Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * fs/direct-io.c | |
3 | * | |
4 | * Copyright (C) 2002, Linus Torvalds. | |
5 | * | |
6 | * O_DIRECT | |
7 | * | |
8 | * 04Jul2002 akpm@zip.com.au | |
9 | * Initial version | |
10 | * 11Sep2002 janetinc@us.ibm.com | |
11 | * added readv/writev support. | |
12 | * 29Oct2002 akpm@zip.com.au | |
13 | * rewrote bio_add_page() support. | |
14 | * 30Oct2002 pbadari@us.ibm.com | |
15 | * added support for non-aligned IO. | |
16 | * 06Nov2002 pbadari@us.ibm.com | |
17 | * added asynchronous IO support. | |
18 | * 21Jul2003 nathans@sgi.com | |
19 | * added IO completion notifier. | |
20 | */ | |
21 | ||
22 | #include <linux/kernel.h> | |
23 | #include <linux/module.h> | |
24 | #include <linux/types.h> | |
25 | #include <linux/fs.h> | |
26 | #include <linux/mm.h> | |
27 | #include <linux/slab.h> | |
28 | #include <linux/highmem.h> | |
29 | #include <linux/pagemap.h> | |
98c4d57d | 30 | #include <linux/task_io_accounting_ops.h> |
1da177e4 LT |
31 | #include <linux/bio.h> |
32 | #include <linux/wait.h> | |
33 | #include <linux/err.h> | |
34 | #include <linux/blkdev.h> | |
35 | #include <linux/buffer_head.h> | |
36 | #include <linux/rwsem.h> | |
37 | #include <linux/uio.h> | |
38 | #include <asm/atomic.h> | |
39 | ||
40 | /* | |
41 | * How many user pages to map in one call to get_user_pages(). This determines | |
42 | * the size of a structure on the stack. | |
43 | */ | |
44 | #define DIO_PAGES 64 | |
45 | ||
46 | /* | |
47 | * This code generally works in units of "dio_blocks". A dio_block is | |
48 | * somewhere between the hard sector size and the filesystem block size. it | |
49 | * is determined on a per-invocation basis. When talking to the filesystem | |
50 | * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity | |
51 | * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted | |
52 | * to bio_block quantities by shifting left by blkfactor. | |
53 | * | |
54 | * If blkfactor is zero then the user's request was aligned to the filesystem's | |
55 | * blocksize. | |
56 | * | |
57 | * lock_type is DIO_LOCKING for regular files on direct-IO-naive filesystems. | |
58 | * This determines whether we need to do the fancy locking which prevents | |
59 | * direct-IO from being able to read uninitialised disk blocks. If its zero | |
1b1dcc1b | 60 | * (blockdev) this locking is not done, and if it is DIO_OWN_LOCKING i_mutex is |
1da177e4 LT |
61 | * not held for the entire direct write (taken briefly, initially, during a |
62 | * direct read though, but its never held for the duration of a direct-IO). | |
63 | */ | |
64 | ||
65 | struct dio { | |
66 | /* BIO submission state */ | |
67 | struct bio *bio; /* bio under assembly */ | |
68 | struct inode *inode; | |
69 | int rw; | |
29504ff3 | 70 | loff_t i_size; /* i_size when submitted */ |
1da177e4 LT |
71 | int lock_type; /* doesn't change */ |
72 | unsigned blkbits; /* doesn't change */ | |
73 | unsigned blkfactor; /* When we're using an alignment which | |
74 | is finer than the filesystem's soft | |
75 | blocksize, this specifies how much | |
76 | finer. blkfactor=2 means 1/4-block | |
77 | alignment. Does not change */ | |
78 | unsigned start_zero_done; /* flag: sub-blocksize zeroing has | |
79 | been performed at the start of a | |
80 | write */ | |
81 | int pages_in_io; /* approximate total IO pages */ | |
82 | size_t size; /* total request size (doesn't change)*/ | |
83 | sector_t block_in_file; /* Current offset into the underlying | |
84 | file in dio_block units. */ | |
85 | unsigned blocks_available; /* At block_in_file. changes */ | |
86 | sector_t final_block_in_request;/* doesn't change */ | |
87 | unsigned first_block_in_page; /* doesn't change, Used only once */ | |
88 | int boundary; /* prev block is at a boundary */ | |
89 | int reap_counter; /* rate limit reaping */ | |
1d8fa7a2 | 90 | get_block_t *get_block; /* block mapping function */ |
1da177e4 LT |
91 | dio_iodone_t *end_io; /* IO completion function */ |
92 | sector_t final_block_in_bio; /* current final block in bio + 1 */ | |
93 | sector_t next_block_for_io; /* next block to be put under IO, | |
94 | in dio_blocks units */ | |
1d8fa7a2 | 95 | struct buffer_head map_bh; /* last get_block() result */ |
1da177e4 LT |
96 | |
97 | /* | |
98 | * Deferred addition of a page to the dio. These variables are | |
99 | * private to dio_send_cur_page(), submit_page_section() and | |
100 | * dio_bio_add_page(). | |
101 | */ | |
102 | struct page *cur_page; /* The page */ | |
103 | unsigned cur_page_offset; /* Offset into it, in bytes */ | |
104 | unsigned cur_page_len; /* Nr of bytes at cur_page_offset */ | |
105 | sector_t cur_page_block; /* Where it starts */ | |
106 | ||
107 | /* | |
108 | * Page fetching state. These variables belong to dio_refill_pages(). | |
109 | */ | |
110 | int curr_page; /* changes */ | |
111 | int total_pages; /* doesn't change */ | |
112 | unsigned long curr_user_address;/* changes */ | |
113 | ||
114 | /* | |
115 | * Page queue. These variables belong to dio_refill_pages() and | |
116 | * dio_get_page(). | |
117 | */ | |
118 | struct page *pages[DIO_PAGES]; /* page buffer */ | |
119 | unsigned head; /* next page to process */ | |
120 | unsigned tail; /* last valid page + 1 */ | |
121 | int page_errors; /* errno from get_user_pages() */ | |
122 | ||
123 | /* BIO completion state */ | |
0273201e | 124 | atomic_t refcount; /* direct_io_worker() and bios */ |
1da177e4 | 125 | spinlock_t bio_lock; /* protects BIO fields below */ |
1da177e4 LT |
126 | struct bio *bio_list; /* singly linked via bi_private */ |
127 | struct task_struct *waiter; /* waiting task (NULL if none) */ | |
128 | ||
129 | /* AIO related stuff */ | |
130 | struct kiocb *iocb; /* kiocb */ | |
131 | int is_async; /* is IO async ? */ | |
174e27c6 | 132 | int io_error; /* IO error in completion path */ |
1da177e4 LT |
133 | ssize_t result; /* IO result */ |
134 | }; | |
135 | ||
136 | /* | |
137 | * How many pages are in the queue? | |
138 | */ | |
