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() 48 * Flags for dio_complete()
51 */ 49 */
52 #define DIO_COMPLETE_ASYNC 0x01 50 #define DIO_COMPLETE_ASYNC 0x01 /* This is async IO */
53 #define DIO_COMPLETE_INVALIDATE 0x02 51 #define DIO_COMPLETE_INVALIDATE 0x02 /* Can invalidate pages */
54 52
55 /* 53 /*
56 * This code generally works in units of "dio_ 54 * This code generally works in units of "dio_blocks". A dio_block is
57 * somewhere between the hard sector size and 55 * somewhere between the hard sector size and the filesystem block size. it
58 * is determined on a per-invocation basis. 56 * is determined on a per-invocation basis. When talking to the filesystem
59 * we need to convert dio_blocks to fs_blocks 57 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
60 * down by dio->blkfactor. Similarly, fs-bloc 58 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
61 * to bio_block quantities by shifting left by 59 * to bio_block quantities by shifting left by blkfactor.
62 * 60 *
63 * If blkfactor is zero then the user's reques 61 * If blkfactor is zero then the user's request was aligned to the filesystem's
64 * blocksize. 62 * blocksize.
65 */ 63 */
66 64
67 /* dio_state only used in the submission path 65 /* dio_state only used in the submission path */
68 66
69 struct dio_submit { 67 struct dio_submit {
70 struct bio *bio; /* bio 68 struct bio *bio; /* bio under assembly */
71 unsigned blkbits; /* doe 69 unsigned blkbits; /* doesn't change */
72 unsigned blkfactor; /* Whe 70 unsigned blkfactor; /* When we're using an alignment which
73 is 71 is finer than the filesystem's soft
74 blo 72 blocksize, this specifies how much
75 fin 73 finer. blkfactor=2 means 1/4-block
76 ali 74 alignment. Does not change */
77 unsigned start_zero_done; /* fla 75 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
78 bee 76 been performed at the start of a
79 wri 77 write */
80 int pages_in_io; /* app 78 int pages_in_io; /* approximate total IO pages */
81 sector_t block_in_file; /* Cur 79 sector_t block_in_file; /* Current offset into the underlying
82 fil 80 file in dio_block units. */
83 unsigned blocks_available; /* At 81 unsigned blocks_available; /* At block_in_file. changes */
84 int reap_counter; /* rat 82 int reap_counter; /* rate limit reaping */
85 sector_t final_block_in_request;/* doe 83 sector_t final_block_in_request;/* doesn't change */
86 int boundary; /* pre 84 int boundary; /* prev block is at a boundary */
87 get_block_t *get_block; /* blo 85 get_block_t *get_block; /* block mapping function */
>> 86 dio_submit_t *submit_io; /* IO submition function */
88 87
89 loff_t logical_offset_in_bio; /* cur 88 loff_t logical_offset_in_bio; /* current first logical block in bio */
90 sector_t final_block_in_bio; /* cur 89 sector_t final_block_in_bio; /* current final block in bio + 1 */
91 sector_t next_block_for_io; /* nex 90 sector_t next_block_for_io; /* next block to be put under IO,
92 in 91 in dio_blocks units */
93 92
94 /* 93 /*
95 * Deferred addition of a page to the 94 * Deferred addition of a page to the dio. These variables are
96 * private to dio_send_cur_page(), sub 95 * private to dio_send_cur_page(), submit_page_section() and
97 * dio_bio_add_page(). 96 * dio_bio_add_page().
98 */ 97 */
99 struct page *cur_page; /* The 98 struct page *cur_page; /* The page */
100 unsigned cur_page_offset; /* Off 99 unsigned cur_page_offset; /* Offset into it, in bytes */
101 unsigned cur_page_len; /* Nr 100 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
102 sector_t cur_page_block; /* Whe 101 sector_t cur_page_block; /* Where it starts */
103 loff_t cur_page_fs_offset; /* Off 102 loff_t cur_page_fs_offset; /* Offset in file */
104 103
105 struct iov_iter *iter; 104 struct iov_iter *iter;
106 /* 105 /*
107 * Page queue. These variables belong 106 * Page queue. These variables belong to dio_refill_pages() and
108 * dio_get_page(). 107 * dio_get_page().
109 */ 108 */
110 unsigned head; /* nex 109 unsigned head; /* next page to process */
111 unsigned tail; /* las 110 unsigned tail; /* last valid page + 1 */
112 size_t from, to; 111 size_t from, to;
113 }; 112 };
114 113
115 /* dio_state communicated between submission p 114 /* dio_state communicated between submission path and end_io */
116 struct dio { 115 struct dio {
117 int flags; /* doe 116 int flags; /* doesn't change */
118 blk_opf_t opf; /* req !! 117 int op;
>> 118 int op_flags;
>> 119 blk_qc_t bio_cookie;
119 struct gendisk *bio_disk; 120 struct gendisk *bio_disk;
120 struct inode *inode; 121 struct inode *inode;
121 loff_t i_size; /* i_s 122 loff_t i_size; /* i_size when submitted */
122 dio_iodone_t *end_io; /* IO 123 dio_iodone_t *end_io; /* IO completion function */
123 bool is_pinned; /* T i <<
124 124
125 void *private; /* cop 125 void *private; /* copy from map_bh.b_private */
126 126
127 /* BIO completion state */ 127 /* BIO completion state */
128 spinlock_t bio_lock; /* pro 128 spinlock_t bio_lock; /* protects BIO fields below */
129 int page_errors; /* err !! 129 int page_errors; /* errno from get_user_pages() */
130 int is_async; /* is 130 int is_async; /* is IO async ? */
131 bool defer_completion; /* def 131 bool defer_completion; /* defer AIO completion to workqueue? */
132 bool should_dirty; /* if 132 bool should_dirty; /* if pages should be dirtied */
133 int io_error; /* IO 133 int io_error; /* IO error in completion path */
134 unsigned long refcount; /* dir 134 unsigned long refcount; /* direct_io_worker() and bios */
135 struct bio *bio_list; /* sin 135 struct bio *bio_list; /* singly linked via bi_private */
136 struct task_struct *waiter; /* wai 136 struct task_struct *waiter; /* waiting task (NULL if none) */
137 137
138 /* AIO related stuff */ 138 /* AIO related stuff */
139 struct kiocb *iocb; /* kio 139 struct kiocb *iocb; /* kiocb */
140 ssize_t result; /* IO 140 ssize_t result; /* IO result */
141 141
142 /* 142 /*
143 * pages[] (and any fields placed afte 143 * pages[] (and any fields placed after it) are not zeroed out at
144 * allocation time. Don't add new fie 144 * allocation time. Don't add new fields after pages[] unless you
145 * wish that they not be zeroed. 145 * wish that they not be zeroed.
146 */ 146 */
147 union { 147 union {
148 struct page *pages[DIO_PAGES]; 148 struct page *pages[DIO_PAGES]; /* page buffer */
149 struct work_struct complete_wo 149 struct work_struct complete_work;/* deferred AIO completion */
150 }; 150 };
151 } ____cacheline_aligned_in_smp; 151 } ____cacheline_aligned_in_smp;
152 152
153 static struct kmem_cache *dio_cache __ro_after !! 153 static struct kmem_cache *dio_cache __read_mostly;
154 154
155 /* 155 /*
156 * How many pages are in the queue? 156 * How many pages are in the queue?
157 */ 157 */
158 static inline unsigned dio_pages_present(struc 158 static inline unsigned dio_pages_present(struct dio_submit *sdio)
159 { 159 {
160 return sdio->tail - sdio->head; 160 return sdio->tail - sdio->head;
161 } 161 }
162 162
163 /* 163 /*
164 * Go grab and pin some userspace pages. Typ 164 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
165 */ 165 */
166 static inline int dio_refill_pages(struct dio 166 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
167 { 167 {
168 struct page **pages = dio->pages; <<
169 const enum req_op dio_op = dio->opf & <<
170 ssize_t ret; 168 ssize_t ret;
171 169
172 ret = iov_iter_extract_pages(sdio->ite !! 170 ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
173 DIO_PAGES !! 171 &sdio->from);
174 172
175 if (ret < 0 && sdio->blocks_available !! 173 if (ret < 0 && sdio->blocks_available && (dio->op == REQ_OP_WRITE)) {
>> 174 struct page *page = ZERO_PAGE(0);
176 /* 175 /*
177 * A memory fault, but the fil 176 * A memory fault, but the filesystem has some outstanding
178 * mapped blocks. We need to 177 * mapped blocks. We need to use those blocks up to avoid
179 * leaking stale data in the f 178 * leaking stale data in the file.
180 */ 179 */
181 if (dio->page_errors == 0) 180 if (dio->page_errors == 0)
182 dio->page_errors = ret 181 dio->page_errors = ret;
183 dio->pages[0] = ZERO_PAGE(0); !! 182 get_page(page);
>> 183 dio->pages[0] = page;
184 sdio->head = 0; 184 sdio->head = 0;
185 sdio->tail = 1; 185 sdio->tail = 1;
186 sdio->from = 0; 186 sdio->from = 0;
187 sdio->to = PAGE_SIZE; 187 sdio->to = PAGE_SIZE;
188 return 0; 188 return 0;
189 } 189 }
190 190
191 if (ret >= 0) { 191 if (ret >= 0) {
>> 192 iov_iter_advance(sdio->iter, ret);
192 ret += sdio->from; 193 ret += sdio->from;
193 sdio->head = 0; 194 sdio->head = 0;
194 sdio->tail = (ret + PAGE_SIZE 195 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
195 sdio->to = ((ret - 1) & (PAGE_ 196 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
196 return 0; 197 return 0;
197 } 198 }
198 return ret; 199 return ret;
199 } 200 }
200 201
201 /* 202 /*
202 * Get another userspace page. Returns an ERR 203 * Get another userspace page. Returns an ERR_PTR on error. Pages are
203 * buffered inside the dio so that we can call !! 204 * buffered inside the dio so that we can call get_user_pages() against a
204 * against a decent number of pages, less freq !! 205 * decent number of pages, less frequently. To provide nicer use of the
205 * the L1 cache. !! 206 * L1 cache.
