~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/fs/direct-io.c

Version: ~ [ linux-5.15-rc5 ] ~ [ linux-5.14.11 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.72 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.152 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.210 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.250 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.286 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.288 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 // SPDX-License-Identifier: GPL-2.0-only
  2 /*
  3  * fs/direct-io.c
  4  *
  5  * Copyright (C) 2002, Linus Torvalds.
  6  *
  7  * O_DIRECT
  8  *
  9  * 04Jul2002    Andrew Morton
 10  *              Initial version
 11  * 11Sep2002    janetinc@us.ibm.com
 12  *              added readv/writev support.
 13  * 29Oct2002    Andrew Morton
 14  *              rewrote bio_add_page() support.
 15  * 30Oct2002    pbadari@us.ibm.com
 16  *              added support for non-aligned IO.
 17  * 06Nov2002    pbadari@us.ibm.com
 18  *              added asynchronous IO support.
 19  * 21Jul2003    nathans@sgi.com
 20  *              added IO completion notifier.
 21  */
 22 
 23 #include <linux/kernel.h>
 24 #include <linux/module.h>
 25 #include <linux/types.h>
 26 #include <linux/fs.h>
 27 #include <linux/mm.h>
 28 #include <linux/slab.h>
 29 #include <linux/highmem.h>
 30 #include <linux/pagemap.h>
 31 #include <linux/task_io_accounting_ops.h>
 32 #include <linux/bio.h>
 33 #include <linux/wait.h>
 34 #include <linux/err.h>
 35 #include <linux/blkdev.h>
 36 #include <linux/buffer_head.h>
 37 #include <linux/rwsem.h>
 38 #include <linux/uio.h>
 39 #include <linux/atomic.h>
 40 #include <linux/prefetch.h>
 41 
 42 #include "internal.h"
 43 
 44 /*
 45  * How many user pages to map in one call to get_user_pages().  This determines
 46  * the size of a structure in the slab cache
 47  */
 48 #define DIO_PAGES       64
 49 
 50 /*
 51  * Flags for dio_complete()
 52  */
 53 #define DIO_COMPLETE_ASYNC              0x01    /* This is async IO */
 54 #define DIO_COMPLETE_INVALIDATE         0x02    /* Can invalidate pages */
 55 
 56 /*
 57  * This code generally works in units of "dio_blocks".  A dio_block is
 58  * somewhere between the hard sector size and the filesystem block size.  it
 59  * is determined on a per-invocation basis.   When talking to the filesystem
 60  * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
 61  * down by dio->blkfactor.  Similarly, fs-blocksize quantities are converted
 62  * to bio_block quantities by shifting left by blkfactor.
 63  *
 64  * If blkfactor is zero then the user's request was aligned to the filesystem's
 65  * blocksize.
 66  */
 67 
 68 /* dio_state only used in the submission path */
 69 
 70 struct dio_submit {
 71         struct bio *bio;                /* bio under assembly */
 72         unsigned blkbits;               /* doesn't change */
 73         unsigned blkfactor;             /* When we're using an alignment which
 74                                            is finer than the filesystem's soft
 75                                            blocksize, this specifies how much
 76                                            finer.  blkfactor=2 means 1/4-block
 77                                            alignment.  Does not change */
 78         unsigned start_zero_done;       /* flag: sub-blocksize zeroing has
 79                                            been performed at the start of a
 80                                            write */
 81         int pages_in_io;                /* approximate total IO pages */
 82         sector_t block_in_file;         /* Current offset into the underlying
 83                                            file in dio_block units. */
 84         unsigned blocks_available;      /* At block_in_file.  changes */
 85         int reap_counter;               /* rate limit reaping */
 86         sector_t final_block_in_request;/* doesn't change */
 87         int boundary;                   /* prev block is at a boundary */
 88         get_block_t *get_block;         /* block mapping function */
 89         dio_submit_t *submit_io;        /* IO submition function */
 90 
 91         loff_t logical_offset_in_bio;   /* current first logical block in bio */
 92         sector_t final_block_in_bio;    /* current final block in bio + 1 */
 93         sector_t next_block_for_io;     /* next block to be put under IO,
 94                                            in dio_blocks units */
 95 
 96         /*
 97          * Deferred addition of a page to the dio.  These variables are
 98          * private to dio_send_cur_page(), submit_page_section() and
 99          * dio_bio_add_page().
100          */
101         struct page *cur_page;          /* The page */
102         unsigned cur_page_offset;       /* Offset into it, in bytes */
103         unsigned cur_page_len;          /* Nr of bytes at cur_page_offset */
104         sector_t cur_page_block;        /* Where it starts */
105         loff_t cur_page_fs_offset;      /* Offset in file */
106 
107         struct iov_iter *iter;
108         /*
109          * Page queue.  These variables belong to dio_refill_pages() and
110          * dio_get_page().
