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 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
>> 319 inode_dio_end(dio->inode);
303 320
304 if (flags & DIO_COMPLETE_ASYNC) { 321 if (flags & DIO_COMPLETE_ASYNC) {
305 /* 322 /*
306 * generic_write_sync expects 323 * generic_write_sync expects ki_pos to have been updated
307 * already, but the submission 324 * already, but the submission path only does this for
308 * synchronous I/O. 325 * synchronous I/O.
309 */ 326 */
310 dio->iocb->ki_pos += transferr 327 dio->iocb->ki_pos += transferred;
311 328
312 if (ret > 0 && dio_op == REQ_O !! 329 if (ret > 0 && dio->op == REQ_OP_WRITE)
313 ret = generic_write_sy 330 ret = generic_write_sync(dio->iocb, ret);
314 dio->iocb->ki_complete(dio->io !! 331 dio->iocb->ki_complete(dio->iocb, ret, 0);
315 } 332 }
316 333
317 kmem_cache_free(dio_cache, dio); 334 kmem_cache_free(dio_cache, dio);
318 return ret; 335 return ret;
319 } 336 }
320 337
321 static void dio_aio_complete_work(struct work_ 338 static void dio_aio_complete_work(struct work_struct *work)
322 { 339 {
323 struct dio *dio = container_of(work, s 340 struct dio *dio = container_of(work, struct dio, complete_work);
324 341
325 dio_complete(dio, 0, DIO_COMPLETE_ASYN 342 dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
326 } 343 }
327 344
328 static blk_status_t dio_bio_complete(struct di 345 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
329 346
330 /* 347 /*
331 * Asynchronous IO callback. 348 * Asynchronous IO callback.
332 */ 349 */
333 static void dio_bio_end_aio(struct bio *bio) 350 static void dio_bio_end_aio(struct bio *bio)
334 { 351 {
335 struct dio *dio = bio->bi_private; 352 struct dio *dio = bio->bi_private;
336 const enum req_op dio_op = dio->opf & <<
337 unsigned long remaining; 353 unsigned long remaining;
338 unsigned long flags; 354 unsigned long flags;
339 bool defer_completion = false; 355 bool defer_completion = false;
340 356
341 /* cleanup the bio */ 357 /* cleanup the bio */
342 dio_bio_complete(dio, bio); 358 dio_bio_complete(dio, bio);
343 359
344 spin_lock_irqsave(&dio->bio_lock, flag 360 spin_lock_irqsave(&dio->bio_lock, flags);
345 remaining = --dio->refcount; 361 remaining = --dio->refcount;
346 if (remaining == 1 && dio->waiter) 362 if (remaining == 1 && dio->waiter)
347 wake_up_process(dio->waiter); 363 wake_up_process(dio->waiter);
348 spin_unlock_irqrestore(&dio->bio_lock, 364 spin_unlock_irqrestore(&dio->bio_lock, flags);
349 365
350 if (remaining == 0) { 366 if (remaining == 0) {
351 /* 367 /*
352 * Defer completion when defer 368 * Defer completion when defer_completion is set or
353 * when the inode has pages ma 369 * when the inode has pages mapped and this is AIO write.
354 * We need to invalidate those 370 * We need to invalidate those pages because there is a
355 * chance they contain stale d 371 * chance they contain stale data in the case buffered IO
356 * went in between AIO submiss 372 * went in between AIO submission and completion into the
357 * same region. 373 * same region.
358 */ 374 */
359 if (dio->result) 375 if (dio->result)
360 defer_completion = dio 376 defer_completion = dio->defer_completion ||
361 (di !! 377 (dio->op == REQ_OP_WRITE &&
362 di 378 dio->inode->i_mapping->nrpages);
363 if (defer_completion) { 379 if (defer_completion) {
364 INIT_WORK(&dio->comple 380 INIT_WORK(&dio->complete_work, dio_aio_complete_work);
365 queue_work(dio->inode- 381 queue_work(dio->inode->i_sb->s_dio_done_wq,
366 &dio->compl 382 &dio->complete_work);
367 } else { 383 } else {
368 dio_complete(dio, 0, D 384 dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
369 } 385 }
370 } 386 }
371 } 387 }
372 388
373 /* 389 /*
374 * The BIO completion handler simply queues th 390 * The BIO completion handler simply queues the BIO up for the process-context
375 * handler. 391 * handler.
376 * 392 *
377 * During I/O bi_private points at the dio. A 393 * 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 394 * implement a singly-linked list of completed BIOs, at dio->bio_list.
379 */ 395 */
380 static void dio_bio_end_io(struct bio *bio) 396 static void dio_bio_end_io(struct bio *bio)
381 { 397 {
382 struct dio *dio = bio->bi_private; 398 struct dio *dio = bio->bi_private;
383 unsigned long flags; 399 unsigned long flags;
384 400
385 spin_lock_irqsave(&dio->bio_lock, flag 401 spin_lock_irqsave(&dio->bio_lock, flags);
386 bio->bi_private = dio->bio_list; 402 bio->bi_private = dio->bio_list;
387 dio->bio_list = bio; 403 dio->bio_list = bio;
388 if (--dio->refcount == 1 && dio->waite 404 if (--dio->refcount == 1 && dio->waiter)
389 wake_up_process(dio->waiter); 405 wake_up_process(dio->waiter);
390 spin_unlock_irqrestore(&dio->bio_lock, 406 spin_unlock_irqrestore(&dio->bio_lock, flags);
391 } 407 }
392 408
>> 409 /**
>> 410 * dio_end_io - handle the end io action for the given bio
>> 411 * @bio: The direct io bio thats being completed
>> 412 *
>> 413 * This is meant to be called by any filesystem that uses their own dio_submit_t
>> 414 * so that the DIO specific endio actions are dealt with after the filesystem
>> 415 * has done it's completion work.
>> 416 */
>> 417 void dio_end_io(struct bio *bio)
>> 418 {
>> 419 struct dio *dio = bio->bi_private;
>> 420
>> 421 if (dio->is_async)
>> 422 dio_bio_end_aio(bio);
>> 423 else
>> 424 dio_bio_end_io(bio);
>> 425 }
>> 426 EXPORT_SYMBOL_GPL(dio_end_io);
>> 427
393 static inline void 428 static inline void
394 dio_bio_alloc(struct dio *dio, struct dio_subm 429 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
395 struct block_device *bdev, 430 struct block_device *bdev,
396 sector_t first_sector, int nr_ve 431 sector_t first_sector, int nr_vecs)
397 { 432 {
398 struct bio *bio; 433 struct bio *bio;
399 434
400 /* 435 /*
401 * bio_alloc() is guaranteed to return !! 436 * bio_alloc() is guaranteed to return a bio when called with
402 * we request a valid number of vector !! 437 * __GFP_RECLAIM and we request a valid number of vectors.
403 */ 438 */
404 bio = bio_alloc(bdev, nr_vecs, dio->op !! 439 bio = bio_alloc(GFP_KERNEL, nr_vecs);
>> 440
>> 441 bio_set_dev(bio, bdev);
405 bio->bi_iter.bi_sector = first_sector; 442 bio->bi_iter.bi_sector = first_sector;
>> 443 bio_set_op_attrs(bio, dio->op, dio->op_flags);
406 if (dio->is_async) 444 if (dio->is_async)
407 bio->bi_end_io = dio_bio_end_a 445 bio->bi_end_io = dio_bio_end_aio;
408 else 446 else
409 bio->bi_end_io = dio_bio_end_i 447 bio->bi_end_io = dio_bio_end_io;
410 if (dio->is_pinned) !! 448
411 bio_set_flag(bio, BIO_PAGE_PIN !! 449 bio->bi_write_hint = dio->iocb->ki_hint;
412 bio->bi_write_hint = file_inode(dio->i <<
413 450
414 sdio->bio = bio; 451 sdio->bio = bio;
415 sdio->logical_offset_in_bio = sdio->cu 452 sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
416 } 453 }
417 454
418 /* 455 /*
419 * In the AIO read case we speculatively dirty 456 * In the AIO read case we speculatively dirty the pages before starting IO.
420 * During IO completion, any of these pages wh 457 * During IO completion, any of these pages which happen to have been written
421 * back will be redirtied by bio_check_pages_d 458 * back will be redirtied by bio_check_pages_dirty().
