeneb28be22011-08-01 21:38:03 -0700832submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
Andi Kleen18772642011-08-01 21:38:07 -0700833 unsigned offset, unsigned len, sector_t blocknr,
834 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835{
836 int ret = 0;
837
Mike Christie8a4c1e42016-06-05 14:31:50 -0500838 if (dio->op == REQ_OP_WRITE) {
Andrew Morton98c4d572006-12-10 02:19:47 -0800839 /*
840 * Read accounting is performed in submit_bio()
841 */
842 task_io_account_write(len);
843 }
844
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845 /*
846 * Can we just grow the current page's presence in the dio?
847 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700848 if (sdio->cur_page == page &&
849 sdio->cur_page_offset + sdio->cur_page_len == offset &&
850 sdio->cur_page_block +
851 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
852 sdio->cur_page_len += len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853 goto out;
854 }
855
856 /*
857 * If there's a deferred page already there then send it.
858 */
Andi Kleeneb28be22011-08-01 21:38:03 -0700859 if (sdio->cur_page) {
Andi Kleen18772642011-08-01 21:38:07 -0700860 ret = dio_send_cur_page(dio, sdio, map_bh);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300861 put_page(sdio->cur_page);
Andi Kleeneb28be22011-08-01 21:38:03 -0700862 sdio->cur_page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863 if (ret)
Jan Karab1058b92013-04-29 15:06:18 -0700864 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700865 }
866
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300867 get_page(page); /* It is in dio */
Andi Kleeneb28be22011-08-01 21:38:03 -0700868 sdio->cur_page = page;
869 sdio->cur_page_offset = offset;
870 sdio->cur_page_len = len;
871 sdio->cur_page_block = blocknr;
872 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700873out:
Jan Karab1058b92013-04-29 15:06:18 -0700874 /*
875 * If sdio->boundary then we want to schedule the IO now to
876 * avoid metadata seeks.
877 */
878 if (sdio->boundary) {
879 ret = dio_send_cur_page(dio, sdio, map_bh);
Andreas Gruenbacher899f0422017-10-09 11:13:18 +0200880 if (sdio->bio)
881 dio_bio_submit(dio, sdio);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300882 put_page(sdio->cur_page);
Jan Karab1058b92013-04-29 15:06:18 -0700883 sdio->cur_page = NULL;
884 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885 return ret;
886}
887
888/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889 * If we are not writing the entire block and get_block() allocated
890 * the block for us, we need to fill-in the unused portion of the
891 * block with zeros. This happens only if user-buffer, fileoffset or
892 * io length is not filesystem block-size multiple.
893 *
894 * `end' is zero if we're doing the start of the IO, 1 at the end of the
895 * IO.
896 */
Andi Kleenba253fb2011-08-01 21:38:08 -0700897static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
898 int end, struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899{
900 unsigned dio_blocks_per_fs_block;
901 unsigned this_chunk_blocks; /* In dio_blocks */
902 unsigned this_chunk_bytes;
903 struct page *page;
904
Andi Kleeneb28be22011-08-01 21:38:03 -0700905 sdio->start_zero_done = 1;
Andi Kleen18772642011-08-01 21:38:07 -0700906 if (!sdio->blkfactor || !buffer_new(map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700907 return;
908
Andi Kleeneb28be22011-08-01 21:38:03 -0700909 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
910 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911
912 if (!this_chunk_blocks)
913 return;
914
915 /*
916 * We need to zero out part of an fs block. It is either at the
917 * beginning or the end of the fs block.
918 */
919 if (end)
920 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
921
Andi Kleeneb28be22011-08-01 21:38:03 -0700922 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923
Nick Piggin557ed1f2007-10-16 01:24:40 -0700924 page = ZERO_PAGE(0);
Andi Kleeneb28be22011-08-01 21:38:03 -0700925 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -0700926 sdio->next_block_for_io, map_bh))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927 return;
928
Andi Kleeneb28be22011-08-01 21:38:03 -0700929 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930}
931
932/*
933 * Walk the user pages, and the file, mapping blocks to disk and generating
934 * a sequence of (page,offset,len,block) mappings. These mappings are injected
935 * into submit_page_section(), which takes care of the next stage of submission
936 *
937 * Direct IO against a blockdev is different from a file. Because we can
938 * happily perform page-sized but 512-byte aligned IOs. It is important that
939 * blockdev IO be able to have fine alignment and large sizes.
