]>
Commit | Line | Data |
---|---|---|
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 | 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 | static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio) | |
168 | { | |
169 | ssize_t ret; | |
170 | ||
171 | ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES, | |
172 | &sdio->from); | |
173 | ||
174 | if (ret < 0 && sdio->blocks_available && (dio->op == REQ_OP_WRITE)) { | |
175 | 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 | if (dio->page_errors == 0) | |
182 | dio->page_errors = ret; | |
183 | get_page(page); | |
184 | dio->pages[0] = page; | |
185 | sdio->head = 0; | |
186 | sdio->tail = 1; | |
187 | sdio->from = 0; | |
188 | sdio->to = PAGE_SIZE; | |
189 | return 0; | |
190 | } | |
191 | ||
192 | if (ret >= 0) { | |
193 | iov_iter_advance(sdio->iter, ret); | |
194 | ret += sdio->from; | |
195 | sdio->head = 0; | |
196 | sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE; | |
197 | sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1; | |
198 | 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 | static inline struct page *dio_get_page(struct dio *dio, | |
210 | struct dio_submit *sdio) | |
211 | { | |
212 | if (dio_pages_present(sdio) == 0) { | |
213 | int ret; | |
214 | ||
215 | ret = dio_refill_pages(dio, sdio); | |
216 | if (ret) | |
217 | return ERR_PTR(ret); | |
218 | BUG_ON(dio_pages_present(sdio) == 0); | |
219 | } | |
220 | return dio->pages[sdio->head]; | |
221 | } | |
222 | ||
223 | /* | |
224 | * Warn about a page cache invalidation failure during a direct io write. | |
225 | */ | |
226 | 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 | errseq_set(&inode->i_mapping->wb_err, -EIO); | |
234 | if (__ratelimit(&_rs)) { | |
235 | path = file_path(filp, pathname, sizeof(pathname)); | |
236 | if (IS_ERR(path)) | |
237 | path = "(unknown)"; | |
238 | pr_crit("Page cache invalidation failure on direct I/O. Possible data corruption due to collision with buffered I/O!\n"); | |
239 | pr_crit("File: %s PID: %d Comm: %.20s\n", path, current->pid, | |
240 | current->comm); | |
241 | } | |
242 | } | |
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 | static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags) | |
256 | { | |
257 | 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 | if (ret == -EIOCBQUEUED) | |
268 | ret = 0; | |
269 | ||
270 | if (dio->result) { | |
271 | transferred = dio->result; | |
272 | ||
273 | /* Check for short read case */ | |
274 | if ((dio->op == REQ_OP_READ) && | |
275 | ((offset + transferred) > dio->i_size)) | |
276 | transferred = dio->i_size - offset; | |
277 | /* ignore EFAULT if some IO has been done */ | |
278 | if (unlikely(ret == -EFAULT) && transferred) | |
279 | ret = 0; | |
280 | } | |
281 | ||
282 | if (ret == 0) | |
283 | ret = dio->page_errors; | |
284 | if (ret == 0) | |
285 | ret = dio->io_error; | |
286 | if (ret == 0) | |
287 | ret = transferred; | |
288 | ||
289 | if (dio->end_io) { | |
290 | // XXX: ki_pos?? | |
291 | err = dio->end_io(dio->iocb, offset, ret, dio->private); | |
292 | 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 | if (flags & DIO_COMPLETE_INVALIDATE && | |
309 | ret > 0 && dio->op == REQ_OP_WRITE && | |
310 | dio->inode->i_mapping->nrpages) { | |
311 | err = invalidate_inode_pages2_range(dio->inode->i_mapping, | |
312 | offset >> PAGE_SHIFT, | |
313 | (offset + ret - 1) >> PAGE_SHIFT); | |
314 | if (err) | |
315 | dio_warn_stale_pagecache(dio->iocb->ki_filp); | |
316 | } | |
317 | ||
318 | inode_dio_end(dio->inode); | |
319 | ||
320 | 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 | dio->iocb->ki_pos += transferred; | |
327 | ||
328 | if (ret > 0 && dio->op == REQ_OP_WRITE) | |
329 | ret = generic_write_sync(dio->iocb, ret); | |
330 | dio->iocb->ki_complete(dio->iocb, ret, 0); | |
331 | } | |
332 | ||
333 | kmem_cache_free(dio_cache, dio); | |
334 | return ret; | |
335 | } | |
336 | ||
337 | static void dio_aio_complete_work(struct work_struct *work) | |
338 | { | |
339 | struct dio *dio = container_of(work, struct dio, complete_work); | |
340 | ||
341 | dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE); | |
342 | } | |
343 | ||
344 | static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio); | |
345 | ||
346 | /* | |
347 | * Asynchronous IO callback. | |
348 | */ | |
349 | static void dio_bio_end_aio(struct bio *bio) | |
350 | { | |
351 | 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 | dio_bio_complete(dio, bio); | |