206 */ 207 */
207 static inline struct page *dio_get_page(struct 208 static inline struct page *dio_get_page(struct dio *dio,
208 struct 209 struct dio_submit *sdio)
209 { 210 {
210 if (dio_pages_present(sdio) == 0) { 211 if (dio_pages_present(sdio) == 0) {
211 int ret; 212 int ret;
212 213
213 ret = dio_refill_pages(dio, sd 214 ret = dio_refill_pages(dio, sdio);
214 if (ret) 215 if (ret)
215 return ERR_PTR(ret); 216 return ERR_PTR(ret);
216 BUG_ON(dio_pages_present(sdio) 217 BUG_ON(dio_pages_present(sdio) == 0);
217 } 218 }
218 return dio->pages[sdio->head]; 219 return dio->pages[sdio->head];
219 } 220 }
220 221
221 static void dio_pin_page(struct dio *dio, stru !! 222 /*
>> 223 * Warn about a page cache invalidation failure during a direct io write.
>> 224 */
>> 225 void dio_warn_stale_pagecache(struct file *filp)
222 { 226 {
223 if (dio->is_pinned) !! 227 static DEFINE_RATELIMIT_STATE(_rs, 86400 * HZ, DEFAULT_RATELIMIT_BURST);
224 folio_add_pin(page_folio(page) !! 228 char pathname[128];
225 } !! 229 struct inode *inode = file_inode(filp);
>> 230 char *path;
226 231
227 static void dio_unpin_page(struct dio *dio, st !! 232 errseq_set(&inode->i_mapping->wb_err, -EIO);
228 { !! 233 if (__ratelimit(&_rs)) {
229 if (dio->is_pinned) !! 234 path = file_path(filp, pathname, sizeof(pathname));
230 unpin_user_page(page); !! 235 if (IS_ERR(path))
>> 236 path = "(unknown)";
>> 237 pr_crit("Page cache invalidation failure on direct I/O. Possible data corruption due to collision with buffered I/O!\n");
>> 238 pr_crit("File: %s PID: %d Comm: %.20s\n", path, current->pid,
>> 239 current->comm);
>> 240 }
231 } 241 }
232 242
233 /* !! 243 /**
234 * dio_complete() - called when all DIO BIO I/ 244 * dio_complete() - called when all DIO BIO I/O has been completed
>> 245 * @offset: the byte offset in the file of the completed operation
235 * 246 *
236 * This drops i_dio_count, lets interested par 247 * This drops i_dio_count, lets interested parties know that a DIO operation
237 * has completed, and calculates the resulting 248 * has completed, and calculates the resulting return code for the operation.
238 * 249 *
239 * It lets the filesystem know if it registere 250 * It lets the filesystem know if it registered an interest earlier via
240 * get_block. Pass the private field of the m 251 * get_block. Pass the private field of the map buffer_head so that
241 * filesystems can use it to hold additional s 252 * filesystems can use it to hold additional state between get_block calls and
242 * dio_complete. 253 * dio_complete.
243 */ 254 */
244 static ssize_t dio_complete(struct dio *dio, s 255 static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags)
245 { 256 {
246 const enum req_op dio_op = dio->opf & <<
247 loff_t offset = dio->iocb->ki_pos; 257 loff_t offset = dio->iocb->ki_pos;
248 ssize_t transferred = 0; 258 ssize_t transferred = 0;
249 int err; 259 int err;
250 260
251 /* 261 /*
252 * AIO submission can race with bio co 262 * AIO submission can race with bio completion to get here while
253 * expecting to have the last io compl 263 * expecting to have the last io completed by bio completion.
254 * In that case -EIOCBQUEUED is in fac 264 * In that case -EIOCBQUEUED is in fact not an error we want
255 * to preserve through this call. 265 * to preserve through this call.
256 */ 266 */
257 if (ret == -EIOCBQUEUED) 267 if (ret == -EIOCBQUEUED)
258 ret = 0; 268 ret = 0;
259 269
260 if (dio->result) { 270 if (dio->result) {
261 transferred = dio->result; 271 transferred = dio->result;
262 272
263 /* Check for short read case * 273 /* Check for short read case */
264 if (dio_op == REQ_OP_READ && !! 274 if ((dio->op == REQ_OP_READ) &&
265 ((offset + transferred) > 275 ((offset + transferred) > dio->i_size))
266 transferred = dio->i_s 276 transferred = dio->i_size - offset;
267 /* ignore EFAULT if some IO ha 277 /* ignore EFAULT if some IO has been done */
268 if (unlikely(ret == -EFAULT) & 278 if (unlikely(ret == -EFAULT) && transferred)
269 ret = 0; 279 ret = 0;
270 } 280 }
271 281
272 if (ret == 0) 282 if (ret == 0)
273 ret = dio->page_errors; 283 ret = dio->page_errors;
274 if (ret == 0) 284 if (ret == 0)
275 ret = dio->io_error; 285 ret = dio->io_error;
276 if (ret == 0) 286 if (ret == 0)
277 ret = transferred; 287 ret = transferred;
278 288
279 if (dio->end_io) { 289 if (dio->end_io) {
280 // XXX: ki_pos?? 290 // XXX: ki_pos??
281 err = dio->end_io(dio->iocb, o 291 err = dio->end_io(dio->iocb, offset, ret, dio->private);
282 if (err) 292 if (err)
283 ret = err; 293 ret = err;
284 } 294 }
285 295
286 /* 296 /*
287 * Try again to invalidate clean pages 297 * Try again to invalidate clean pages which might have been cached by
288 * non-direct readahead, or faulted in 298 * non-direct readahead, or faulted in by get_user_pages() if the source
289 * of the write was an mmap'ed region 299 * of the write was an mmap'ed region of the file we're writing. Either
290 * one is a pretty crazy thing to do, 300 * one is a pretty crazy thing to do, so we don't support it 100%. If
291 * this invalidation fails, tough, the 301 * this invalidation fails, tough, the write still worked...
292 * 302 *
293 * And this page cache invalidation ha 303 * And this page cache invalidation has to be after dio->end_io(), as
294 * some filesystems convert unwritten 304 * some filesystems convert unwritten extents to real allocations in
295 * end_io() when necessary, otherwise 305 * end_io() when necessary, otherwise a racing buffer read would cache
296 * zeros from unwritten extents. 306 * zeros from unwritten extents.
297 */ 307 */
298 if (flags & DIO_COMPLETE_INVALIDATE && 308 if (flags & DIO_COMPLETE_INVALIDATE &&
299 ret > 0 && dio_op == REQ_OP_WRITE) !! 309 ret > 0 && dio->op == REQ_OP_WRITE &&
300 kiocb_invalidate_post_direct_w !! 310 dio->inode->i_mapping->nrpages) {
>> 311 err = invalidate_inode_pages2_range(dio->inode->i_mapping,
>> 312 offset >> PAGE_SHIFT,
>> 313 (offset + ret - 1) >> PAGE_SHIFT);
>> 314 if (err)
>> 315 dio_warn_stale_pagecache(dio->iocb->ki_filp);
>> 316 }
301 317
302 inode_dio_end(dio->inode); 318 inode_dio_end(dio->inode);
303 319
304 if (flags & DIO_COMPLETE_ASYNC) { 320 if (flags & DIO_COMPLETE_ASYNC) {
305 /* 321 /*
306 * generic_write_sync expects 322 * generic_write_sync expects ki_pos to have been updated
307 * already, but the submission 323 * already, but the submission path only does this for
308 * synchronous I/O. 324 * synchronous I/O.
309 */ 325 */
310 dio->iocb->ki_pos += transferr 326 dio->iocb->ki_pos += transferred;
311 327
312 if (ret > 0 && dio_op == REQ_O !! 328 if (dio->op == REQ_OP_WRITE)
313 ret = generic_write_sy !! 329 ret = generic_write_sync(dio->iocb, transferred);
314 dio->iocb->ki_complete(dio->io !! 330 dio->iocb->ki_complete(dio->iocb, ret, 0);
315 } 331 }
316 332
317 kmem_cache_free(dio_cache, dio); 333 kmem_cache_free(dio_cache, dio);
318 return ret; 334 return ret;
319 } 335 }
320 336
321 static void dio_aio_complete_work(struct work_ 337 static void dio_aio_complete_work(struct work_struct *work)
322 { 338 {
323 struct dio *dio = container_of(work, s 339 struct dio *dio = container_of(work, struct dio, complete_work);
324 340
325 dio_complete(dio, 0, DIO_COMPLETE_ASYN 341 dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
326 } 342 }
327 343
328 static blk_status_t dio_bio_complete(struct di 344 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
329 345
330 /* 346 /*
331 * Asynchronous IO callback. 347 * Asynchronous IO callback.
332 */ 348 */
333 static void dio_bio_end_aio(struct bio *bio) 349 static void dio_bio_end_aio(struct bio *bio)
334 { 350 {
335 struct dio *dio = bio->bi_private; 351 struct dio *dio = bio->bi_private;
336 const enum req_op dio_op = dio->opf & <<
337 unsigned long remaining; 352 unsigned long remaining;
338 unsigned long flags; 353 unsigned long flags;
339 bool defer_completion = false; 354 bool defer_completion = false;
340 355
341 /* cleanup the bio */ 356 /* cleanup the bio */
342 dio_bio_complete(dio, bio); 357 dio_bio_complete(dio, bio);
343 358
344 spin_lock_irqsave(&dio->bio_lock, flag 359 spin_lock_irqsave(&dio->bio_lock, flags);
345 remaining = --dio->refcount; 360 remaining = --dio->refcount;
346 if (remaining == 1 && dio->waiter) 361 if (remaining == 1 && dio->waiter)
347 wake_up_process(dio->waiter); 362 wake_up_process(dio->waiter);
348 spin_unlock_irqrestore(&dio->bio_lock, 363 spin_unlock_irqrestore(&dio->bio_lock, flags);
349 364
350 if (remaining == 0) { 365 if (remaining == 0) {
351 /* 366 /*
352 * Defer completion when defer 367 * Defer completion when defer_completion is set or
353 * when the inode has pages ma 368 * when the inode has pages mapped and this is AIO write.
354 * We need to invalidate those 369 * We need to invalidate those pages because there is a
355 * chance they contain stale d 370 * chance they contain stale data in the case buffered IO
356 * went in between AIO submiss 371 * went in between AIO submission and completion into the
357 * same region. 372 * same region.