139 | static inline unsigned dio_pages_present(struct dio *dio) | |
140 | { | |
141 | return dio->tail - dio->head; | |
142 | } | |
143 | ||
144 | /* | |
145 | * Go grab and pin some userspace pages. Typically we'll get 64 at a time. | |
146 | */ | |
147 | static int dio_refill_pages(struct dio *dio) | |
148 | { | |
149 | int ret; | |
150 | int nr_pages; | |
151 | ||
152 | nr_pages = min(dio->total_pages - dio->curr_page, DIO_PAGES); | |
153 | down_read(¤t->mm->mmap_sem); | |
154 | ret = get_user_pages( | |
155 | current, /* Task for fault acounting */ | |
156 | current->mm, /* whose pages? */ | |
157 | dio->curr_user_address, /* Where from? */ | |
158 | nr_pages, /* How many pages? */ | |
159 | dio->rw == READ, /* Write to memory? */ | |
160 | 0, /* force (?) */ | |
161 | &dio->pages[0], | |
162 | NULL); /* vmas */ | |
163 | up_read(¤t->mm->mmap_sem); | |
164 | ||
b31dc66a | 165 | if (ret < 0 && dio->blocks_available && (dio->rw & WRITE)) { |
b5810039 | 166 | struct page *page = ZERO_PAGE(dio->curr_user_address); |
1da177e4 LT |
167 | /* |
168 | * A memory fault, but the filesystem has some outstanding | |
169 | * mapped blocks. We need to use those blocks up to avoid | |
170 | * leaking stale data in the file. | |
171 | */ | |
172 | if (dio->page_errors == 0) | |
173 | dio->page_errors = ret; | |
b5810039 NP |
174 | page_cache_get(page); |
175 | dio->pages[0] = page; | |
1da177e4 LT |
176 | dio->head = 0; |
177 | dio->tail = 1; | |
178 | ret = 0; | |
179 | goto out; | |
180 | } | |
181 | ||
182 | if (ret >= 0) { | |
183 | dio->curr_user_address += ret * PAGE_SIZE; | |
184 | dio->curr_page += ret; | |
185 | dio->head = 0; | |
186 | dio->tail = ret; | |
187 | ret = 0; | |
188 | } | |
189 | out: | |
190 | return ret; | |
191 | } | |
192 | ||
193 | /* | |
194 | * Get another userspace page. Returns an ERR_PTR on error. Pages are | |
195 | * buffered inside the dio so that we can call get_user_pages() against a | |
196 | * decent number of pages, less frequently. To provide nicer use of the | |
197 | * L1 cache. | |
198 | */ | |
199 | static struct page *dio_get_page(struct dio *dio) | |
200 | { | |
201 | if (dio_pages_present(dio) == 0) { | |
202 | int ret; | |
203 | ||
204 | ret = dio_refill_pages(dio); | |
205 | if (ret) | |
206 | return ERR_PTR(ret); | |
207 | BUG_ON(dio_pages_present(dio) == 0); | |
208 | } | |
209 | return dio->pages[dio->head++]; | |
210 | } | |
211 | ||
6d544bb4 ZB |
212 | /** |
213 | * dio_complete() - called when all DIO BIO I/O has been completed | |
214 | * @offset: the byte offset in the file of the completed operation | |
215 | * | |
216 | * This releases locks as dictated by the locking type, lets interested parties | |
217 | * know that a DIO operation has completed, and calculates the resulting return | |
218 | * code for the operation. | |
219 | * | |
220 | * It lets the filesystem know if it registered an interest earlier via | |
221 | * get_block. Pass the private field of the map buffer_head so that | |
222 | * filesystems can use it to hold additional state between get_block calls and | |
223 | * dio_complete. | |
1da177e4 | 224 | */ |
6d544bb4 | 225 | static int dio_complete(struct dio *dio, loff_t offset, int ret) |
1da177e4 | 226 | { |
6d544bb4 ZB |
227 | ssize_t transferred = 0; |
228 | ||
229 | if (dio->result) { | |
230 | transferred = dio->result; | |
231 | ||
232 | /* Check for short read case */ | |
233 | if ((dio->rw == READ) && ((offset + transferred) > dio->i_size)) | |
234 | transferred = dio->i_size - offset; | |
235 | } | |
236 | ||
1da177e4 | 237 | if (dio->end_io && dio->result) |
6d544bb4 ZB |
238 | dio->end_io(dio->iocb, offset, transferred, |
239 | dio->map_bh.b_private); | |
1da177e4 | 240 | if (dio->lock_type == DIO_LOCKING) |
d8aa905b IM |
241 | /* lockdep: non-owner release */ |
242 | up_read_non_owner(&dio->inode->i_alloc_sem); | |
6d544bb4 ZB |
243 | |
244 | if (ret == 0) | |
245 | ret = dio->page_errors; | |
246 | if (ret == 0) | |
247 | ret = dio->io_error; | |
248 | if (ret == 0) | |
249 | ret = transferred; | |
250 | ||
251 | return ret; | |
1da177e4 LT |
252 | } |
253 | ||
254 | /* | |
255 | * Called when a BIO has been processed. If the count goes to zero then IO is | |
256 | * complete and we can signal this to the AIO layer. | |
257 | */ | |
0273201e | 258 | static void dio_complete_aio(struct dio *dio) |
1da177e4 LT |
259 | { |
260 | unsigned long flags; | |
0273201e | 261 | int ret; |
1da177e4 | 262 | |
0273201e | 263 | ret = dio_complete(dio, dio->iocb->ki_pos, 0); |
29504ff3 | 264 | |
0273201e ZB |
265 | /* Complete AIO later if falling back to buffered i/o */ |
266 | if (dio->result == dio->size || | |
267 | ((dio->rw == READ) && dio->result)) { | |
268 | aio_complete(dio->iocb, ret, 0); | |
269 | kfree(dio); | |
270 | } else { | |
271 | /* | |
272 | * Falling back to buffered | |
273 | */ | |
274 | spin_lock_irqsave(&dio->bio_lock, flags); | |
275 | if (dio->waiter) | |
276 | wake_up_process(dio->waiter); | |
277 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
1da177e4 | 278 | } |
1da177e4 LT |
279 | } |
280 | ||
281 | static int dio_bio_complete(struct dio *dio, struct bio *bio); | |
282 | /* | |
283 | * Asynchronous IO callback. | |
284 | */ | |
285 | static int dio_bio_end_aio(struct bio *bio, unsigned int bytes_done, int error) | |
286 | { | |
287 | struct dio *dio = bio->bi_private; | |
288 | ||
289 | if (bio->bi_size) | |
290 | return 1; | |
291 | ||
292 | /* cleanup the bio */ | |
293 | dio_bio_complete(dio, bio); | |
0273201e ZB |
294 | |
295 | if (atomic_dec_and_test(&dio->refcount)) | |
296 | dio_complete_aio(dio); | |
297 | ||
1da177e4 LT |
298 | return 0; |
299 | } | |
300 | ||
301 | /* | |
302 | * The BIO completion handler simply queues the BIO up for the process-context | |
303 | * handler. | |
304 | * | |
305 | * During I/O bi_private points at the dio. After I/O, bi_private is used to | |
306 | * implement a singly-linked list of completed BIOs, at dio->bio_list. | |
307 | */ | |