111          */
112         unsigned head;                  /* next page to process */
113         unsigned tail;                  /* last valid page + 1 */
114         size_t from, to;
115 };
116 
117 /* dio_state communicated between submission path and end_io */
118 struct dio {
119         int flags;                      /* doesn't change */
120         int op;
121         int op_flags;
122         blk_qc_t bio_cookie;
123         struct gendisk *bio_disk;
124         struct inode *inode;
125         loff_t i_size;                  /* i_size when submitted */
126         dio_iodone_t *end_io;           /* IO completion function */
127 
128         void *private;                  /* copy from map_bh.b_private */
129 
130         /* BIO completion state */
131         spinlock_t bio_lock;            /* protects BIO fields below */
132         int page_errors;                /* errno from get_user_pages() */
133         int is_async;                   /* is IO async ? */
134         bool defer_completion;          /* defer AIO completion to workqueue? */
135         bool should_dirty;              /* if pages should be dirtied */
136         int io_error;                   /* IO error in completion path */
137         unsigned long refcount;         /* direct_io_worker() and bios */
138         struct bio *bio_list;           /* singly linked via bi_private */
139         struct task_struct *waiter;     /* waiting task (NULL if none) */
140 
141         /* AIO related stuff */
142         struct kiocb *iocb;             /* kiocb */
143         ssize_t result;                 /* IO result */
144 
145         /*
146          * pages[] (and any fields placed after it) are not zeroed out at
147          * allocation time.  Don't add new fields after pages[] unless you
148          * wish that they not be zeroed.
149          */
150         union {
151                 struct page *pages[DIO_PAGES];  /* page buffer */
152                 struct work_struct complete_work;/* deferred AIO completion */
153         };
154 } ____cacheline_aligned_in_smp;
155 
156 static struct kmem_cache *dio_cache __read_mostly;
157 
158 /*
159  * How many pages are in the queue?
160  */
161 static inline unsigned dio_pages_present(struct dio_submit *sdio)
162 {
163         return sdio->tail - sdio->head;
164 }
165 
166 /*
167  * Go grab and pin some userspace pages.   Typically we'll get 64 at a time.
168  */
169 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
170 {
171         ssize_t ret;
172 
173         ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
174                                 &sdio->from);
175 
176         if (ret < 0 && sdio->blocks_available && (dio->op == REQ_OP_WRITE)) {
177                 struct page *page = ZERO_PAGE(0);
178                 /*
179                  * A memory fault, but the filesystem has some outstanding
180                  * mapped blocks.  We need to use those blocks up to avoid
181                  * leaking stale data in the file.
182                  */
183                 if (dio->page_errors == 0)
184                         dio->page_errors = ret;
185                 get_page(page);
186                 dio->pages[0] = page;
187                 sdio->head = 0;
188                 sdio->tail = 1;
189                 sdio->from = 0;
190                 sdio->to = PAGE_SIZE;
191                 return 0;
192         }
193 
194         if (ret >= 0) {
195                 iov_iter_advance(sdio->iter, ret);
196                 ret += sdio->from;
197                 sdio->head = 0;
198                 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
199                 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
200                 return 0;
201         }
202         return ret;     
203 }
204 
205 /*
206  * Get another userspace page.  Returns an ERR_PTR on error.  Pages are
207  * buffered inside the dio so that we can call get_user_pages() against a
208  * decent number of pages, less frequently.  To provide nicer use of the
209  * L1 cache.
210  */
211 static inline struct page *dio_get_page(struct dio *dio,
212                                         struct dio_submit *sdio)
213 {
214         if (dio_pages_present(sdio) == 0) {
215                 int ret;
216 
217                 ret = dio_refill_pages(dio, sdio);
218                 if (ret)
219                         return ERR_PTR(ret);
220                 BUG_ON(dio_pages_present(sdio) == 0);
221         }
222         return dio->pages[sdio->head];
223 }
224 
225 /*
226  * dio_complete() - called when all DIO BIO I/O has been completed
227  *
228  * This drops i_dio_count, lets interested parties know that a DIO operation
229  * has completed, and calculates the resulting return code for the operation.
230  *
231  * It lets the filesystem know if it registered an interest earlier via
232  * get_block.  Pass the private field of the map buffer_head so that
233  * filesystems can use it to hold additional state between get_block calls and
234  * dio_complete.
235  */
236 static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags)
237 {
238         loff_t offset = dio->iocb->ki_pos;
239         ssize_t transferred = 0;
240         int err;
241 
242         /*
243          * AIO submission can race with bio completion to get here while
244          * expecting to have the last io completed by bio completion.