940 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800941 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds1da177e2005-04-16 15:20:36 -0700942 * with the size of IO which is permitted at this offset and this i_blkbits.
943 *
944 * For best results, the blockdev should be set up with 512-byte i_blkbits and
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800945 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946 * fine alignment but still allows this function to work in PAGE_SIZE units.
947 */
Andi Kleen18772642011-08-01 21:38:07 -0700948static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
949 struct buffer_head *map_bh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950{
Andi Kleeneb28be22011-08-01 21:38:03 -0700951 const unsigned blkbits = sdio->blkbits;
Chandan Rajendradd545b52017-01-10 13:29:54 -0700952 const unsigned i_blkbits = blkbits + sdio->blkfactor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953 int ret = 0;
954
Andi Kleeneb28be22011-08-01 21:38:03 -0700955 while (sdio->block_in_file < sdio->final_block_in_request) {
Al Viro7b2c99d2014-03-15 04:05:57 -0400956 struct page *page;
957 size_t from, to;
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300958
959 page = dio_get_page(dio, sdio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960 if (IS_ERR(page)) {
961 ret = PTR_ERR(page);
962 goto out;
963 }
Boaz Harrosh6fcc5422014-07-20 12:09:04 +0300964 from = sdio->head ? 0 : sdio->from;
965 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
966 sdio->head++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967
Al Viro7b2c99d2014-03-15 04:05:57 -0400968 while (from < to) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969 unsigned this_chunk_bytes; /* # of bytes mapped */
970 unsigned this_chunk_blocks; /* # of blocks */
971 unsigned u;
972
Andi Kleeneb28be22011-08-01 21:38:03 -0700973 if (sdio->blocks_available == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974 /*
975 * Need to go and map some more disk
976 */
977 unsigned long blkmask;
978 unsigned long dio_remainder;
979
Andi Kleen18772642011-08-01 21:38:07 -0700980 ret = get_more_blocks(dio, sdio, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981 if (ret) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300982 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983 goto out;
984 }
985 if (!buffer_mapped(map_bh))
986 goto do_holes;
987
Andi Kleeneb28be22011-08-01 21:38:03 -0700988 sdio->blocks_available =
Jan Karaf734c892016-11-04 18:08:12 +0100989 map_bh->b_size >> blkbits;
Andi Kleeneb28be22011-08-01 21:38:03 -0700990 sdio->next_block_for_io =
991 map_bh->b_blocknr << sdio->blkfactor;
Jan Karaf734c892016-11-04 18:08:12 +0100992 if (buffer_new(map_bh)) {
993 clean_bdev_aliases(
994 map_bh->b_bdev,
995 map_bh->b_blocknr,
Chandan Rajendradd545b52017-01-10 13:29:54 -0700996 map_bh->b_size >> i_blkbits);
Jan Karaf734c892016-11-04 18:08:12 +0100997 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998
Andi Kleeneb28be22011-08-01 21:38:03 -0700999 if (!sdio->blkfactor)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000 goto do_holes;
1001
Andi Kleeneb28be22011-08-01 21:38:03 -07001002 blkmask = (1 << sdio->blkfactor) - 1;
1003 dio_remainder = (sdio->block_in_file & blkmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004
1005 /*
1006 * If we are at the start of IO and that IO
1007 * starts partway into a fs-block,
1008 * dio_remainder will be non-zero. If the IO
1009 * is a read then we can simply advance the IO
1010 * cursor to the first block which is to be
1011 * read. But if the IO is a write and the
1012 * block was newly allocated we cannot do that;
1013 * the start of the fs block must be zeroed out
1014 * on-disk
1015 */
1016 if (!buffer_new(map_bh))
Andi Kleeneb28be22011-08-01 21:38:03 -07001017 sdio->next_block_for_io += dio_remainder;
1018 sdio->blocks_available -= dio_remainder;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001019 }
1020do_holes:
1021 /* Handle holes */
1022 if (!buffer_mapped(map_bh)) {
Jeff Moyer35dc8162006-02-03 03:04:27 -08001023 loff_t i_size_aligned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001024
1025 /* AKPM: eargh, -ENOTBLK is a hack */
Mike Christie8a4c1e42016-06-05 14:31:50 -05001026 if (dio->op == REQ_OP_WRITE) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001027 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001028 return -ENOTBLK;
1029 }
1030