308 | static int dio_bio_end_io(struct bio *bio, unsigned int bytes_done, int error) | |
309 | { | |
310 | struct dio *dio = bio->bi_private; | |
311 | unsigned long flags; | |
312 | ||
313 | if (bio->bi_size) | |
314 | return 1; | |
315 | ||
316 | spin_lock_irqsave(&dio->bio_lock, flags); | |
317 | bio->bi_private = dio->bio_list; | |
318 | dio->bio_list = bio; | |
0273201e | 319 | if ((atomic_sub_return(1, &dio->refcount) == 1) && dio->waiter) |
1da177e4 LT |
320 | wake_up_process(dio->waiter); |
321 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
322 | return 0; | |
323 | } | |
324 | ||
325 | static int | |
326 | dio_bio_alloc(struct dio *dio, struct block_device *bdev, | |
327 | sector_t first_sector, int nr_vecs) | |
328 | { | |
329 | struct bio *bio; | |
330 | ||
331 | bio = bio_alloc(GFP_KERNEL, nr_vecs); | |
332 | if (bio == NULL) | |
333 | return -ENOMEM; | |
334 | ||
335 | bio->bi_bdev = bdev; | |
336 | bio->bi_sector = first_sector; | |
337 | if (dio->is_async) | |
338 | bio->bi_end_io = dio_bio_end_aio; | |
339 | else | |
340 | bio->bi_end_io = dio_bio_end_io; | |
341 | ||
342 | dio->bio = bio; | |
343 | return 0; | |
344 | } | |
345 | ||
346 | /* | |
347 | * In the AIO read case we speculatively dirty the pages before starting IO. | |
348 | * During IO completion, any of these pages which happen to have been written | |
349 | * back will be redirtied by bio_check_pages_dirty(). | |
0273201e ZB |
350 | * |
351 | * bios hold a dio reference between submit_bio and ->end_io. | |
1da177e4 LT |
352 | */ |
353 | static void dio_bio_submit(struct dio *dio) | |
354 | { | |
355 | struct bio *bio = dio->bio; | |
1da177e4 LT |
356 | |
357 | bio->bi_private = dio; | |
0273201e | 358 | atomic_inc(&dio->refcount); |
1da177e4 LT |
359 | if (dio->is_async && dio->rw == READ) |
360 | bio_set_pages_dirty(bio); | |
361 | submit_bio(dio->rw, bio); | |
362 | ||
363 | dio->bio = NULL; | |
364 | dio->boundary = 0; | |
365 | } | |
366 | ||
367 | /* | |
368 | * Release any resources in case of a failure | |
369 | */ | |
370 | static void dio_cleanup(struct dio *dio) | |
371 | { | |
372 | while (dio_pages_present(dio)) | |
373 | page_cache_release(dio_get_page(dio)); | |
374 | } | |
375 | ||
0273201e ZB |
376 | static int wait_for_more_bios(struct dio *dio) |
377 | { | |
378 | assert_spin_locked(&dio->bio_lock); | |
379 | ||
380 | return (atomic_read(&dio->refcount) > 1) && (dio->bio_list == NULL); | |
381 | } | |
382 | ||
1da177e4 | 383 | /* |
0273201e ZB |
384 | * Wait for the next BIO to complete. Remove it and return it. NULL is |
385 | * returned once all BIOs have been completed. This must only be called once | |
386 | * all bios have been issued so that dio->refcount can only decrease. This | |
387 | * requires that that the caller hold a reference on the dio. | |
1da177e4 LT |
388 | */ |
389 | static struct bio *dio_await_one(struct dio *dio) | |
390 | { | |
391 | unsigned long flags; | |
0273201e | 392 | struct bio *bio = NULL; |
1da177e4 LT |
393 | |
394 | spin_lock_irqsave(&dio->bio_lock, flags); | |
0273201e | 395 | while (wait_for_more_bios(dio)) { |
1da177e4 | 396 | set_current_state(TASK_UNINTERRUPTIBLE); |
0273201e | 397 | if (wait_for_more_bios(dio)) { |
1da177e4 LT |
398 | dio->waiter = current; |
399 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
1da177e4 LT |
400 | io_schedule(); |
401 | spin_lock_irqsave(&dio->bio_lock, flags); | |
402 | dio->waiter = NULL; | |
403 | } | |
404 | set_current_state(TASK_RUNNING); | |
405 | } | |
0273201e ZB |
406 | if (dio->bio_list) { |
407 | bio = dio->bio_list; | |
408 | dio->bio_list = bio->bi_private; | |
409 | } | |
1da177e4 LT |
410 | spin_unlock_irqrestore(&dio->bio_lock, flags); |
411 | return bio; | |
412 | } | |
413 | ||
414 | /* | |
415 | * Process one completed BIO. No locks are held. | |
416 | */ | |
417 | static int dio_bio_complete(struct dio *dio, struct bio *bio) | |
418 | { | |
419 | const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); | |
420 | struct bio_vec *bvec = bio->bi_io_vec; | |
421 | int page_no; | |
422 | ||
423 | if (!uptodate) | |
174e27c6 | 424 | dio->io_error = -EIO; |
1da177e4 LT |
425 | |
426 | if (dio->is_async && dio->rw == READ) { | |
427 | bio_check_pages_dirty(bio); /* transfers ownership */ | |
428 | } else { | |
429 | for (page_no = 0; page_no < bio->bi_vcnt; page_no++) { | |
430 | struct page *page = bvec[page_no].bv_page; | |
431 | ||
432 | if (dio->rw == READ && !PageCompound(page)) | |
433 | set_page_dirty_lock(page); | |
434 | page_cache_release(page); | |
435 | } | |
436 | bio_put(bio); | |
437 | } | |
1da177e4 LT |
438 | return uptodate ? 0 : -EIO; |
439 | } | |
440 | ||
441 | /* | |
0273201e ZB |
442 | * Wait on and process all in-flight BIOs. This must only be called once |
443 | * all bios have been issued so that the refcount can only decrease. | |
444 | * This just waits for all bios to make it through dio_bio_complete. IO | |
445 | * errors are propogated through dio->io_error and should be propogated via | |
446 | * dio_complete(). | |
1da177e4 | 447 | */ |
6d544bb4 | 448 | static void dio_await_completion(struct dio *dio) |
1da177e4 | 449 | { |
0273201e ZB |
450 | struct bio *bio; |
451 | do { | |
452 | bio = dio_await_one(dio); | |
453 | if (bio) | |
454 | dio_bio_complete(dio, bio); | |
455 | } while (bio); | |
1da177e4 LT |
456 | } |
457 | ||
458 | /* | |
459 | * A really large O_DIRECT read or write can generate a lot of BIOs. So | |
460 | * to keep the memory consumption sane we periodically reap any completed BIOs | |
461 | * during the BIO generation phase. | |
462 | * | |
463 | * This also helps to limit the peak amount of pinned userspace memory. | |
464 | */ | |
465 | static int dio_bio_reap(struct dio *dio) | |
466 | { | |
467 | int ret = 0; | |
468 | ||
469 | if (dio->reap_counter++ >= 64) { | |
470 | while (dio->bio_list) { | |
471 | unsigned long flags; | |
472 | struct bio *bio; | |
473 | int ret2; | |
474 | ||
475 | spin_lock_irqsave(&dio->bio_lock, flags); | |