358 | ||
359 | spin_lock_irqsave(&dio->bio_lock, flags); | |
360 | remaining = --dio->refcount; | |
361 | if (remaining == 1 && dio->waiter) | |
362 | wake_up_process(dio->waiter); | |
363 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
364 | ||
365 | 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 | if (dio->result) | |
375 | defer_completion = dio->defer_completion || | |
376 | (dio->op == REQ_OP_WRITE && | |
377 | dio->inode->i_mapping->nrpages); | |
378 | if (defer_completion) { | |
379 | INIT_WORK(&dio->complete_work, dio_aio_complete_work); | |
380 | queue_work(dio->inode->i_sb->s_dio_done_wq, | |
381 | &dio->complete_work); | |
382 | } else { | |
383 | dio_complete(dio, 0, DIO_COMPLETE_ASYNC); | |
384 | } | |
385 | } | |
386 | } | |
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 | static void dio_bio_end_io(struct bio *bio) | |
396 | { | |
397 | struct dio *dio = bio->bi_private; | |
398 | unsigned long flags; | |
399 | ||
400 | spin_lock_irqsave(&dio->bio_lock, flags); | |
401 | bio->bi_private = dio->bio_list; | |
402 | dio->bio_list = bio; | |
403 | if (--dio->refcount == 1 && dio->waiter) | |
404 | wake_up_process(dio->waiter); | |
405 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
406 | } | |
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 | void dio_end_io(struct bio *bio) | |
417 | { | |
418 | struct dio *dio = bio->bi_private; | |
419 | ||
420 | if (dio->is_async) | |
421 | dio_bio_end_aio(bio); | |
422 | else | |
423 | dio_bio_end_io(bio); | |
424 | } | |
425 | EXPORT_SYMBOL_GPL(dio_end_io); | |
426 | ||
427 | static inline void | |
428 | 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 | bio = bio_alloc(GFP_KERNEL, nr_vecs); | |
439 | ||
440 | bio_set_dev(bio, bdev); | |
441 | bio->bi_iter.bi_sector = first_sector; | |
442 | bio_set_op_attrs(bio, dio->op, dio->op_flags); | |
443 | if (dio->is_async) | |
444 | bio->bi_end_io = dio_bio_end_aio; | |
445 | else | |
446 | bio->bi_end_io = dio_bio_end_io; | |
447 | ||
448 | bio->bi_write_hint = dio->iocb->ki_hint; | |
449 | ||
450 | sdio->bio = bio; | |
451 | sdio->logical_offset_in_bio = sdio->cur_page_fs_offset; | |
452 | } | |
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 | static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio) | |
462 | { | |
463 | struct bio *bio = sdio->bio; | |
464 | unsigned long flags; | |
465 | ||
466 | bio->bi_private = dio; | |
467 | ||
468 | spin_lock_irqsave(&dio->bio_lock, flags); | |
469 | dio->refcount++; | |
470 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
471 | ||
472 | if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty) | |
473 | bio_set_pages_dirty(bio); | |
474 | ||
475 | dio->bio_disk = bio->bi_disk; | |
476 | ||
477 | if (sdio->submit_io) { | |
478 | sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio); | |
479 | dio->bio_cookie = BLK_QC_T_NONE; | |
480 | } else | |
481 | dio->bio_cookie = submit_bio(bio); | |
482 | ||
483 | sdio->bio = NULL; | |
484 | sdio->boundary = 0; | |
485 | sdio->logical_offset_in_bio = 0; | |
486 | } | |
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 | while (sdio->head < sdio->tail) | |
494 | 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 | static struct bio *dio_await_one(struct dio *dio) | |
504 | { | |
505 | unsigned long flags; | |
506 | struct bio *bio = NULL; | |
507 | ||
508 | 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 | while (dio->refcount > 1 && dio->bio_list == NULL) { | |
517 | __set_current_state(TASK_UNINTERRUPTIBLE); | |
518 | dio->waiter = current; | |
519 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
520 | if (!(dio->iocb->ki_flags & IOCB_HIPRI) || | |
521 | !blk_poll(dio->bio_disk->queue, dio->bio_cookie, true)) | |
522 | io_schedule(); | |
523 | /* wake up sets us TASK_RUNNING */ | |
524 | spin_lock_irqsave(&dio->bio_lock, flags); | |
525 | dio->waiter = NULL; | |
526 | } | |
527 | if (dio->bio_list) { | |
528 | bio = dio->bio_list; | |
529 | dio->bio_list = bio->bi_private; | |
530 | } | |
531 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
532 | return bio; | |
533 | } | |
534 | ||
535 | /* | |
536 | * Process one completed BIO. No locks are held. | |
537 | */ | |
538 | static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio) | |
539 | { | |
540 | blk_status_t err = bio->bi_status; | |
541 | bool should_dirty = dio->op == REQ_OP_READ && dio->should_dirty; | |
542 | ||
543 | if (err) { | |
544 | if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT)) | |