Jeff Moyer35dc8162006-02-03 03:04:27 -08001031 /*
1032 * Be sure to account for a partial block as the
1033 * last block in the file
1034 */
1035 i_size_aligned = ALIGN(i_size_read(dio->inode),
1036 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -07001037 if (sdio->block_in_file >=
Jeff Moyer35dc8162006-02-03 03:04:27 -08001038 i_size_aligned >> blkbits) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001039 /* We hit eof */
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001040 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001041 goto out;
1042 }
Al Viro7b2c99d2014-03-15 04:05:57 -04001043 zero_user(page, from, 1 << blkbits);
Andi Kleeneb28be22011-08-01 21:38:03 -07001044 sdio->block_in_file++;
Al Viro7b2c99d2014-03-15 04:05:57 -04001045 from += 1 << blkbits;
Al Viro3320c602014-03-10 02:30:55 -04001046 dio->result += 1 << blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001047 goto next_block;
1048 }
1049
1050 /*
1051 * If we're performing IO which has an alignment which
1052 * is finer than the underlying fs, go check to see if
1053 * we must zero out the start of this block.
1054 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001055 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
Andi Kleen18772642011-08-01 21:38:07 -07001056 dio_zero_block(dio, sdio, 0, map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001057
1058 /*
1059 * Work out, in this_chunk_blocks, how much disk we
1060 * can add to this page
1061 */
Andi Kleeneb28be22011-08-01 21:38:03 -07001062 this_chunk_blocks = sdio->blocks_available;
Al Viro7b2c99d2014-03-15 04:05:57 -04001063 u = (to - from) >> blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064 if (this_chunk_blocks > u)
1065 this_chunk_blocks = u;
Andi Kleeneb28be22011-08-01 21:38:03 -07001066 u = sdio->final_block_in_request - sdio->block_in_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001067 if (this_chunk_blocks > u)
1068 this_chunk_blocks = u;
1069 this_chunk_bytes = this_chunk_blocks << blkbits;
1070 BUG_ON(this_chunk_bytes == 0);
1071
Jan Kara092c8d42013-04-29 15:06:17 -07001072 if (this_chunk_blocks == sdio->blocks_available)
1073 sdio->boundary = buffer_boundary(map_bh);
Andi Kleeneb28be22011-08-01 21:38:03 -07001074 ret = submit_page_section(dio, sdio, page,
Al Viro7b2c99d2014-03-15 04:05:57 -04001075 from,
Andi Kleeneb28be22011-08-01 21:38:03 -07001076 this_chunk_bytes,
Andi Kleen18772642011-08-01 21:38:07 -07001077 sdio->next_block_for_io,
1078 map_bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001079 if (ret) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001080 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081 goto out;
1082 }
Andi Kleeneb28be22011-08-01 21:38:03 -07001083 sdio->next_block_for_io += this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084
Andi Kleeneb28be22011-08-01 21:38:03 -07001085 sdio->block_in_file += this_chunk_blocks;
Al Viro7b2c99d2014-03-15 04:05:57 -04001086 from += this_chunk_bytes;
1087 dio->result += this_chunk_bytes;
Andi Kleeneb28be22011-08-01 21:38:03 -07001088 sdio->blocks_available -= this_chunk_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089next_block:
Andi Kleeneb28be22011-08-01 21:38:03 -07001090 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1091 if (sdio->block_in_file == sdio->final_block_in_request)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092 break;
1093 }
1094
1095 /* Drop the ref which was taken in get_user_pages() */
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001096 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097 }
1098out:
1099 return ret;
1100}
1101
Andi Kleen847cc632011-08-01 21:38:09 -07001102static inline int drop_refcount(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001103{
Andi Kleen847cc632011-08-01 21:38:09 -07001104 int ret2;
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001105 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106
Zach Brown8459d862006-12-10 02:21:05 -08001107 /*
1108 * Sync will always be dropping the final ref and completing the
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001109 * operation. AIO can if it was a broken operation described above or
1110 * in fact if all the bios race to complete before we get here. In
1111 * that case dio_complete() translates the EIOCBQUEUED into the proper
Christoph Hellwig04b2fa92015-02-02 14:49:06 +01001112 * return code that the caller will hand to ->complete().