476 | bio = dio->bio_list; | |
477 | dio->bio_list = bio->bi_private; | |
478 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
479 | ret2 = dio_bio_complete(dio, bio); | |
480 | if (ret == 0) | |
481 | ret = ret2; | |
482 | } | |
483 | dio->reap_counter = 0; | |
484 | } | |
485 | return ret; | |
486 | } | |
487 | ||
488 | /* | |
489 | * Call into the fs to map some more disk blocks. We record the current number | |
490 | * of available blocks at dio->blocks_available. These are in units of the | |
491 | * fs blocksize, (1 << inode->i_blkbits). | |
492 | * | |
493 | * The fs is allowed to map lots of blocks at once. If it wants to do that, | |
494 | * it uses the passed inode-relative block number as the file offset, as usual. | |
495 | * | |
1d8fa7a2 | 496 | * get_block() is passed the number of i_blkbits-sized blocks which direct_io |
1da177e4 LT |
497 | * has remaining to do. The fs should not map more than this number of blocks. |
498 | * | |
499 | * If the fs has mapped a lot of blocks, it should populate bh->b_size to | |
500 | * indicate how much contiguous disk space has been made available at | |
501 | * bh->b_blocknr. | |
502 | * | |
503 | * If *any* of the mapped blocks are new, then the fs must set buffer_new(). | |
504 | * This isn't very efficient... | |
505 | * | |
506 | * In the case of filesystem holes: the fs may return an arbitrarily-large | |
507 | * hole by returning an appropriate value in b_size and by clearing | |
508 | * buffer_mapped(). However the direct-io code will only process holes one | |
1d8fa7a2 | 509 | * block at a time - it will repeatedly call get_block() as it walks the hole. |
1da177e4 LT |
510 | */ |
511 | static int get_more_blocks(struct dio *dio) | |
512 | { | |
513 | int ret; | |
514 | struct buffer_head *map_bh = &dio->map_bh; | |
515 | sector_t fs_startblk; /* Into file, in filesystem-sized blocks */ | |
516 | unsigned long fs_count; /* Number of filesystem-sized blocks */ | |
517 | unsigned long dio_count;/* Number of dio_block-sized blocks */ | |
518 | unsigned long blkmask; | |
519 | int create; | |
520 | ||
521 | /* | |
522 | * If there was a memory error and we've overwritten all the | |
523 | * mapped blocks then we can now return that memory error | |
524 | */ | |
525 | ret = dio->page_errors; | |
526 | if (ret == 0) { | |
1da177e4 LT |
527 | BUG_ON(dio->block_in_file >= dio->final_block_in_request); |
528 | fs_startblk = dio->block_in_file >> dio->blkfactor; | |
529 | dio_count = dio->final_block_in_request - dio->block_in_file; | |
530 | fs_count = dio_count >> dio->blkfactor; | |
531 | blkmask = (1 << dio->blkfactor) - 1; | |
532 | if (dio_count & blkmask) | |
533 | fs_count++; | |
534 | ||
3c674e74 NS |
535 | map_bh->b_state = 0; |
536 | map_bh->b_size = fs_count << dio->inode->i_blkbits; | |
537 | ||
b31dc66a | 538 | create = dio->rw & WRITE; |
1da177e4 LT |
539 | if (dio->lock_type == DIO_LOCKING) { |
540 | if (dio->block_in_file < (i_size_read(dio->inode) >> | |
541 | dio->blkbits)) | |
542 | create = 0; | |
543 | } else if (dio->lock_type == DIO_NO_LOCKING) { | |
544 | create = 0; | |
545 | } | |
3c674e74 | 546 | |
1da177e4 LT |
547 | /* |
548 | * For writes inside i_size we forbid block creations: only | |
549 | * overwrites are permitted. We fall back to buffered writes | |
550 | * at a higher level for inside-i_size block-instantiating | |
551 | * writes. | |
552 | */ | |
1d8fa7a2 | 553 | ret = (*dio->get_block)(dio->inode, fs_startblk, |
1da177e4 LT |
554 | map_bh, create); |
555 | } | |
556 | return ret; | |
557 | } | |
558 | ||
559 | /* | |
560 | * There is no bio. Make one now. | |
561 | */ | |
562 | static int dio_new_bio(struct dio *dio, sector_t start_sector) | |
563 | { | |
564 | sector_t sector; | |
565 | int ret, nr_pages; | |
566 | ||
567 | ret = dio_bio_reap(dio); | |
568 | if (ret) | |
569 | goto out; | |
570 | sector = start_sector << (dio->blkbits - 9); | |
571 | nr_pages = min(dio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev)); | |
572 | BUG_ON(nr_pages <= 0); | |
573 | ret = dio_bio_alloc(dio, dio->map_bh.b_bdev, sector, nr_pages); | |
574 | dio->boundary = 0; | |
575 | out: | |
576 | return ret; | |
577 | } | |
578 | ||
579 | /* | |
580 | * Attempt to put the current chunk of 'cur_page' into the current BIO. If | |
581 | * that was successful then update final_block_in_bio and take a ref against | |
582 | * the just-added page. | |
583 | * | |
584 | * Return zero on success. Non-zero means the caller needs to start a new BIO. | |
585 | */ | |
586 | static int dio_bio_add_page(struct dio *dio) | |
587 | { | |
588 | int ret; | |
589 | ||
590 | ret = bio_add_page(dio->bio, dio->cur_page, | |
591 | dio->cur_page_len, dio->cur_page_offset); | |
592 | if (ret == dio->cur_page_len) { | |
593 | /* | |
594 | * Decrement count only, if we are done with this page | |
595 | */ | |
596 | if ((dio->cur_page_len + dio->cur_page_offset) == PAGE_SIZE) | |
597 | dio->pages_in_io--; | |
598 | page_cache_get(dio->cur_page); | |
599 | dio->final_block_in_bio = dio->cur_page_block + | |
600 | (dio->cur_page_len >> dio->blkbits); | |
601 | ret = 0; | |
602 | } else { | |
603 | ret = 1; | |
604 | } | |
605 | return ret; | |
606 | } | |
607 | ||
608 | /* | |
609 | * Put cur_page under IO. The section of cur_page which is described by | |
610 | * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page | |
611 | * starts on-disk at cur_page_block. | |
612 | * | |
613 | * We take a ref against the page here (on behalf of its presence in the bio). | |
614 | * | |
615 | * The caller of this function is responsible for removing cur_page from the | |
616 | * dio, and for dropping the refcount which came from that presence. | |
617 | */ | |
618 | static int dio_send_cur_page(struct dio *dio) | |
619 | { | |
620 | int ret = 0; | |
621 | ||
622 | if (dio->bio) { | |
623 | /* | |
624 | * See whether this new request is contiguous with the old | |
625 | */ | |
626 | if (dio->final_block_in_bio != dio->cur_page_block) | |
627 | dio_bio_submit(dio); | |