422 * 459 *
423 * bios hold a dio reference between submit_bi 460 * bios hold a dio reference between submit_bio and ->end_io.
424 */ 461 */
425 static inline void dio_bio_submit(struct dio * 462 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
426 { 463 {
427 const enum req_op dio_op = dio->opf & <<
428 struct bio *bio = sdio->bio; 464 struct bio *bio = sdio->bio;
429 unsigned long flags; 465 unsigned long flags;
430 466
431 bio->bi_private = dio; 467 bio->bi_private = dio;
432 468
433 spin_lock_irqsave(&dio->bio_lock, flag 469 spin_lock_irqsave(&dio->bio_lock, flags);
434 dio->refcount++; 470 dio->refcount++;
435 spin_unlock_irqrestore(&dio->bio_lock, 471 spin_unlock_irqrestore(&dio->bio_lock, flags);
436 472
437 if (dio->is_async && dio_op == REQ_OP_ !! 473 if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty)
438 bio_set_pages_dirty(bio); 474 bio_set_pages_dirty(bio);
439 475
440 dio->bio_disk = bio->bi_bdev->bd_disk; !! 476 dio->bio_disk = bio->bi_disk;
441 477
442 submit_bio(bio); !! 478 if (sdio->submit_io) {
>> 479 sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio);
>> 480 dio->bio_cookie = BLK_QC_T_NONE;
>> 481 } else
>> 482 dio->bio_cookie = submit_bio(bio);
443 483
444 sdio->bio = NULL; 484 sdio->bio = NULL;
445 sdio->boundary = 0; 485 sdio->boundary = 0;
446 sdio->logical_offset_in_bio = 0; 486 sdio->logical_offset_in_bio = 0;
447 } 487 }
448 488
449 /* 489 /*
450 * Release any resources in case of a failure 490 * Release any resources in case of a failure
451 */ 491 */
452 static inline void dio_cleanup(struct dio *dio 492 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
453 { 493 {
454 if (dio->is_pinned) !! 494 while (sdio->head < sdio->tail)
455 unpin_user_pages(dio->pages + !! 495 put_page(dio->pages[sdio->head++]);
456 sdio->tail - <<
457 sdio->head = sdio->tail; <<
458 } 496 }
459 497
460 /* 498 /*
461 * Wait for the next BIO to complete. Remove 499 * Wait for the next BIO to complete. Remove it and return it. NULL is
462 * returned once all BIOs have been completed. 500 * returned once all BIOs have been completed. This must only be called once
463 * all bios have been issued so that dio->refc 501 * all bios have been issued so that dio->refcount can only decrease. This
464 * requires that the caller hold a reference o !! 502 * requires that that the caller hold a reference on the dio.
465 */ 503 */
466 static struct bio *dio_await_one(struct dio *d 504 static struct bio *dio_await_one(struct dio *dio)
467 { 505 {
468 unsigned long flags; 506 unsigned long flags;
469 struct bio *bio = NULL; 507 struct bio *bio = NULL;
470 508
471 spin_lock_irqsave(&dio->bio_lock, flag 509 spin_lock_irqsave(&dio->bio_lock, flags);
472 510
473 /* 511 /*
474 * Wait as long as the list is empty a 512 * Wait as long as the list is empty and there are bios in flight. bio
475 * completion drops the count, maybe a 513 * completion drops the count, maybe adds to the list, and wakes while
476 * holding the bio_lock so we don't ne 514 * holding the bio_lock so we don't need set_current_state()'s barrier
477 * and can call it after testing our c 515 * and can call it after testing our condition.
478 */ 516 */
479 while (dio->refcount > 1 && dio->bio_l 517 while (dio->refcount > 1 && dio->bio_list == NULL) {
480 __set_current_state(TASK_UNINT 518 __set_current_state(TASK_UNINTERRUPTIBLE);
481 dio->waiter = current; 519 dio->waiter = current;
482 spin_unlock_irqrestore(&dio->b 520 spin_unlock_irqrestore(&dio->bio_lock, flags);
483 blk_io_schedule(); !! 521 if (!(dio->iocb->ki_flags & IOCB_HIPRI) ||
>> 522 !blk_mq_poll(dio->bio_disk->queue, dio->bio_cookie))
>> 523 io_schedule();
484 /* wake up sets us TASK_RUNNIN 524 /* wake up sets us TASK_RUNNING */
485 spin_lock_irqsave(&dio->bio_lo 525 spin_lock_irqsave(&dio->bio_lock, flags);
486 dio->waiter = NULL; 526 dio->waiter = NULL;
487 } 527 }
488 if (dio->bio_list) { 528 if (dio->bio_list) {
489 bio = dio->bio_list; 529 bio = dio->bio_list;
490 dio->bio_list = bio->bi_privat 530 dio->bio_list = bio->bi_private;
491 } 531 }
492 spin_unlock_irqrestore(&dio->bio_lock, 532 spin_unlock_irqrestore(&dio->bio_lock, flags);
493 return bio; 533 return bio;
494 } 534 }
495 535
496 /* 536 /*
497 * Process one completed BIO. No locks are he 537 * Process one completed BIO. No locks are held.
498 */ 538 */
499 static blk_status_t dio_bio_complete(struct di 539 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
500 { 540 {
>> 541 struct bio_vec *bvec;
>> 542 unsigned i;
501 blk_status_t err = bio->bi_status; 543 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 <<
504 544
505 if (err) { 545 if (err) {
506 if (err == BLK_STS_AGAIN && (b 546 if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
507 dio->io_error = -EAGAI 547 dio->io_error = -EAGAIN;
508 else 548 else
509 dio->io_error = -EIO; 549 dio->io_error = -EIO;
510 } 550 }
511 551
512 if (dio->is_async && should_dirty) { !! 552 if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty) {
513 bio_check_pages_dirty(bio); 553 bio_check_pages_dirty(bio); /* transfers ownership */
514 } else { 554 } else {
515 bio_release_pages(bio, should_ !! 555 bio_for_each_segment_all(bvec, bio, i) {
>> 556 struct page *page = bvec->bv_page;
>> 557
>> 558 if (dio->op == REQ_OP_READ && !PageCompound(page) &&
>> 559 dio->should_dirty)
>> 560 set_page_dirty_lock(page);
>> 561 put_page(page);
>> 562 }
516 bio_put(bio); 563 bio_put(bio);
517 } 564 }
518 return err; 565 return err;
519 } 566 }
520 567
521 /* 568 /*
522 * Wait on and process all in-flight BIOs. Th 569 * Wait on and process all in-flight BIOs. This must only be called once
523 * all bios have been issued so that the refco 570 * all bios have been issued so that the refcount can only decrease.
524 * This just waits for all bios to make it thr 571 * This just waits for all bios to make it through dio_bio_complete. IO
525 * errors are propagated through dio->io_error 572 * errors are propagated through dio->io_error and should be propagated via
526 * dio_complete(). 573 * dio_complete().
527 */ 574 */
528 static void dio_await_completion(struct dio *d 575 static void dio_await_completion(struct dio *dio)
529 { 576 {
530 struct bio *bio; 577 struct bio *bio;
531 do { 578 do {
532 bio = dio_await_one(dio); 579 bio = dio_await_one(dio);
533 if (bio) 580 if (bio)
534 dio_bio_complete(dio, 581 dio_bio_complete(dio, bio);
535 } while (bio); 582 } while (bio);
536 } 583 }
537 584
538 /* 585 /*
539 * A really large O_DIRECT read or write can g 586 * A really large O_DIRECT read or write can generate a lot of BIOs. So
540 * to keep the memory consumption sane we peri 587 * to keep the memory consumption sane we periodically reap any completed BIOs
541 * during the BIO generation phase. 588 * during the BIO generation phase.
542 * 589 *
543 * This also helps to limit the peak amount of 590 * This also helps to limit the peak amount of pinned userspace memory.