245          * In that case -EIOCBQUEUED is in fact not an error we want
246          * to preserve through this call.
247          */
248         if (ret == -EIOCBQUEUED)
249                 ret = 0;
250 
251         if (dio->result) {
252                 transferred = dio->result;
253 
254                 /* Check for short read case */
255                 if ((dio->op == REQ_OP_READ) &&
256                     ((offset + transferred) > dio->i_size))
257                         transferred = dio->i_size - offset;
258                 /* ignore EFAULT if some IO has been done */
259                 if (unlikely(ret == -EFAULT) && transferred)
260                         ret = 0;
261         }
262 
263         if (ret == 0)
264                 ret = dio->page_errors;
265         if (ret == 0)
266                 ret = dio->io_error;
267         if (ret == 0)
268                 ret = transferred;
269 
270         if (dio->end_io) {
271                 // XXX: ki_pos??
272                 err = dio->end_io(dio->iocb, offset, ret, dio->private);
273                 if (err)
274                         ret = err;
275         }
276 
277         /*
278          * Try again to invalidate clean pages which might have been cached by
279          * non-direct readahead, or faulted in by get_user_pages() if the source
280          * of the write was an mmap'ed region of the file we're writing.  Either
281          * one is a pretty crazy thing to do, so we don't support it 100%.  If
282          * this invalidation fails, tough, the write still worked...
283          *
284          * And this page cache invalidation has to be after dio->end_io(), as
285          * some filesystems convert unwritten extents to real allocations in
286          * end_io() when necessary, otherwise a racing buffer read would cache
287          * zeros from unwritten extents.
288          */
289         if (flags & DIO_COMPLETE_INVALIDATE &&
290             ret > 0 && dio->op == REQ_OP_WRITE &&
291             dio->inode->i_mapping->nrpages) {
292                 err = invalidate_inode_pages2_range(dio->inode->i_mapping,
293                                         offset >> PAGE_SHIFT,
294                                         (offset + ret - 1) >> PAGE_SHIFT);
295                 if (err)
296                         dio_warn_stale_pagecache(dio->iocb->ki_filp);
297         }
298 
299         inode_dio_end(dio->inode);
300 
301         if (flags & DIO_COMPLETE_ASYNC) {
302                 /*
303                  * generic_write_sync expects ki_pos to have been updated
304                  * already, but the submission path only does this for
305                  * synchronous I/O.
306                  */
307                 dio->iocb->ki_pos += transferred;
308 
309                 if (ret > 0 && dio->op == REQ_OP_WRITE)
310                         ret = generic_write_sync(dio->iocb, ret);
311                 dio->iocb->ki_complete(dio->iocb, ret, 0);
312         }
313 
314         kmem_cache_free(dio_cache, dio);
315         return ret;
316 }
317 
318 static void dio_aio_complete_work(struct work_struct *work)
319 {
320         struct dio *dio = container_of(work, struct dio, complete_work);
321 
322         dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
323 }
324 
325 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
326 
327 /*
328  * Asynchronous IO callback. 
329  */
330 static void dio_bio_end_aio(struct bio *bio)
331 {
332         struct dio *dio = bio->bi_private;
333         unsigned long remaining;
334         unsigned long flags;
335         bool defer_completion = false;
336 
337         /* cleanup the bio */
338         dio_bio_complete(dio, bio);
339 
340         spin_lock_irqsave(&dio->bio_lock, flags);
341         remaining = --dio->refcount;
342         if (remaining == 1 && dio->waiter)
343                 wake_up_process(dio->waiter);
344         spin_unlock_irqrestore(&dio->bio_lock, flags);
345 
346         if (remaining == 0) {
347                 /*
348                  * Defer completion when defer_completion is set or
349                  * when the inode has pages mapped and this is AIO write.
350                  * We need to invalidate those pages because there is a
351                  * chance they contain stale data in the case buffered IO
352                  * went in between AIO submission and completion into the
353                  * same region.
354                  */
355                 if (dio->result)
356                         defer_completion = dio->defer_completion ||
357                                            (dio->op == REQ_OP_WRITE &&
358                                             dio->inode->i_mapping->nrpages);
359                 if (defer_completion) {
360                         INIT_WORK(&dio->complete_work, dio_aio_complete_work);
361                         queue_work(dio->inode->i_sb->s_dio_done_wq,
362                                    &dio->complete_work);
363                 } else {
364                         dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
365                 }
366         }
367 }
368 
369 /*
370  * The BIO completion handler simply queues the BIO up for the process-context
371  * handler.
372  *
373  * During I/O bi_private points at the dio.  After I/O, bi_private is used to
374  * implement a singly-linked list of completed BIOs, at dio->bio_list.