545 | dio->io_error = -EAGAIN; | |
546 | else | |
547 | dio->io_error = -EIO; | |
548 | } | |
549 | ||
550 | if (dio->is_async && should_dirty) { | |
551 | bio_check_pages_dirty(bio); /* transfers ownership */ | |
552 | } else { | |
553 | bio_release_pages(bio, should_dirty); | |
554 | bio_put(bio); | |
555 | } | |
556 | 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 | static void dio_await_completion(struct dio *dio) | |
567 | { | |
568 | struct bio *bio; | |
569 | do { | |
570 | bio = dio_await_one(dio); | |
571 | if (bio) | |
572 | dio_bio_complete(dio, bio); | |
573 | } while (bio); | |
574 | } | |
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 | static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio) | |
584 | { | |
585 | int ret = 0; | |
586 | ||
587 | if (sdio->reap_counter++ >= 64) { | |
588 | while (dio->bio_list) { | |
589 | unsigned long flags; | |
590 | struct bio *bio; | |
591 | int ret2; | |
592 | ||
593 | spin_lock_irqsave(&dio->bio_lock, flags); | |
594 | bio = dio->bio_list; | |
595 | dio->bio_list = bio->bi_private; | |
596 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
597 | ret2 = blk_status_to_errno(dio_bio_complete(dio, bio)); | |
598 | if (ret == 0) | |
599 | ret = ret2; | |
600 | } | |
601 | sdio->reap_counter = 0; | |
602 | } | |
603 | 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 | int sb_init_dio_done_wq(struct super_block *sb) | |
613 | { | |
614 | struct workqueue_struct *old; | |
615 | struct workqueue_struct *wq = alloc_workqueue("dio/%s", | |
616 | WQ_MEM_RECLAIM, 0, | |
617 | sb->s_id); | |
618 | if (!wq) | |
619 | return -ENOMEM; | |
620 | /* | |
621 | * This has to be atomic as more DIOs can race to create the workqueue | |
622 | */ | |
623 | old = cmpxchg(&sb->s_dio_done_wq, NULL, wq); | |
624 | /* Someone created workqueue before us? Free ours... */ | |
625 | if (old) | |
626 | destroy_workqueue(wq); | |
627 | return 0; | |
628 | } | |
629 | ||
630 | static int dio_set_defer_completion(struct dio *dio) | |
631 | { | |
632 | struct super_block *sb = dio->inode->i_sb; | |
633 | ||
634 | if (dio->defer_completion) | |
635 | return 0; | |
636 | dio->defer_completion = true; | |
637 | if (!sb->s_dio_done_wq) | |
638 | 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 | 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 | 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 | ret = dio->page_errors; | |
681 | if (ret == 0) { | |
682 | BUG_ON(sdio->block_in_file >= sdio->final_block_in_request); | |
683 | fs_startblk = sdio->block_in_file >> sdio->blkfactor; | |
684 | fs_endblk = (sdio->final_block_in_request - 1) >> | |
685 | sdio->blkfactor; | |
686 | fs_count = fs_endblk - fs_startblk + 1; | |
687 | ||
688 | map_bh->b_state = 0; | |
689 | 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 | create = dio->op == REQ_OP_WRITE; | |
703 | if (dio->flags & DIO_SKIP_HOLES) { | |
704 | i_size = i_size_read(dio->inode); | |
705 | if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits) | |
706 | create = 0; | |
707 | } | |
708 | ||
709 | ret = (*sdio->get_block)(dio->inode, fs_startblk, | |
710 | map_bh, create); | |
711 | ||
712 | /* Store for completion */ | |
713 | dio->private = map_bh->b_private; | |
714 | ||
715 | if (ret == 0 && buffer_defer_completion(map_bh)) | |
716 | ret = dio_set_defer_completion(dio); | |
717 | } | |
718 | return ret; | |
719 | } | |
720 | ||
721 | /* | |
722 | * There is no bio. Make one now. | |
723 | */ | |
724 | 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 | ret = dio_bio_reap(dio, sdio); | |
731 | if (ret) | |
732 | goto out; | |
733 | sector = start_sector << (sdio->blkbits - 9); | |
734 | nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES); | |
735 | BUG_ON(nr_pages <= 0); | |
736 | dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages); | |
737 | sdio->boundary = 0; | |
738 | out: | |
739 | 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 | static inline int dio_bio_add_page(struct dio_submit *sdio) | |
750 | { | |
751 | int ret; | |
752 | ||
753 | ret = bio_add_page(sdio->bio, sdio->cur_page, | |
754 | sdio->cur_page_len, sdio->cur_page_offset); | |
755 | if (ret == sdio->cur_page_len) { | |
756 | /* | |
757 | * Decrement count only, if we are done with this page | |
758 | */ | |
759 | if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE) | |
760 | sdio->pages_in_io--; | |
761 | get_page(sdio->cur_page); | |
762 | sdio->final_block_in_bio = sdio->cur_page_block + | |