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001113 *
1114 * This is managed by the bio_lock instead of being an atomic_t so that
1115 * completion paths can drop their ref and use the remaining count to
1116 * decide to wake the submission path atomically.
Zach Brown8459d862006-12-10 02:21:05 -08001117 */
Zach Brown5eb6c7a2006-12-10 02:21:07 -08001118 spin_lock_irqsave(&dio->bio_lock, flags);
1119 ret2 = --dio->refcount;
1120 spin_unlock_irqrestore(&dio->bio_lock, flags);
Andi Kleen847cc632011-08-01 21:38:09 -07001121 return ret2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122}
1123
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001124/*
1125 * This is a library function for use by filesystem drivers.
1126 *
1127 * The locking rules are governed by the flags parameter:
1128 * - if the flags value contains DIO_LOCKING we use a fancy locking
1129 * scheme for dumb filesystems.
1130 * For writes this function is called under i_mutex and returns with
1131 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1132 * taken and dropped again before returning.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001133 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1134 * internal locking but rather rely on the filesystem to synchronize
1135 * direct I/O reads/writes versus each other and truncate.
Christoph Hellwigdf2d6f22011-06-24 14:29:46 -04001136 *
1137 * To help with locking against truncate we incremented the i_dio_count
1138 * counter before starting direct I/O, and decrement it once we are done.
1139 * Truncate can wait for it to reach zero to provide exclusion. It is
1140 * expected that filesystem provide exclusion between new direct I/O
1141 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1142 * but other filesystems need to take care of this on their own.
Andi Kleenba253fb2011-08-01 21:38:08 -07001143 *
1144 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1145 * is always inlined. Otherwise gcc is unable to split the structure into
1146 * individual fields and will generate much worse code. This is important
1147 * for the whole file.
Christoph Hellwigeafdc7d2010-06-04 11:29:53 +02001148 */
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001149static inline ssize_t
Omar Sandoval17f8c842015-03-16 04:33:50 -07001150do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1151 struct block_device *bdev, struct iov_iter *iter,
Christoph Hellwigc8b8e322016-04-07 08:51:58 -07001152 get_block_t get_block, dio_iodone_t end_io,
Omar Sandoval17f8c842015-03-16 04:33:50 -07001153 dio_submit_t submit_io, int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154{
Mark Rutland6aa7de02017-10-23 14:07:29 -07001155 unsigned i_blkbits = READ_ONCE(inode->i_blkbits);
Linus Torvaldsab738572012-11-29 12:27:00 -08001156 unsigned blkbits = i_blkbits;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157 unsigned blocksize_mask = (1 << blkbits) - 1;
1158 ssize_t retval = -EINVAL;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001159 size_t count = iov_iter_count(iter);
Christoph Hellwigc8b8e322016-04-07 08:51:58 -07001160 loff_t offset = iocb->ki_pos;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001161 loff_t end = offset + count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162 struct dio *dio;
Andi Kleeneb28be22011-08-01 21:38:03 -07001163 struct dio_submit sdio = { 0, };
Andi Kleen847cc632011-08-01 21:38:09 -07001164 struct buffer_head map_bh = { 0, };
Fengguang Wu647d1e42012-08-09 15:23:09 +02001165 struct blk_plug plug;
Al Viro886a3912014-03-05 13:50:45 -05001166 unsigned long align = offset | iov_iter_alignment(iter);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001168 /*
1169 * Avoid references to bdev if not absolutely needed to give
1170 * the early prefetch in the caller enough time.