358 */ 373 */
359 if (dio->result) 374 if (dio->result)
360 defer_completion = dio 375 defer_completion = dio->defer_completion ||
361 (di !! 376 (dio->op == REQ_OP_WRITE &&
362 di 377 dio->inode->i_mapping->nrpages);
363 if (defer_completion) { 378 if (defer_completion) {
364 INIT_WORK(&dio->comple 379 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
365 queue_work(dio->inode- 380 queue_work(dio->inode->i_sb->s_dio_done_wq,
366 &dio->compl 381 &dio->complete_work);
367 } else { 382 } else {
368 dio_complete(dio, 0, D 383 dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
369 } 384 }
370 } 385 }
371 } 386 }
372 387
373 /* 388 /*
374 * The BIO completion handler simply queues th 389 * The BIO completion handler simply queues the BIO up for the process-context
375 * handler. 390 * handler.
376 * 391 *
377 * During I/O bi_private points at the dio. A 392 * 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 393 * implement a singly-linked list of completed BIOs, at dio->bio_list.
379 */ 394 */
380 static void dio_bio_end_io(struct bio *bio) 395 static void dio_bio_end_io(struct bio *bio)
381 { 396 {
382 struct dio *dio = bio->bi_private; 397 struct dio *dio = bio->bi_private;
383 unsigned long flags; 398 unsigned long flags;
384 399
385 spin_lock_irqsave(&dio->bio_lock, flag 400 spin_lock_irqsave(&dio->bio_lock, flags);
386 bio->bi_private = dio->bio_list; 401 bio->bi_private = dio->bio_list;
387 dio->bio_list = bio; 402 dio->bio_list = bio;
388 if (--dio->refcount == 1 && dio->waite 403 if (--dio->refcount == 1 && dio->waiter)
389 wake_up_process(dio->waiter); 404 wake_up_process(dio->waiter);
390 spin_unlock_irqrestore(&dio->bio_lock, 405 spin_unlock_irqrestore(&dio->bio_lock, flags);
391 } 406 }
392 407
>> 408 /**
>> 409 * dio_end_io - handle the end io action for the given bio
>> 410 * @bio: The direct io bio thats being completed
>> 411 *
>> 412 * This is meant to be called by any filesystem that uses their own dio_submit_t
>> 413 * so that the DIO specific endio actions are dealt with after the filesystem
>> 414 * has done it's completion work.
>> 415 */
>> 416 void dio_end_io(struct bio *bio)
>> 417 {
>> 418 struct dio *dio = bio->bi_private;
>> 419
>> 420 if (dio->is_async)
>> 421 dio_bio_end_aio(bio);
>> 422 else
>> 423 dio_bio_end_io(bio);
>> 424 }
>> 425 EXPORT_SYMBOL_GPL(dio_end_io);
>> 426
393 static inline void 427 static inline void
394 dio_bio_alloc(struct dio *dio, struct dio_subm 428 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
395 struct block_device *bdev, 429 struct block_device *bdev,
396 sector_t first_sector, int nr_ve 430 sector_t first_sector, int nr_vecs)
397 { 431 {
398 struct bio *bio; 432 struct bio *bio;
399 433
400 /* 434 /*
401 * bio_alloc() is guaranteed to return !! 435 * bio_alloc() is guaranteed to return a bio when called with
402 * we request a valid number of vector !! 436 * __GFP_RECLAIM and we request a valid number of vectors.
403 */ 437 */
404 bio = bio_alloc(bdev, nr_vecs, dio->op !! 438 bio = bio_alloc(GFP_KERNEL, nr_vecs);
>> 439
>> 440 bio_set_dev(bio, bdev);
405 bio->bi_iter.bi_sector = first_sector; 441 bio->bi_iter.bi_sector = first_sector;
>> 442 bio_set_op_attrs(bio, dio->op, dio->op_flags);
406 if (dio->is_async) 443 if (dio->is_async)
407 bio->bi_end_io = dio_bio_end_a 444 bio->bi_end_io = dio_bio_end_aio;
408 else 445 else
409 bio->bi_end_io = dio_bio_end_i 446 bio->bi_end_io = dio_bio_end_io;
410 if (dio->is_pinned) !! 447
411 bio_set_flag(bio, BIO_PAGE_PIN !! 448 bio->bi_write_hint = dio->iocb->ki_hint;
412 bio->bi_write_hint = file_inode(dio->i <<
413 449
414 sdio->bio = bio; 450 sdio->bio = bio;
415 sdio->logical_offset_in_bio = sdio->cu 451 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
416 } 452 }
417 453
418 /* 454 /*
419 * In the AIO read case we speculatively dirty 455 * In the AIO read case we speculatively dirty the pages before starting IO.
420 * During IO completion, any of these pages wh 456 * During IO completion, any of these pages which happen to have been written
421 * back will be redirtied by bio_check_pages_d 457 * back will be redirtied by bio_check_pages_dirty().
422 * 458 *
423 * bios hold a dio reference between submit_bi 459 * bios hold a dio reference between submit_bio and ->end_io.
424 */ 460 */
425 static inline void dio_bio_submit(struct dio * 461 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
426 { 462 {
427 const enum req_op dio_op = dio->opf & <<
428 struct bio *bio = sdio->bio; 463 struct bio *bio = sdio->bio;
429 unsigned long flags; 464 unsigned long flags;
430 465
431 bio->bi_private = dio; 466 bio->bi_private = dio;
432 467
433 spin_lock_irqsave(&dio->bio_lock, flag 468 spin_lock_irqsave(&dio->bio_lock, flags);
434 dio->refcount++; 469 dio->refcount++;
435 spin_unlock_irqrestore(&dio->bio_lock, 470 spin_unlock_irqrestore(&dio->bio_lock, flags);
436 471
437 if (dio->is_async && dio_op == REQ_OP_ !! 472 if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty)
438 bio_set_pages_dirty(bio); 473 bio_set_pages_dirty(bio);
439 474
440 dio->bio_disk = bio->bi_bdev->bd_disk; !! 475 dio->bio_disk = bio->bi_disk;
441 476
442 submit_bio(bio); !! 477 if (sdio->submit_io) {
>> 478 sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio);
>> 479 dio->bio_cookie = BLK_QC_T_NONE;
>> 480 } else
>> 481 dio->bio_cookie = submit_bio(bio);
443 482
444 sdio->bio = NULL; 483 sdio->bio = NULL;
445 sdio->boundary = 0; 484 sdio->boundary = 0;
446 sdio->logical_offset_in_bio = 0; 485 sdio->logical_offset_in_bio = 0;
447 } 486 }
448 487
449 /* 488 /*
450 * Release any resources in case of a failure 489 * Release any resources in case of a failure
451 */ 490 */
452 static inline void dio_cleanup(struct dio *dio 491 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
453 { 492 {
454 if (dio->is_pinned) !! 493 while (sdio->head < sdio->tail)
455 unpin_user_pages(dio->pages + !! 494 put_page(dio->pages[sdio->head++]);
456 sdio->tail - <<
457 sdio->head = sdio->tail; <<
458 } 495 }
459 496
460 /* 497 /*
461 * Wait for the next BIO to complete. Remove 498 * Wait for the next BIO to complete. Remove it and return it. NULL is
462 * returned once all BIOs have been completed. 499 * returned once all BIOs have been completed. This must only be called once
463 * all bios have been issued so that dio->refc 500 * all bios have been issued so that dio->refcount can only decrease. This
464 * requires that the caller hold a reference o !! 501 * requires that that the caller hold a reference on the dio.
465 */ 502 */
466 static struct bio *dio_await_one(struct dio *d 503 static struct bio *dio_await_one(struct dio *dio)
467 { 504 {
468 unsigned long flags; 505 unsigned long flags;
469 struct bio *bio = NULL; 506 struct bio *bio = NULL;
470 507
471 spin_lock_irqsave(&dio->bio_lock, flag 508 spin_lock_irqsave(&dio->bio_lock, flags);
472 509
473 /* 510 /*
474 * Wait as long as the list is empty a 511 * Wait as long as the list is empty and there are bios in flight. bio
475 * completion drops the count, maybe a 512 * completion drops the count, maybe adds to the list, and wakes while
476 * holding the bio_lock so we don't ne 513 * holding the bio_lock so we don't need set_current_state()'s barrier
477 * and can call it after testing our c 514 * and can call it after testing our condition.
478 */ 515 */
479 while (dio->refcount > 1 && dio->bio_l 516 while (dio->refcount > 1 && dio->bio_list == NULL) {
480 __set_current_state(TASK_UNINT 517 __set_current_state(TASK_UNINTERRUPTIBLE);
481 dio->waiter = current; 518 dio->waiter = current;
482 spin_unlock_irqrestore(&dio->b 519 spin_unlock_irqrestore(&dio->bio_lock, flags);
483 blk_io_schedule(); !! 520 if (!(dio->iocb->ki_flags & IOCB_HIPRI) ||
>> 521 !blk_poll(dio->bio_disk->queue, dio->bio_cookie))
>> 522 io_schedule();
484 /* wake up sets us TASK_RUNNIN 523 /* wake up sets us TASK_RUNNING */
485 spin_lock_irqsave(&dio->bio_lo 524 spin_lock_irqsave(&dio->bio_lock, flags);
486 dio->waiter = NULL; 525 dio->waiter = NULL;
487 } 526 }
488 if (dio->bio_list) { 527 if (dio->bio_list) {
489 bio = dio->bio_list; 528 bio = dio->bio_list;
490 dio->bio_list = bio->bi_privat 529 dio->bio_list = bio->bi_private;
491 } 530 }
492 spin_unlock_irqrestore(&dio->bio_lock, 531 spin_unlock_irqrestore(&dio->bio_lock, flags);
493 return bio; 532 return bio;
494 } 533 }
495 534
496 /* 535 /*
497 * Process one completed BIO. No locks are he 536 * Process one completed BIO. No locks are held.