628 | /* | |
629 | * Submit now if the underlying fs is about to perform a | |
630 | * metadata read | |
631 | */ | |
632 | if (dio->boundary) | |
633 | dio_bio_submit(dio); | |
634 | } | |
635 | ||
636 | if (dio->bio == NULL) { | |
637 | ret = dio_new_bio(dio, dio->cur_page_block); | |
638 | if (ret) | |
639 | goto out; | |
640 | } | |
641 | ||
642 | if (dio_bio_add_page(dio) != 0) { | |
643 | dio_bio_submit(dio); | |
644 | ret = dio_new_bio(dio, dio->cur_page_block); | |
645 | if (ret == 0) { | |
646 | ret = dio_bio_add_page(dio); | |
647 | BUG_ON(ret != 0); | |
648 | } | |
649 | } | |
650 | out: | |
651 | return ret; | |
652 | } | |
653 | ||
654 | /* | |
655 | * An autonomous function to put a chunk of a page under deferred IO. | |
656 | * | |
657 | * The caller doesn't actually know (or care) whether this piece of page is in | |
658 | * a BIO, or is under IO or whatever. We just take care of all possible | |
659 | * situations here. The separation between the logic of do_direct_IO() and | |
660 | * that of submit_page_section() is important for clarity. Please don't break. | |
661 | * | |
662 | * The chunk of page starts on-disk at blocknr. | |
663 | * | |
664 | * We perform deferred IO, by recording the last-submitted page inside our | |
665 | * private part of the dio structure. If possible, we just expand the IO | |
666 | * across that page here. | |
667 | * | |
668 | * If that doesn't work out then we put the old page into the bio and add this | |
669 | * page to the dio instead. | |
670 | */ | |
671 | static int | |
672 | submit_page_section(struct dio *dio, struct page *page, | |
673 | unsigned offset, unsigned len, sector_t blocknr) | |
674 | { | |
675 | int ret = 0; | |
676 | ||
98c4d57d AM |
677 | if (dio->rw & WRITE) { |
678 | /* | |
679 | * Read accounting is performed in submit_bio() | |
680 | */ | |
681 | task_io_account_write(len); | |
682 | } | |
683 | ||
1da177e4 LT |
684 | /* |
685 | * Can we just grow the current page's presence in the dio? | |
686 | */ | |
687 | if ( (dio->cur_page == page) && | |
688 | (dio->cur_page_offset + dio->cur_page_len == offset) && | |
689 | (dio->cur_page_block + | |
690 | (dio->cur_page_len >> dio->blkbits) == blocknr)) { | |
691 | dio->cur_page_len += len; | |
692 | ||
693 | /* | |
694 | * If dio->boundary then we want to schedule the IO now to | |
695 | * avoid metadata seeks. | |
696 | */ | |
697 | if (dio->boundary) { | |
698 | ret = dio_send_cur_page(dio); | |
699 | page_cache_release(dio->cur_page); | |
700 | dio->cur_page = NULL; | |
701 | } | |
702 | goto out; | |
703 | } | |
704 | ||
705 | /* | |
706 | * If there's a deferred page already there then send it. | |
707 | */ | |
708 | if (dio->cur_page) { | |
709 | ret = dio_send_cur_page(dio); | |
710 | page_cache_release(dio->cur_page); | |
711 | dio->cur_page = NULL; | |
712 | if (ret) | |
713 | goto out; | |
714 | } | |
715 | ||
716 | page_cache_get(page); /* It is in dio */ | |
717 | dio->cur_page = page; | |
718 | dio->cur_page_offset = offset; | |
719 | dio->cur_page_len = len; | |
720 | dio->cur_page_block = blocknr; | |
721 | out: | |
722 | return ret; | |
723 | } | |
724 | ||
725 | /* | |
726 | * Clean any dirty buffers in the blockdev mapping which alias newly-created | |
727 | * file blocks. Only called for S_ISREG files - blockdevs do not set | |
728 | * buffer_new | |
729 | */ | |
730 | static void clean_blockdev_aliases(struct dio *dio) | |
731 | { | |
732 | unsigned i; | |
733 | unsigned nblocks; | |
734 | ||
735 | nblocks = dio->map_bh.b_size >> dio->inode->i_blkbits; | |
736 | ||
737 | for (i = 0; i < nblocks; i++) { | |
738 | unmap_underlying_metadata(dio->map_bh.b_bdev, | |
739 | dio->map_bh.b_blocknr + i); | |
740 | } | |
741 | } | |
742 | ||
743 | /* | |
744 | * If we are not writing the entire block and get_block() allocated | |
745 | * the block for us, we need to fill-in the unused portion of the | |
746 | * block with zeros. This happens only if user-buffer, fileoffset or | |
747 | * io length is not filesystem block-size multiple. | |
748 | * | |
749 | * `end' is zero if we're doing the start of the IO, 1 at the end of the | |
750 | * IO. | |
751 | */ | |
752 | static void dio_zero_block(struct dio *dio, int end) | |
753 | { | |
754 | unsigned dio_blocks_per_fs_block; | |
755 | unsigned this_chunk_blocks; /* In dio_blocks */ | |
756 | unsigned this_chunk_bytes; | |
757 | struct page *page; | |
758 | ||
759 | dio->start_zero_done = 1; | |
760 | if (!dio->blkfactor || !buffer_new(&dio->map_bh)) | |
761 | return; | |
762 | ||
763 | dio_blocks_per_fs_block = 1 << dio->blkfactor; | |
764 | this_chunk_blocks = dio->block_in_file & (dio_blocks_per_fs_block - 1); | |
765 | ||
766 | if (!this_chunk_blocks) | |
767 | return; | |
768 | ||
769 | /* | |
770 | * We need to zero out part of an fs block. It is either at the | |
771 | * beginning or the end of the fs block. | |
772 | */ | |
773 | if (end) | |
774 | this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks; | |
775 | ||
776 | this_chunk_bytes = this_chunk_blocks << dio->blkbits; | |
777 | ||
778 | page = ZERO_PAGE(dio->curr_user_address); | |
779 | if (submit_page_section(dio, page, 0, this_chunk_bytes, | |
780 | dio->next_block_for_io)) | |
781 | return; | |
782 | ||
783 | dio->next_block_for_io += this_chunk_blocks; | |
784 | } | |
785 | ||
786 | /* | |
787 | * Walk the user pages, and the file, mapping blocks to disk and generating | |
788 | * a sequence of (page,offset,len,block) mappings. These mappings are injected | |
789 | * into submit_page_section(), which takes care of the next stage of submission | |
790 | * | |
791 | * Direct IO against a blockdev is different from a file. Because we can | |
792 | * happily perform page-sized but 512-byte aligned IOs. It is important that | |
793 | * blockdev IO be able to have fine alignment and large sizes. | |
794 | * | |
1d8fa7a2 | 795 | * So what we do is to permit the ->get_block function to populate bh.b_size |
1da177e4 LT |