1171 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172
Al Viro886a3912014-03-05 13:50:45 -05001173 if (align & blocksize_mask) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174 if (bdev)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001175 blkbits = blksize_bits(bdev_logical_block_size(bdev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001176 blocksize_mask = (1 << blkbits) - 1;
Al Viro886a3912014-03-05 13:50:45 -05001177 if (align & blocksize_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178 goto out;
1179 }
1180
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001181 /* watch out for a 0 len io from a tricksy fs */
Omar Sandoval17f8c842015-03-16 04:33:50 -07001182 if (iov_iter_rw(iter) == READ && !iov_iter_count(iter))
Christoph Hellwigf9b55702011-06-24 14:29:42 -04001183 return 0;
1184
Andi Kleen6e8267f2011-08-01 21:38:06 -07001185 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186 retval = -ENOMEM;
1187 if (!dio)
1188 goto out;
Jeff Moyer23aee092009-12-15 16:47:49 -08001189 /*
1190 * Believe it or not, zeroing out the page array caused a .5%
1191 * performance regression in a database benchmark. So, we take
1192 * care to only zero out what's needed.
1193 */
1194 memset(dio, 0, offsetof(struct dio, pages));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195
Christoph Hellwig5fe878ae2009-12-15 16:47:50 -08001196 dio->flags = flags;
1197 if (dio->flags & DIO_LOCKING) {
Omar Sandoval17f8c842015-03-16 04:33:50 -07001198 if (iov_iter_rw(iter) == READ) {
Christoph Hellwig5fe878ae2009-12-15 16:47:50 -08001199 struct address_space *mapping =
1200 iocb->ki_filp->f_mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001201
Christoph Hellwig5fe878ae2009-12-15 16:47:50 -08001202 /* will be released by direct_io_worker */
Al Viro59551022016-01-22 15:40:57 -05001203 inode_lock(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001204
1205 retval = filemap_write_and_wait_range(mapping, offset,
1206 end - 1);
1207 if (retval) {
Al Viro59551022016-01-22 15:40:57 -05001208 inode_unlock(inode);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001209 kmem_cache_free(dio_cache, dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001210 goto out;
1211 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001212 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213 }
1214
Jan Kara74cedf92015-11-30 10:15:42 -07001215 /* Once we sampled i_size check for reads beyond EOF */
1216 dio->i_size = i_size_read(inode);
1217 if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1218 if (dio->flags & DIO_LOCKING)
Al Viro59551022016-01-22 15:40:57 -05001219 inode_unlock(inode);
Jan Kara74cedf92015-11-30 10:15:42 -07001220 kmem_cache_free(dio_cache, dio);
Al Viro2d4594a2015-12-08 12:22:47 -05001221 retval = 0;
Jan Kara74cedf92015-11-30 10:15:42 -07001222 goto out;
1223 }
1224
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225 /*
Christoph Hellwig60392572014-02-10 10:27:11 +11001226 * For file extending writes updating i_size before data writeouts
1227 * complete can expose uninitialized blocks in dumb filesystems.
1228 * In that case we need to wait for I/O completion even if asked
1229 * for an asynchronous write.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230 */
Christoph Hellwig60392572014-02-10 10:27:11 +11001231 if (is_sync_kiocb(iocb))
1232 dio->is_async = false;
1233 else if (!(dio->flags & DIO_ASYNC_EXTEND) &&
Omar Sandoval17f8c842015-03-16 04:33:50 -07001234 iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
Christoph Hellwig60392572014-02-10 10:27:11 +11001235 dio->is_async = false;
1236 else
1237 dio->is_async = true;
1238
Andi Kleen847cc632011-08-01 21:38:09 -07001239 dio->inode = inode;
Mike Christie8a4c1e42016-06-05 14:31:50 -05001240 if (iov_iter_rw(iter) == WRITE) {
1241 dio->op = REQ_OP_WRITE;
Christoph Hellwig70fd7612016-11-01 07:40:10 -06001242 dio->op_flags = REQ_SYNC | REQ_IDLE;
Goldwyn Rodrigues03a07c92017-06-20 07:05:46 -05001243 if (iocb->ki_flags & IOCB_NOWAIT)
1244 dio->op_flags |= REQ_NOWAIT;
Mike Christie8a4c1e42016-06-05 14:31:50 -05001245 } else {
1246 dio->op = REQ_OP_READ;
1247 }
Christoph Hellwig02afc272013-09-04 15:04:40 +02001248
1249 /*
1250 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1251 * so that we can call ->fsync.