498 */ 537 */
499 static blk_status_t dio_bio_complete(struct di 538 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
500 { 539 {
>> 540 struct bio_vec *bvec;
>> 541 unsigned i;
501 blk_status_t err = bio->bi_status; 542 blk_status_t err = bio->bi_status;
502 const enum req_op dio_op = dio->opf & <<
503 bool should_dirty = dio_op == REQ_OP_R <<
504 543
505 if (err) { 544 if (err) {
506 if (err == BLK_STS_AGAIN && (b 545 if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
507 dio->io_error = -EAGAI 546 dio->io_error = -EAGAIN;
508 else 547 else
509 dio->io_error = -EIO; 548 dio->io_error = -EIO;
510 } 549 }
511 550
512 if (dio->is_async && should_dirty) { !! 551 if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty) {
513 bio_check_pages_dirty(bio); 552 bio_check_pages_dirty(bio); /* transfers ownership */
514 } else { 553 } else {
515 bio_release_pages(bio, should_ !! 554 bio_for_each_segment_all(bvec, bio, i) {
>> 555 struct page *page = bvec->bv_page;
>> 556
>> 557 if (dio->op == REQ_OP_READ && !PageCompound(page) &&
>> 558 dio->should_dirty)
>> 559 set_page_dirty_lock(page);
>> 560 put_page(page);
>> 561 }
516 bio_put(bio); 562 bio_put(bio);
517 } 563 }
518 return err; 564 return err;
519 } 565 }
520 566
521 /* 567 /*
522 * Wait on and process all in-flight BIOs. Th 568 * Wait on and process all in-flight BIOs. This must only be called once
523 * all bios have been issued so that the refco 569 * all bios have been issued so that the refcount can only decrease.
524 * This just waits for all bios to make it thr 570 * This just waits for all bios to make it through dio_bio_complete. IO
525 * errors are propagated through dio->io_error 571 * errors are propagated through dio->io_error and should be propagated via
526 * dio_complete(). 572 * dio_complete().
527 */ 573 */
528 static void dio_await_completion(struct dio *d 574 static void dio_await_completion(struct dio *dio)
529 { 575 {
530 struct bio *bio; 576 struct bio *bio;
531 do { 577 do {
532 bio = dio_await_one(dio); 578 bio = dio_await_one(dio);
533 if (bio) 579 if (bio)
534 dio_bio_complete(dio, 580 dio_bio_complete(dio, bio);
535 } while (bio); 581 } while (bio);
536 } 582 }
537 583
538 /* 584 /*
539 * A really large O_DIRECT read or write can g 585 * A really large O_DIRECT read or write can generate a lot of BIOs. So
540 * to keep the memory consumption sane we peri 586 * to keep the memory consumption sane we periodically reap any completed BIOs
541 * during the BIO generation phase. 587 * during the BIO generation phase.
542 * 588 *
543 * This also helps to limit the peak amount of 589 * This also helps to limit the peak amount of pinned userspace memory.
544 */ 590 */
545 static inline int dio_bio_reap(struct dio *dio 591 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
546 { 592 {
547 int ret = 0; 593 int ret = 0;
548 594
549 if (sdio->reap_counter++ >= 64) { 595 if (sdio->reap_counter++ >= 64) {
550 while (dio->bio_list) { 596 while (dio->bio_list) {
551 unsigned long flags; 597 unsigned long flags;
552 struct bio *bio; 598 struct bio *bio;
553 int ret2; 599 int ret2;
554 600
555 spin_lock_irqsave(&dio 601 spin_lock_irqsave(&dio->bio_lock, flags);
556 bio = dio->bio_list; 602 bio = dio->bio_list;
557 dio->bio_list = bio->b 603 dio->bio_list = bio->bi_private;
558 spin_unlock_irqrestore 604 spin_unlock_irqrestore(&dio->bio_lock, flags);
559 ret2 = blk_status_to_e 605 ret2 = blk_status_to_errno(dio_bio_complete(dio, bio));
560 if (ret == 0) 606 if (ret == 0)
561 ret = ret2; 607 ret = ret2;
562 } 608 }
563 sdio->reap_counter = 0; 609 sdio->reap_counter = 0;
564 } 610 }
565 return ret; 611 return ret;
566 } 612 }
567 613
>> 614 /*
>> 615 * Create workqueue for deferred direct IO completions. We allocate the
>> 616 * workqueue when it's first needed. This avoids creating workqueue for
>> 617 * filesystems that don't need it and also allows us to create the workqueue
>> 618 * late enough so the we can include s_id in the name of the workqueue.
>> 619 */
>> 620 int sb_init_dio_done_wq(struct super_block *sb)
>> 621 {
>> 622 struct workqueue_struct *old;
>> 623 struct workqueue_struct *wq = alloc_workqueue("dio/%s",
>> 624 WQ_MEM_RECLAIM, 0,
>> 625 sb->s_id);
>> 626 if (!wq)
>> 627 return -ENOMEM;
>> 628 /*
>> 629 * This has to be atomic as more DIOs can race to create the workqueue
>> 630 */
>> 631 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
>> 632 /* Someone created workqueue before us? Free ours... */
>> 633 if (old)
>> 634 destroy_workqueue(wq);
>> 635 return 0;
>> 636 }
>> 637
568 static int dio_set_defer_completion(struct dio 638 static int dio_set_defer_completion(struct dio *dio)
569 { 639 {
570 struct super_block *sb = dio->inode->i 640 struct super_block *sb = dio->inode->i_sb;
571 641
572 if (dio->defer_completion) 642 if (dio->defer_completion)
573 return 0; 643 return 0;
574 dio->defer_completion = true; 644 dio->defer_completion = true;
575 if (!sb->s_dio_done_wq) 645 if (!sb->s_dio_done_wq)
576 return sb_init_dio_done_wq(sb) 646 return sb_init_dio_done_wq(sb);
577 return 0; 647 return 0;
578 } 648 }
579 649
580 /* 650 /*
581 * Call into the fs to map some more disk bloc 651 * Call into the fs to map some more disk blocks. We record the current number
582 * of available blocks at sdio->blocks_availab 652 * of available blocks at sdio->blocks_available. These are in units of the
583 * fs blocksize, i_blocksize(inode). 653 * fs blocksize, i_blocksize(inode).
584 * 654 *
585 * The fs is allowed to map lots of blocks at 655 * 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 656 * it uses the passed inode-relative block number as the file offset, as usual.
587 * 657 *
588 * get_block() is passed the number of i_blkbi 658 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
589 * has remaining to do. The fs should not map 659 * has remaining to do. The fs should not map more than this number of blocks.
590 * 660 *
591 * If the fs has mapped a lot of blocks, it sh 661 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
592 * indicate how much contiguous disk space has 662 * indicate how much contiguous disk space has been made available at
593 * bh->b_blocknr. 663 * bh->b_blocknr.
594 * 664 *
595 * If *any* of the mapped blocks are new, then 665 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
596 * This isn't very efficient... 666 * This isn't very efficient...
597 * 667 *
598 * In the case of filesystem holes: the fs may 668 * In the case of filesystem holes: the fs may return an arbitrarily-large
599 * hole by returning an appropriate value in b 669 * hole by returning an appropriate value in b_size and by clearing
600 * buffer_mapped(). However the direct-io cod 670 * buffer_mapped(). However the direct-io code will only process holes one
601 * block at a time - it will repeatedly call g 671 * block at a time - it will repeatedly call get_block() as it walks the hole.
602 */ 672 */
603 static int get_more_blocks(struct dio *dio, st 673 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
604 struct buffer_head 674 struct buffer_head *map_bh)
605 { 675 {
606 const enum req_op dio_op = dio->opf & <<
607 int ret; 676 int ret;
608 sector_t fs_startblk; /* Into file, 677 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
609 sector_t fs_endblk; /* Into file, 678 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
610 unsigned long fs_count; /* Number of f 679 unsigned long fs_count; /* Number of filesystem-sized blocks */
611 int create; 680 int create;
612 unsigned int i_blkbits = sdio->blkbits 681 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
613 loff_t i_size; <<
614 682
615 /* 683 /*
616 * If there was a memory error and we' 684 * If there was a memory error and we've overwritten all the
617 * mapped blocks then we can now retur 685 * mapped blocks then we can now return that memory error
618 */ 686 */
619 ret = dio->page_errors; 687 ret = dio->page_errors;
620 if (ret == 0) { 688 if (ret == 0) {
621 BUG_ON(sdio->block_in_file >= 689 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
622 fs_startblk = sdio->block_in_f 690 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
623 fs_endblk = (sdio->final_block 691 fs_endblk = (sdio->final_block_in_request - 1) >>
624 sdio-> 692 sdio->blkfactor;
625 fs_count = fs_endblk - fs_star 693 fs_count = fs_endblk - fs_startblk + 1;
626 694
627 map_bh->b_state = 0; 695 map_bh->b_state = 0;
628 map_bh->b_size = fs_count << i 696 map_bh->b_size = fs_count << i_blkbits;
629 697
630 /* 698 /*
631 * For writes that could fill 699 * For writes that could fill holes inside i_size on a
632 * DIO_SKIP_HOLES filesystem w 700 * DIO_SKIP_HOLES filesystem we forbid block creations: only
633 * overwrites are permitted. W 701 * overwrites are permitted. We will return early to the caller
634 * once we see an unmapped buf 702 * once we see an unmapped buffer head returned, and the caller
635 * will fall back to buffered 703 * will fall back to buffered I/O.
636 * 704 *
637 * Otherwise the decision is l 705 * Otherwise the decision is left to the get_blocks method,
638 * which may decide to handle 706 * which may decide to handle it or also return an unmapped
639 * buffer head. 707 * buffer head.
640 */ 708 */
641 create = dio_op == REQ_OP_WRIT !! 709 create = dio->op == REQ_OP_WRITE;
642 if (dio->flags & DIO_SKIP_HOLE 710 if (dio->flags & DIO_SKIP_HOLES) {
643 i_size = i_size_read(d !! 711 if (fs_startblk <= ((i_size_read(dio->inode) - 1) >>
644 if (i_size && fs_start !! 712 i_blkbits))
645 create = 0; 713 create = 0;
646 } 714 }
647 715
648 ret = (*sdio->get_block)(dio-> 716 ret = (*sdio->get_block)(dio->inode, fs_startblk,
649 717 map_bh, create);
650 718
651 /* Store for completion */ 719 /* Store for completion */
652 dio->private = map_bh->b_priva 720 dio->private = map_bh->b_private;
653 721
654 if (ret == 0 && buffer_defer_c 722 if (ret == 0 && buffer_defer_completion(map_bh))
655 ret = dio_set_defer_co 723 ret = dio_set_defer_completion(dio);
656 } 724 }
657 return ret; 725 return ret;
658 } 726 }
659 727
660 /* 728 /*
661 * There is no bio. Make one now. 729 * There is no bio. Make one now.
