Branch data Line data Source code
1 : : // SPDX-License-Identifier: GPL-2.0-only
2 : : /*
3 : : * fs/direct-io.c
4 : : *
5 : : * Copyright (C) 2002, Linus Torvalds.
6 : : *
7 : : * O_DIRECT
8 : : *
9 : : * 04Jul2002 Andrew Morton
10 : : * Initial version
11 : : * 11Sep2002 janetinc@us.ibm.com
12 : : * added readv/writev support.
13 : : * 29Oct2002 Andrew Morton
14 : : * rewrote bio_add_page() support.
15 : : * 30Oct2002 pbadari@us.ibm.com
16 : : * added support for non-aligned IO.
17 : : * 06Nov2002 pbadari@us.ibm.com
18 : : * added asynchronous IO support.
19 : : * 21Jul2003 nathans@sgi.com
20 : : * added IO completion notifier.
21 : : */
22 : :
23 : : #include <linux/kernel.h>
24 : : #include <linux/module.h>
25 : : #include <linux/types.h>
26 : : #include <linux/fs.h>
27 : : #include <linux/mm.h>
28 : : #include <linux/slab.h>
29 : : #include <linux/highmem.h>
30 : : #include <linux/pagemap.h>
31 : : #include <linux/task_io_accounting_ops.h>
32 : : #include <linux/bio.h>
33 : : #include <linux/wait.h>
34 : : #include <linux/err.h>
35 : : #include <linux/blkdev.h>
36 : : #include <linux/buffer_head.h>
37 : : #include <linux/rwsem.h>
38 : : #include <linux/uio.h>
39 : : #include <linux/atomic.h>
40 : : #include <linux/prefetch.h>
41 : :
42 : : /*
43 : : * How many user pages to map in one call to get_user_pages(). This determines
44 : : * the size of a structure in the slab cache
45 : : */
46 : : #define DIO_PAGES 64
47 : :
48 : : /*
49 : : * Flags for dio_complete()
50 : : */
51 : : #define DIO_COMPLETE_ASYNC 0x01 /* This is async IO */
52 : : #define DIO_COMPLETE_INVALIDATE 0x02 /* Can invalidate pages */
53 : :
54 : : /*
55 : : * This code generally works in units of "dio_blocks". A dio_block is
56 : : * somewhere between the hard sector size and the filesystem block size. it
57 : : * is determined on a per-invocation basis. When talking to the filesystem
58 : : * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
59 : : * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
60 : : * to bio_block quantities by shifting left by blkfactor.
61 : : *
62 : : * If blkfactor is zero then the user's request was aligned to the filesystem's
63 : : * blocksize.
64 : : */
65 : :
66 : : /* dio_state only used in the submission path */
67 : :
68 : : struct dio_submit {
69 : : struct bio *bio; /* bio under assembly */
70 : : unsigned blkbits; /* doesn't change */
71 : : unsigned blkfactor; /* When we're using an alignment which
72 : : is finer than the filesystem's soft
73 : : blocksize, this specifies how much
74 : : finer. blkfactor=2 means 1/4-block
75 : : alignment. Does not change */
76 : : unsigned start_zero_done; /* flag: sub-blocksize zeroing has
77 : : been performed at the start of a
78 : : write */
79 : : int pages_in_io; /* approximate total IO pages */
80 : : sector_t block_in_file; /* Current offset into the underlying
81 : : file in dio_block units. */
82 : : unsigned blocks_available; /* At block_in_file. changes */
83 : : int reap_counter; /* rate limit reaping */
84 : : sector_t final_block_in_request;/* doesn't change */
85 : : int boundary; /* prev block is at a boundary */
86 : : get_block_t *get_block; /* block mapping function */
87 : : dio_submit_t *submit_io; /* IO submition function */
88 : :
89 : : loff_t logical_offset_in_bio; /* current first logical block in bio */
90 : : sector_t final_block_in_bio; /* current final block in bio + 1 */
91 : : sector_t next_block_for_io; /* next block to be put under IO,
92 : : in dio_blocks units */
93 : :
94 : : /*
95 : : * Deferred addition of a page to the dio. These variables are
96 : : * private to dio_send_cur_page(), submit_page_section() and
97 : : * dio_bio_add_page().
98 : : */
99 : : struct page *cur_page; /* The page */
100 : : unsigned cur_page_offset; /* Offset into it, in bytes */
101 : : unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
102 : : sector_t cur_page_block; /* Where it starts */
103 : : loff_t cur_page_fs_offset; /* Offset in file */
104 : :
105 : : struct iov_iter *iter;
106 : : /*
107 : : * Page queue. These variables belong to dio_refill_pages() and
108 : : * dio_get_page().
109 : : */
110 : : unsigned head; /* next page to process */
111 : : unsigned tail; /* last valid page + 1 */
112 : : size_t from, to;
113 : : };
114 : :
115 : : /* dio_state communicated between submission path and end_io */
116 : : struct dio {
117 : : int flags; /* doesn't change */
118 : : int op;
119 : : int op_flags;
120 : : blk_qc_t bio_cookie;
121 : : struct gendisk *bio_disk;
122 : : struct inode *inode;
123 : : loff_t i_size; /* i_size when submitted */
124 : : dio_iodone_t *end_io; /* IO completion function */
125 : :
126 : : void *private; /* copy from map_bh.b_private */
127 : :
128 : : /* BIO completion state */
129 : : spinlock_t bio_lock; /* protects BIO fields below */
130 : : int page_errors; /* errno from get_user_pages() */
131 : : int is_async; /* is IO async ? */
132 : : bool defer_completion; /* defer AIO completion to workqueue? */
133 : : bool should_dirty; /* if pages should be dirtied */
134 : : int io_error; /* IO error in completion path */
135 : : unsigned long refcount; /* direct_io_worker() and bios */
136 : : struct bio *bio_list; /* singly linked via bi_private */
137 : : struct task_struct *waiter; /* waiting task (NULL if none) */
138 : :
139 : : /* AIO related stuff */
140 : : struct kiocb *iocb; /* kiocb */
141 : : ssize_t result; /* IO result */
142 : :
143 : : /*
144 : : * pages[] (and any fields placed after it) are not zeroed out at
145 : : * allocation time. Don't add new fields after pages[] unless you
146 : : * wish that they not be zeroed.
147 : : */
148 : : union {
149 : : struct page *pages[DIO_PAGES]; /* page buffer */
150 : : struct work_struct complete_work;/* deferred AIO completion */
151 : : };
152 : : } ____cacheline_aligned_in_smp;
153 : :
154 : : static struct kmem_cache *dio_cache __read_mostly;
155 : :
156 : : /*
157 : : * How many pages are in the queue?
158 : : */
159 : : static inline unsigned dio_pages_present(struct dio_submit *sdio)
160 : : {
161 : 0 : return sdio->tail - sdio->head;
162 : : }
163 : :
164 : : /*
165 : : * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
166 : : */
167 : 0 : static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
168 : : {
169 : : ssize_t ret;
170 : :
171 : 0 : ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
172 : : &sdio->from);
173 : :
174 : 0 : if (ret < 0 && sdio->blocks_available && (dio->op == REQ_OP_WRITE)) {
175 : 0 : struct page *page = ZERO_PAGE(0);
176 : : /*
177 : : * A memory fault, but the filesystem has some outstanding
178 : : * mapped blocks. We need to use those blocks up to avoid
179 : : * leaking stale data in the file.