796 | * with the size of IO which is permitted at this offset and this i_blkbits. |
797 | * | |
798 | * For best results, the blockdev should be set up with 512-byte i_blkbits and | |
1d8fa7a2 | 799 | * it should set b_size to PAGE_SIZE or more inside get_block(). This gives |
1da177e4 LT |
800 | * fine alignment but still allows this function to work in PAGE_SIZE units. |
801 | */ | |
802 | static int do_direct_IO(struct dio *dio) | |
803 | { | |
804 | const unsigned blkbits = dio->blkbits; | |
805 | const unsigned blocks_per_page = PAGE_SIZE >> blkbits; | |
806 | struct page *page; | |
807 | unsigned block_in_page; | |
808 | struct buffer_head *map_bh = &dio->map_bh; | |
809 | int ret = 0; | |
810 | ||
811 | /* The I/O can start at any block offset within the first page */ | |
812 | block_in_page = dio->first_block_in_page; | |
813 | ||
814 | while (dio->block_in_file < dio->final_block_in_request) { | |
815 | page = dio_get_page(dio); | |
816 | if (IS_ERR(page)) { | |
817 | ret = PTR_ERR(page); | |
818 | goto out; | |
819 | } | |
820 | ||
821 | while (block_in_page < blocks_per_page) { | |
822 | unsigned offset_in_page = block_in_page << blkbits; | |
823 | unsigned this_chunk_bytes; /* # of bytes mapped */ | |
824 | unsigned this_chunk_blocks; /* # of blocks */ | |
825 | unsigned u; | |
826 | ||
827 | if (dio->blocks_available == 0) { | |
828 | /* | |
829 | * Need to go and map some more disk | |
830 | */ | |
831 | unsigned long blkmask; | |
832 | unsigned long dio_remainder; | |
833 | ||
834 | ret = get_more_blocks(dio); | |
835 | if (ret) { | |
836 | page_cache_release(page); | |
837 | goto out; | |
838 | } | |
839 | if (!buffer_mapped(map_bh)) | |
840 | goto do_holes; | |
841 | ||
842 | dio->blocks_available = | |
843 | map_bh->b_size >> dio->blkbits; | |
844 | dio->next_block_for_io = | |
845 | map_bh->b_blocknr << dio->blkfactor; | |
846 | if (buffer_new(map_bh)) | |
847 | clean_blockdev_aliases(dio); | |
848 | ||
849 | if (!dio->blkfactor) | |
850 | goto do_holes; | |
851 | ||
852 | blkmask = (1 << dio->blkfactor) - 1; | |
853 | dio_remainder = (dio->block_in_file & blkmask); | |
854 | ||
855 | /* | |
856 | * If we are at the start of IO and that IO | |
857 | * starts partway into a fs-block, | |
858 | * dio_remainder will be non-zero. If the IO | |
859 | * is a read then we can simply advance the IO | |
860 | * cursor to the first block which is to be | |
861 | * read. But if the IO is a write and the | |
862 | * block was newly allocated we cannot do that; | |
863 | * the start of the fs block must be zeroed out | |
864 | * on-disk | |
865 | */ | |
866 | if (!buffer_new(map_bh)) | |
867 | dio->next_block_for_io += dio_remainder; | |
868 | dio->blocks_available -= dio_remainder; | |
869 | } | |
870 | do_holes: | |
871 | /* Handle holes */ | |
872 | if (!buffer_mapped(map_bh)) { | |
873 | char *kaddr; | |
35dc8161 | 874 | loff_t i_size_aligned; |
1da177e4 LT |
875 | |
876 | /* AKPM: eargh, -ENOTBLK is a hack */ | |
b31dc66a | 877 | if (dio->rw & WRITE) { |
1da177e4 LT |
878 | page_cache_release(page); |
879 | return -ENOTBLK; | |
880 | } | |
881 | ||
35dc8161 JM |
882 | /* |
883 | * Be sure to account for a partial block as the | |
884 | * last block in the file | |
885 | */ | |
886 | i_size_aligned = ALIGN(i_size_read(dio->inode), | |
887 | 1 << blkbits); | |
1da177e4 | 888 | if (dio->block_in_file >= |
35dc8161 | 889 | i_size_aligned >> blkbits) { |
1da177e4 LT |
890 | /* We hit eof */ |
891 | page_cache_release(page); | |
892 | goto out; | |
893 | } | |
894 | kaddr = kmap_atomic(page, KM_USER0); | |
895 | memset(kaddr + (block_in_page << blkbits), | |
896 | 0, 1 << blkbits); | |
897 | flush_dcache_page(page); | |
898 | kunmap_atomic(kaddr, KM_USER0); | |
899 | dio->block_in_file++; | |
900 | block_in_page++; | |
901 | goto next_block; | |
902 | } | |
903 | ||
904 | /* | |
905 | * If we're performing IO which has an alignment which | |
906 | * is finer than the underlying fs, go check to see if | |
907 | * we must zero out the start of this block. | |
908 | */ | |
909 | if (unlikely(dio->blkfactor && !dio->start_zero_done)) | |
910 | dio_zero_block(dio, 0); | |
911 | ||
912 | /* | |
913 | * Work out, in this_chunk_blocks, how much disk we | |
914 | * can add to this page | |
915 | */ | |
916 | this_chunk_blocks = dio->blocks_available; | |
917 | u = (PAGE_SIZE - offset_in_page) >> blkbits; | |
918 | if (this_chunk_blocks > u) | |
919 | this_chunk_blocks = u; | |
920 | u = dio->final_block_in_request - dio->block_in_file; | |
921 | if (this_chunk_blocks > u) | |
922 | this_chunk_blocks = u; | |
923 | this_chunk_bytes = this_chunk_blocks << blkbits; | |
924 | BUG_ON(this_chunk_bytes == 0); | |
925 | ||
926 | dio->boundary = buffer_boundary(map_bh); | |
927 | ret = submit_page_section(dio, page, offset_in_page, | |
928 | this_chunk_bytes, dio->next_block_for_io); | |
929 | if (ret) { | |
930 | page_cache_release(page); | |
931 | goto out; | |
932 | } | |
933 | dio->next_block_for_io += this_chunk_blocks; | |
934 | ||
935 | dio->block_in_file += this_chunk_blocks; | |
936 | block_in_page += this_chunk_blocks; | |
937 | dio->blocks_available -= this_chunk_blocks; | |
938 | next_block: | |
d4569d2e | 939 | BUG_ON(dio->block_in_file > dio->final_block_in_request); |
1da177e4 LT |
940 | if (dio->block_in_file == dio->final_block_in_request) |
941 | break; | |
942 | } | |
943 | ||
944 | /* Drop the ref which was taken in get_user_pages() */ | |
945 | page_cache_release(page); | |
946 | block_in_page = 0; | |
947 | } | |
948 | out: | |
949 | return ret; | |
950 | } | |
951 | ||
952 | /* | |
1b1dcc1b | 953 | * Releases both i_mutex and i_alloc_sem |
1da177e4 LT |
954 | */ |
955 | static ssize_t | |
956 | direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode, | |
957 | const struct iovec *iov, loff_t offset, unsigned long nr_segs, | |
1d8fa7a2 | 958 | unsigned blkbits, get_block_t get_block, dio_iodone_t end_io, |
1da177e4 LT |
959 | struct dio *dio) |
960 | { | |
961 | unsigned long user_addr; | |