1252 */
Lukas Czerner332391a2017-09-21 08:16:29 -06001253 if (dio->is_async && iov_iter_rw(iter) == WRITE) {
1254 retval = 0;
1255 if ((iocb->ki_filp->f_flags & O_DSYNC) ||
1256 IS_SYNC(iocb->ki_filp->f_mapping->host))
1257 retval = dio_set_defer_completion(dio);
1258 else if (!dio->inode->i_sb->s_dio_done_wq) {
1259 /*
1260 * In case of AIO write racing with buffered read we
1261 * need to defer completion. We can't decide this now,
1262 * however the workqueue needs to be initialized here.
1263 */
1264 retval = sb_init_dio_done_wq(dio->inode->i_sb);
1265 }
Christoph Hellwig02afc272013-09-04 15:04:40 +02001266 if (retval) {
1267 /*
1268 * We grab i_mutex only for reads so we don't have
1269 * to release it here
1270 */
1271 kmem_cache_free(dio_cache, dio);
1272 goto out;
1273 }
1274 }
1275
1276 /*
1277 * Will be decremented at I/O completion time.
1278 */
Jens Axboefe0f07d2015-04-15 17:05:48 -06001279 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
1280 inode_dio_begin(inode);
Christoph Hellwig02afc272013-09-04 15:04:40 +02001281
1282 retval = 0;
Andi Kleen847cc632011-08-01 21:38:09 -07001283 sdio.blkbits = blkbits;
Linus Torvaldsab738572012-11-29 12:27:00 -08001284 sdio.blkfactor = i_blkbits - blkbits;
Andi Kleen847cc632011-08-01 21:38:09 -07001285 sdio.block_in_file = offset >> blkbits;
1286
1287 sdio.get_block = get_block;
1288 dio->end_io = end_io;
1289 sdio.submit_io = submit_io;
1290 sdio.final_block_in_bio = -1;
1291 sdio.next_block_for_io = -1;
1292
1293 dio->iocb = iocb;
Andi Kleen847cc632011-08-01 21:38:09 -07001294
1295 spin_lock_init(&dio->bio_lock);
1296 dio->refcount = 1;
1297
Ming Lei53cbf3b2015-08-17 10:31:46 +08001298 dio->should_dirty = (iter->type == ITER_IOVEC);
Al Viro7b2c99d2014-03-15 04:05:57 -04001299 sdio.iter = iter;
1300 sdio.final_block_in_request =
1301 (offset + iov_iter_count(iter)) >> blkbits;
1302
Andi Kleen847cc632011-08-01 21:38:09 -07001303 /*
1304 * In case of non-aligned buffers, we may need 2 more
1305 * pages since we need to zero out first and last block.
1306 */
1307 if (unlikely(sdio.blkfactor))
1308 sdio.pages_in_io = 2;
1309
Al Virof67da302014-03-19 01:16:16 -04001310 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
Andi Kleen847cc632011-08-01 21:38:09 -07001311
Fengguang Wu647d1e42012-08-09 15:23:09 +02001312 blk_start_plug(&plug);
1313
Al Viro7b2c99d2014-03-15 04:05:57 -04001314 retval = do_direct_IO(dio, &sdio, &map_bh);
1315 if (retval)
1316 dio_cleanup(dio, &sdio);
Andi Kleen847cc632011-08-01 21:38:09 -07001317
1318 if (retval == -ENOTBLK) {
1319 /*
1320 * The remaining part of the request will be
1321 * be handled by buffered I/O when we return
1322 */
1323 retval = 0;
1324 }
1325 /*
1326 * There may be some unwritten disk at the end of a part-written
1327 * fs-block-sized block. Go zero that now.
