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