544 */ 591 */
545 static inline int dio_bio_reap(struct dio *dio 592 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
546 { 593 {
547 int ret = 0; 594 int ret = 0;
548 595
549 if (sdio->reap_counter++ >= 64) { 596 if (sdio->reap_counter++ >= 64) {
550 while (dio->bio_list) { 597 while (dio->bio_list) {
551 unsigned long flags; 598 unsigned long flags;
552 struct bio *bio; 599 struct bio *bio;
553 int ret2; 600 int ret2;
554 601
555 spin_lock_irqsave(&dio 602 spin_lock_irqsave(&dio->bio_lock, flags);
556 bio = dio->bio_list; 603 bio = dio->bio_list;
557 dio->bio_list = bio->b 604 dio->bio_list = bio->bi_private;
558 spin_unlock_irqrestore 605 spin_unlock_irqrestore(&dio->bio_lock, flags);
559 ret2 = blk_status_to_e 606 ret2 = blk_status_to_errno(dio_bio_complete(dio, bio));
560 if (ret == 0) 607 if (ret == 0)
561 ret = ret2; 608 ret = ret2;
562 } 609 }
563 sdio->reap_counter = 0; 610 sdio->reap_counter = 0;
564 } 611 }
565 return ret; 612 return ret;
566 } 613 }
567 614
>> 615 /*
>> 616 * Create workqueue for deferred direct IO completions. We allocate the
>> 617 * workqueue when it's first needed. This avoids creating workqueue for
>> 618 * filesystems that don't need it and also allows us to create the workqueue
>> 619 * late enough so the we can include s_id in the name of the workqueue.
>> 620 */
>> 621 int sb_init_dio_done_wq(struct super_block *sb)
>> 622 {
>> 623 struct workqueue_struct *old;
>> 624 struct workqueue_struct *wq = alloc_workqueue("dio/%s",
>> 625 WQ_MEM_RECLAIM, 0,
>> 626 sb->s_id);
>> 627 if (!wq)
>> 628 return -ENOMEM;
>> 629 /*
>> 630 * This has to be atomic as more DIOs can race to create the workqueue
>> 631 */
>> 632 old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
>> 633 /* Someone created workqueue before us? Free ours... */
>> 634 if (old)
>> 635 destroy_workqueue(wq);
>> 636 return 0;
>> 637 }
>> 638
568 static int dio_set_defer_completion(struct dio 639 static int dio_set_defer_completion(struct dio *dio)
569 { 640 {
570 struct super_block *sb = dio->inode->i 641 struct super_block *sb = dio->inode->i_sb;
571 642
572 if (dio->defer_completion) 643 if (dio->defer_completion)
573 return 0; 644 return 0;
574 dio->defer_completion = true; 645 dio->defer_completion = true;
575 if (!sb->s_dio_done_wq) 646 if (!sb->s_dio_done_wq)
576 return sb_init_dio_done_wq(sb) 647 return sb_init_dio_done_wq(sb);
577 return 0; 648 return 0;
578 } 649 }
579 650
580 /* 651 /*
581 * Call into the fs to map some more disk bloc 652 * Call into the fs to map some more disk blocks. We record the current number
582 * of available blocks at sdio->blocks_availab 653 * of available blocks at sdio->blocks_available. These are in units of the
583 * fs blocksize, i_blocksize(inode). 654 * fs blocksize, i_blocksize(inode).
584 * 655 *
585 * The fs is allowed to map lots of blocks at 656 * 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 657 * it uses the passed inode-relative block number as the file offset, as usual.
587 * 658 *
588 * get_block() is passed the number of i_blkbi 659 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
589 * has remaining to do. The fs should not map 660 * has remaining to do. The fs should not map more than this number of blocks.
590 * 661 *
591 * If the fs has mapped a lot of blocks, it sh 662 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
592 * indicate how much contiguous disk space has 663 * indicate how much contiguous disk space has been made available at
593 * bh->b_blocknr. 664 * bh->b_blocknr.
594 * 665 *
595 * If *any* of the mapped blocks are new, then 666 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
596 * This isn't very efficient... 667 * This isn't very efficient...
597 * 668 *
598 * In the case of filesystem holes: the fs may 669 * In the case of filesystem holes: the fs may return an arbitrarily-large
599 * hole by returning an appropriate value in b 670 * hole by returning an appropriate value in b_size and by clearing
600 * buffer_mapped(). However the direct-io cod 671 * buffer_mapped(). However the direct-io code will only process holes one
601 * block at a time - it will repeatedly call g 672 * block at a time - it will repeatedly call get_block() as it walks the hole.
602 */ 673 */
603 static int get_more_blocks(struct dio *dio, st 674 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
604 struct buffer_head 675 struct buffer_head *map_bh)
605 { 676 {
606 const enum req_op dio_op = dio->opf & <<
607 int ret; 677 int ret;
608 sector_t fs_startblk; /* Into file, 678 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
609 sector_t fs_endblk; /* Into file, 679 sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
610 unsigned long fs_count; /* Number of f 680 unsigned long fs_count; /* Number of filesystem-sized blocks */
611 int create; 681 int create;
612 unsigned int i_blkbits = sdio->blkbits 682 unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
613 loff_t i_size; 683 loff_t i_size;
614 684
615 /* 685 /*
616 * If there was a memory error and we' 686 * If there was a memory error and we've overwritten all the
617 * mapped blocks then we can now retur 687 * mapped blocks then we can now return that memory error
618 */ 688 */
619 ret = dio->page_errors; 689 ret = dio->page_errors;
620 if (ret == 0) { 690 if (ret == 0) {
621 BUG_ON(sdio->block_in_file >= 691 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
622 fs_startblk = sdio->block_in_f 692 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
623 fs_endblk = (sdio->final_block 693 fs_endblk = (sdio->final_block_in_request - 1) >>
624 sdio-> 694 sdio->blkfactor;
625 fs_count = fs_endblk - fs_star 695 fs_count = fs_endblk - fs_startblk + 1;
626 696
627 map_bh->b_state = 0; 697 map_bh->b_state = 0;
628 map_bh->b_size = fs_count << i 698 map_bh->b_size = fs_count << i_blkbits;
629 699
630 /* 700 /*
631 * For writes that could fill 701 * For writes that could fill holes inside i_size on a
632 * DIO_SKIP_HOLES filesystem w 702 * DIO_SKIP_HOLES filesystem we forbid block creations: only
633 * overwrites are permitted. W 703 * overwrites are permitted. We will return early to the caller
634 * once we see an unmapped buf 704 * once we see an unmapped buffer head returned, and the caller
635 * will fall back to buffered 705 * will fall back to buffered I/O.
636 * 706 *
637 * Otherwise the decision is l 707 * Otherwise the decision is left to the get_blocks method,
638 * which may decide to handle 708 * which may decide to handle it or also return an unmapped
639 * buffer head. 709 * buffer head.
640 */ 710 */
641 create = dio_op == REQ_OP_WRIT !! 711 create = dio->op == REQ_OP_WRITE;
642 if (dio->flags & DIO_SKIP_HOLE 712 if (dio->flags & DIO_SKIP_HOLES) {
643 i_size = i_size_read(d 713 i_size = i_size_read(dio->inode);
644 if (i_size && fs_start 714 if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits)
645 create = 0; 715 create = 0;
646 } 716 }
647 717
648 ret = (*sdio->get_block)(dio-> 718 ret = (*sdio->get_block)(dio->inode, fs_startblk,
649 719 map_bh, create);
650 720
651 /* Store for completion */ 721 /* Store for completion */
652 dio->private = map_bh->b_priva 722 dio->private = map_bh->b_private;
653 723
654 if (ret == 0 && buffer_defer_c 724 if (ret == 0 && buffer_defer_completion(map_bh))
655 ret = dio_set_defer_co 725 ret = dio_set_defer_completion(dio);
656 } 726 }
657 return ret; 727 return ret;
658 } 728 }
659 729
660 /* 730 /*
661 * There is no bio. Make one now. 731 * There is no bio. Make one now.
662 */ 732 */
663 static inline int dio_new_bio(struct dio *dio, 733 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
664 sector_t start_sector, struct 734 sector_t start_sector, struct buffer_head *map_bh)
665 { 735 {
666 sector_t sector; 736 sector_t sector;
667 int ret, nr_pages; 737 int ret, nr_pages;
668 738
669 ret = dio_bio_reap(dio, sdio); 739 ret = dio_bio_reap(dio, sdio);
670 if (ret) 740 if (ret)
671 goto out; 741 goto out;
672 sector = start_sector << (sdio->blkbit 742 sector = start_sector << (sdio->blkbits - 9);
673 nr_pages = bio_max_segs(sdio->pages_in !! 743 nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES);
674 BUG_ON(nr_pages <= 0); 744 BUG_ON(nr_pages <= 0);
675 dio_bio_alloc(dio, sdio, map_bh->b_bde 745 dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
676 sdio->boundary = 0; 746 sdio->boundary = 0;
677 out: 747 out:
678 return ret; 748 return ret;
679 } 749 }
680 750
681 /* 751 /*
682 * Attempt to put the current chunk of 'cur_pa 752 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
683 * that was successful then update final_block 753 * that was successful then update final_block_in_bio and take a ref against
684 * the just-added page. 754 * the just-added page.