763 | (sdio->cur_page_len >> sdio->blkbits); | |
764 | ret = 0; | |
765 | } else { | |
766 | ret = 1; | |
767 | } | |
768 | 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 | 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 | if (sdio->bio) { | |
787 | loff_t cur_offset = sdio->cur_page_fs_offset; | |
788 | loff_t bio_next_offset = sdio->logical_offset_in_bio + | |
789 | 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 | if (sdio->final_block_in_bio != sdio->cur_page_block || | |
806 | cur_offset != bio_next_offset) | |
807 | dio_bio_submit(dio, sdio); | |
808 | } | |
809 | ||
810 | if (sdio->bio == NULL) { | |
811 | ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh); | |
812 | if (ret) | |
813 | goto out; | |
814 | } | |
815 | ||
816 | if (dio_bio_add_page(sdio) != 0) { | |
817 | dio_bio_submit(dio, sdio); | |
818 | ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh); | |
819 | if (ret == 0) { | |
820 | ret = dio_bio_add_page(sdio); | |
821 | BUG_ON(ret != 0); | |
822 | } | |
823 | } | |
824 | out: | |
825 | 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 | 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 | int boundary = sdio->boundary; /* dio_send_cur_page may clear it */ | |
852 | ||
853 | if (dio->op == REQ_OP_WRITE) { | |
854 | /* | |
855 | * Read accounting is performed in submit_bio() | |
856 | */ | |
857 | task_io_account_write(len); | |
858 | } | |
859 | ||
860 | /* | |
861 | * Can we just grow the current page's presence in the dio? | |
862 | */ | |
863 | if (sdio->cur_page == page && | |
864 | sdio->cur_page_offset + sdio->cur_page_len == offset && | |
865 | sdio->cur_page_block + | |
866 | (sdio->cur_page_len >> sdio->blkbits) == blocknr) { | |
867 | sdio->cur_page_len += len; | |
868 | goto out; | |
869 | } | |
870 | ||
871 | /* | |
872 | * If there's a deferred page already there then send it. | |
873 | */ | |
874 | if (sdio->cur_page) { | |
875 | ret = dio_send_cur_page(dio, sdio, map_bh); | |
876 | put_page(sdio->cur_page); | |
877 | sdio->cur_page = NULL; | |
878 | if (ret) | |
879 | return ret; | |
880 | } | |
881 | ||
882 | get_page(page); /* It is in dio */ | |
883 | sdio->cur_page = page; | |
884 | sdio->cur_page_offset = offset; | |
885 | sdio->cur_page_len = len; | |
886 | sdio->cur_page_block = blocknr; | |
887 | sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits; | |
888 | out: | |
889 | /* | |
890 | * If boundary then we want to schedule the IO now to | |
891 | * avoid metadata seeks. | |
892 | */ | |
893 | if (boundary) { | |
894 | ret = dio_send_cur_page(dio, sdio, map_bh); | |
895 | if (sdio->bio) | |
896 | dio_bio_submit(dio, sdio); | |
897 | put_page(sdio->cur_page); | |
898 | sdio->cur_page = NULL; | |
899 | } | |
900 | return ret; | |
901 | } | |
902 | ||
903 | /* | |
904 | * If we are not writing the entire block and get_block() allocated | |
905 | * the block for us, we need to fill-in the unused portion of the | |
906 | * block with zeros. This happens only if user-buffer, fileoffset or | |
907 | * io length is not filesystem block-size multiple. | |
908 | * | |
909 | * `end' is zero if we're doing the start of the IO, 1 at the end of the | |
910 | * IO. | |
911 | */ | |
912 | static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio, | |
913 | int end, struct buffer_head *map_bh) | |
914 | { | |
915 | unsigned dio_blocks_per_fs_block; | |
916 | unsigned this_chunk_blocks; /* In dio_blocks */ | |
917 | unsigned this_chunk_bytes; | |
918 | struct page *page; | |
919 | ||
920 | sdio->start_zero_done = 1; | |
921 | if (!sdio->blkfactor || !buffer_new(map_bh)) | |
922 | return; | |
923 | ||
924 | dio_blocks_per_fs_block = 1 << sdio->blkfactor; | |
925 | this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1); | |
926 | ||
927 | if (!this_chunk_blocks) | |
928 | return; | |
929 | ||
930 | /* | |
931 | * We need to zero out part of an fs block. It is either at the | |
932 | * beginning or the end of the fs block. | |
933 | */ | |
934 | if (end) | |
935 | this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks; | |
936 | ||
937 | this_chunk_bytes = this_chunk_blocks << sdio->blkbits; | |
938 | ||
939 | page = ZERO_PAGE(0); | |
940 | if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes, | |
941 | sdio->next_block_for_io, map_bh)) | |
942 | return; | |
943 | ||
944 | sdio->next_block_for_io += this_chunk_blocks; | |
945 | } | |
946 | ||
947 | /* | |
948 | * Walk the user pages, and the file, mapping blocks to disk and generating | |