180 : : */
181 : 0 : if (dio->page_errors == 0)
182 : 0 : dio->page_errors = ret;
183 : 0 : get_page(page);
184 : 0 : dio->pages[0] = page;
185 : 0 : sdio->head = 0;
186 : 0 : sdio->tail = 1;
187 : 0 : sdio->from = 0;
188 : 0 : sdio->to = PAGE_SIZE;
189 : 0 : return 0;
190 : : }
191 : :
192 : 0 : if (ret >= 0) {
193 : 0 : iov_iter_advance(sdio->iter, ret);
194 : 0 : ret += sdio->from;
195 : 0 : sdio->head = 0;
196 : 0 : sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
197 : 0 : sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
198 : 0 : return 0;
199 : : }
200 : : return ret;
201 : : }
202 : :
203 : : /*
204 : : * Get another userspace page. Returns an ERR_PTR on error. Pages are
205 : : * buffered inside the dio so that we can call get_user_pages() against a
206 : : * decent number of pages, less frequently. To provide nicer use of the
207 : : * L1 cache.
208 : : */
209 : 0 : static inline struct page *dio_get_page(struct dio *dio,
210 : : struct dio_submit *sdio)
211 : : {
212 : 0 : if (dio_pages_present(sdio) == 0) {
213 : : int ret;
214 : :
215 : 0 : ret = dio_refill_pages(dio, sdio);
216 : 0 : if (ret)
217 : 0 : return ERR_PTR(ret);
218 : 0 : BUG_ON(dio_pages_present(sdio) == 0);
219 : : }
220 : 0 : return dio->pages[sdio->head];
221 : : }
222 : :
223 : : /*
224 : : * Warn about a page cache invalidation failure during a direct io write.
225 : : */
226 : 0 : void dio_warn_stale_pagecache(struct file *filp)
227 : : {
228 : : static DEFINE_RATELIMIT_STATE(_rs, 86400 * HZ, DEFAULT_RATELIMIT_BURST);
229 : : char pathname[128];
230 : : struct inode *inode = file_inode(filp);
231 : : char *path;
232 : :
233 : 0 : errseq_set(&inode->i_mapping->wb_err, -EIO);
234 : 0 : if (__ratelimit(&_rs)) {
235 : 0 : path = file_path(filp, pathname, sizeof(pathname));
236 : 0 : if (IS_ERR(path))
237 : : path = "(unknown)";
238 : 0 : pr_crit("Page cache invalidation failure on direct I/O. Possible data corruption due to collision with buffered I/O!\n");
239 : 0 : pr_crit("File: %s PID: %d Comm: %.20s\n", path, current->pid,
240 : : current->comm);
241 : : }
242 : 0 : }
243 : :
244 : : /*
245 : : * dio_complete() - called when all DIO BIO I/O has been completed
246 : : *
247 : : * This drops i_dio_count, lets interested parties know that a DIO operation
248 : : * has completed, and calculates the resulting return code for the operation.
249 : : *
250 : : * It lets the filesystem know if it registered an interest earlier via
251 : : * get_block. Pass the private field of the map buffer_head so that
252 : : * filesystems can use it to hold additional state between get_block calls and
253 : : * dio_complete.
254 : : */
255 : 0 : static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags)
256 : : {
257 : 0 : loff_t offset = dio->iocb->ki_pos;
258 : : ssize_t transferred = 0;
259 : : int err;
260 : :
261 : : /*
262 : : * AIO submission can race with bio completion to get here while
263 : : * expecting to have the last io completed by bio completion.
264 : : * In that case -EIOCBQUEUED is in fact not an error we want
265 : : * to preserve through this call.
266 : : */
267 : 0 : if (ret == -EIOCBQUEUED)
268 : : ret = 0;
269 : :
270 : 0 : if (dio->result) {
271 : : transferred = dio->result;
272 : :
273 : : /* Check for short read case */
274 : 0 : if ((dio->op == REQ_OP_READ) &&
275 : 0 : ((offset + transferred) > dio->i_size))
276 : 0 : transferred = dio->i_size - offset;
277 : : /* ignore EFAULT if some IO has been done */
278 : 0 : if (unlikely(ret == -EFAULT) && transferred)
279 : : ret = 0;
280 : : }
281 : :
282 : 0 : if (ret == 0)
283 : 0 : ret = dio->page_errors;
284 : 0 : if (ret == 0)
285 : 0 : ret = dio->io_error;
286 : 0 : if (ret == 0)
287 : : ret = transferred;
288 : :
289 : 0 : if (dio->end_io) {
290 : : // XXX: ki_pos??
291 : 0 : err = dio->end_io(dio->iocb, offset, ret, dio->private);
292 : 0 : if (err)
293 : : ret = err;
294 : : }
295 : :
296 : : /*
297 : : * Try again to invalidate clean pages which might have been cached by
298 : : * non-direct readahead, or faulted in by get_user_pages() if the source
299 : : * of the write was an mmap'ed region of the file we're writing. Either
300 : : * one is a pretty crazy thing to do, so we don't support it 100%. If
301 : : * this invalidation fails, tough, the write still worked...
302 : : *
303 : : * And this page cache invalidation has to be after dio->end_io(), as
304 : : * some filesystems convert unwritten extents to real allocations in
305 : : * end_io() when necessary, otherwise a racing buffer read would cache
306 : : * zeros from unwritten extents.
307 : : */
308 : 0 : if (flags & DIO_COMPLETE_INVALIDATE &&
309 : 0 : ret > 0 && dio->op == REQ_OP_WRITE &&
310 : 0 : dio->inode->i_mapping->nrpages) {
311 : 0 : err = invalidate_inode_pages2_range(dio->inode->i_mapping,
312 : 0 : offset >> PAGE_SHIFT,
313 : 0 : (offset + ret - 1) >> PAGE_SHIFT);
314 : 0 : if (err)
315 : 0 : dio_warn_stale_pagecache(dio->iocb->ki_filp);
316 : : }
317 : :
318 : 0 : inode_dio_end(dio->inode);
319 : :
320 : 0 : if (flags & DIO_COMPLETE_ASYNC) {
321 : : /*
322 : : * generic_write_sync expects ki_pos to have been updated
323 : : * already, but the submission path only does this for
324 : : * synchronous I/O.
325 : : */
326 : 0 : dio->iocb->ki_pos += transferred;
327 : :
328 : 0 : if (ret > 0 && dio->op == REQ_OP_WRITE)
329 : 0 : ret = generic_write_sync(dio->iocb, ret);
330 : 0 : dio->iocb->ki_complete(dio->iocb, ret, 0);
331 : : }
332 : :
333 : 0 : kmem_cache_free(dio_cache, dio);
334 : 0 : return ret;
335 : : }
336 : :
337 : 0 : static void dio_aio_complete_work(struct work_struct *work)
338 : : {
339 : 0 : struct dio *dio = container_of(work, struct dio, complete_work);
340 : :
341 : 0 : dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
342 : 0 : }
343 : :
344 : : static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
345 : :
346 : : /*
347 : : * Asynchronous IO callback.