962 | int seg; | |
963 | ssize_t ret = 0; | |
964 | ssize_t ret2; | |
965 | size_t bytes; | |
966 | ||
967 | dio->bio = NULL; | |
968 | dio->inode = inode; | |
969 | dio->rw = rw; | |
970 | dio->blkbits = blkbits; | |
971 | dio->blkfactor = inode->i_blkbits - blkbits; | |
972 | dio->start_zero_done = 0; | |
973 | dio->size = 0; | |
974 | dio->block_in_file = offset >> blkbits; | |
975 | dio->blocks_available = 0; | |
976 | dio->cur_page = NULL; | |
977 | ||
978 | dio->boundary = 0; | |
979 | dio->reap_counter = 0; | |
1d8fa7a2 | 980 | dio->get_block = get_block; |
1da177e4 LT |
981 | dio->end_io = end_io; |
982 | dio->map_bh.b_private = NULL; | |
983 | dio->final_block_in_bio = -1; | |
984 | dio->next_block_for_io = -1; | |
985 | ||
986 | dio->page_errors = 0; | |
174e27c6 | 987 | dio->io_error = 0; |
1da177e4 LT |
988 | dio->result = 0; |
989 | dio->iocb = iocb; | |
29504ff3 | 990 | dio->i_size = i_size_read(inode); |
1da177e4 | 991 | |
0273201e | 992 | atomic_set(&dio->refcount, 1); |
1da177e4 LT |
993 | spin_lock_init(&dio->bio_lock); |
994 | dio->bio_list = NULL; | |
995 | dio->waiter = NULL; | |
996 | ||
997 | /* | |
998 | * In case of non-aligned buffers, we may need 2 more | |
999 | * pages since we need to zero out first and last block. | |
1000 | */ | |
1001 | if (unlikely(dio->blkfactor)) | |
1002 | dio->pages_in_io = 2; | |
1003 | else | |
1004 | dio->pages_in_io = 0; | |
1005 | ||
1006 | for (seg = 0; seg < nr_segs; seg++) { | |
1007 | user_addr = (unsigned long)iov[seg].iov_base; | |
1008 | dio->pages_in_io += | |
1009 | ((user_addr+iov[seg].iov_len +PAGE_SIZE-1)/PAGE_SIZE | |
1010 | - user_addr/PAGE_SIZE); | |
1011 | } | |
1012 | ||
1013 | for (seg = 0; seg < nr_segs; seg++) { | |
1014 | user_addr = (unsigned long)iov[seg].iov_base; | |
1015 | dio->size += bytes = iov[seg].iov_len; | |
1016 | ||
1017 | /* Index into the first page of the first block */ | |
1018 | dio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits; | |
1019 | dio->final_block_in_request = dio->block_in_file + | |
1020 | (bytes >> blkbits); | |
1021 | /* Page fetching state */ | |
1022 | dio->head = 0; | |
1023 | dio->tail = 0; | |
1024 | dio->curr_page = 0; | |
1025 | ||
1026 | dio->total_pages = 0; | |
1027 | if (user_addr & (PAGE_SIZE-1)) { | |
1028 | dio->total_pages++; | |
1029 | bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1)); | |
1030 | } | |
1031 | dio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE; | |
1032 | dio->curr_user_address = user_addr; | |
1033 | ||
1034 | ret = do_direct_IO(dio); | |
1035 | ||
1036 | dio->result += iov[seg].iov_len - | |
1037 | ((dio->final_block_in_request - dio->block_in_file) << | |
1038 | blkbits); | |
1039 | ||
1040 | if (ret) { | |
1041 | dio_cleanup(dio); | |
1042 | break; | |
1043 | } | |
1044 | } /* end iovec loop */ | |
1045 | ||
b31dc66a | 1046 | if (ret == -ENOTBLK && (rw & WRITE)) { |
1da177e4 LT |
1047 | /* |
1048 | * The remaining part of the request will be | |
1049 | * be handled by buffered I/O when we return | |
1050 | */ | |
1051 | ret = 0; | |
1052 | } | |
1053 | /* | |
1054 | * There may be some unwritten disk at the end of a part-written | |
1055 | * fs-block-sized block. Go zero that now. | |
1056 | */ | |
1057 | dio_zero_block(dio, 1); | |
1058 | ||
1059 | if (dio->cur_page) { | |
1060 | ret2 = dio_send_cur_page(dio); | |
1061 | if (ret == 0) | |
1062 | ret = ret2; | |
1063 | page_cache_release(dio->cur_page); | |
1064 | dio->cur_page = NULL; | |
1065 | } | |
1066 | if (dio->bio) | |
1067 | dio_bio_submit(dio); | |
1068 | ||
17a7b1d7 ZB |
1069 | /* All IO is now issued, send it on its way */ |
1070 | blk_run_address_space(inode->i_mapping); | |
1071 | ||
1da177e4 LT |
1072 | /* |
1073 | * It is possible that, we return short IO due to end of file. | |
1074 | * In that case, we need to release all the pages we got hold on. | |
1075 | */ | |
1076 | dio_cleanup(dio); | |
1077 | ||
1078 | /* | |
1079 | * All block lookups have been performed. For READ requests | |
1b1dcc1b | 1080 | * we can let i_mutex go now that its achieved its purpose |
1da177e4 LT |
1081 | * of protecting us from looking up uninitialized blocks. |
1082 | */ | |
1083 | if ((rw == READ) && (dio->lock_type == DIO_LOCKING)) | |
1b1dcc1b | 1084 | mutex_unlock(&dio->inode->i_mutex); |
1da177e4 LT |
1085 | |
1086 | /* | |
1087 | * OK, all BIOs are submitted, so we can decrement bio_count to truly | |
1088 | * reflect the number of to-be-processed BIOs. | |
1089 | */ | |
1090 | if (dio->is_async) { | |
1091 | int should_wait = 0; | |
1092 | ||
b31dc66a | 1093 | if (dio->result < dio->size && (rw & WRITE)) { |
1da177e4 LT |
1094 | dio->waiter = current; |
1095 | should_wait = 1; | |
1096 | } | |
1097 | if (ret == 0) | |
1098 | ret = dio->result; | |
0273201e ZB |
1099 | |
1100 | /* this can free the dio */ | |
1101 | if (atomic_dec_and_test(&dio->refcount)) | |
1102 | dio_complete_aio(dio); | |
1103 | ||
1da177e4 LT |
1104 | if (should_wait) { |
1105 | unsigned long flags; | |
1106 | /* | |
1107 | * Wait for already issued I/O to drain out and | |
1108 | * release its references to user-space pages | |
1109 | * before returning to fallback on buffered I/O | |
1110 | */ | |
1111 | ||
1112 | spin_lock_irqsave(&dio->bio_lock, flags); | |
1113 | set_current_state(TASK_UNINTERRUPTIBLE); | |
0273201e | 1114 | while (atomic_read(&dio->refcount)) { |
1da177e4 LT |
1115 | spin_unlock_irqrestore(&dio->bio_lock, flags); |
1116 | io_schedule(); | |
1117 | spin_lock_irqsave(&dio->bio_lock, flags); | |
1118 | set_current_state(TASK_UNINTERRUPTIBLE); | |
1119 | } | |
1120 | spin_unlock_irqrestore(&dio->bio_lock, flags); | |
1121 | set_current_state(TASK_RUNNING); | |
1122 | kfree(dio); | |
1123 | } | |
1124 | } else { | |
6d544bb4 | 1125 | dio_await_completion(dio); |
1da177e4 | 1126 | |
6d544bb4 | 1127 | ret = dio_complete(dio, offset, ret); |
1da177e4 LT |
1128 | |
1129 | /* We could have also come here on an AIO file extend */ | |
b31dc66a | 1130 | if (!is_sync_kiocb(iocb) && (rw & WRITE) && |