662 */ 730 */
663 static inline int dio_new_bio(struct dio *dio, 731 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
664 sector_t start_sector, struct 732 sector_t start_sector, struct buffer_head *map_bh)
665 { 733 {
666 sector_t sector; 734 sector_t sector;
667 int ret, nr_pages; 735 int ret, nr_pages;
668 736
669 ret = dio_bio_reap(dio, sdio); 737 ret = dio_bio_reap(dio, sdio);
670 if (ret) 738 if (ret)
671 goto out; 739 goto out;
672 sector = start_sector << (sdio->blkbit 740 sector = start_sector << (sdio->blkbits - 9);
673 nr_pages = bio_max_segs(sdio->pages_in !! 741 nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES);
674 BUG_ON(nr_pages <= 0); 742 BUG_ON(nr_pages <= 0);
675 dio_bio_alloc(dio, sdio, map_bh->b_bde 743 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
676 sdio->boundary = 0; 744 sdio->boundary = 0;
677 out: 745 out:
678 return ret; 746 return ret;
679 } 747 }
680 748
681 /* 749 /*
682 * Attempt to put the current chunk of 'cur_pa 750 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
683 * that was successful then update final_block 751 * that was successful then update final_block_in_bio and take a ref against
684 * the just-added page. 752 * the just-added page.
685 * 753 *
686 * Return zero on success. Non-zero means the 754 * Return zero on success. Non-zero means the caller needs to start a new BIO.
687 */ 755 */
688 static inline int dio_bio_add_page(struct dio !! 756 static inline int dio_bio_add_page(struct dio_submit *sdio)
689 { 757 {
690 int ret; 758 int ret;
691 759
692 ret = bio_add_page(sdio->bio, sdio->cu 760 ret = bio_add_page(sdio->bio, sdio->cur_page,
693 sdio->cur_page_len, sd 761 sdio->cur_page_len, sdio->cur_page_offset);
694 if (ret == sdio->cur_page_len) { 762 if (ret == sdio->cur_page_len) {
695 /* 763 /*
696 * Decrement count only, if we 764 * Decrement count only, if we are done with this page
697 */ 765 */
698 if ((sdio->cur_page_len + sdio 766 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
699 sdio->pages_in_io--; 767 sdio->pages_in_io--;
700 dio_pin_page(dio, sdio->cur_pa !! 768 get_page(sdio->cur_page);
701 sdio->final_block_in_bio = sdi 769 sdio->final_block_in_bio = sdio->cur_page_block +
702 (sdio->cur_page_len >> 770 (sdio->cur_page_len >> sdio->blkbits);
703 ret = 0; 771 ret = 0;
704 } else { 772 } else {
705 ret = 1; 773 ret = 1;
706 } 774 }
707 return ret; 775 return ret;
708 } 776 }
709 777
710 /* 778 /*
711 * Put cur_page under IO. The section of cur_ 779 * 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 780 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
713 * starts on-disk at cur_page_block. 781 * starts on-disk at cur_page_block.
714 * 782 *
715 * We take a ref against the page here (on beh 783 * We take a ref against the page here (on behalf of its presence in the bio).
716 * 784 *
717 * The caller of this function is responsible 785 * The caller of this function is responsible for removing cur_page from the
718 * dio, and for dropping the refcount which ca 786 * dio, and for dropping the refcount which came from that presence.
719 */ 787 */
720 static inline int dio_send_cur_page(struct dio 788 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
721 struct buffer_head *map_bh) 789 struct buffer_head *map_bh)
722 { 790 {
723 int ret = 0; 791 int ret = 0;
724 792
725 if (sdio->bio) { 793 if (sdio->bio) {
726 loff_t cur_offset = sdio->cur_ 794 loff_t cur_offset = sdio->cur_page_fs_offset;
727 loff_t bio_next_offset = sdio- 795 loff_t bio_next_offset = sdio->logical_offset_in_bio +
728 sdio->bio->bi_iter.bi_ 796 sdio->bio->bi_iter.bi_size;
729 797
730 /* 798 /*
731 * See whether this new reques 799 * See whether this new request is contiguous with the old.
732 * 800 *
733 * Btrfs cannot handle having 801 * Btrfs cannot handle having logically non-contiguous requests
734 * submitted. For example if 802 * submitted. For example if you have
735 * 803 *
736 * Logical: [0-4095][HOLE][81 804 * Logical: [0-4095][HOLE][8192-12287]
737 * Physical: [0-4095] [40 805 * Physical: [0-4095] [4096-8191]
738 * 806 *
739 * We cannot submit those page 807 * We cannot submit those pages together as one BIO. So if our
740 * current logical offset in t 808 * current logical offset in the file does not equal what would
741 * be the next logical offset 809 * be the next logical offset in the bio, submit the bio we
742 * have. 810 * have.
743 */ 811 */
744 if (sdio->final_block_in_bio ! 812 if (sdio->final_block_in_bio != sdio->cur_page_block ||
745 cur_offset != bio_next_off 813 cur_offset != bio_next_offset)
746 dio_bio_submit(dio, sd 814 dio_bio_submit(dio, sdio);
747 } 815 }
748 816
749 if (sdio->bio == NULL) { 817 if (sdio->bio == NULL) {
750 ret = dio_new_bio(dio, sdio, s 818 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
751 if (ret) 819 if (ret)
752 goto out; 820 goto out;
753 } 821 }
754 822
755 if (dio_bio_add_page(dio, sdio) != 0) !! 823 if (dio_bio_add_page(sdio) != 0) {
756 dio_bio_submit(dio, sdio); 824 dio_bio_submit(dio, sdio);
757 ret = dio_new_bio(dio, sdio, s 825 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
758 if (ret == 0) { 826 if (ret == 0) {
759 ret = dio_bio_add_page !! 827 ret = dio_bio_add_page(sdio);
760 BUG_ON(ret != 0); 828 BUG_ON(ret != 0);
761 } 829 }
762 } 830 }
763 out: 831 out:
764 return ret; 832 return ret;
765 } 833 }
766 834
767 /* 835 /*
768 * An autonomous function to put a chunk of a 836 * An autonomous function to put a chunk of a page under deferred IO.
769 * 837 *
770 * The caller doesn't actually know (or care) 838 * 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 839 * a BIO, or is under IO or whatever. We just take care of all possible
772 * situations here. The separation between th 840 * situations here. The separation between the logic of do_direct_IO() and
773 * that of submit_page_section() is important 841 * that of submit_page_section() is important for clarity. Please don't break.
774 * 842 *
775 * The chunk of page starts on-disk at blocknr 843 * The chunk of page starts on-disk at blocknr.
776 * 844 *
777 * We perform deferred IO, by recording the la 845 * We perform deferred IO, by recording the last-submitted page inside our
778 * private part of the dio structure. If poss 846 * private part of the dio structure. If possible, we just expand the IO
779 * across that page here. 847 * across that page here.
780 * 848 *
781 * If that doesn't work out then we put the ol 849 * If that doesn't work out then we put the old page into the bio and add this
782 * page to the dio instead. 850 * page to the dio instead.
783 */ 851 */
784 static inline int 852 static inline int
785 submit_page_section(struct dio *dio, struct di 853 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
786 unsigned offset, unsigned 854 unsigned offset, unsigned len, sector_t blocknr,
787 struct buffer_head *map_bh 855 struct buffer_head *map_bh)
788 { 856 {
789 const enum req_op dio_op = dio->opf & <<
790 int ret = 0; 857 int ret = 0;
791 int boundary = sdio->boundary; /* dio <<
792 858
793 if (dio_op == REQ_OP_WRITE) { !! 859 if (dio->op == REQ_OP_WRITE) {
794 /* 860 /*
795 * Read accounting is performe 861 * Read accounting is performed in submit_bio()
796 */ 862 */
797 task_io_account_write(len); 863 task_io_account_write(len);
798 } 864 }
799 865
800 /* 866 /*
801 * Can we just grow the current page's 867 * Can we just grow the current page's presence in the dio?
802 */ 868 */
803 if (sdio->cur_page == page && 869 if (sdio->cur_page == page &&
804 sdio->cur_page_offset + sdio->cur_ 870 sdio->cur_page_offset + sdio->cur_page_len == offset &&
805 sdio->cur_page_block + 871 sdio->cur_page_block +
806 (sdio->cur_page_len >> sdio->blkbi 872 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
807 sdio->cur_page_len += len; 873 sdio->cur_page_len += len;
808 goto out; 874 goto out;
809 } 875 }
810 876
811 /* 877 /*
812 * If there's a deferred page already 878 * If there's a deferred page already there then send it.
813 */ 879 */
814 if (sdio->cur_page) { 880 if (sdio->cur_page) {
815 ret = dio_send_cur_page(dio, s 881 ret = dio_send_cur_page(dio, sdio, map_bh);
816 dio_unpin_page(dio, sdio->cur_ !! 882 put_page(sdio->cur_page);
817 sdio->cur_page = NULL; 883 sdio->cur_page = NULL;
818 if (ret) 884 if (ret)
819 return ret; 885 return ret;
820 } 886 }
821 887
822 dio_pin_page(dio, page); !! 888 get_page(page); /* It is in dio */
823 sdio->cur_page = page; 889 sdio->cur_page = page;
824 sdio->cur_page_offset = offset; 890 sdio->cur_page_offset = offset;
825 sdio->cur_page_len = len; 891 sdio->cur_page_len = len;
826 sdio->cur_page_block = blocknr; 892 sdio->cur_page_block = blocknr;
827 sdio->cur_page_fs_offset = sdio->block 893 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
828 out: 894 out:
829 /* 895 /*
830 * If boundary then we want to schedul !! 896 * If sdio->boundary then we want to schedule the IO now to
831 * avoid metadata seeks. 897 * avoid metadata seeks.
832 */ 898 */
833 if (boundary) { !! 899 if (sdio->boundary) {
834 ret = dio_send_cur_page(dio, s 900 ret = dio_send_cur_page(dio, sdio, map_bh);
835 if (sdio->bio) 901 if (sdio->bio)
836 dio_bio_submit(dio, sd 902 dio_bio_submit(dio, sdio);
837 dio_unpin_page(dio, sdio->cur_ !! 903 put_page(sdio->cur_page);
838 sdio->cur_page = NULL; 904 sdio->cur_page = NULL;
839 } 905 }
840 return ret; 906 return ret;
841 } 907 }
842 908
843 /* 909 /*
844 * If we are not writing the entire block and 910 * If we are not writing the entire block and get_block() allocated
845 * the block for us, we need to fill-in the un 911 * the block for us, we need to fill-in the unused portion of the
846 * block with zeros. This happens only if user 912 * block with zeros. This happens only if user-buffer, fileoffset or
847 * io length is not filesystem block-size mult 913 * io length is not filesystem block-size multiple.