685 * 755 *
686 * Return zero on success. Non-zero means the 756 * Return zero on success. Non-zero means the caller needs to start a new BIO.
687 */ 757 */
688 static inline int dio_bio_add_page(struct dio !! 758 static inline int dio_bio_add_page(struct dio_submit *sdio)
689 { 759 {
690 int ret; 760 int ret;
691 761
692 ret = bio_add_page(sdio->bio, sdio->cu 762 ret = bio_add_page(sdio->bio, sdio->cur_page,
693 sdio->cur_page_len, sd 763 sdio->cur_page_len, sdio->cur_page_offset);
694 if (ret == sdio->cur_page_len) { 764 if (ret == sdio->cur_page_len) {
695 /* 765 /*
696 * Decrement count only, if we 766 * Decrement count only, if we are done with this page
697 */ 767 */
698 if ((sdio->cur_page_len + sdio 768 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
699 sdio->pages_in_io--; 769 sdio->pages_in_io--;
700 dio_pin_page(dio, sdio->cur_pa !! 770 get_page(sdio->cur_page);
701 sdio->final_block_in_bio = sdi 771 sdio->final_block_in_bio = sdio->cur_page_block +
702 (sdio->cur_page_len >> 772 (sdio->cur_page_len >> sdio->blkbits);
703 ret = 0; 773 ret = 0;
704 } else { 774 } else {
705 ret = 1; 775 ret = 1;
706 } 776 }
707 return ret; 777 return ret;
708 } 778 }
709 779
710 /* 780 /*
711 * Put cur_page under IO. The section of cur_ 781 * 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 782 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
713 * starts on-disk at cur_page_block. 783 * starts on-disk at cur_page_block.
714 * 784 *
715 * We take a ref against the page here (on beh 785 * We take a ref against the page here (on behalf of its presence in the bio).
716 * 786 *
717 * The caller of this function is responsible 787 * The caller of this function is responsible for removing cur_page from the
718 * dio, and for dropping the refcount which ca 788 * dio, and for dropping the refcount which came from that presence.
719 */ 789 */
720 static inline int dio_send_cur_page(struct dio 790 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
721 struct buffer_head *map_bh) 791 struct buffer_head *map_bh)
722 { 792 {
723 int ret = 0; 793 int ret = 0;
724 794
725 if (sdio->bio) { 795 if (sdio->bio) {
726 loff_t cur_offset = sdio->cur_ 796 loff_t cur_offset = sdio->cur_page_fs_offset;
727 loff_t bio_next_offset = sdio- 797 loff_t bio_next_offset = sdio->logical_offset_in_bio +
728 sdio->bio->bi_iter.bi_ 798 sdio->bio->bi_iter.bi_size;
729 799
730 /* 800 /*
731 * See whether this new reques 801 * See whether this new request is contiguous with the old.
732 * 802 *
733 * Btrfs cannot handle having 803 * Btrfs cannot handle having logically non-contiguous requests
734 * submitted. For example if 804 * submitted. For example if you have
735 * 805 *
736 * Logical: [0-4095][HOLE][81 806 * Logical: [0-4095][HOLE][8192-12287]
737 * Physical: [0-4095] [40 807 * Physical: [0-4095] [4096-8191]
738 * 808 *
739 * We cannot submit those page 809 * We cannot submit those pages together as one BIO. So if our
740 * current logical offset in t 810 * current logical offset in the file does not equal what would
741 * be the next logical offset 811 * be the next logical offset in the bio, submit the bio we
742 * have. 812 * have.
743 */ 813 */
744 if (sdio->final_block_in_bio ! 814 if (sdio->final_block_in_bio != sdio->cur_page_block ||
745 cur_offset != bio_next_off 815 cur_offset != bio_next_offset)
746 dio_bio_submit(dio, sd 816 dio_bio_submit(dio, sdio);
747 } 817 }
748 818
749 if (sdio->bio == NULL) { 819 if (sdio->bio == NULL) {
750 ret = dio_new_bio(dio, sdio, s 820 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
751 if (ret) 821 if (ret)
752 goto out; 822 goto out;
753 } 823 }
754 824
755 if (dio_bio_add_page(dio, sdio) != 0) !! 825 if (dio_bio_add_page(sdio) != 0) {
756 dio_bio_submit(dio, sdio); 826 dio_bio_submit(dio, sdio);
757 ret = dio_new_bio(dio, sdio, s 827 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
758 if (ret == 0) { 828 if (ret == 0) {
759 ret = dio_bio_add_page !! 829 ret = dio_bio_add_page(sdio);
760 BUG_ON(ret != 0); 830 BUG_ON(ret != 0);
761 } 831 }
762 } 832 }
763 out: 833 out:
764 return ret; 834 return ret;
765 } 835 }
766 836
767 /* 837 /*
768 * An autonomous function to put a chunk of a 838 * An autonomous function to put a chunk of a page under deferred IO.
769 * 839 *
770 * The caller doesn't actually know (or care) 840 * 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 841 * a BIO, or is under IO or whatever. We just take care of all possible
772 * situations here. The separation between th 842 * situations here. The separation between the logic of do_direct_IO() and
773 * that of submit_page_section() is important 843 * that of submit_page_section() is important for clarity. Please don't break.
774 * 844 *
775 * The chunk of page starts on-disk at blocknr 845 * The chunk of page starts on-disk at blocknr.
776 * 846 *
777 * We perform deferred IO, by recording the la 847 * We perform deferred IO, by recording the last-submitted page inside our
778 * private part of the dio structure. If poss 848 * private part of the dio structure. If possible, we just expand the IO
779 * across that page here. 849 * across that page here.
780 * 850 *
781 * If that doesn't work out then we put the ol 851 * If that doesn't work out then we put the old page into the bio and add this
782 * page to the dio instead. 852 * page to the dio instead.
783 */ 853 */
784 static inline int 854 static inline int
785 submit_page_section(struct dio *dio, struct di 855 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
786 unsigned offset, unsigned 856 unsigned offset, unsigned len, sector_t blocknr,
787 struct buffer_head *map_bh 857 struct buffer_head *map_bh)
788 { 858 {
789 const enum req_op dio_op = dio->opf & <<
790 int ret = 0; 859 int ret = 0;
791 int boundary = sdio->boundary; /* dio 860 int boundary = sdio->boundary; /* dio_send_cur_page may clear it */
792 861
793 if (dio_op == REQ_OP_WRITE) { !! 862 if (dio->op == REQ_OP_WRITE) {
794 /* 863 /*
795 * Read accounting is performe 864 * Read accounting is performed in submit_bio()
796 */ 865 */
797 task_io_account_write(len); 866 task_io_account_write(len);
798 } 867 }
799 868
800 /* 869 /*
801 * Can we just grow the current page's 870 * Can we just grow the current page's presence in the dio?
802 */ 871 */
803 if (sdio->cur_page == page && 872 if (sdio->cur_page == page &&
804 sdio->cur_page_offset + sdio->cur_ 873 sdio->cur_page_offset + sdio->cur_page_len == offset &&
805 sdio->cur_page_block + 874 sdio->cur_page_block +
806 (sdio->cur_page_len >> sdio->blkbi 875 (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
807 sdio->cur_page_len += len; 876 sdio->cur_page_len += len;
808 goto out; 877 goto out;
809 } 878 }
810 879
811 /* 880 /*
812 * If there's a deferred page already 881 * If there's a deferred page already there then send it.