348 : : */
349 : 0 : static void dio_bio_end_aio(struct bio *bio)
350 : : {
351 : 0 : struct dio *dio = bio->bi_private;
352 : : unsigned long remaining;
353 : : unsigned long flags;
354 : : bool defer_completion = false;
355 : :
356 : : /* cleanup the bio */
357 : 0 : dio_bio_complete(dio, bio);
358 : :
359 : 0 : spin_lock_irqsave(&dio->bio_lock, flags);
360 : 0 : remaining = --dio->refcount;
361 : 0 : if (remaining == 1 && dio->waiter)
362 : 0 : wake_up_process(dio->waiter);
363 : : spin_unlock_irqrestore(&dio->bio_lock, flags);
364 : :
365 : 0 : if (remaining == 0) {
366 : : /*
367 : : * Defer completion when defer_completion is set or
368 : : * when the inode has pages mapped and this is AIO write.
369 : : * We need to invalidate those pages because there is a
370 : : * chance they contain stale data in the case buffered IO
371 : : * went in between AIO submission and completion into the
372 : : * same region.
373 : : */
374 : 0 : if (dio->result)
375 : 0 : defer_completion = dio->defer_completion ||
376 : 0 : (dio->op == REQ_OP_WRITE &&
377 : 0 : dio->inode->i_mapping->nrpages);
378 : 0 : if (defer_completion) {
379 : 0 : INIT_WORK(&dio->complete_work, dio_aio_complete_work);
380 : 0 : queue_work(dio->inode->i_sb->s_dio_done_wq,
381 : : &dio->complete_work);
382 : : } else {
383 : 0 : dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
384 : : }
385 : : }
386 : 0 : }
387 : :
388 : : /*
389 : : * The BIO completion handler simply queues the BIO up for the process-context
390 : : * handler.
391 : : *
392 : : * During I/O bi_private points at the dio. After I/O, bi_private is used to
393 : : * implement a singly-linked list of completed BIOs, at dio->bio_list.
394 : : */
395 : 0 : static void dio_bio_end_io(struct bio *bio)
396 : : {
397 : 0 : struct dio *dio = bio->bi_private;
398 : : unsigned long flags;
399 : :
400 : 0 : spin_lock_irqsave(&dio->bio_lock, flags);
401 : 0 : bio->bi_private = dio->bio_list;
402 : 0 : dio->bio_list = bio;
403 : 0 : if (--dio->refcount == 1 && dio->waiter)
404 : 0 : wake_up_process(dio->waiter);
405 : : spin_unlock_irqrestore(&dio->bio_lock, flags);
406 : 0 : }
407 : :
408 : : /**
409 : : * dio_end_io - handle the end io action for the given bio
410 : : * @bio: The direct io bio thats being completed
411 : : *
412 : : * This is meant to be called by any filesystem that uses their own dio_submit_t
413 : : * so that the DIO specific endio actions are dealt with after the filesystem
414 : : * has done it's completion work.
415 : : */
416 : 0 : void dio_end_io(struct bio *bio)
417 : : {
418 : 0 : struct dio *dio = bio->bi_private;
419 : :
420 : 0 : if (dio->is_async)
421 : 0 : dio_bio_end_aio(bio);
422 : : else
423 : 0 : dio_bio_end_io(bio);
424 : 0 : }
425 : : EXPORT_SYMBOL_GPL(dio_end_io);
426 : :
427 : : static inline void
428 : 0 : dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
429 : : struct block_device *bdev,
430 : : sector_t first_sector, int nr_vecs)
431 : : {
432 : : struct bio *bio;
433 : :
434 : : /*
435 : : * bio_alloc() is guaranteed to return a bio when allowed to sleep and
436 : : * we request a valid number of vectors.
437 : : */
438 : 0 : bio = bio_alloc(GFP_KERNEL, nr_vecs);
439 : :
440 : 0 : bio_set_dev(bio, bdev);
441 : 0 : bio->bi_iter.bi_sector = first_sector;
442 : 0 : bio_set_op_attrs(bio, dio->op, dio->op_flags);
443 : 0 : if (dio->is_async)
444 : 0 : bio->bi_end_io = dio_bio_end_aio;
445 : : else
446 : 0 : bio->bi_end_io = dio_bio_end_io;
447 : :
448 : 0 : bio->bi_write_hint = dio->iocb->ki_hint;
449 : :
450 : 0 : sdio->bio = bio;
451 : 0 : sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
452 : 0 : }
453 : :
454 : : /*
455 : : * In the AIO read case we speculatively dirty the pages before starting IO.
456 : : * During IO completion, any of these pages which happen to have been written
457 : : * back will be redirtied by bio_check_pages_dirty().
458 : : *
459 : : * bios hold a dio reference between submit_bio and ->end_io.
460 : : */
461 : 0 : static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
462 : : {
463 : 0 : struct bio *bio = sdio->bio;
464 : : unsigned long flags;
465 : :
466 : 0 : bio->bi_private = dio;
467 : :
468 : 0 : spin_lock_irqsave(&dio->bio_lock, flags);
469 : 0 : dio->refcount++;
470 : : spin_unlock_irqrestore(&dio->bio_lock, flags);
471 : :
472 : 0 : if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty)
473 : 0 : bio_set_pages_dirty(bio);
474 : :
475 : 0 : dio->bio_disk = bio->bi_disk;
476 : :
477 : 0 : if (sdio->submit_io) {
478 : 0 : sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio);
479 : 0 : dio->bio_cookie = BLK_QC_T_NONE;
480 : : } else
481 : 0 : dio->bio_cookie = submit_bio(bio);
482 : :
483 : 0 : sdio->bio = NULL;
484 : 0 : sdio->boundary = 0;
485 : 0 : sdio->logical_offset_in_bio = 0;
486 : 0 : }
487 : :
488 : : /*
489 : : * Release any resources in case of a failure
490 : : */
491 : : static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
492 : : {
493 : 0 : while (sdio->head < sdio->tail)
494 : 0 : put_page(dio->pages[sdio->head++]);
495 : : }
496 : :
497 : : /*
498 : : * Wait for the next BIO to complete. Remove it and return it. NULL is
499 : : * returned once all BIOs have been completed. This must only be called once
500 : : * all bios have been issued so that dio->refcount can only decrease. This
501 : : * requires that that the caller hold a reference on the dio.
502 : : */
503 : 0 : static struct bio *dio_await_one(struct dio *dio)
504 : : {
505 : : unsigned long flags;
506 : : struct bio *bio = NULL;
507 : :
508 : 0 : spin_lock_irqsave(&dio->bio_lock, flags);
509 : :
510 : : /*
511 : : * Wait as long as the list is empty and there are bios in flight. bio
512 : : * completion drops the count, maybe adds to the list, and wakes while
513 : : * holding the bio_lock so we don't need set_current_state()'s barrier
514 : : * and can call it after testing our condition.
515 : : */
516 : 0 : while (dio->refcount > 1 && dio->bio_list == NULL) {
517 : 0 : __set_current_state(TASK_UNINTERRUPTIBLE);
518 : 0 : dio->waiter = current;
519 : : spin_unlock_irqrestore(&dio->bio_lock, flags);
520 : 0 : if (!(dio->iocb->ki_flags & IOCB_HIPRI) ||
521 : 0 : !blk_poll(dio->bio_disk->queue, dio->bio_cookie, true))
522 : 0 : io_schedule();
523 : : /* wake up sets us TASK_RUNNING */
524 : 0 : spin_lock_irqsave(&dio->bio_lock, flags);
525 : 0 : dio->waiter = NULL;
526 : : }
527 : 0 : if (dio->bio_list) {
528 : : bio = dio->bio_list;
529 : 0 : dio->bio_list = bio->bi_private;
530 : : }
531 : : spin_unlock_irqrestore(&dio->bio_lock, flags);
532 : 0 : return bio;
533 : : }
534 : :
535 : : /*
536 : : * Process one completed BIO. No locks are held.