949 | * a sequence of (page,offset,len,block) mappings. These mappings are injected | |
950 | * into submit_page_section(), which takes care of the next stage of submission | |
951 | * | |
952 | * Direct IO against a blockdev is different from a file. Because we can | |
953 | * happily perform page-sized but 512-byte aligned IOs. It is important that | |
954 | * blockdev IO be able to have fine alignment and large sizes. | |
955 | * | |
956 | * So what we do is to permit the ->get_block function to populate bh.b_size | |
957 | * with the size of IO which is permitted at this offset and this i_blkbits. | |
958 | * | |
959 | * For best results, the blockdev should be set up with 512-byte i_blkbits and | |
960 | * it should set b_size to PAGE_SIZE or more inside get_block(). This gives | |
961 | * fine alignment but still allows this function to work in PAGE_SIZE units. | |
962 | */ | |
963 | static int do_direct_IO(struct dio *dio, struct dio_submit *sdio, | |
964 | struct buffer_head *map_bh) | |
965 | { | |
966 | const unsigned blkbits = sdio->blkbits; | |
967 | const unsigned i_blkbits = blkbits + sdio->blkfactor; | |
968 | int ret = 0; | |
969 | ||
970 | while (sdio->block_in_file < sdio->final_block_in_request) { | |
971 | struct page *page; | |
972 | size_t from, to; | |
973 | ||
974 | page = dio_get_page(dio, sdio); | |
975 | if (IS_ERR(page)) { | |
976 | ret = PTR_ERR(page); | |
977 | goto out; | |
978 | } | |
979 | from = sdio->head ? 0 : sdio->from; | |
980 | to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE; | |
981 | sdio->head++; | |
982 | ||
983 | while (from < to) { | |
984 | unsigned this_chunk_bytes; /* # of bytes mapped */ | |
985 | unsigned this_chunk_blocks; /* # of blocks */ | |
986 | unsigned u; | |
987 | ||
988 | if (sdio->blocks_available == 0) { | |
989 | /* | |
990 | * Need to go and map some more disk | |
991 | */ | |
992 | unsigned long blkmask; | |
993 | unsigned long dio_remainder; | |
994 | ||
995 | ret = get_more_blocks(dio, sdio, map_bh); | |
996 | if (ret) { | |
997 | put_page(page); | |
998 | goto out; | |
999 | } | |
1000 | if (!buffer_mapped(map_bh)) | |
1001 | goto do_holes; | |
1002 | ||
1003 | sdio->blocks_available = | |
1004 | map_bh->b_size >> blkbits; | |
1005 | sdio->next_block_for_io = | |
1006 | map_bh->b_blocknr << sdio->blkfactor; | |
1007 | if (buffer_new(map_bh)) { | |
1008 | clean_bdev_aliases( | |
1009 | map_bh->b_bdev, | |
1010 | map_bh->b_blocknr, | |
1011 | map_bh->b_size >> i_blkbits); | |
1012 | } | |
1013 | ||
1014 | if (!sdio->blkfactor) | |
1015 | goto do_holes; | |
1016 | ||
1017 | blkmask = (1 << sdio->blkfactor) - 1; | |
1018 | dio_remainder = (sdio->block_in_file & blkmask); | |
1019 | ||
1020 | /* | |
1021 | * If we are at the start of IO and that IO | |
1022 | * starts partway into a fs-block, | |
1023 | * dio_remainder will be non-zero. If the IO | |
1024 | * is a read then we can simply advance the IO | |
1025 | * cursor to the first block which is to be | |
1026 | * read. But if the IO is a write and the | |
1027 | * block was newly allocated we cannot do that; | |
1028 | * the start of the fs block must be zeroed out | |
1029 | * on-disk | |
1030 | */ | |
1031 | if (!buffer_new(map_bh)) | |
1032 | sdio->next_block_for_io += dio_remainder; | |
1033 | sdio->blocks_available -= dio_remainder; | |
1034 | } | |
1035 | do_holes: | |
1036 | /* Handle holes */ | |
1037 | if (!buffer_mapped(map_bh)) { | |
1038 | loff_t i_size_aligned; | |
1039 | ||
1040 | /* AKPM: eargh, -ENOTBLK is a hack */ | |
1041 | if (dio->op == REQ_OP_WRITE) { | |
1042 | put_page(page); | |
1043 | return -ENOTBLK; | |
1044 | } | |
1045 | ||
1046 | /* | |
1047 | * Be sure to account for a partial block as the | |
1048 | * last block in the file | |
1049 | */ | |
1050 | i_size_aligned = ALIGN(i_size_read(dio->inode), | |
1051 | 1 << blkbits); | |
1052 | if (sdio->block_in_file >= | |
1053 | i_size_aligned >> blkbits) { | |
1054 | /* We hit eof */ | |
1055 | put_page(page); | |
1056 | goto out; | |
1057 | } | |
1058 | zero_user(page, from, 1 << blkbits); | |
1059 | sdio->block_in_file++; | |
1060 | from += 1 << blkbits; | |
1061 | dio->result += 1 << blkbits; | |
1062 | goto next_block; | |
1063 | } | |
1064 | ||
1065 | /* | |
1066 | * If we're performing IO which has an alignment which | |
1067 | * is finer than the underlying fs, go check to see if | |
1068 | * we must zero out the start of this block. | |
1069 | */ | |