813 */ 882 */
814 if (sdio->cur_page) { 883 if (sdio->cur_page) {
815 ret = dio_send_cur_page(dio, s 884 ret = dio_send_cur_page(dio, sdio, map_bh);
816 dio_unpin_page(dio, sdio->cur_ !! 885 put_page(sdio->cur_page);
817 sdio->cur_page = NULL; 886 sdio->cur_page = NULL;
818 if (ret) 887 if (ret)
819 return ret; 888 return ret;
820 } 889 }
821 890
822 dio_pin_page(dio, page); !! 891 get_page(page); /* It is in dio */
823 sdio->cur_page = page; 892 sdio->cur_page = page;
824 sdio->cur_page_offset = offset; 893 sdio->cur_page_offset = offset;
825 sdio->cur_page_len = len; 894 sdio->cur_page_len = len;
826 sdio->cur_page_block = blocknr; 895 sdio->cur_page_block = blocknr;
827 sdio->cur_page_fs_offset = sdio->block 896 sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
828 out: 897 out:
829 /* 898 /*
830 * If boundary then we want to schedul 899 * If boundary then we want to schedule the IO now to
831 * avoid metadata seeks. 900 * avoid metadata seeks.
832 */ 901 */
833 if (boundary) { 902 if (boundary) {
834 ret = dio_send_cur_page(dio, s 903 ret = dio_send_cur_page(dio, sdio, map_bh);
835 if (sdio->bio) 904 if (sdio->bio)
836 dio_bio_submit(dio, sd 905 dio_bio_submit(dio, sdio);
837 dio_unpin_page(dio, sdio->cur_ !! 906 put_page(sdio->cur_page);
838 sdio->cur_page = NULL; 907 sdio->cur_page = NULL;
839 } 908 }
840 return ret; 909 return ret;
841 } 910 }
842 911
843 /* 912 /*
844 * If we are not writing the entire block and 913 * If we are not writing the entire block and get_block() allocated
845 * the block for us, we need to fill-in the un 914 * the block for us, we need to fill-in the unused portion of the
846 * block with zeros. This happens only if user 915 * block with zeros. This happens only if user-buffer, fileoffset or
847 * io length is not filesystem block-size mult 916 * io length is not filesystem block-size multiple.
848 * 917 *
849 * `end' is zero if we're doing the start of t 918 * `end' is zero if we're doing the start of the IO, 1 at the end of the
850 * IO. 919 * IO.
851 */ 920 */
852 static inline void dio_zero_block(struct dio * 921 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
853 int end, struct buffer_head *m 922 int end, struct buffer_head *map_bh)
854 { 923 {
855 unsigned dio_blocks_per_fs_block; 924 unsigned dio_blocks_per_fs_block;
856 unsigned this_chunk_blocks; /* In 925 unsigned this_chunk_blocks; /* In dio_blocks */
857 unsigned this_chunk_bytes; 926 unsigned this_chunk_bytes;
858 struct page *page; 927 struct page *page;
859 928
860 sdio->start_zero_done = 1; 929 sdio->start_zero_done = 1;
861 if (!sdio->blkfactor || !buffer_new(ma 930 if (!sdio->blkfactor || !buffer_new(map_bh))
862 return; 931 return;
863 932
864 dio_blocks_per_fs_block = 1 << sdio->b 933 dio_blocks_per_fs_block = 1 << sdio->blkfactor;
865 this_chunk_blocks = sdio->block_in_fil 934 this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
866 935
867 if (!this_chunk_blocks) 936 if (!this_chunk_blocks)
868 return; 937 return;
869 938
870 /* 939 /*
871 * We need to zero out part of an fs b 940 * We need to zero out part of an fs block. It is either at the
872 * beginning or the end of the fs bloc 941 * beginning or the end of the fs block.
873 */ 942 */
874 if (end) 943 if (end)
875 this_chunk_blocks = dio_blocks 944 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
876 945
877 this_chunk_bytes = this_chunk_blocks < 946 this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
878 947
879 page = ZERO_PAGE(0); 948 page = ZERO_PAGE(0);
880 if (submit_page_section(dio, sdio, pag 949 if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
881 sdio->next_blo 950 sdio->next_block_for_io, map_bh))
882 return; 951 return;
883 952
884 sdio->next_block_for_io += this_chunk_ 953 sdio->next_block_for_io += this_chunk_blocks;
885 } 954 }
886 955
887 /* 956 /*
888 * Walk the user pages, and the file, mapping 957 * Walk the user pages, and the file, mapping blocks to disk and generating
889 * a sequence of (page,offset,len,block) mappi 958 * a sequence of (page,offset,len,block) mappings. These mappings are injected
890 * into submit_page_section(), which takes car 959 * into submit_page_section(), which takes care of the next stage of submission
891 * 960 *
892 * Direct IO against a blockdev is different f 961 * Direct IO against a blockdev is different from a file. Because we can
893 * happily perform page-sized but 512-byte ali 962 * happily perform page-sized but 512-byte aligned IOs. It is important that
894 * blockdev IO be able to have fine alignment 963 * blockdev IO be able to have fine alignment and large sizes.
895 * 964 *
896 * So what we do is to permit the ->get_block 965 * 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 966 * with the size of IO which is permitted at this offset and this i_blkbits.
898 * 967 *
899 * For best results, the blockdev should be se 968 * 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 969 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
901 * fine alignment but still allows this functi 970 * fine alignment but still allows this function to work in PAGE_SIZE units.
902 */ 971 */
903 static int do_direct_IO(struct dio *dio, struc 972 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
904 struct buffer_head *ma 973 struct buffer_head *map_bh)
905 { 974 {
906 const enum req_op dio_op = dio->opf & <<
907 const unsigned blkbits = sdio->blkbits 975 const unsigned blkbits = sdio->blkbits;
908 const unsigned i_blkbits = blkbits + s 976 const unsigned i_blkbits = blkbits + sdio->blkfactor;
909 int ret = 0; 977 int ret = 0;
910 978
911 while (sdio->block_in_file < sdio->fin 979 while (sdio->block_in_file < sdio->final_block_in_request) {
912 struct page *page; 980 struct page *page;
913 size_t from, to; 981 size_t from, to;
914 982
915 page = dio_get_page(dio, sdio) 983 page = dio_get_page(dio, sdio);
916 if (IS_ERR(page)) { 984 if (IS_ERR(page)) {
917 ret = PTR_ERR(page); 985 ret = PTR_ERR(page);
918 goto out; 986 goto out;
919 } 987 }
920 from = sdio->head ? 0 : sdio-> 988 from = sdio->head ? 0 : sdio->from;
921 to = (sdio->head == sdio->tail 989 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
922 sdio->head++; 990 sdio->head++;
923 991
924 while (from < to) { 992 while (from < to) {
925 unsigned this_chunk_by 993 unsigned this_chunk_bytes; /* # of bytes mapped */
926 unsigned this_chunk_bl 994 unsigned this_chunk_blocks; /* # of blocks */
927 unsigned u; 995 unsigned u;
928 996
929 if (sdio->blocks_avail 997 if (sdio->blocks_available == 0) {
930 /* 998 /*
931 * Need to go 999 * Need to go and map some more disk
932 */ 1000 */
933 unsigned long 1001 unsigned long blkmask;
934 unsigned long 1002 unsigned long dio_remainder;
935 1003
936 ret = get_more 1004 ret = get_more_blocks(dio, sdio, map_bh);
937 if (ret) { 1005 if (ret) {
938 dio_un !! 1006 put_page(page);
939 goto o 1007 goto out;
940 } 1008 }
941 if (!buffer_ma 1009 if (!buffer_mapped(map_bh))
942 goto d 1010 goto do_holes;
943 1011
944 sdio->blocks_a 1012 sdio->blocks_available =
945 1013 map_bh->b_size >> blkbits;
946 sdio->next_blo 1014 sdio->next_block_for_io =
947 map_bh 1015 map_bh->b_blocknr << sdio->blkfactor;
948 if (buffer_new 1016 if (buffer_new(map_bh)) {
949 clean_ 1017 clean_bdev_aliases(
950 1018 map_bh->b_bdev,
951 1019 map_bh->b_blocknr,
952 1020 map_bh->b_size >> i_blkbits);
953 } 1021 }
954 1022
955 if (!sdio->blk 1023 if (!sdio->blkfactor)
956 goto d 1024 goto do_holes;
957 1025
958 blkmask = (1 < 1026 blkmask = (1 << sdio->blkfactor) - 1;
959 dio_remainder 1027 dio_remainder = (sdio->block_in_file & blkmask);
960 1028
961 /* 1029 /*
962 * If we are a 1030 * If we are at the start of IO and that IO
963 * starts part 1031 * starts partway into a fs-block,
964 * dio_remaind 1032 * dio_remainder will be non-zero. If the IO
965 * is a read t 1033 * is a read then we can simply advance the IO
966 * cursor to t 1034 * cursor to the first block which is to be
967 * read. But 1035 * read. But if the IO is a write and the
968 * block was n 1036 * block was newly allocated we cannot do that;
969 * the start o 1037 * the start of the fs block must be zeroed out
970 * on-disk 1038 * on-disk
971 */ 1039 */
972 if (!buffer_ne 1040 if (!buffer_new(map_bh))
973 sdio-> 1041 sdio->next_block_for_io += dio_remainder;
974 sdio->blocks_a 1042 sdio->blocks_available -= dio_remainder;
975 } 1043 }
976 do_holes: 1044 do_holes:
977 /* Handle holes */ 1045 /* Handle holes */
978 if (!buffer_mapped(map 1046 if (!buffer_mapped(map_bh)) {
979 loff_t i_size_ 1047 loff_t i_size_aligned;
980 1048
981 /* AKPM: eargh 1049 /* AKPM: eargh, -ENOTBLK is a hack */
982 if (dio_op == !! 1050 if (dio->op == REQ_OP_WRITE) {
983 dio_un !! 1051 put_page(page);
984 return 1052 return -ENOTBLK;
985 } 1053 }
986 1054
987 /* 1055 /*
988 * Be sure to 1056 * Be sure to account for a partial block as the
989 * last block 1057 * last block in the file
990 */ 1058 */
991 i_size_aligned 1059 i_size_aligned = ALIGN(i_size_read(dio->inode),
992 1060 1 << blkbits);
993 if (sdio->bloc 1061 if (sdio->block_in_file >=
994 1062 i_size_aligned >> blkbits) {
995 /* We 1063 /* We hit eof */
996 dio_un !! 1064 put_page(page);
997 goto o 1065 goto out;
998 } 1066 }
999 zero_user(page 1067 zero_user(page, from, 1 << blkbits);
1000 sdio->block_i 1068 sdio->block_in_file++;
1001 from += 1 << 1069 from += 1 << blkbits;
1002 dio->result + 1070 dio->result += 1 << blkbits;
1003 goto next_blo 1071 goto next_block;
1004 } 1072 }
1005 1073
1006 /* 1074 /*
1007 * If we're performin 1075 * If we're performing IO which has an alignment which
1008 * is finer than the 1076 * is finer than the underlying fs, go check to see if
1009 * we must zero out t 1077 * we must zero out the start of this block.