848 * 914 *
849 * `end' is zero if we're doing the start of t 915 * `end' is zero if we're doing the start of the IO, 1 at the end of the
850 * IO. 916 * IO.
851 */ 917 */
852 static inline void dio_zero_block(struct dio * 918 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
853 int end, struct buffer_head *m 919 int end, struct buffer_head *map_bh)
854 { 920 {
855 unsigned dio_blocks_per_fs_block; 921 unsigned dio_blocks_per_fs_block;
856 unsigned this_chunk_blocks; /* In 922 unsigned this_chunk_blocks; /* In dio_blocks */
857 unsigned this_chunk_bytes; 923 unsigned this_chunk_bytes;
858 struct page *page; 924 struct page *page;
859 925
860 sdio->start_zero_done = 1; 926 sdio->start_zero_done = 1;
861 if (!sdio->blkfactor || !buffer_new(ma 927 if (!sdio->blkfactor || !buffer_new(map_bh))
862 return; 928 return;
863 929
864 dio_blocks_per_fs_block = 1 << sdio->b 930 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
865 this_chunk_blocks = sdio->block_in_fil 931 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
866 932
867 if (!this_chunk_blocks) 933 if (!this_chunk_blocks)
868 return; 934 return;
869 935
870 /* 936 /*
871 * We need to zero out part of an fs b 937 * We need to zero out part of an fs block. It is either at the
872 * beginning or the end of the fs bloc 938 * beginning or the end of the fs block.
873 */ 939 */
874 if (end) 940 if (end)
875 this_chunk_blocks = dio_blocks 941 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
876 942
877 this_chunk_bytes = this_chunk_blocks < 943 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
878 944
879 page = ZERO_PAGE(0); 945 page = ZERO_PAGE(0);
880 if (submit_page_section(dio, sdio, pag 946 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
881 sdio->next_blo 947 sdio->next_block_for_io, map_bh))
882 return; 948 return;
883 949
884 sdio->next_block_for_io += this_chunk_ 950 sdio->next_block_for_io += this_chunk_blocks;
885 } 951 }
886 952
887 /* 953 /*
888 * Walk the user pages, and the file, mapping 954 * Walk the user pages, and the file, mapping blocks to disk and generating
889 * a sequence of (page,offset,len,block) mappi 955 * a sequence of (page,offset,len,block) mappings. These mappings are injected
890 * into submit_page_section(), which takes car 956 * into submit_page_section(), which takes care of the next stage of submission
891 * 957 *
892 * Direct IO against a blockdev is different f 958 * Direct IO against a blockdev is different from a file. Because we can
893 * happily perform page-sized but 512-byte ali 959 * happily perform page-sized but 512-byte aligned IOs. It is important that
894 * blockdev IO be able to have fine alignment 960 * blockdev IO be able to have fine alignment and large sizes.
895 * 961 *
896 * So what we do is to permit the ->get_block 962 * 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 963 * with the size of IO which is permitted at this offset and this i_blkbits.
898 * 964 *
899 * For best results, the blockdev should be se 965 * 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 966 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
901 * fine alignment but still allows this functi 967 * fine alignment but still allows this function to work in PAGE_SIZE units.
902 */ 968 */
903 static int do_direct_IO(struct dio *dio, struc 969 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
904 struct buffer_head *ma 970 struct buffer_head *map_bh)
905 { 971 {
906 const enum req_op dio_op = dio->opf & <<
907 const unsigned blkbits = sdio->blkbits 972 const unsigned blkbits = sdio->blkbits;
908 const unsigned i_blkbits = blkbits + s 973 const unsigned i_blkbits = blkbits + sdio->blkfactor;
909 int ret = 0; 974 int ret = 0;
910 975
911 while (sdio->block_in_file < sdio->fin 976 while (sdio->block_in_file < sdio->final_block_in_request) {
912 struct page *page; 977 struct page *page;
913 size_t from, to; 978 size_t from, to;
914 979
915 page = dio_get_page(dio, sdio) 980 page = dio_get_page(dio, sdio);
916 if (IS_ERR(page)) { 981 if (IS_ERR(page)) {
917 ret = PTR_ERR(page); 982 ret = PTR_ERR(page);
918 goto out; 983 goto out;
919 } 984 }
920 from = sdio->head ? 0 : sdio-> 985 from = sdio->head ? 0 : sdio->from;
921 to = (sdio->head == sdio->tail 986 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
922 sdio->head++; 987 sdio->head++;
923 988
924 while (from < to) { 989 while (from < to) {
925 unsigned this_chunk_by 990 unsigned this_chunk_bytes; /* # of bytes mapped */
926 unsigned this_chunk_bl 991 unsigned this_chunk_blocks; /* # of blocks */
927 unsigned u; 992 unsigned u;
928 993
929 if (sdio->blocks_avail 994 if (sdio->blocks_available == 0) {
930 /* 995 /*
931 * Need to go 996 * Need to go and map some more disk
932 */ 997 */
933 unsigned long 998 unsigned long blkmask;
934 unsigned long 999 unsigned long dio_remainder;
935 1000
936 ret = get_more 1001 ret = get_more_blocks(dio, sdio, map_bh);
937 if (ret) { 1002 if (ret) {
938 dio_un !! 1003 put_page(page);
939 goto o 1004 goto out;
940 } 1005 }
941 if (!buffer_ma 1006 if (!buffer_mapped(map_bh))
942 goto d 1007 goto do_holes;
943 1008
944 sdio->blocks_a 1009 sdio->blocks_available =
945 1010 map_bh->b_size >> blkbits;
946 sdio->next_blo 1011 sdio->next_block_for_io =
947 map_bh 1012 map_bh->b_blocknr << sdio->blkfactor;
948 if (buffer_new 1013 if (buffer_new(map_bh)) {
949 clean_ 1014 clean_bdev_aliases(
950 1015 map_bh->b_bdev,
951 1016 map_bh->b_blocknr,
952 1017 map_bh->b_size >> i_blkbits);
953 } 1018 }
954 1019
955 if (!sdio->blk 1020 if (!sdio->blkfactor)
956 goto d 1021 goto do_holes;
957 1022
958 blkmask = (1 < 1023 blkmask = (1 << sdio->blkfactor) - 1;
959 dio_remainder 1024 dio_remainder = (sdio->block_in_file & blkmask);
960 1025
961 /* 1026 /*
962 * If we are a 1027 * If we are at the start of IO and that IO
963 * starts part 1028 * starts partway into a fs-block,
964 * dio_remaind 1029 * dio_remainder will be non-zero. If the IO
965 * is a read t 1030 * is a read then we can simply advance the IO
966 * cursor to t 1031 * cursor to the first block which is to be
967 * read. But 1032 * read. But if the IO is a write and the
968 * block was n 1033 * block was newly allocated we cannot do that;
969 * the start o 1034 * the start of the fs block must be zeroed out
970 * on-disk 1035 * on-disk
971 */ 1036 */
972 if (!buffer_ne 1037 if (!buffer_new(map_bh))
973 sdio-> 1038 sdio->next_block_for_io += dio_remainder;
974 sdio->blocks_a 1039 sdio->blocks_available -= dio_remainder;
975 } 1040 }
976 do_holes: 1041 do_holes:
977 /* Handle holes */ 1042 /* Handle holes */
978 if (!buffer_mapped(map 1043 if (!buffer_mapped(map_bh)) {
979 loff_t i_size_ 1044 loff_t i_size_aligned;
980 1045
981 /* AKPM: eargh 1046 /* AKPM: eargh, -ENOTBLK is a hack */
982 if (dio_op == !! 1047 if (dio->op == REQ_OP_WRITE) {
983 dio_un !! 1048 put_page(page);
984 return 1049 return -ENOTBLK;
985 } 1050 }
986 1051
987 /* 1052 /*
988 * Be sure to 1053 * Be sure to account for a partial block as the
989 * last block 1054 * last block in the file
990 */ 1055 */
991 i_size_aligned 1056 i_size_aligned = ALIGN(i_size_read(dio->inode),
992 1057 1 << blkbits);
993 if (sdio->bloc 1058 if (sdio->block_in_file >=
994 1059 i_size_aligned >> blkbits) {
995 /* We 1060 /* We hit eof */
996 dio_un !! 1061 put_page(page);
997 goto o 1062 goto out;
998 } 1063 }
999 zero_user(page 1064 zero_user(page, from, 1 << blkbits);
1000 sdio->block_i 1065 sdio->block_in_file++;
1001 from += 1 << 1066 from += 1 << blkbits;
1002 dio->result + 1067 dio->result += 1 << blkbits;
1003 goto next_blo 1068 goto next_block;
1004 } 1069 }
1005 1070
1006 /* 1071 /*
1007 * If we're performin 1072 * If we're performing IO which has an alignment which
1008 * is finer than the 1073 * is finer than the underlying fs, go check to see if
1009 * we must zero out t 1074 * we must zero out the start of this block.