1da177e4 LT |
1131 | ret >= 0 && dio->result == dio->size) |
1132 | /* | |
1133 | * For AIO writes where we have completed the | |
1134 | * i/o, we have to mark the the aio complete. | |
1135 | */ | |
1136 | aio_complete(iocb, ret, 0); | |
0273201e ZB |
1137 | |
1138 | if (atomic_dec_and_test(&dio->refcount)) | |
1139 | kfree(dio); | |
1140 | else | |
1141 | BUG(); | |
1da177e4 LT |
1142 | } |
1143 | return ret; | |
1144 | } | |
1145 | ||
1146 | /* | |
1147 | * This is a library function for use by filesystem drivers. | |
1148 | * The locking rules are governed by the dio_lock_type parameter. | |
1149 | * | |
1150 | * DIO_NO_LOCKING (no locking, for raw block device access) | |
1b1dcc1b | 1151 | * For writes, i_mutex is not held on entry; it is never taken. |
1da177e4 LT |
1152 | * |
1153 | * DIO_LOCKING (simple locking for regular files) | |
3fb962bd NS |
1154 | * For writes we are called under i_mutex and return with i_mutex held, even |
1155 | * though it is internally dropped. | |
1b1dcc1b | 1156 | * For reads, i_mutex is not held on entry, but it is taken and dropped before |
1da177e4 LT |
1157 | * returning. |
1158 | * | |
1159 | * DIO_OWN_LOCKING (filesystem provides synchronisation and handling of | |
1160 | * uninitialised data, allowing parallel direct readers and writers) | |
1b1dcc1b | 1161 | * For writes we are called without i_mutex, return without it, never touch it. |
3fb962bd NS |
1162 | * For reads we are called under i_mutex and return with i_mutex held, even |
1163 | * though it may be internally dropped. | |
1da177e4 LT |
1164 | * |
1165 | * Additional i_alloc_sem locking requirements described inline below. | |
1166 | */ | |
1167 | ssize_t | |
1168 | __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode, | |
1169 | struct block_device *bdev, const struct iovec *iov, loff_t offset, | |
1d8fa7a2 | 1170 | unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io, |
1da177e4 LT |
1171 | int dio_lock_type) |
1172 | { | |
1173 | int seg; | |
1174 | size_t size; | |
1175 | unsigned long addr; | |
1176 | unsigned blkbits = inode->i_blkbits; | |
1177 | unsigned bdev_blkbits = 0; | |
1178 | unsigned blocksize_mask = (1 << blkbits) - 1; | |
1179 | ssize_t retval = -EINVAL; | |
1180 | loff_t end = offset; | |
1181 | struct dio *dio; | |
3fb962bd NS |
1182 | int release_i_mutex = 0; |
1183 | int acquire_i_mutex = 0; | |
1da177e4 LT |
1184 | |
1185 | if (rw & WRITE) | |
b31dc66a | 1186 | rw = WRITE_SYNC; |
1da177e4 LT |
1187 | |
1188 | if (bdev) | |
1189 | bdev_blkbits = blksize_bits(bdev_hardsect_size(bdev)); | |
1190 | ||
1191 | if (offset & blocksize_mask) { | |
1192 | if (bdev) | |
1193 | blkbits = bdev_blkbits; | |
1194 | blocksize_mask = (1 << blkbits) - 1; | |
1195 | if (offset & blocksize_mask) | |
1196 | goto out; | |
1197 | } | |
1198 | ||
1199 | /* Check the memory alignment. Blocks cannot straddle pages */ | |
1200 | for (seg = 0; seg < nr_segs; seg++) { | |
1201 | addr = (unsigned long)iov[seg].iov_base; | |
1202 | size = iov[seg].iov_len; | |
1203 | end += size; | |
1204 | if ((addr & blocksize_mask) || (size & blocksize_mask)) { | |
1205 | if (bdev) | |
1206 | blkbits = bdev_blkbits; | |
1207 | blocksize_mask = (1 << blkbits) - 1; | |
1208 | if ((addr & blocksize_mask) || (size & blocksize_mask)) | |
1209 | goto out; | |
1210 | } | |
1211 | } | |
1212 | ||
1213 | dio = kmalloc(sizeof(*dio), GFP_KERNEL); | |
1214 | retval = -ENOMEM; | |
1215 | if (!dio) | |
1216 | goto out; | |
1217 | ||
1218 | /* | |
1219 | * For block device access DIO_NO_LOCKING is used, | |
1220 | * neither readers nor writers do any locking at all | |
1221 | * For regular files using DIO_LOCKING, | |
1b1dcc1b JS |
1222 | * readers need to grab i_mutex and i_alloc_sem |
1223 | * writers need to grab i_alloc_sem only (i_mutex is already held) | |
1da177e4 LT |
1224 | * For regular files using DIO_OWN_LOCKING, |
1225 | * neither readers nor writers take any locks here | |
1da177e4 LT |
1226 | */ |
1227 | dio->lock_type = dio_lock_type; | |
1228 | if (dio_lock_type != DIO_NO_LOCKING) { | |
1229 | /* watch out for a 0 len io from a tricksy fs */ | |
1230 | if (rw == READ && end > offset) { | |
1231 | struct address_space *mapping; | |
1232 | ||
1233 | mapping = iocb->ki_filp->f_mapping; | |
1234 | if (dio_lock_type != DIO_OWN_LOCKING) { | |
1b1dcc1b | 1235 | mutex_lock(&inode->i_mutex); |
3fb962bd | 1236 | release_i_mutex = 1; |
1da177e4 LT |
1237 | } |
1238 | ||
1239 | retval = filemap_write_and_wait_range(mapping, offset, | |
1240 | end - 1); | |
1241 | if (retval) { | |
1242 | kfree(dio); | |
1243 | goto out; | |
1244 | } | |
1245 | ||
1246 | if (dio_lock_type == DIO_OWN_LOCKING) { | |
1b1dcc1b | 1247 | mutex_unlock(&inode->i_mutex); |
3fb962bd | 1248 | acquire_i_mutex = 1; |
1da177e4 LT |
1249 | } |
1250 | } | |
1251 | ||
1252 | if (dio_lock_type == DIO_LOCKING) | |
d8aa905b IM |
1253 | /* lockdep: not the owner will release it */ |
1254 | down_read_non_owner(&inode->i_alloc_sem); | |
1da177e4 LT |
1255 | } |
1256 | ||
1257 | /* | |
1258 | * For file extending writes updating i_size before data | |
1259 | * writeouts complete can expose uninitialized blocks. So | |
1260 | * even for AIO, we need to wait for i/o to complete before | |
1261 | * returning in this case. | |
1262 | */ | |
b31dc66a | 1263 | dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) && |
1da177e4 LT |
1264 | (end > i_size_read(inode))); |
1265 | ||
1266 | retval = direct_io_worker(rw, iocb, inode, iov, offset, | |
1d8fa7a2 | 1267 | nr_segs, blkbits, get_block, end_io, dio); |
1da177e4 LT |
1268 | |
1269 | if (rw == READ && dio_lock_type == DIO_LOCKING) | |
3fb962bd | 1270 | release_i_mutex = 0; |
1da177e4 LT |
1271 | |
1272 | out: | |
3fb962bd | 1273 | if (release_i_mutex) |
1b1dcc1b | 1274 | mutex_unlock(&inode->i_mutex); |
3fb962bd NS |
1275 | else if (acquire_i_mutex) |
1276 | mutex_lock(&inode->i_mutex); | |
1da177e4 LT |
1277 | return retval; |
1278 | } | |
1279 | EXPORT_SYMBOL(__blockdev_direct_IO); |