1010 */ 1078 */
1011 if (unlikely(sdio->bl 1079 if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1012 dio_zero_bloc 1080 dio_zero_block(dio, sdio, 0, map_bh);
1013 1081
1014 /* 1082 /*
1015 * Work out, in this_ 1083 * Work out, in this_chunk_blocks, how much disk we
1016 * can add to this pa 1084 * can add to this page
1017 */ 1085 */
1018 this_chunk_blocks = s 1086 this_chunk_blocks = sdio->blocks_available;
1019 u = (to - from) >> bl 1087 u = (to - from) >> blkbits;
1020 if (this_chunk_blocks 1088 if (this_chunk_blocks > u)
1021 this_chunk_bl 1089 this_chunk_blocks = u;
1022 u = sdio->final_block 1090 u = sdio->final_block_in_request - sdio->block_in_file;
1023 if (this_chunk_blocks 1091 if (this_chunk_blocks > u)
1024 this_chunk_bl 1092 this_chunk_blocks = u;
1025 this_chunk_bytes = th 1093 this_chunk_bytes = this_chunk_blocks << blkbits;
1026 BUG_ON(this_chunk_byt 1094 BUG_ON(this_chunk_bytes == 0);
1027 1095
1028 if (this_chunk_blocks 1096 if (this_chunk_blocks == sdio->blocks_available)
1029 sdio->boundar 1097 sdio->boundary = buffer_boundary(map_bh);
1030 ret = submit_page_sec 1098 ret = submit_page_section(dio, sdio, page,
1031 1099 from,
1032 1100 this_chunk_bytes,
1033 1101 sdio->next_block_for_io,
1034 1102 map_bh);
1035 if (ret) { 1103 if (ret) {
1036 dio_unpin_pag !! 1104 put_page(page);
1037 goto out; 1105 goto out;
1038 } 1106 }
1039 sdio->next_block_for_ 1107 sdio->next_block_for_io += this_chunk_blocks;
1040 1108
1041 sdio->block_in_file + 1109 sdio->block_in_file += this_chunk_blocks;
1042 from += this_chunk_by 1110 from += this_chunk_bytes;
1043 dio->result += this_c 1111 dio->result += this_chunk_bytes;
1044 sdio->blocks_availabl 1112 sdio->blocks_available -= this_chunk_blocks;
1045 next_block: 1113 next_block:
1046 BUG_ON(sdio->block_in 1114 BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1047 if (sdio->block_in_fi 1115 if (sdio->block_in_file == sdio->final_block_in_request)
1048 break; 1116 break;
1049 } 1117 }
1050 1118
1051 /* Drop the pin which was tak !! 1119 /* Drop the ref which was taken in get_user_pages() */
1052 dio_unpin_page(dio, page); !! 1120 put_page(page);
1053 } 1121 }
1054 out: 1122 out:
1055 return ret; 1123 return ret;
1056 } 1124 }
1057 1125
1058 static inline int drop_refcount(struct dio *d 1126 static inline int drop_refcount(struct dio *dio)
1059 { 1127 {
1060 int ret2; 1128 int ret2;
1061 unsigned long flags; 1129 unsigned long flags;
1062 1130
1063 /* 1131 /*
1064 * Sync will always be dropping the f 1132 * Sync will always be dropping the final ref and completing the
1065 * operation. AIO can if it was a br 1133 * operation. AIO can if it was a broken operation described above or
1066 * in fact if all the bios race to co 1134 * in fact if all the bios race to complete before we get here. In
1067 * that case dio_complete() translate 1135 * that case dio_complete() translates the EIOCBQUEUED into the proper
1068 * return code that the caller will h 1136 * return code that the caller will hand to ->complete().
1069 * 1137 *
1070 * This is managed by the bio_lock in 1138 * This is managed by the bio_lock instead of being an atomic_t so that
1071 * completion paths can drop their re 1139 * completion paths can drop their ref and use the remaining count to
1072 * decide to wake the submission path 1140 * decide to wake the submission path atomically.
1073 */ 1141 */
1074 spin_lock_irqsave(&dio->bio_lock, fla 1142 spin_lock_irqsave(&dio->bio_lock, flags);
1075 ret2 = --dio->refcount; 1143 ret2 = --dio->refcount;
1076 spin_unlock_irqrestore(&dio->bio_lock 1144 spin_unlock_irqrestore(&dio->bio_lock, flags);
1077 return ret2; 1145 return ret2;
1078 } 1146 }
1079 1147
1080 /* 1148 /*
1081 * This is a library function for use by file 1149 * This is a library function for use by filesystem drivers.
1082 * 1150 *
1083 * The locking rules are governed by the flag 1151 * The locking rules are governed by the flags parameter:
1084 * - if the flags value contains DIO_LOCKING 1152 * - if the flags value contains DIO_LOCKING we use a fancy locking
1085 * scheme for dumb filesystems. 1153 * scheme for dumb filesystems.
1086 * For writes this function is called unde 1154 * For writes this function is called under i_mutex and returns with
1087 * i_mutex held, for reads, i_mutex is not 1155 * i_mutex held, for reads, i_mutex is not held on entry, but it is
1088 * taken and dropped again before returnin 1156 * taken and dropped again before returning.
1089 * - if the flags value does NOT contain DIO 1157 * - if the flags value does NOT contain DIO_LOCKING we don't use any
1090 * internal locking but rather rely on the 1158 * internal locking but rather rely on the filesystem to synchronize
1091 * direct I/O reads/writes versus each oth 1159 * direct I/O reads/writes versus each other and truncate.
1092 * 1160 *
1093 * To help with locking against truncate we i 1161 * To help with locking against truncate we incremented the i_dio_count
1094 * counter before starting direct I/O, and de 1162 * counter before starting direct I/O, and decrement it once we are done.
1095 * Truncate can wait for it to reach zero to 1163 * Truncate can wait for it to reach zero to provide exclusion. It is
1096 * expected that filesystem provide exclusion 1164 * expected that filesystem provide exclusion between new direct I/O
1097 * and truncates. For DIO_LOCKING filesystem 1165 * and truncates. For DIO_LOCKING filesystems this is done by i_mutex,
1098 * but other filesystems need to take care of 1166 * but other filesystems need to take care of this on their own.