1010 */ 1075 */
1011 if (unlikely(sdio->bl 1076 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1012 dio_zero_bloc 1077 dio_zero_block(dio, sdio, 0, map_bh);
1013 1078
1014 /* 1079 /*
1015 * Work out, in this_ 1080 * Work out, in this_chunk_blocks, how much disk we
1016 * can add to this pa 1081 * can add to this page
1017 */ 1082 */
1018 this_chunk_blocks = s 1083 this_chunk_blocks = sdio->blocks_available;
1019 u = (to - from) >> bl 1084 u = (to - from) >> blkbits;
1020 if (this_chunk_blocks 1085 if (this_chunk_blocks > u)
1021 this_chunk_bl 1086 this_chunk_blocks = u;
1022 u = sdio->final_block 1087 u = sdio->final_block_in_request - sdio->block_in_file;
1023 if (this_chunk_blocks 1088 if (this_chunk_blocks > u)
1024 this_chunk_bl 1089 this_chunk_blocks = u;
1025 this_chunk_bytes = th 1090 this_chunk_bytes = this_chunk_blocks << blkbits;
1026 BUG_ON(this_chunk_byt 1091 BUG_ON(this_chunk_bytes == 0);
1027 1092
1028 if (this_chunk_blocks 1093 if (this_chunk_blocks == sdio->blocks_available)
1029 sdio->boundar 1094 sdio->boundary = buffer_boundary(map_bh);
1030 ret = submit_page_sec 1095 ret = submit_page_section(dio, sdio, page,
1031 1096 from,
1032 1097 this_chunk_bytes,
1033 1098 sdio->next_block_for_io,
1034 1099 map_bh);
1035 if (ret) { 1100 if (ret) {
1036 dio_unpin_pag !! 1101 put_page(page);
1037 goto out; 1102 goto out;
1038 } 1103 }
1039 sdio->next_block_for_ 1104 sdio->next_block_for_io += this_chunk_blocks;
1040 1105
1041 sdio->block_in_file + 1106 sdio->block_in_file += this_chunk_blocks;
1042 from += this_chunk_by 1107 from += this_chunk_bytes;
1043 dio->result += this_c 1108 dio->result += this_chunk_bytes;
1044 sdio->blocks_availabl 1109 sdio->blocks_available -= this_chunk_blocks;
1045 next_block: 1110 next_block:
1046 BUG_ON(sdio->block_in 1111 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1047 if (sdio->block_in_fi 1112 if (sdio->block_in_file == sdio->final_block_in_request)
1048 break; 1113 break;
1049 } 1114 }
1050 1115
1051 /* Drop the pin which was tak !! 1116 /* Drop the ref which was taken in get_user_pages() */
1052 dio_unpin_page(dio, page); !! 1117 put_page(page);
1053 } 1118 }
1054 out: 1119 out:
1055 return ret; 1120 return ret;
1056 } 1121 }
1057 1122
1058 static inline int drop_refcount(struct dio *d 1123 static inline int drop_refcount(struct dio *dio)
1059 { 1124 {
1060 int ret2; 1125 int ret2;
1061 unsigned long flags; 1126 unsigned long flags;
1062 1127
1063 /* 1128 /*
1064 * Sync will always be dropping the f 1129 * Sync will always be dropping the final ref and completing the
1065 * operation. AIO can if it was a br 1130 * operation. AIO can if it was a broken operation described above or
1066 * in fact if all the bios race to co 1131 * in fact if all the bios race to complete before we get here. In
1067 * that case dio_complete() translate 1132 * that case dio_complete() translates the EIOCBQUEUED into the proper
1068 * return code that the caller will h 1133 * return code that the caller will hand to ->complete().
1069 * 1134 *
1070 * This is managed by the bio_lock in 1135 * This is managed by the bio_lock instead of being an atomic_t so that
1071 * completion paths can drop their re 1136 * completion paths can drop their ref and use the remaining count to
1072 * decide to wake the submission path 1137 * decide to wake the submission path atomically.
1073 */ 1138 */
1074 spin_lock_irqsave(&dio->bio_lock, fla 1139 spin_lock_irqsave(&dio->bio_lock, flags);
1075 ret2 = --dio->refcount; 1140 ret2 = --dio->refcount;
1076 spin_unlock_irqrestore(&dio->bio_lock 1141 spin_unlock_irqrestore(&dio->bio_lock, flags);
1077 return ret2; 1142 return ret2;
1078 } 1143 }
1079 1144
1080 /* 1145 /*
1081 * This is a library function for use by file 1146 * This is a library function for use by filesystem drivers.
1082 * 1147 *
1083 * The locking rules are governed by the flag 1148 * The locking rules are governed by the flags parameter:
1084 * - if the flags value contains DIO_LOCKING 1149 * - if the flags value contains DIO_LOCKING we use a fancy locking
1085 * scheme for dumb filesystems. 1150 * scheme for dumb filesystems.
1086 * For writes this function is called unde 1151 * For writes this function is called under i_mutex and returns with
1087 * i_mutex held, for reads, i_mutex is not 1152 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1088 * taken and dropped again before returnin 1153 * taken and dropped again before returning.
1089 * - if the flags value does NOT contain DIO 1154 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1090 * internal locking but rather rely on the 1155 * internal locking but rather rely on the filesystem to synchronize
1091 * direct I/O reads/writes versus each oth 1156 * direct I/O reads/writes versus each other and truncate.
1092 * 1157 *
1093 * To help with locking against truncate we i 1158 * To help with locking against truncate we incremented the i_dio_count
1094 * counter before starting direct I/O, and de 1159 * counter before starting direct I/O, and decrement it once we are done.
1095 * Truncate can wait for it to reach zero to 1160 * Truncate can wait for it to reach zero to provide exclusion. It is
1096 * expected that filesystem provide exclusion 1161 * expected that filesystem provide exclusion between new direct I/O
1097 * and truncates. For DIO_LOCKING filesystem 1162 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1098 * but other filesystems need to take care of 1163 * but other filesystems need to take care of this on their own.
1099 * 1164 *
1100 * NOTE: if you pass "sdio" to anything by po 1165 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1101 * is always inlined. Otherwise gcc is unable 1166 * is always inlined. Otherwise gcc is unable to split the structure into
1102 * individual fields and will generate much w 1167 * individual fields and will generate much worse code. This is important
1103 * for the whole file. 1168 * for the whole file.
1104 */ 1169 */
1105 ssize_t __blockdev_direct_IO(struct kiocb *io !! 1170 static inline ssize_t
1106 struct block_device *bdev, st !! 1171 do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1107 get_block_t get_block, dio_io !! 1172 struct block_device *bdev, struct iov_iter *iter,
1108 int flags) !! 1173 get_block_t get_block, dio_iodone_t end_io,
>> 1174 dio_submit_t submit_io, int flags)
1109 { 1175 {
1110 unsigned i_blkbits = READ_ONCE(inode- 1176 unsigned i_blkbits = READ_ONCE(inode->i_blkbits);
1111 unsigned blkbits = i_blkbits; 1177 unsigned blkbits = i_blkbits;
1112 unsigned blocksize_mask = (1 << blkbi 1178 unsigned blocksize_mask = (1 << blkbits) - 1;
1113 ssize_t retval = -EINVAL; 1179 ssize_t retval = -EINVAL;
1114 const size_t count = iov_iter_count(i 1180 const size_t count = iov_iter_count(iter);
1115 loff_t offset = iocb->ki_pos; 1181 loff_t offset = iocb->ki_pos;
1116 const loff_t end = offset + count; 1182 const loff_t end = offset + count;
1117 struct dio *dio; 1183 struct dio *dio;
1118 struct dio_submit sdio = { NULL, }; !! 1184 struct dio_submit sdio = { 0, };
1119 struct buffer_head map_bh = { 0, }; 1185 struct buffer_head map_bh = { 0, };
1120 struct blk_plug plug; 1186 struct blk_plug plug;
1121 unsigned long align = offset | iov_it 1187 unsigned long align = offset | iov_iter_alignment(iter);
1122 1188
>> 1189 /*
>> 1190 * Avoid references to bdev if not absolutely needed to give
>> 1191 * the early prefetch in the caller enough time.
>> 1192 */
>> 1193
>> 1194 if (align & blocksize_mask) {
>> 1195 if (bdev)
>> 1196 blkbits = blksize_bits(bdev_logical_block_size(bdev));
>> 1197 blocksize_mask = (1 << blkbits) - 1;
>> 1198 if (align & blocksize_mask)
>> 1199 goto out;
>> 1200 }
>> 1201
1123 /* watch out for a 0 len io from a tr 1202 /* watch out for a 0 len io from a tricksy fs */
1124 if (iov_iter_rw(iter) == READ && !cou 1203 if (iov_iter_rw(iter) == READ && !count)
1125 return 0; 1204 return 0;
1126 1205
1127 dio = kmem_cache_alloc(dio_cache, GFP 1206 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
>> 1207 retval = -ENOMEM;
1128 if (!dio) 1208 if (!dio)
1129 return -ENOMEM; !! 1209 goto out;
1130 /* 1210 /*
1131 * Believe it or not, zeroing out the 1211 * Believe it or not, zeroing out the page array caused a .5%
1132 * performance regression in a databa 1212 * performance regression in a database benchmark. So, we take
1133 * care to only zero out what's neede 1213 * care to only zero out what's needed.
1134 */ 1214 */
1135 memset(dio, 0, offsetof(struct dio, p 1215 memset(dio, 0, offsetof(struct dio, pages));
1136 1216
1137 dio->flags = flags; 1217 dio->flags = flags;
1138 if (dio->flags & DIO_LOCKING && iov_i !! 1218 if (dio->flags & DIO_LOCKING) {
1139 /* will be released by direct !! 1219 if (iov_iter_rw(iter) == READ) {
1140 inode_lock(inode); !! 1220 struct address_space *mapping =
>> 1221 iocb->ki_filp->f_mapping;
>> 1222
>> 1223 /* will be released by direct_io_worker */
>> 1224 inode_lock(inode);
>> 1225
>> 1226 retval = filemap_write_and_wait_range(mapping, offset,
>> 1227 end - 1);
>> 1228 if (retval) {
>> 1229 inode_unlock(inode);
>> 1230 kmem_cache_free(dio_cache, dio);
>> 1231 goto out;
>> 1232 }
>> 1233 }
1141 } 1234 }
1142 dio->is_pinned = iov_iter_extract_wil <<
1143 1235
1144 /* Once we sampled i_size check for r 1236 /* Once we sampled i_size check for reads beyond EOF */
1145 dio->i_size = i_size_read(inode); 1237 dio->i_size = i_size_read(inode);
1146 if (iov_iter_rw(iter) == READ && offs 1238 if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
>> 1239 if (dio->flags & DIO_LOCKING)
>> 1240 inode_unlock(inode);
>> 1241 kmem_cache_free(dio_cache, dio);
1147 retval = 0; 1242 retval = 0;
1148 goto fail_dio; !! 1243 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 } 1244 }
1166 1245
1167 /* 1246 /*
1168 * For file extending writes updating 1247 * For file extending writes updating i_size before data writeouts
1169 * complete can expose uninitialized 1248 * complete can expose uninitialized blocks in dumb filesystems.
1170 * In that case we need to wait for I 1249 * In that case we need to wait for I/O completion even if asked
1171 * for an asynchronous write. 1250 * for an asynchronous write.