1070 | if (unlikely(sdio->blkfactor && !sdio->start_zero_done)) | |
1071 | dio_zero_block(dio, sdio, 0, map_bh); | |
1072 | ||
1073 | /* | |
1074 | * Work out, in this_chunk_blocks, how much disk we | |
1075 | * can add to this page | |
1076 | */ | |
1077 | this_chunk_blocks = sdio->blocks_available; | |
1078 | u = (to - from) >> blkbits; | |
1079 | if (this_chunk_blocks > u) | |
1080 | this_chunk_blocks = u; | |
1081 | u = sdio->final_block_in_request - sdio->block_in_file; | |
1082 | if (this_chunk_blocks > u) | |
1083 | this_chunk_blocks = u; | |
1084 | this_chunk_bytes = this_chunk_blocks << blkbits; | |
1085 | BUG_ON(this_chunk_bytes == 0); | |
1086 | ||
1087 | if (this_chunk_blocks == sdio->blocks_available) | |
1088 | sdio->boundary = buffer_boundary(map_bh); | |
1089 | ret = submit_page_section(dio, sdio, page, | |
1090 | from, | |
1091 | this_chunk_bytes, | |
1092 | sdio->next_block_for_io, | |
1093 | map_bh); | |
1094 | if (ret) { | |
1095 | put_page(page); | |
1096 | goto out; | |
1097 | } | |
1098 | sdio->next_block_for_io += this_chunk_blocks; | |
1099 | ||
1100 | sdio->block_in_file += this_chunk_blocks; | |
1101 | from += this_chunk_bytes; | |
1102 | dio->result += this_chunk_bytes; | |
1103 | sdio->blocks_available -= this_chunk_blocks; | |
1104 | next_block: | |
1105 | BUG_ON(sdio->block_in_file > sdio->final_block_in_request); | |
1106 | if (sdio->block_in_file == sdio->final_block_in_request) | |
1107 | break; | |
1108 | } | |
1109 | ||
1110 | /* Drop the ref which was taken in get_user_pages() */ | |
1111 | put_page(page); | |
1112 | } | |
1113 | out: | |
1114 | return ret; | |
1115 | } | |
1116 | ||
1117 | static inline int drop_refcount(struct dio *dio) | |
1118 | { | |
1119 | int ret2; | |
1120 | unsigned long flags; | |
1121 | ||
1122 | /* | |
1123 | * Sync will always be dropping the final ref and completing the | |
1124 | * operation. AIO can if it was a broken operation described above or | |
1125 | * in fact if all the bios race to complete before we get here. In | |
1126 | * that case dio_complete() translates the EIOCBQUEUED into the proper | |
1127 | * return code that the caller will hand to ->complete(). | |
1128 | * | |
1129 | * This is managed by the bio_lock instead of being an atomic_t so that | |
1130 | * completion paths can drop their ref and use the remaining count to | |
1131 | * decide to wake the submission path atomically. | |
1132 | */ | |
1133 | spin_lock_irqsave(&dio->bio_lock, flags); | |
1134 | ret2 = --dio->refcount; | |
1135 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
1136 | return ret2; | |
1137 | } | |
1138 | ||
1139 | /* | |
1140 | * This is a library function for use by filesystem drivers. | |
1141 | * | |
1142 | * The locking rules are governed by the flags parameter: | |
1143 | * - if the flags value contains DIO_LOCKING we use a fancy locking | |
1144 | * scheme for dumb filesystems. | |
1145 | * For writes this function is called under i_mutex and returns with | |
1146 | * i_mutex held, for reads, i_mutex is not held on entry, but it is | |
1147 | * taken and dropped again before returning. | |
1148 | * - if the flags value does NOT contain DIO_LOCKING we don't use any | |
1149 | * internal locking but rather rely on the filesystem to synchronize | |
1150 | * direct I/O reads/writes versus each other and truncate. | |
1151 | * | |
1152 | * To help with locking against truncate we incremented the i_dio_count | |
1153 | * counter before starting direct I/O, and decrement it once we are done. | |
1154 | * Truncate can wait for it to reach zero to provide exclusion. It is | |
1155 | * expected that filesystem provide exclusion between new direct I/O | |
1156 | * and truncates. For DIO_LOCKING filesystems this is done by i_mutex, | |
1157 | * but other filesystems need to take care of this on their own. | |
1158 | * | |
1159 | * NOTE: if you pass "sdio" to anything by pointer make sure that function | |
1160 | * is always inlined. Otherwise gcc is unable to split the structure into | |
1161 | * individual fields and will generate much worse code. This is important | |
1162 | * for the whole file. | |
1163 | */ | |
1164 | static inline ssize_t | |
1165 | do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode, | |
1166 | struct block_device *bdev, struct iov_iter *iter, | |
1167 | get_block_t get_block, dio_iodone_t end_io, | |
1168 | dio_submit_t submit_io, int flags) | |
1169 | { | |
1170 | unsigned i_blkbits = READ_ONCE(inode->i_blkbits); | |
1171 | unsigned blkbits = i_blkbits; | |