537 : : */
538 : 0 : static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
539 : : {
540 : 0 : blk_status_t err = bio->bi_status;
541 : 0 : bool should_dirty = dio->op == REQ_OP_READ && dio->should_dirty;
542 : :
543 : 0 : if (err) {
544 : 0 : if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
545 : 0 : dio->io_error = -EAGAIN;
546 : : else
547 : 0 : dio->io_error = -EIO;
548 : : }
549 : :
550 : 0 : if (dio->is_async && should_dirty) {
551 : 0 : bio_check_pages_dirty(bio); /* transfers ownership */
552 : : } else {
553 : 0 : bio_release_pages(bio, should_dirty);
554 : 0 : bio_put(bio);
555 : : }
556 : 0 : return err;
557 : : }
558 : :
559 : : /*
560 : : * Wait on and process all in-flight BIOs. This must only be called once
561 : : * all bios have been issued so that the refcount can only decrease.
562 : : * This just waits for all bios to make it through dio_bio_complete. IO
563 : : * errors are propagated through dio->io_error and should be propagated via
564 : : * dio_complete().
565 : : */
566 : 0 : static void dio_await_completion(struct dio *dio)
567 : : {
568 : : struct bio *bio;
569 : : do {
570 : 0 : bio = dio_await_one(dio);
571 : 0 : if (bio)
572 : 0 : dio_bio_complete(dio, bio);
573 : 0 : } while (bio);
574 : 0 : }
575 : :
576 : : /*
577 : : * A really large O_DIRECT read or write can generate a lot of BIOs. So
578 : : * to keep the memory consumption sane we periodically reap any completed BIOs
579 : : * during the BIO generation phase.
580 : : *
581 : : * This also helps to limit the peak amount of pinned userspace memory.
582 : : */
583 : 0 : static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
584 : : {
585 : : int ret = 0;
586 : :
587 : 0 : if (sdio->reap_counter++ >= 64) {
588 : 0 : while (dio->bio_list) {
589 : : unsigned long flags;
590 : : struct bio *bio;
591 : : int ret2;
592 : :
593 : 0 : spin_lock_irqsave(&dio->bio_lock, flags);
594 : 0 : bio = dio->bio_list;
595 : 0 : dio->bio_list = bio->bi_private;
596 : : spin_unlock_irqrestore(&dio->bio_lock, flags);
597 : 0 : ret2 = blk_status_to_errno(dio_bio_complete(dio, bio));
598 : 0 : if (ret == 0)
599 : : ret = ret2;
600 : : }
601 : 0 : sdio->reap_counter = 0;
602 : : }
603 : 0 : return ret;
604 : : }
605 : :
606 : : /*
607 : : * Create workqueue for deferred direct IO completions. We allocate the
608 : : * workqueue when it's first needed. This avoids creating workqueue for
609 : : * filesystems that don't need it and also allows us to create the workqueue
610 : : * late enough so the we can include s_id in the name of the workqueue.
611 : : */
612 : 0 : int sb_init_dio_done_wq(struct super_block *sb)
613 : : {
614 : : struct workqueue_struct *old;
615 : 0 : struct workqueue_struct *wq = alloc_workqueue("dio/%s",
616 : : WQ_MEM_RECLAIM, 0,
617 : 0 : sb->s_id);
618 : 0 : if (!wq)
619 : : return -ENOMEM;
620 : : /*
621 : : * This has to be atomic as more DIOs can race to create the workqueue
622 : : */
623 : 0 : old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
624 : : /* Someone created workqueue before us? Free ours... */
625 : 0 : if (old)
626 : 0 : destroy_workqueue(wq);
627 : : return 0;
628 : : }
629 : :
630 : 0 : static int dio_set_defer_completion(struct dio *dio)
631 : : {
632 : 0 : struct super_block *sb = dio->inode->i_sb;
633 : :
634 : 0 : if (dio->defer_completion)
635 : : return 0;
636 : 0 : dio->defer_completion = true;
637 : 0 : if (!sb->s_dio_done_wq)
638 : 0 : return sb_init_dio_done_wq(sb);
639 : : return 0;
640 : : }
641 : :
642 : : /*
643 : : * Call into the fs to map some more disk blocks. We record the current number
644 : : * of available blocks at sdio->blocks_available. These are in units of the
645 : : * fs blocksize, i_blocksize(inode).
646 : : *
647 : : * The fs is allowed to map lots of blocks at once. If it wants to do that,
648 : : * it uses the passed inode-relative block number as the file offset, as usual.
649 : : *
650 : : * get_block() is passed the number of i_blkbits-sized blocks which direct_io
651 : : * has remaining to do. The fs should not map more than this number of blocks.
652 : : *
653 : : * If the fs has mapped a lot of blocks, it should populate bh->b_size to
654 : : * indicate how much contiguous disk space has been made available at
655 : : * bh->b_blocknr.
656 : : *
657 : : * If *any* of the mapped blocks are new, then the fs must set buffer_new().
658 : : * This isn't very efficient...
659 : : *
660 : : * In the case of filesystem holes: the fs may return an arbitrarily-large
661 : : * hole by returning an appropriate value in b_size and by clearing
662 : : * buffer_mapped(). However the direct-io code will only process holes one
663 : : * block at a time - it will repeatedly call get_block() as it walks the hole.
664 : : */
665 : 0 : static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
666 : : struct buffer_head *map_bh)
667 : : {
668 : : int ret;
669 : : sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
670 : : sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
671 : : unsigned long fs_count; /* Number of filesystem-sized blocks */
672 : : int create;
673 : 0 : unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
674 : : loff_t i_size;
675 : :
676 : : /*
677 : : * If there was a memory error and we've overwritten all the
678 : : * mapped blocks then we can now return that memory error
679 : : */
680 : 0 : ret = dio->page_errors;
681 : 0 : if (ret == 0) {
682 : 0 : BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
683 : 0 : fs_startblk = sdio->block_in_file >> sdio->blkfactor;
684 : 0 : fs_endblk = (sdio->final_block_in_request - 1) >>
685 : : sdio->blkfactor;
686 : 0 : fs_count = fs_endblk - fs_startblk + 1;
687 : :
688 : 0 : map_bh->b_state = 0;
689 : 0 : map_bh->b_size = fs_count << i_blkbits;
690 : :
691 : : /*
692 : : * For writes that could fill holes inside i_size on a
693 : : * DIO_SKIP_HOLES filesystem we forbid block creations: only
694 : : * overwrites are permitted. We will return early to the caller
695 : : * once we see an unmapped buffer head returned, and the caller
696 : : * will fall back to buffered I/O.
697 : : *
698 : : * Otherwise the decision is left to the get_blocks method,
699 : : * which may decide to handle it or also return an unmapped
700 : : * buffer head.