1099 * 1167 *
1100 * NOTE: if you pass "sdio" to anything by po 1168 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1101 * is always inlined. Otherwise gcc is unable 1169 * is always inlined. Otherwise gcc is unable to split the structure into
1102 * individual fields and will generate much w 1170 * individual fields and will generate much worse code. This is important
1103 * for the whole file. 1171 * for the whole file.
1104 */ 1172 */
1105 ssize_t __blockdev_direct_IO(struct kiocb *io !! 1173 static inline ssize_t
1106 struct block_device *bdev, st !! 1174 do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1107 get_block_t get_block, dio_io !! 1175 struct block_device *bdev, struct iov_iter *iter,
1108 int flags) !! 1176 get_block_t get_block, dio_iodone_t end_io,
>> 1177 dio_submit_t submit_io, int flags)
1109 { 1178 {
1110 unsigned i_blkbits = READ_ONCE(inode- !! 1179 unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1111 unsigned blkbits = i_blkbits; 1180 unsigned blkbits = i_blkbits;
1112 unsigned blocksize_mask = (1 << blkbi 1181 unsigned blocksize_mask = (1 << blkbits) - 1;
1113 ssize_t retval = -EINVAL; 1182 ssize_t retval = -EINVAL;
1114 const size_t count = iov_iter_count(i !! 1183 size_t count = iov_iter_count(iter);
1115 loff_t offset = iocb->ki_pos; 1184 loff_t offset = iocb->ki_pos;
1116 const loff_t end = offset + count; !! 1185 loff_t end = offset + count;
1117 struct dio *dio; 1186 struct dio *dio;
1118 struct dio_submit sdio = { NULL, }; !! 1187 struct dio_submit sdio = { 0, };
1119 struct buffer_head map_bh = { 0, }; 1188 struct buffer_head map_bh = { 0, };
1120 struct blk_plug plug; 1189 struct blk_plug plug;
1121 unsigned long align = offset | iov_it 1190 unsigned long align = offset | iov_iter_alignment(iter);
1122 1191
>> 1192 /*
>> 1193 * Avoid references to bdev if not absolutely needed to give
>> 1194 * the early prefetch in the caller enough time.
>> 1195 */
>> 1196
>> 1197 if (align & blocksize_mask) {
>> 1198 if (bdev)
>> 1199 blkbits = blksize_bits(bdev_logical_block_size(bdev));
>> 1200 blocksize_mask = (1 << blkbits) - 1;
>> 1201 if (align & blocksize_mask)
>> 1202 goto out;
>> 1203 }
>> 1204
1123 /* watch out for a 0 len io from a tr 1205 /* watch out for a 0 len io from a tricksy fs */
1124 if (iov_iter_rw(iter) == READ && !cou !! 1206 if (iov_iter_rw(iter) == READ && !iov_iter_count(iter))
1125 return 0; 1207 return 0;
1126 1208
1127 dio = kmem_cache_alloc(dio_cache, GFP 1209 dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
>> 1210 retval = -ENOMEM;
1128 if (!dio) 1211 if (!dio)
1129 return -ENOMEM; !! 1212 goto out;
1130 /* 1213 /*
1131 * Believe it or not, zeroing out the 1214 * Believe it or not, zeroing out the page array caused a .5%
1132 * performance regression in a databa 1215 * performance regression in a database benchmark. So, we take
1133 * care to only zero out what's neede 1216 * care to only zero out what's needed.
1134 */ 1217 */
1135 memset(dio, 0, offsetof(struct dio, p 1218 memset(dio, 0, offsetof(struct dio, pages));
1136 1219
1137 dio->flags = flags; 1220 dio->flags = flags;
1138 if (dio->flags & DIO_LOCKING && iov_i !! 1221 if (dio->flags & DIO_LOCKING) {
1139 /* will be released by direct !! 1222 if (iov_iter_rw(iter) == READ) {
1140 inode_lock(inode); !! 1223 struct address_space *mapping =
>> 1224 iocb->ki_filp->f_mapping;
>> 1225
>> 1226 /* will be released by direct_io_worker */
>> 1227 inode_lock(inode);
>> 1228
>> 1229 retval = filemap_write_and_wait_range(mapping, offset,
>> 1230 end - 1);
>> 1231 if (retval) {
>> 1232 inode_unlock(inode);
>> 1233 kmem_cache_free(dio_cache, dio);
>> 1234 goto out;
>> 1235 }
>> 1236 }
1141 } 1237 }
1142 dio->is_pinned = iov_iter_extract_wil <<
1143 1238
1144 /* Once we sampled i_size check for r 1239 /* Once we sampled i_size check for reads beyond EOF */
1145 dio->i_size = i_size_read(inode); 1240 dio->i_size = i_size_read(inode);
1146 if (iov_iter_rw(iter) == READ && offs 1241 if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
>> 1242 if (dio->flags & DIO_LOCKING)
>> 1243 inode_unlock(inode);
>> 1244 kmem_cache_free(dio_cache, dio);
1147 retval = 0; 1245 retval = 0;
1148 goto fail_dio; !! 1246 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 } 1247 }
1166 1248
1167 /* 1249 /*
1168 * For file extending writes updating 1250 * For file extending writes updating i_size before data writeouts
1169 * complete can expose uninitialized 1251 * complete can expose uninitialized blocks in dumb filesystems.
1170 * In that case we need to wait for I 1252 * In that case we need to wait for I/O completion even if asked
1171 * for an asynchronous write. 1253 * for an asynchronous write.
1172 */ 1254 */
1173 if (is_sync_kiocb(iocb)) 1255 if (is_sync_kiocb(iocb))
1174 dio->is_async = false; 1256 dio->is_async = false;
1175 else if (iov_iter_rw(iter) == WRITE & !! 1257 else if (!(dio->flags & DIO_ASYNC_EXTEND) &&
>> 1258 iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
1176 dio->is_async = false; 1259 dio->is_async = false;
1177 else 1260 else
1178 dio->is_async = true; 1261 dio->is_async = true;
1179 1262
1180 dio->inode = inode; 1263 dio->inode = inode;
1181 if (iov_iter_rw(iter) == WRITE) { 1264 if (iov_iter_rw(iter) == WRITE) {
1182 dio->opf = REQ_OP_WRITE | REQ !! 1265 dio->op = REQ_OP_WRITE;
>> 1266 dio->op_flags = REQ_SYNC | REQ_IDLE;
1183 if (iocb->ki_flags & IOCB_NOW 1267 if (iocb->ki_flags & IOCB_NOWAIT)
1184 dio->opf |= REQ_NOWAI !! 1268 dio->op_flags |= REQ_NOWAIT;
1185 } else { 1269 } else {
1186 dio->opf = REQ_OP_READ; !! 1270 dio->op = REQ_OP_READ;
1187 } 1271 }
1188 1272
1189 /* 1273 /*
1190 * For AIO O_(D)SYNC writes we need t 1274 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1191 * so that we can call ->fsync. 1275 * so that we can call ->fsync.
1192 */ 1276 */
1193 if (dio->is_async && iov_iter_rw(iter 1277 if (dio->is_async && iov_iter_rw(iter) == WRITE) {
1194 retval = 0; 1278 retval = 0;
1195 if (iocb_is_dsync(iocb)) !! 1279 if (iocb->ki_flags & IOCB_DSYNC)
1196 retval = dio_set_defe 1280 retval = dio_set_defer_completion(dio);
1197 else if (!dio->inode->i_sb->s 1281 else if (!dio->inode->i_sb->s_dio_done_wq) {
1198 /* 1282 /*
1199 * In case of AIO wri 1283 * In case of AIO write racing with buffered read we
1200 * need to defer comp 1284 * need to defer completion. We can't decide this now,
1201 * however the workqu 1285 * however the workqueue needs to be initialized here.
1202 */ 1286 */
1203 retval = sb_init_dio_ 1287 retval = sb_init_dio_done_wq(dio->inode->i_sb);
1204 } 1288 }
1205 if (retval) !! 1289 if (retval) {
1206 goto fail_dio; !! 1290 /*
>> 1291 * We grab i_mutex only for reads so we don't have
>> 1292 * to release it here
>> 1293 */
>> 1294 kmem_cache_free(dio_cache, dio);
>> 1295 goto out;
>> 1296 }
1207 } 1297 }
1208 1298
1209 /* 1299 /*
1210 * Will be decremented at I/O complet 1300 * Will be decremented at I/O completion time.