1172 | unsigned blocksize_mask = (1 << blkbits) - 1; | |
1173 | ssize_t retval = -EINVAL; | |
1174 | const size_t count = iov_iter_count(iter); | |
1175 | loff_t offset = iocb->ki_pos; | |
1176 | const loff_t end = offset + count; | |
1177 | struct dio *dio; | |
1178 | struct dio_submit sdio = { 0, }; | |
1179 | struct buffer_head map_bh = { 0, }; | |
1180 | struct blk_plug plug; | |
1181 | unsigned long align = offset | iov_iter_alignment(iter); | |
1182 | ||
1183 | /* | |
1184 | * Avoid references to bdev if not absolutely needed to give | |
1185 | * the early prefetch in the caller enough time. | |
1186 | */ | |
1187 | ||
1188 | if (align & blocksize_mask) { | |
1189 | if (bdev) | |
1190 | blkbits = blksize_bits(bdev_logical_block_size(bdev)); | |
1191 | blocksize_mask = (1 << blkbits) - 1; | |
1192 | if (align & blocksize_mask) | |
1193 | goto out; | |
1194 | } | |
1195 | ||
1196 | /* watch out for a 0 len io from a tricksy fs */ | |
1197 | if (iov_iter_rw(iter) == READ && !count) | |
1198 | return 0; | |
1199 | ||
1200 | dio = kmem_cache_alloc(dio_cache, GFP_KERNEL); | |
1201 | retval = -ENOMEM; | |
1202 | if (!dio) | |
1203 | goto out; | |
1204 | /* | |
1205 | * Believe it or not, zeroing out the page array caused a .5% | |
1206 | * performance regression in a database benchmark. So, we take | |
1207 | * care to only zero out what's needed. | |
1208 | */ | |
1209 | memset(dio, 0, offsetof(struct dio, pages)); | |
1210 | ||
1211 | dio->flags = flags; | |
1212 | if (dio->flags & DIO_LOCKING) { | |
1213 | if (iov_iter_rw(iter) == READ) { | |
1214 | struct address_space *mapping = | |
1215 | iocb->ki_filp->f_mapping; | |
1216 | ||
1217 | /* will be released by direct_io_worker */ | |
1218 | inode_lock(inode); | |
1219 | ||
1220 | retval = filemap_write_and_wait_range(mapping, offset, | |
1221 | end - 1); | |
1222 | if (retval) { | |
1223 | inode_unlock(inode); | |
1224 | kmem_cache_free(dio_cache, dio); | |
1225 | goto out; | |
1226 | } | |
1227 | } | |
1228 | } | |
1229 | ||
1230 | /* Once we sampled i_size check for reads beyond EOF */ | |
1231 | dio->i_size = i_size_read(inode); | |
1232 | if (iov_iter_rw(iter) == READ && offset >= dio->i_size) { | |
1233 | if (dio->flags & DIO_LOCKING) | |
1234 | inode_unlock(inode); | |
1235 | kmem_cache_free(dio_cache, dio); | |
1236 | retval = 0; | |
1237 | goto out; | |
1238 | } | |
1239 | ||
1240 | /* | |
1241 | * For file extending writes updating i_size before data writeouts | |
1242 | * complete can expose uninitialized blocks in dumb filesystems. | |
1243 | * In that case we need to wait for I/O completion even if asked | |
1244 | * for an asynchronous write. | |
1245 | */ | |
1246 | if (is_sync_kiocb(iocb)) | |
1247 | dio->is_async = false; | |
1248 | else if (iov_iter_rw(iter) == WRITE && end > i_size_read(inode)) | |
1249 | dio->is_async = false; | |
1250 | else | |
1251 | dio->is_async = true; | |
1252 | ||
1253 | dio->inode = inode; | |
1254 | if (iov_iter_rw(iter) == WRITE) { | |
1255 | dio->op = REQ_OP_WRITE; | |
1256 | dio->op_flags = REQ_SYNC | REQ_IDLE; | |
1257 | if (iocb->ki_flags & IOCB_NOWAIT) | |
1258 | dio->op_flags |= REQ_NOWAIT; | |
1259 | } else { | |
1260 | dio->op = REQ_OP_READ; | |
1261 | } | |
1262 | if (iocb->ki_flags & IOCB_HIPRI) | |
1263 | dio->op_flags |= REQ_HIPRI; | |
1264 | ||
1265 | /* | |
1266 | * For AIO O_(D)SYNC writes we need to defer completions to a workqueue | |
1267 | * so that we can call ->fsync. | |
1268 | */ | |
1269 | if (dio->is_async && iov_iter_rw(iter) == WRITE) { | |
1270 | retval = 0; | |
1271 | if (iocb->ki_flags & IOCB_DSYNC) | |
1272 | retval = dio_set_defer_completion(dio); | |
1273 | else if (!dio->inode->i_sb->s_dio_done_wq) { | |
1274 | /* | |
1275 | * In case of AIO write racing with buffered read we | |
1276 | * need to defer completion. We can't decide this now, | |
1277 | * however the workqueue needs to be initialized here. | |
1278 | */ | |
1279 | retval = sb_init_dio_done_wq(dio->inode->i_sb); | |
1280 | } | |
1281 | if (retval) { | |
1282 | /* | |
1283 | * We grab i_mutex only for reads so we don't have | |
1284 | * to release it here | |
1285 | */ | |
1286 | kmem_cache_free(dio_cache, dio); | |
1287 | goto out; | |
1288 | } | |
1289 | } | |
1290 | ||
1291 | /* | |
1292 | * Will be decremented at I/O completion time. | |
1293 | */ | |
1294 | inode_dio_begin(inode); | |
1295 | ||
1296 | retval = 0; | |
1297 | sdio.blkbits = blkbits; | |
1298 | sdio.blkfactor = i_blkbits - blkbits; | |