701 : : */
702 : 0 : create = dio->op == REQ_OP_WRITE;
703 : 0 : if (dio->flags & DIO_SKIP_HOLES) {
704 : 0 : i_size = i_size_read(dio->inode);
705 : 0 : if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits)
706 : : create = 0;
707 : : }
708 : :
709 : 0 : ret = (*sdio->get_block)(dio->inode, fs_startblk,
710 : : map_bh, create);
711 : :
712 : : /* Store for completion */
713 : 0 : dio->private = map_bh->b_private;
714 : :
715 : 0 : if (ret == 0 && buffer_defer_completion(map_bh))
716 : 0 : ret = dio_set_defer_completion(dio);
717 : : }
718 : 0 : return ret;
719 : : }
720 : :
721 : : /*
722 : : * There is no bio. Make one now.
723 : : */
724 : 0 : static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
725 : : sector_t start_sector, struct buffer_head *map_bh)
726 : : {
727 : : sector_t sector;
728 : : int ret, nr_pages;
729 : :
730 : 0 : ret = dio_bio_reap(dio, sdio);
731 : 0 : if (ret)
732 : : goto out;
733 : 0 : sector = start_sector << (sdio->blkbits - 9);
734 : 0 : nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES);
735 : 0 : BUG_ON(nr_pages <= 0);
736 : 0 : dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
737 : 0 : sdio->boundary = 0;
738 : : out:
739 : 0 : return ret;
740 : : }
741 : :
742 : : /*
743 : : * Attempt to put the current chunk of 'cur_page' into the current BIO. If
744 : : * that was successful then update final_block_in_bio and take a ref against
745 : : * the just-added page.
746 : : *
747 : : * Return zero on success. Non-zero means the caller needs to start a new BIO.
748 : : */
749 : 0 : static inline int dio_bio_add_page(struct dio_submit *sdio)
750 : : {
751 : : int ret;
752 : :
753 : 0 : ret = bio_add_page(sdio->bio, sdio->cur_page,
754 : : sdio->cur_page_len, sdio->cur_page_offset);
755 : 0 : if (ret == sdio->cur_page_len) {
756 : : /*
757 : : * Decrement count only, if we are done with this page
758 : : */
759 : 0 : if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
760 : 0 : sdio->pages_in_io--;
761 : 0 : get_page(sdio->cur_page);
762 : 0 : sdio->final_block_in_bio = sdio->cur_page_block +
763 : 0 : (sdio->cur_page_len >> sdio->blkbits);
764 : : ret = 0;
765 : : } else {
766 : : ret = 1;
767 : : }
768 : 0 : return ret;
769 : : }
770 : :
771 : : /*
772 : : * Put cur_page under IO. The section of cur_page which is described by
773 : : * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
774 : : * starts on-disk at cur_page_block.
775 : : *
776 : : * We take a ref against the page here (on behalf of its presence in the bio).
777 : : *
778 : : * The caller of this function is responsible for removing cur_page from the
779 : : * dio, and for dropping the refcount which came from that presence.
780 : : */
781 : 0 : static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
782 : : struct buffer_head *map_bh)
783 : : {
784 : : int ret = 0;
785 : :
786 : 0 : if (sdio->bio) {
787 : 0 : loff_t cur_offset = sdio->cur_page_fs_offset;
788 : 0 : loff_t bio_next_offset = sdio->logical_offset_in_bio +
789 : 0 : sdio->bio->bi_iter.bi_size;
790 : :
791 : : /*
792 : : * See whether this new request is contiguous with the old.
793 : : *
794 : : * Btrfs cannot handle having logically non-contiguous requests
795 : : * submitted. For example if you have
796 : : *
797 : : * Logical: [0-4095][HOLE][8192-12287]
798 : : * Physical: [0-4095] [4096-8191]
799 : : *
800 : : * We cannot submit those pages together as one BIO. So if our
801 : : * current logical offset in the file does not equal what would
802 : : * be the next logical offset in the bio, submit the bio we
803 : : * have.
804 : : */
805 : 0 : if (sdio->final_block_in_bio != sdio->cur_page_block ||
806 : : cur_offset != bio_next_offset)
807 : 0 : dio_bio_submit(dio, sdio);
808 : : }
809 : :
810 : 0 : if (sdio->bio == NULL) {
811 : 0 : ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
812 : 0 : if (ret)
813 : : goto out;
814 : : }
815 : :
816 : 0 : if (dio_bio_add_page(sdio) != 0) {
817 : 0 : dio_bio_submit(dio, sdio);
818 : 0 : ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
819 : 0 : if (ret == 0) {
820 : 0 : ret = dio_bio_add_page(sdio);
821 : 0 : BUG_ON(ret != 0);
822 : : }
823 : : }
824 : : out:
825 : 0 : return ret;
826 : : }
827 : :
828 : : /*
829 : : * An autonomous function to put a chunk of a page under deferred IO.
830 : : *
831 : : * The caller doesn't actually know (or care) whether this piece of page is in
832 : : * a BIO, or is under IO or whatever. We just take care of all possible
833 : : * situations here. The separation between the logic of do_direct_IO() and
834 : : * that of submit_page_section() is important for clarity. Please don't break.
835 : : *
836 : : * The chunk of page starts on-disk at blocknr.
837 : : *
838 : : * We perform deferred IO, by recording the last-submitted page inside our
839 : : * private part of the dio structure. If possible, we just expand the IO
840 : : * across that page here.
841 : : *
842 : : * If that doesn't work out then we put the old page into the bio and add this
843 : : * page to the dio instead.
844 : : */
845 : : static inline int
846 : 0 : submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
847 : : unsigned offset, unsigned len, sector_t blocknr,
848 : : struct buffer_head *map_bh)
849 : : {
850 : : int ret = 0;
851 : :
852 : 0 : if (dio->op == REQ_OP_WRITE) {
853 : : /*
854 : : * Read accounting is performed in submit_bio()
855 : : */
856 : : task_io_account_write(len);
857 : : }
858 : :
859 : : /*
860 : : * Can we just grow the current page's presence in the dio?
861 : : */
862 : 0 : if (sdio->cur_page == page &&
863 : 0 : sdio->cur_page_offset + sdio->cur_page_len == offset &&
864 : 0 : sdio->cur_page_block +
865 : 0 : (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
866 : 0 : sdio->cur_page_len += len;
867 : 0 : goto out;
868 : : }
869 : :
870 : : /*
871 : : * If there's a deferred page already there then send it.
872 : : */
873 : 0 : if (sdio->cur_page) {
874 : 0 : ret = dio_send_cur_page(dio, sdio, map_bh);
875 : 0 : put_page(sdio->cur_page);
876 : 0 : sdio->cur_page = NULL;
877 : 0 : if (ret)
878 : : return ret;
879 : : }
880 : :
881 : 0 : get_page(page); /* It is in dio */
882 : 0 : sdio->cur_page = page;
883 : 0 : sdio->cur_page_offset = offset;
884 : 0 : sdio->cur_page_len = len;
885 : 0 : sdio->cur_page_block = blocknr;
886 : 0 : sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
887 : : out:
888 : : /*
889 : : * If sdio->boundary then we want to schedule the IO now to
890 : : * avoid metadata seeks.
891 : : */
892 : 0 : if (sdio->boundary) {
893 : 0 : ret = dio_send_cur_page(dio, sdio, map_bh);
894 : 0 : if (sdio->bio)
895 : 0 : dio_bio_submit(dio, sdio);
896 : 0 : put_page(sdio->cur_page);
897 : 0 : sdio->cur_page = NULL;
898 : : }
899 : 0 : return ret;
900 : : }
901 : :
902 : : /*
903 : : * If we are not writing the entire block and get_block() allocated
904 : : * the block for us, we need to fill-in the unused portion of the
905 : : * block with zeros. This happens only if user-buffer, fileoffset or
906 : : * io length is not filesystem block-size multiple.