1211 */ 1301 */
1212 inode_dio_begin(inode); !! 1302 if (!(dio->flags & DIO_SKIP_DIO_COUNT))
>> 1303 inode_dio_begin(inode);
1213 1304
>> 1305 retval = 0;
1214 sdio.blkbits = blkbits; 1306 sdio.blkbits = blkbits;
1215 sdio.blkfactor = i_blkbits - blkbits; 1307 sdio.blkfactor = i_blkbits - blkbits;
1216 sdio.block_in_file = offset >> blkbit 1308 sdio.block_in_file = offset >> blkbits;
1217 1309
1218 sdio.get_block = get_block; 1310 sdio.get_block = get_block;
1219 dio->end_io = end_io; 1311 dio->end_io = end_io;
>> 1312 sdio.submit_io = submit_io;
1220 sdio.final_block_in_bio = -1; 1313 sdio.final_block_in_bio = -1;
1221 sdio.next_block_for_io = -1; 1314 sdio.next_block_for_io = -1;
1222 1315
1223 dio->iocb = iocb; 1316 dio->iocb = iocb;
1224 1317
1225 spin_lock_init(&dio->bio_lock); 1318 spin_lock_init(&dio->bio_lock);
1226 dio->refcount = 1; 1319 dio->refcount = 1;
1227 1320
1228 dio->should_dirty = user_backed_iter( !! 1321 dio->should_dirty = (iter->type == ITER_IOVEC);
1229 sdio.iter = iter; 1322 sdio.iter = iter;
1230 sdio.final_block_in_request = end >> !! 1323 sdio.final_block_in_request =
>> 1324 (offset + iov_iter_count(iter)) >> blkbits;
1231 1325
1232 /* 1326 /*
1233 * In case of non-aligned buffers, we 1327 * In case of non-aligned buffers, we may need 2 more
1234 * pages since we need to zero out fi 1328 * pages since we need to zero out first and last block.
1235 */ 1329 */
1236 if (unlikely(sdio.blkfactor)) 1330 if (unlikely(sdio.blkfactor))
1237 sdio.pages_in_io = 2; 1331 sdio.pages_in_io = 2;
1238 1332
1239 sdio.pages_in_io += iov_iter_npages(i 1333 sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
1240 1334
1241 blk_start_plug(&plug); 1335 blk_start_plug(&plug);
1242 1336
1243 retval = do_direct_IO(dio, &sdio, &ma 1337 retval = do_direct_IO(dio, &sdio, &map_bh);
1244 if (retval) 1338 if (retval)
1245 dio_cleanup(dio, &sdio); 1339 dio_cleanup(dio, &sdio);
1246 1340
1247 if (retval == -ENOTBLK) { 1341 if (retval == -ENOTBLK) {
1248 /* 1342 /*
1249 * The remaining part of the 1343 * The remaining part of the request will be
1250 * handled by buffered I/O wh !! 1344 * be handled by buffered I/O when we return
1251 */ 1345 */
1252 retval = 0; 1346 retval = 0;
1253 } 1347 }
1254 /* 1348 /*
1255 * There may be some unwritten disk a 1349 * There may be some unwritten disk at the end of a part-written
1256 * fs-block-sized block. Go zero tha 1350 * fs-block-sized block. Go zero that now.
1257 */ 1351 */
1258 dio_zero_block(dio, &sdio, 1, &map_bh 1352 dio_zero_block(dio, &sdio, 1, &map_bh);
1259 1353
1260 if (sdio.cur_page) { 1354 if (sdio.cur_page) {
1261 ssize_t ret2; 1355 ssize_t ret2;
1262 1356
1263 ret2 = dio_send_cur_page(dio, 1357 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1264 if (retval == 0) 1358 if (retval == 0)
1265 retval = ret2; 1359 retval = ret2;
1266 dio_unpin_page(dio, sdio.cur_ !! 1360 put_page(sdio.cur_page);
1267 sdio.cur_page = NULL; 1361 sdio.cur_page = NULL;
1268 } 1362 }
1269 if (sdio.bio) 1363 if (sdio.bio)
1270 dio_bio_submit(dio, &sdio); 1364 dio_bio_submit(dio, &sdio);
1271 1365
1272 blk_finish_plug(&plug); 1366 blk_finish_plug(&plug);
1273 1367
1274 /* 1368 /*
1275 * It is possible that, we return sho 1369 * It is possible that, we return short IO due to end of file.
1276 * In that case, we need to release a 1370 * In that case, we need to release all the pages we got hold on.
1277 */ 1371 */
1278 dio_cleanup(dio, &sdio); 1372 dio_cleanup(dio, &sdio);
1279 1373
1280 /* 1374 /*
1281 * All block lookups have been perfor 1375 * All block lookups have been performed. For READ requests
1282 * we can let i_mutex go now that its 1376 * we can let i_mutex go now that its achieved its purpose
1283 * of protecting us from looking up u 1377 * of protecting us from looking up uninitialized blocks.
1284 */ 1378 */
1285 if (iov_iter_rw(iter) == READ && (dio 1379 if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
1286 inode_unlock(dio->inode); 1380 inode_unlock(dio->inode);
1287 1381
1288 /* 1382 /*
1289 * The only time we want to leave bio 1383 * The only time we want to leave bios in flight is when a successful
1290 * partial aio read or full aio write 1384 * partial aio read or full aio write have been setup. In that case
1291 * bio completion will call aio_compl 1385 * bio completion will call aio_complete. The only time it's safe to
1292 * call aio_complete is when we retur 1386 * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1293 * This had *better* be the only plac 1387 * This had *better* be the only place that raises -EIOCBQUEUED.
1294 */ 1388 */
1295 BUG_ON(retval == -EIOCBQUEUED); 1389 BUG_ON(retval == -EIOCBQUEUED);
1296 if (dio->is_async && retval == 0 && d 1390 if (dio->is_async && retval == 0 && dio->result &&
1297 (iov_iter_rw(iter) == READ || dio 1391 (iov_iter_rw(iter) == READ || dio->result == count))
1298 retval = -EIOCBQUEUED; 1392 retval = -EIOCBQUEUED;
1299 else 1393 else
1300 dio_await_completion(dio); 1394 dio_await_completion(dio);
1301 1395
1302 if (drop_refcount(dio) == 0) { 1396 if (drop_refcount(dio) == 0) {
1303 retval = dio_complete(dio, re 1397 retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE);
1304 } else 1398 } else
1305 BUG_ON(retval != -EIOCBQUEUED 1399 BUG_ON(retval != -EIOCBQUEUED);
1306 1400
>> 1401 out:
1307 return retval; 1402 return retval;
>> 1403 }
1308 1404
1309 fail_dio: !! 1405 ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1310 if (dio->flags & DIO_LOCKING && iov_i !! 1406 struct block_device *bdev, struct iov_iter *iter,
1311 inode_unlock(inode); !! 1407 get_block_t get_block,
>> 1408 dio_iodone_t end_io, dio_submit_t submit_io,
>> 1409 int flags)
>> 1410 {
>> 1411 /*
>> 1412 * The block device state is needed in the end to finally
>> 1413 * submit everything. Since it's likely to be cache cold
>> 1414 * prefetch it here as first thing to hide some of the
>> 1415 * latency.
>> 1416 *
>> 1417 * Attempt to prefetch the pieces we likely need later.
>> 1418 */
>> 1419 prefetch(&bdev->bd_disk->part_tbl);
>> 1420 prefetch(bdev->bd_queue);
>> 1421 prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1312 1422
1313 kmem_cache_free(dio_cache, dio); !! 1423 return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block,
1314 return retval; !! 1424 end_io, submit_io, flags);
1315 } 1425 }
>> 1426
1316 EXPORT_SYMBOL(__blockdev_direct_IO); 1427 EXPORT_SYMBOL(__blockdev_direct_IO);
1317 1428
1318 static __init int dio_init(void) 1429 static __init int dio_init(void)
1319 { 1430 {
1320 dio_cache = KMEM_CACHE(dio, SLAB_PANI 1431 dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1321 return 0; 1432 return 0;
1322 } 1433 }
1323 module_init(dio_init) 1434 module_init(dio_init)
1324 1435
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.