1299 | sdio.block_in_file = offset >> blkbits; | |
1300 | ||
1301 | sdio.get_block = get_block; | |
1302 | dio->end_io = end_io; | |
1303 | sdio.submit_io = submit_io; | |
1304 | sdio.final_block_in_bio = -1; | |
1305 | sdio.next_block_for_io = -1; | |
1306 | ||
1307 | dio->iocb = iocb; | |
1308 | ||
1309 | spin_lock_init(&dio->bio_lock); | |
1310 | dio->refcount = 1; | |
1311 | ||
1312 | dio->should_dirty = iter_is_iovec(iter) && iov_iter_rw(iter) == READ; | |
1313 | sdio.iter = iter; | |
1314 | sdio.final_block_in_request = end >> blkbits; | |
1315 | ||
1316 | /* | |
1317 | * In case of non-aligned buffers, we may need 2 more | |
1318 | * pages since we need to zero out first and last block. | |
1319 | */ | |
1320 | if (unlikely(sdio.blkfactor)) | |
1321 | sdio.pages_in_io = 2; | |
1322 | ||
1323 | sdio.pages_in_io += iov_iter_npages(iter, INT_MAX); | |
1324 | ||
1325 | blk_start_plug(&plug); | |
1326 | ||
1327 | retval = do_direct_IO(dio, &sdio, &map_bh); | |
1328 | if (retval) | |
1329 | dio_cleanup(dio, &sdio); | |
1330 | ||
1331 | if (retval == -ENOTBLK) { | |
1332 | /* | |
1333 | * The remaining part of the request will be | |
1334 | * be handled by buffered I/O when we return | |
1335 | */ | |
1336 | retval = 0; | |
1337 | } | |
1338 | /* | |
1339 | * There may be some unwritten disk at the end of a part-written | |
1340 | * fs-block-sized block. Go zero that now. | |
1341 | */ | |
1342 | dio_zero_block(dio, &sdio, 1, &map_bh); | |
1343 | ||
1344 | if (sdio.cur_page) { | |
1345 | ssize_t ret2; | |
1346 | ||
1347 | ret2 = dio_send_cur_page(dio, &sdio, &map_bh); | |
1348 | if (retval == 0) | |
1349 | retval = ret2; | |
1350 | put_page(sdio.cur_page); | |
1351 | sdio.cur_page = NULL; | |
1352 | } | |
1353 | if (sdio.bio) | |
1354 | dio_bio_submit(dio, &sdio); | |
1355 | ||
1356 | blk_finish_plug(&plug); | |
1357 | ||
1358 | /* | |
1359 | * It is possible that, we return short IO due to end of file. | |
1360 | * In that case, we need to release all the pages we got hold on. | |
1361 | */ | |
1362 | dio_cleanup(dio, &sdio); | |
1363 | ||
1364 | /* | |
1365 | * All block lookups have been performed. For READ requests | |
1366 | * we can let i_mutex go now that its achieved its purpose | |
1367 | * of protecting us from looking up uninitialized blocks. | |
1368 | */ | |
1369 | if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING)) | |
1370 | inode_unlock(dio->inode); | |
1371 | ||
1372 | /* | |
1373 | * The only time we want to leave bios in flight is when a successful | |
1374 | * partial aio read or full aio write have been setup. In that case | |
1375 | * bio completion will call aio_complete. The only time it's safe to | |
1376 | * call aio_complete is when we return -EIOCBQUEUED, so we key on that. | |
1377 | * This had *better* be the only place that raises -EIOCBQUEUED. | |
1378 | */ | |
1379 | BUG_ON(retval == -EIOCBQUEUED); | |
1380 | if (dio->is_async && retval == 0 && dio->result && | |
1381 | (iov_iter_rw(iter) == READ || dio->result == count)) | |
1382 | retval = -EIOCBQUEUED; | |
1383 | else | |
1384 | dio_await_completion(dio); | |
1385 | ||
1386 | if (drop_refcount(dio) == 0) { | |
1387 | retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE); | |
1388 | } else | |
1389 | BUG_ON(retval != -EIOCBQUEUED); | |
1390 | ||
1391 | out: | |
1392 | return retval; | |
1393 | } | |
1394 | ||
1395 | ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode, | |
1396 | struct block_device *bdev, struct iov_iter *iter, | |
1397 | get_block_t get_block, | |
1398 | dio_iodone_t end_io, dio_submit_t submit_io, | |
1399 | int flags) | |
1400 | { | |
1401 | /* | |
1402 | * The block device state is needed in the end to finally | |
1403 | * submit everything. Since it's likely to be cache cold | |
1404 | * prefetch it here as first thing to hide some of the | |
1405 | * latency. | |
1406 | * | |
1407 | * Attempt to prefetch the pieces we likely need later. | |
1408 | */ | |
1409 | prefetch(&bdev->bd_disk->part_tbl); | |
1410 | prefetch(bdev->bd_queue); | |
1411 | prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES); | |
1412 | ||
1413 | return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block, | |
1414 | end_io, submit_io, flags); | |
1415 | } | |
1416 | ||
1417 | EXPORT_SYMBOL(__blockdev_direct_IO); | |
1418 | ||
1419 | static __init int dio_init(void) | |
1420 | { | |
1421 | dio_cache = KMEM_CACHE(dio, SLAB_PANIC); | |
1422 | return 0; | |
1423 | } | |
1424 | module_init(dio_init) |