907 : : *
908 : : * `end' is zero if we're doing the start of the IO, 1 at the end of the
909 : : * IO.
910 : : */
911 : 0 : static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
912 : : int end, struct buffer_head *map_bh)
913 : : {
914 : : unsigned dio_blocks_per_fs_block;
915 : : unsigned this_chunk_blocks; /* In dio_blocks */
916 : : unsigned this_chunk_bytes;
917 : : struct page *page;
918 : :
919 : 0 : sdio->start_zero_done = 1;
920 : 0 : if (!sdio->blkfactor || !buffer_new(map_bh))
921 : : return;
922 : :
923 : 0 : dio_blocks_per_fs_block = 1 << sdio->blkfactor;
924 : 0 : this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
925 : :
926 : 0 : if (!this_chunk_blocks)
927 : : return;
928 : :
929 : : /*
930 : : * We need to zero out part of an fs block. It is either at the
931 : : * beginning or the end of the fs block.
932 : : */
933 : 0 : if (end)
934 : 0 : this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
935 : :
936 : 0 : this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
937 : :
938 : 0 : page = ZERO_PAGE(0);
939 : 0 : if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
940 : : sdio->next_block_for_io, map_bh))
941 : : return;
942 : :
943 : 0 : sdio->next_block_for_io += this_chunk_blocks;
944 : : }
945 : :
946 : : /*
947 : : * Walk the user pages, and the file, mapping blocks to disk and generating
948 : : * a sequence of (page,offset,len,block) mappings. These mappings are injected
949 : : * into submit_page_section(), which takes care of the next stage of submission
950 : : *
951 : : * Direct IO against a blockdev is different from a file. Because we can
952 : : * happily perform page-sized but 512-byte aligned IOs. It is important that
953 : : * blockdev IO be able to have fine alignment and large sizes.
954 : : *
955 : : * So what we do is to permit the ->get_block function to populate bh.b_size
956 : : * with the size of IO which is permitted at this offset and this i_blkbits.
957 : : *
958 : : * For best results, the blockdev should be set up with 512-byte i_blkbits and
959 : : * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
960 : : * fine alignment but still allows this function to work in PAGE_SIZE units.
961 : : */
962 : 0 : static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
963 : : struct buffer_head *map_bh)
964 : : {
965 : 0 : const unsigned blkbits = sdio->blkbits;
966 : 0 : const unsigned i_blkbits = blkbits + sdio->blkfactor;
967 : : int ret = 0;
968 : :
969 : 0 : while (sdio->block_in_file < sdio->final_block_in_request) {
970 : : struct page *page;
971 : : size_t from, to;
972 : :
973 : 0 : page = dio_get_page(dio, sdio);
974 : 0 : if (IS_ERR(page)) {
975 : : ret = PTR_ERR(page);
976 : 0 : goto out;
977 : : }
978 : 0 : from = sdio->head ? 0 : sdio->from;
979 : 0 : to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
980 : 0 : sdio->head++;
981 : :
982 : 0 : while (from < to) {
983 : : unsigned this_chunk_bytes; /* # of bytes mapped */
984 : : unsigned this_chunk_blocks; /* # of blocks */
985 : : unsigned u;
986 : :
987 : 0 : if (sdio->blocks_available == 0) {
988 : : /*
989 : : * Need to go and map some more disk
990 : : */
991 : : unsigned long blkmask;
992 : : unsigned long dio_remainder;
993 : :
994 : 0 : ret = get_more_blocks(dio, sdio, map_bh);
995 : 0 : if (ret) {
996 : 0 : put_page(page);
997 : 0 : goto out;
998 : : }
999 : 0 : if (!buffer_mapped(map_bh))
1000 : : goto do_holes;
1001 : :
1002 : 0 : sdio->blocks_available =
1003 : 0 : map_bh->b_size >> blkbits;
1004 : 0 : sdio->next_block_for_io =
1005 : 0 : map_bh->b_blocknr << sdio->blkfactor;
1006 : 0 : if (buffer_new(map_bh)) {
1007 : 0 : clean_bdev_aliases(
1008 : : map_bh->b_bdev,
1009 : : map_bh->b_blocknr,
1010 : 0 : map_bh->b_size >> i_blkbits);
1011 : : }
1012 : :
1013 : 0 : if (!sdio->blkfactor)
1014 : : goto do_holes;
1015 : :
1016 : 0 : blkmask = (1 << sdio->blkfactor) - 1;
1017 : 0 : dio_remainder = (sdio->block_in_file & blkmask);
1018 : :
1019 : : /*
1020 : : * If we are at the start of IO and that IO
1021 : : * starts partway into a fs-block,
1022 : : * dio_remainder will be non-zero. If the IO
1023 : : * is a read then we can simply advance the IO
1024 : : * cursor to the first block which is to be
1025 : : * read. But if the IO is a write and the
1026 : : * block was newly allocated we cannot do that;
1027 : : * the start of the fs block must be zeroed out
1028 : : * on-disk
1029 : : */
1030 : 0 : if (!buffer_new(map_bh))
1031 : 0 : sdio->next_block_for_io += dio_remainder;
1032 : 0 : sdio->blocks_available -= dio_remainder;
1033 : : }
1034 : : do_holes:
1035 : : /* Handle holes */
1036 : 0 : if (!buffer_mapped(map_bh)) {
1037 : : loff_t i_size_aligned;
1038 : :
1039 : : /* AKPM: eargh, -ENOTBLK is a hack */
1040 : 0 : if (dio->op == REQ_OP_WRITE) {
1041 : 0 : put_page(page);
1042 : 0 : return -ENOTBLK;
1043 : : }
1044 : :
1045 : : /*
1046 : : * Be sure to account for a partial block as the
1047 : : * last block in the file
1048 : : */
1049 : 0 : i_size_aligned = ALIGN(i_size_read(dio->inode),
1050 : : 1 << blkbits);
1051 : 0 : if (sdio->block_in_file >=
1052 : 0 : i_size_aligned >> blkbits) {
1053 : : /* We hit eof */
1054 : 0 : put_page(page);
1055 : 0 : goto out;
1056 : : }
1057 : 0 : zero_user(page, from, 1 << blkbits);
1058 : 0 : sdio->block_in_file++;
1059 : : from += 1 << blkbits;
1060 : 0 : dio->result += 1 << blkbits;
1061 : 0 : goto next_block;
1062 : : }
1063 : :
1064 : : /*
1065 : : * If we're performing IO which has an alignment which
1066 : : * is finer than the underlying fs, go check to see if
1067 : : * we must zero out the start of this block.