1172 */ 1251 */
1173 if (is_sync_kiocb(iocb)) 1252 if (is_sync_kiocb(iocb))
1174 dio->is_async = false; 1253 dio->is_async = false;
1175 else if (iov_iter_rw(iter) == WRITE & 1254 else if (iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
1176 dio->is_async = false; 1255 dio->is_async = false;
1177 else 1256 else
1178 dio->is_async = true; 1257 dio->is_async = true;
1179 1258
1180 dio->inode = inode; 1259 dio->inode = inode;
1181 if (iov_iter_rw(iter) == WRITE) { 1260 if (iov_iter_rw(iter) == WRITE) {
1182 dio->opf = REQ_OP_WRITE | REQ !! 1261 dio->op = REQ_OP_WRITE;
>> 1262 dio->op_flags = REQ_SYNC | REQ_IDLE;
1183 if (iocb->ki_flags & IOCB_NOW 1263 if (iocb->ki_flags & IOCB_NOWAIT)
1184 dio->opf |= REQ_NOWAI !! 1264 dio->op_flags |= REQ_NOWAIT;
1185 } else { 1265 } else {
1186 dio->opf = REQ_OP_READ; !! 1266 dio->op = REQ_OP_READ;
1187 } 1267 }
1188 1268
1189 /* 1269 /*
1190 * For AIO O_(D)SYNC writes we need t 1270 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1191 * so that we can call ->fsync. 1271 * so that we can call ->fsync.
1192 */ 1272 */
1193 if (dio->is_async && iov_iter_rw(iter 1273 if (dio->is_async && iov_iter_rw(iter) == WRITE) {
1194 retval = 0; 1274 retval = 0;
1195 if (iocb_is_dsync(iocb)) !! 1275 if (iocb->ki_flags & IOCB_DSYNC)
1196 retval = dio_set_defe 1276 retval = dio_set_defer_completion(dio);
1197 else if (!dio->inode->i_sb->s 1277 else if (!dio->inode->i_sb->s_dio_done_wq) {
1198 /* 1278 /*
1199 * In case of AIO wri 1279 * In case of AIO write racing with buffered read we
1200 * need to defer comp 1280 * need to defer completion. We can't decide this now,
1201 * however the workqu 1281 * however the workqueue needs to be initialized here.
1202 */ 1282 */
1203 retval = sb_init_dio_ 1283 retval = sb_init_dio_done_wq(dio->inode->i_sb);
1204 } 1284 }
1205 if (retval) !! 1285 if (retval) {
1206 goto fail_dio; !! 1286 /*
>> 1287 * We grab i_mutex only for reads so we don't have
>> 1288 * to release it here
>> 1289 */
>> 1290 kmem_cache_free(dio_cache, dio);
>> 1291 goto out;
>> 1292 }
1207 } 1293 }
1208 1294
1209 /* 1295 /*
1210 * Will be decremented at I/O complet 1296 * Will be decremented at I/O completion time.
1211 */ 1297 */
1212 inode_dio_begin(inode); 1298 inode_dio_begin(inode);
1213 1299
>> 1300 retval = 0;
1214 sdio.blkbits = blkbits; 1301 sdio.blkbits = blkbits;
1215 sdio.blkfactor = i_blkbits - blkbits; 1302 sdio.blkfactor = i_blkbits - blkbits;
1216 sdio.block_in_file = offset >> blkbit 1303 sdio.block_in_file = offset >> blkbits;
1217 1304
1218 sdio.get_block = get_block; 1305 sdio.get_block = get_block;
1219 dio->end_io = end_io; 1306 dio->end_io = end_io;
>> 1307 sdio.submit_io = submit_io;
1220 sdio.final_block_in_bio = -1; 1308 sdio.final_block_in_bio = -1;
1221 sdio.next_block_for_io = -1; 1309 sdio.next_block_for_io = -1;
1222 1310
1223 dio->iocb = iocb; 1311 dio->iocb = iocb;
1224 1312
1225 spin_lock_init(&dio->bio_lock); 1313 spin_lock_init(&dio->bio_lock);
1226 dio->refcount = 1; 1314 dio->refcount = 1;
1227 1315
1228 dio->should_dirty = user_backed_iter( !! 1316 dio->should_dirty = (iter->type == ITER_IOVEC);
1229 sdio.iter = iter; 1317 sdio.iter = iter;
1230 sdio.final_block_in_request = end >> 1318 sdio.final_block_in_request = end >> blkbits;
1231 1319
1232 /* 1320 /*
1233 * In case of non-aligned buffers, we 1321 * In case of non-aligned buffers, we may need 2 more
1234 * pages since we need to zero out fi 1322 * pages since we need to zero out first and last block.
1235 */ 1323 */
1236 if (unlikely(sdio.blkfactor)) 1324 if (unlikely(sdio.blkfactor))
1237 sdio.pages_in_io = 2; 1325 sdio.pages_in_io = 2;
1238 1326
1239 sdio.pages_in_io += iov_iter_npages(i 1327 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
1240 1328
1241 blk_start_plug(&plug); 1329 blk_start_plug(&plug);
1242 1330
1243 retval = do_direct_IO(dio, &sdio, &ma 1331 retval = do_direct_IO(dio, &sdio, &map_bh);
1244 if (retval) 1332 if (retval)
1245 dio_cleanup(dio, &sdio); 1333 dio_cleanup(dio, &sdio);
1246 1334
1247 if (retval == -ENOTBLK) { 1335 if (retval == -ENOTBLK) {
1248 /* 1336 /*
1249 * The remaining part of the 1337 * The remaining part of the request will be
1250 * handled by buffered I/O wh !! 1338 * be handled by buffered I/O when we return
1251 */ 1339 */
1252 retval = 0; 1340 retval = 0;
1253 } 1341 }
1254 /* 1342 /*
1255 * There may be some unwritten disk a 1343 * There may be some unwritten disk at the end of a part-written
1256 * fs-block-sized block. Go zero tha 1344 * fs-block-sized block. Go zero that now.
1257 */ 1345 */
1258 dio_zero_block(dio, &sdio, 1, &map_bh 1346 dio_zero_block(dio, &sdio, 1, &map_bh);
1259 1347
1260 if (sdio.cur_page) { 1348 if (sdio.cur_page) {
1261 ssize_t ret2; 1349 ssize_t ret2;
1262 1350
1263 ret2 = dio_send_cur_page(dio, 1351 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1264 if (retval == 0) 1352 if (retval == 0)
1265 retval = ret2; 1353 retval = ret2;
1266 dio_unpin_page(dio, sdio.cur_ !! 1354 put_page(sdio.cur_page);
1267 sdio.cur_page = NULL; 1355 sdio.cur_page = NULL;
1268 } 1356 }
1269 if (sdio.bio) 1357 if (sdio.bio)
1270 dio_bio_submit(dio, &sdio); 1358 dio_bio_submit(dio, &sdio);
1271 1359
1272 blk_finish_plug(&plug); 1360 blk_finish_plug(&plug);
1273 1361
1274 /* 1362 /*
1275 * It is possible that, we return sho 1363 * It is possible that, we return short IO due to end of file.
1276 * In that case, we need to release a 1364 * In that case, we need to release all the pages we got hold on.
1277 */ 1365 */
1278 dio_cleanup(dio, &sdio); 1366 dio_cleanup(dio, &sdio);
1279 1367
1280 /* 1368 /*
1281 * All block lookups have been perfor 1369 * All block lookups have been performed. For READ requests
1282 * we can let i_mutex go now that its 1370 * we can let i_mutex go now that its achieved its purpose
1283 * of protecting us from looking up u 1371 * of protecting us from looking up uninitialized blocks.
1284 */ 1372 */
1285 if (iov_iter_rw(iter) == READ && (dio 1373 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
1286 inode_unlock(dio->inode); 1374 inode_unlock(dio->inode);
1287 1375
1288 /* 1376 /*
1289 * The only time we want to leave bio 1377 * The only time we want to leave bios in flight is when a successful
1290 * partial aio read or full aio write 1378 * partial aio read or full aio write have been setup. In that case
1291 * bio completion will call aio_compl 1379 * bio completion will call aio_complete. The only time it's safe to
1292 * call aio_complete is when we retur 1380 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1293 * This had *better* be the only plac 1381 * This had *better* be the only place that raises -EIOCBQUEUED.
1294 */ 1382 */
1295 BUG_ON(retval == -EIOCBQUEUED); 1383 BUG_ON(retval == -EIOCBQUEUED);
1296 if (dio->is_async && retval == 0 && d 1384 if (dio->is_async && retval == 0 && dio->result &&
1297 (iov_iter_rw(iter) == READ || dio 1385 (iov_iter_rw(iter) == READ || dio->result == count))
1298 retval = -EIOCBQUEUED; 1386 retval = -EIOCBQUEUED;
1299 else 1387 else
1300 dio_await_completion(dio); 1388 dio_await_completion(dio);
1301 1389
1302 if (drop_refcount(dio) == 0) { 1390 if (drop_refcount(dio) == 0) {
1303 retval = dio_complete(dio, re 1391 retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE);
1304 } else 1392 } else
1305 BUG_ON(retval != -EIOCBQUEUED 1393 BUG_ON(retval != -EIOCBQUEUED);
1306 1394
>> 1395 out:
1307 return retval; 1396 return retval;
>> 1397 }
1308 1398
1309 fail_dio: !! 1399 ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1310 if (dio->flags & DIO_LOCKING && iov_i !! 1400 struct block_device *bdev, struct iov_iter *iter,
1311 inode_unlock(inode); !! 1401 get_block_t get_block,
>> 1402 dio_iodone_t end_io, dio_submit_t submit_io,
>> 1403 int flags)
>> 1404 {
>> 1405 /*
>> 1406 * The block device state is needed in the end to finally
>> 1407 * submit everything. Since it's likely to be cache cold
>> 1408 * prefetch it here as first thing to hide some of the
>> 1409 * latency.
>> 1410 *
>> 1411 * Attempt to prefetch the pieces we likely need later.
>> 1412 */
>> 1413 prefetch(&bdev->bd_disk->part_tbl);
>> 1414 prefetch(bdev->bd_queue);
>> 1415 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1312 1416
1313 kmem_cache_free(dio_cache, dio); !! 1417 return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block,
1314 return retval; !! 1418 end_io, submit_io, flags);
1315 } 1419 }
>> 1420
1316 EXPORT_SYMBOL(__blockdev_direct_IO); 1421 EXPORT_SYMBOL(__blockdev_direct_IO);
1317 1422
1318 static __init int dio_init(void) 1423 static __init int dio_init(void)
1319 { 1424 {
1320 dio_cache = KMEM_CACHE(dio, SLAB_PANI 1425 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1321 return 0; 1426 return 0;
1322 } 1427 }
1323 module_init(dio_init) 1428 module_init(dio_init)
1324 1429
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.