1328 */
1329 dio_zero_block(dio, &sdio, 1, &map_bh);
1330
1331 if (sdio.cur_page) {
1332 ssize_t ret2;
1333
1334 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1335 if (retval == 0)
1336 retval = ret2;
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001337 put_page(sdio.cur_page);
Andi Kleen847cc632011-08-01 21:38:09 -07001338 sdio.cur_page = NULL;
1339 }
1340 if (sdio.bio)
1341 dio_bio_submit(dio, &sdio);
1342
Fengguang Wu647d1e42012-08-09 15:23:09 +02001343 blk_finish_plug(&plug);
1344
Andi Kleen847cc632011-08-01 21:38:09 -07001345 /*
1346 * It is possible that, we return short IO due to end of file.
1347 * In that case, we need to release all the pages we got hold on.
1348 */
1349 dio_cleanup(dio, &sdio);
1350
1351 /*
1352 * All block lookups have been performed. For READ requests
1353 * we can let i_mutex go now that its achieved its purpose
1354 * of protecting us from looking up uninitialized blocks.
1355 */
Omar Sandoval17f8c842015-03-16 04:33:50 -07001356 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
Al Viro59551022016-01-22 15:40:57 -05001357 inode_unlock(dio->inode);
Andi Kleen847cc632011-08-01 21:38:09 -07001358
1359 /*
1360 * The only time we want to leave bios in flight is when a successful
1361 * partial aio read or full aio write have been setup. In that case
1362 * bio completion will call aio_complete. The only time it's safe to
1363 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1364 * This had *better* be the only place that raises -EIOCBQUEUED.
1365 */
1366 BUG_ON(retval == -EIOCBQUEUED);
1367 if (dio->is_async && retval == 0 && dio->result &&
Omar Sandoval17f8c842015-03-16 04:33:50 -07001368 (iov_iter_rw(iter) == READ || dio->result == count))
Andi Kleen847cc632011-08-01 21:38:09 -07001369 retval = -EIOCBQUEUED;
Christoph Hellwigaf436472014-07-30 07:18:48 -04001370 else
Andi Kleen847cc632011-08-01 21:38:09 -07001371 dio_await_completion(dio);
1372
1373 if (drop_refcount(dio) == 0) {
Lukas Czernerffe51f02017-10-17 08:43:09 -06001374 retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE);
Andi Kleen847cc632011-08-01 21:38:09 -07001375 } else
1376 BUG_ON(retval != -EIOCBQUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377
npiggin@suse.de7bb46a62010-05-27 01:05:33 +10001378out:
1379 return retval;
1380}
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001381
Omar Sandoval17f8c842015-03-16 04:33:50 -07001382ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1383 struct block_device *bdev, struct iov_iter *iter,
Christoph Hellwigc8b8e322016-04-07 08:51:58 -07001384 get_block_t get_block,
Omar Sandoval17f8c842015-03-16 04:33:50 -07001385 dio_iodone_t end_io, dio_submit_t submit_io,
1386 int flags)
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001387{
1388 /*
1389 * The block device state is needed in the end to finally
1390 * submit everything. Since it's likely to be cache cold
1391 * prefetch it here as first thing to hide some of the
1392 * latency.
1393 *
1394 * Attempt to prefetch the pieces we likely need later.
1395 */
1396 prefetch(&bdev->bd_disk->part_tbl);
1397 prefetch(bdev->bd_queue);
1398 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1399
Christoph Hellwigc8b8e322016-04-07 08:51:58 -07001400 return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block,
Omar Sandoval17f8c842015-03-16 04:33:50 -07001401 end_io, submit_io, flags);
Andi Kleen65dd2aa2012-01-12 17:20:35 -08001402}
1403
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404EXPORT_SYMBOL(__blockdev_direct_IO);
Andi Kleen6e8267f2011-08-01 21:38:06 -07001405
1406static __init int dio_init(void)
1407{
1408 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1409 return 0;
1410}
1411module_init(dio_init)