1068 : : */
1069 : 0 : if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1070 : 0 : dio_zero_block(dio, sdio, 0, map_bh);
1071 : :
1072 : : /*
1073 : : * Work out, in this_chunk_blocks, how much disk we
1074 : : * can add to this page
1075 : : */
1076 : 0 : this_chunk_blocks = sdio->blocks_available;
1077 : 0 : u = (to - from) >> blkbits;
1078 : 0 : if (this_chunk_blocks > u)
1079 : : this_chunk_blocks = u;
1080 : 0 : u = sdio->final_block_in_request - sdio->block_in_file;
1081 : 0 : if (this_chunk_blocks > u)
1082 : : this_chunk_blocks = u;
1083 : 0 : this_chunk_bytes = this_chunk_blocks << blkbits;
1084 : 0 : BUG_ON(this_chunk_bytes == 0);
1085 : :
1086 : 0 : if (this_chunk_blocks == sdio->blocks_available)
1087 : 0 : sdio->boundary = buffer_boundary(map_bh);
1088 : 0 : ret = submit_page_section(dio, sdio, page,
1089 : : from,
1090 : : this_chunk_bytes,
1091 : : sdio->next_block_for_io,
1092 : : map_bh);
1093 : 0 : if (ret) {
1094 : 0 : put_page(page);
1095 : 0 : goto out;
1096 : : }
1097 : 0 : sdio->next_block_for_io += this_chunk_blocks;
1098 : :
1099 : 0 : sdio->block_in_file += this_chunk_blocks;
1100 : 0 : from += this_chunk_bytes;
1101 : 0 : dio->result += this_chunk_bytes;
1102 : 0 : sdio->blocks_available -= this_chunk_blocks;
1103 : : next_block:
1104 : 0 : BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1105 : 0 : if (sdio->block_in_file == sdio->final_block_in_request)
1106 : : break;
1107 : : }
1108 : :
1109 : : /* Drop the ref which was taken in get_user_pages() */
1110 : 0 : put_page(page);
1111 : : }
1112 : : out:
1113 : 0 : return ret;
1114 : : }
1115 : :
1116 : 0 : static inline int drop_refcount(struct dio *dio)
1117 : : {
1118 : : int ret2;
1119 : : unsigned long flags;
1120 : :
1121 : : /*
1122 : : * Sync will always be dropping the final ref and completing the
1123 : : * operation. AIO can if it was a broken operation described above or
1124 : : * in fact if all the bios race to complete before we get here. In
1125 : : * that case dio_complete() translates the EIOCBQUEUED into the proper
1126 : : * return code that the caller will hand to ->complete().
1127 : : *
1128 : : * This is managed by the bio_lock instead of being an atomic_t so that
1129 : : * completion paths can drop their ref and use the remaining count to
1130 : : * decide to wake the submission path atomically.
1131 : : */
1132 : 0 : spin_lock_irqsave(&dio->bio_lock, flags);
1133 : 0 : ret2 = --dio->refcount;
1134 : : spin_unlock_irqrestore(&dio->bio_lock, flags);
1135 : 0 : return ret2;
1136 : : }
1137 : :
1138 : : /*
1139 : : * This is a library function for use by filesystem drivers.
1140 : : *
1141 : : * The locking rules are governed by the flags parameter:
1142 : : * - if the flags value contains DIO_LOCKING we use a fancy locking
1143 : : * scheme for dumb filesystems.
1144 : : * For writes this function is called under i_mutex and returns with
1145 : : * i_mutex held, for reads, i_mutex is not held on entry, but it is
1146 : : * taken and dropped again before returning.
1147 : : * - if the flags value does NOT contain DIO_LOCKING we don't use any
1148 : : * internal locking but rather rely on the filesystem to synchronize
1149 : : * direct I/O reads/writes versus each other and truncate.
1150 : : *
1151 : : * To help with locking against truncate we incremented the i_dio_count
1152 : : * counter before starting direct I/O, and decrement it once we are done.
1153 : : * Truncate can wait for it to reach zero to provide exclusion. It is
1154 : : * expected that filesystem provide exclusion between new direct I/O
1155 : : * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1156 : : * but other filesystems need to take care of this on their own.
1157 : : *
1158 : : * NOTE: if you pass "sdio" to anything by pointer make sure that function
1159 : : * is always inlined. Otherwise gcc is unable to split the structure into
1160 : : * individual fields and will generate much worse code. This is important
1161 : : * for the whole file.
1162 : : */
1163 : : static inline ssize_t
1164 : 0 : do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1165 : : struct block_device *bdev, struct iov_iter *iter,
1166 : : get_block_t get_block, dio_iodone_t end_io,
1167 : : dio_submit_t submit_io, int flags)
1168 : : {
1169 : 0 : unsigned i_blkbits = READ_ONCE(inode->i_blkbits);
1170 : : unsigned blkbits = i_blkbits;
1171 : 0 : unsigned blocksize_mask = (1 << blkbits) - 1;
1172 : : ssize_t retval = -EINVAL;
1173 : : const size_t count = iov_iter_count(iter);
1174 : 0 : loff_t offset = iocb->ki_pos;
1175 : 0 : const loff_t end = offset + count;
1176 : : struct dio *dio;
1177 : 0 : struct dio_submit sdio = { 0, };
1178 : 0 : struct buffer_head map_bh = { 0, };
1179 : : struct blk_plug plug;
1180 : 0 : unsigned long align = offset | iov_iter_alignment(iter);
1181 : :
1182 : : /*
1183 : : * Avoid references to bdev if not absolutely needed to give
1184 : : * the early prefetch in the caller enough time.
1185 : : */
1186 : :
1187 : 0 : if (align & blocksize_mask) {
1188 : 0 : if (bdev)
1189 : : blkbits = blksize_bits(bdev_logical_block_size(bdev));
1190 : 0 : blocksize_mask = (1 << blkbits) - 1;
1191 : 0 : if (align & blocksize_mask)
1192 : : goto out;
1193 : : }
1194 : :
1195 : : /* watch out for a 0 len io from a tricksy fs */
1196 : 0 : if (iov_iter_rw(iter) == READ && !count)
1197 : : return 0;
1198 : :
1199 : 0 : dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1200 : : retval = -ENOMEM;
1201 : 0 : if (!dio)
1202 : : goto out;
1203 : : /*
1204 : : * Believe it or not, zeroing out the page array caused a .5%
1205 : : * performance regression in a database benchmark. So, we take
1206 : : * care to only zero out what's needed.
1207 : : */
1208 : 0 : memset(dio, 0, offsetof(struct dio, pages));
1209 : :
1210 : 0 : dio->flags = flags;
1211 : 0 : if (dio->flags & DIO_LOCKING) {
1212 : 0 : if (iov_iter_rw(iter) == READ) {
1213 : 0 : struct address_space *mapping =
1214 : 0 : iocb->ki_filp->f_mapping;
1215 : :
1216 : : /* will be released by direct_io_worker */
1217 : : inode_lock(inode);
1218 : :
1219 : 0 : retval = filemap_write_and_wait_range(mapping, offset,
1220 : : end - 1);
1221 : 0 : if (retval) {
1222 : : inode_unlock(inode);
1223 : 0 : kmem_cache_free(dio_cache, dio);
1224 : 0 : goto out;
1225 : : }
1226 : : }
1227 : : }
1228 : :
1229 : : /* Once we sampled i_size check for reads beyond EOF */
1230 : 0 : dio->i_size = i_size_read(inode);
1231 : 0 : if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1232 : 0 : if (dio->flags & DIO_LOCKING)
1233 : : inode_unlock(inode);
1234 : 0 : kmem_cache_free(dio_cache, dio);
1235 : : retval = 0;
1236 : 0 : goto out;
1237 : : }
1238 : :
1239 : : /*
1240 : : * For file extending writes updating i_size before data writeouts
1241 : : * complete can expose uninitialized blocks in dumb filesystems.
1242 : : * In that case we need to wait for I/O completion even if asked
1243 : : * for an asynchronous write.
1244 : : */
1245 : 0 : if (is_sync_kiocb(iocb))
1246 : 0 : dio->is_async = false;
1247 : 0 : else if (iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
1248 : 0 : dio->is_async = false;
1249 : : else
1250 : 0 : dio->is_async = true;
1251 : :
1252 : 0 : dio->inode = inode;
1253 : 0 : if (iov_iter_rw(iter) == WRITE) {
1254 : 0 : dio->op = REQ_OP_WRITE;
1255 : 0 : dio->op_flags = REQ_SYNC | REQ_IDLE;
1256 : 0 : if (iocb->ki_flags & IOCB_NOWAIT)
1257 : 0 : dio->op_flags |= REQ_NOWAIT;
1258 : : } else {
1259 : 0 : dio->op = REQ_OP_READ;
1260 : : }
1261 : 0 : if (iocb->ki_flags & IOCB_HIPRI)
1262 : 0 : dio->op_flags |= REQ_HIPRI;
1263 : :
1264 : : /*
1265 : : * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1266 : : * so that we can call ->fsync.
1267 : : */
1268 : 0 : if (dio->is_async && iov_iter_rw(iter) == WRITE) {
1269 : : retval = 0;
1270 : 0 : if (iocb->ki_flags & IOCB_DSYNC)
1271 : 0 : retval = dio_set_defer_completion(dio);
1272 : 0 : else if (!dio->inode->i_sb->s_dio_done_wq) {
1273 : : /*
1274 : : * In case of AIO write racing with buffered read we
1275 : : * need to defer completion. We can't decide this now,
1276 : : * however the workqueue needs to be initialized here.
1277 : : */
1278 : 0 : retval = sb_init_dio_done_wq(dio->inode->i_sb);
1279 : : }
1280 : 0 : if (retval) {
1281 : : /*
1282 : : * We grab i_mutex only for reads so we don't have
1283 : : * to release it here
1284 : : */
1285 : 0 : kmem_cache_free(dio_cache, dio);
1286 : 0 : goto out;
1287 : : }
1288 : : }
1289 : :
1290 : : /*
1291 : : * Will be decremented at I/O completion time.
1292 : : */
1293 : : inode_dio_begin(inode);
1294 : :
1295 : : retval = 0;
1296 : 0 : sdio.blkbits = blkbits;
1297 : 0 : sdio.blkfactor = i_blkbits - blkbits;
1298 : 0 : sdio.block_in_file = offset >> blkbits;
1299 : :
1300 : 0 : sdio.get_block = get_block;
1301 : 0 : dio->end_io = end_io;
1302 : 0 : sdio.submit_io = submit_io;
1303 : 0 : sdio.final_block_in_bio = -1;
1304 : 0 : sdio.next_block_for_io = -1;
1305 : :
1306 : 0 : dio->iocb = iocb;
1307 : :
1308 : 0 : spin_lock_init(&dio->bio_lock);
1309 : 0 : dio->refcount = 1;
1310 : :
1311 : 0 : dio->should_dirty = iter_is_iovec(iter) && iov_iter_rw(iter) == READ;
1312 : 0 : sdio.iter = iter;
1313 : 0 : sdio.final_block_in_request = end >> blkbits;
1314 : :
1315 : : /*
1316 : : * In case of non-aligned buffers, we may need 2 more
1317 : : * pages since we need to zero out first and last block.
1318 : : */
1319 : 0 : if (unlikely(sdio.blkfactor))
1320 : 0 : sdio.pages_in_io = 2;
1321 : :
1322 : 0 : sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
1323 : :
1324 : 0 : blk_start_plug(&plug);
1325 : :
1326 : 0 : retval = do_direct_IO(dio, &sdio, &map_bh);
1327 : 0 : if (retval)
1328 : : dio_cleanup(dio, &sdio);
1329 : :
1330 : 0 : if (retval == -ENOTBLK) {
1331 : : /*
1332 : : * The remaining part of the request will be
1333 : : * be handled by buffered I/O when we return
1334 : : */
1335 : : retval = 0;
1336 : : }
1337 : : /*
1338 : : * There may be some unwritten disk at the end of a part-written
1339 : : * fs-block-sized block. Go zero that now.
1340 : : */
1341 : 0 : dio_zero_block(dio, &sdio, 1, &map_bh);
1342 : :
1343 : 0 : if (sdio.cur_page) {
1344 : : ssize_t ret2;
1345 : :
1346 : 0 : ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1347 : 0 : if (retval == 0)
1348 : : retval = ret2;
1349 : 0 : put_page(sdio.cur_page);
1350 : 0 : sdio.cur_page = NULL;
1351 : : }
1352 : 0 : if (sdio.bio)
1353 : 0 : dio_bio_submit(dio, &sdio);
1354 : :
1355 : 0 : blk_finish_plug(&plug);
1356 : :
1357 : : /*
1358 : : * It is possible that, we return short IO due to end of file.
1359 : : * In that case, we need to release all the pages we got hold on.
1360 : : */
1361 : : dio_cleanup(dio, &sdio);
1362 : :
1363 : : /*
1364 : : * All block lookups have been performed. For READ requests
1365 : : * we can let i_mutex go now that its achieved its purpose
1366 : : * of protecting us from looking up uninitialized blocks.
1367 : : */
1368 : 0 : if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
1369 : 0 : inode_unlock(dio->inode);
1370 : :
1371 : : /*
1372 : : * The only time we want to leave bios in flight is when a successful
1373 : : * partial aio read or full aio write have been setup. In that case
1374 : : * bio completion will call aio_complete. The only time it's safe to
1375 : : * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1376 : : * This had *better* be the only place that raises -EIOCBQUEUED.
1377 : : */
1378 : 0 : BUG_ON(retval == -EIOCBQUEUED);
1379 : 0 : if (dio->is_async && retval == 0 && dio->result &&
1380 : 0 : (iov_iter_rw(iter) == READ || dio->result == count))
1381 : : retval = -EIOCBQUEUED;
1382 : : else
1383 : 0 : dio_await_completion(dio);
1384 : :
1385 : 0 : if (drop_refcount(dio) == 0) {
1386 : 0 : retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE);
1387 : : } else
1388 : 0 : BUG_ON(retval != -EIOCBQUEUED);
1389 : :
1390 : : out:
1391 : 0 : return retval;
1392 : : }
1393 : :
1394 : 0 : ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1395 : : struct block_device *bdev, struct iov_iter *iter,
1396 : : get_block_t get_block,
1397 : : dio_iodone_t end_io, dio_submit_t submit_io,
1398 : : int flags)
1399 : : {
1400 : : /*
1401 : : * The block device state is needed in the end to finally
1402 : : * submit everything. Since it's likely to be cache cold
1403 : : * prefetch it here as first thing to hide some of the
1404 : : * latency.
1405 : : *
1406 : : * Attempt to prefetch the pieces we likely need later.
1407 : : */
1408 : 0 : prefetch(&bdev->bd_disk->part_tbl);
1409 : 0 : prefetch(bdev->bd_queue);
1410 : 0 : prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1411 : :
1412 : 0 : return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block,
1413 : : end_io, submit_io, flags);
1414 : : }
1415 : :
1416 : : EXPORT_SYMBOL(__blockdev_direct_IO);
1417 : :
1418 : 3 : static __init int dio_init(void)
1419 : : {
1420 : 3 : dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1421 : 3 : return 0;
1422 : : }
1423 : : module_init(dio_init)
|