375  */
376 static void dio_bio_end_io(struct bio *bio)
377 {
378         struct dio *dio = bio->bi_private;
379         unsigned long flags;
380 
381         spin_lock_irqsave(&dio->bio_lock, flags);
382         bio->bi_private = dio->bio_list;
383         dio->bio_list = bio;
384         if (--dio->refcount == 1 && dio->waiter)
385                 wake_up_process(dio->waiter);
386         spin_unlock_irqrestore(&dio->bio_lock, flags);
387 }
388 
389 /**
390  * dio_end_io - handle the end io action for the given bio
391  * @bio: The direct io bio thats being completed
392  *
393  * This is meant to be called by any filesystem that uses their own dio_submit_t
394  * so that the DIO specific endio actions are dealt with after the filesystem
395  * has done it's completion work.
396  */
397 void dio_end_io(struct bio *bio)
398 {
399         struct dio *dio = bio->bi_private;
400 
401         if (dio->is_async)
402                 dio_bio_end_aio(bio);
403         else
404                 dio_bio_end_io(bio);
405 }
406 EXPORT_SYMBOL_GPL(dio_end_io);
407 
408 static inline void
409 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
410               struct block_device *bdev,
411               sector_t first_sector, int nr_vecs)
412 {
413         struct bio *bio;
414 
415         /*
416          * bio_alloc() is guaranteed to return a bio when allowed to sleep and
417          * we request a valid number of vectors.
418          */
419         bio = bio_alloc(GFP_KERNEL, nr_vecs);
420 
421         bio_set_dev(bio, bdev);
422         bio->bi_iter.bi_sector = first_sector;
423         bio_set_op_attrs(bio, dio->op, dio->op_flags);
424         if (dio->is_async)
425                 bio->bi_end_io = dio_bio_end_aio;
426         else
427                 bio->bi_end_io = dio_bio_end_io;
428 
429         bio->bi_write_hint = dio->iocb->ki_hint;
430 
431         sdio->bio = bio;
432         sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
433 }
434 
435 /*
436  * In the AIO read case we speculatively dirty the pages before starting IO.
437  * During IO completion, any of these pages which happen to have been written
438  * back will be redirtied by bio_check_pages_dirty().
439  *
440  * bios hold a dio reference between submit_bio and ->end_io.
441  */
442 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
443 {
444         struct bio *bio = sdio->bio;
445         unsigned long flags;
446 
447         bio->bi_private = dio;
448 
449         spin_lock_irqsave(&dio->bio_lock, flags);
450         dio->refcount++;
451         spin_unlock_irqrestore(&dio->bio_lock, flags);
452 
453         if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty)
454                 bio_set_pages_dirty(bio);
455 
456         dio->bio_disk = bio->bi_disk;
457 
458         if (sdio->submit_io) {
459                 sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio);
460                 dio->bio_cookie = BLK_QC_T_NONE;
461         } else
462                 dio->bio_cookie = submit_bio(bio);
463 
464         sdio->bio = NULL;
465         sdio->boundary = 0;
466         sdio->logical_offset_in_bio = 0;
467 }
468 
469 /*
470  * Release any resources in case of a failure
471  */
472 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
473 {
474         while (sdio->head < sdio->tail)
475                 put_page(dio->pages[sdio->head++]);
476 }
477 
478 /*
479  * Wait for the next BIO to complete.  Remove it and return it.  NULL is
480  * returned once all BIOs have been completed.  This must only be called once
481  * all bios have been issued so that dio->refcount can only decrease.  This
482  * requires that that the caller hold a reference on the dio.
483  */
484 static struct bio *dio_await_one(struct dio *dio)
485 {
486         unsigned long flags;
487         struct bio *bio = NULL;
488 
489         spin_lock_irqsave(&dio->bio_lock, flags);
490 
491         /*
492          * Wait as long as the list is empty and there are bios in flight.  bio
493          * completion drops the count, maybe adds to the list, and wakes while
494          * holding the bio_lock so we don't need set_current_state()'s barrier
495          * and can call it after testing our condition.
496          */
497         while (dio->refcount > 1 && dio->bio_list == NULL) {
498                 __set_current_state(TASK_UNINTERRUPTIBLE);
499                 dio->waiter = current;
500                 spin_unlock_irqrestore(&dio->bio_lock, flags);
501                 if (!(dio->iocb->ki_flags & IOCB_HIPRI) ||
502                     !blk_poll(dio->bio_disk->queue, dio->bio_cookie, true))
503                         io_schedule();
504                 /* wake up sets us TASK_RUNNING */
505                 spin_lock_irqsave(&dio->bio_lock, flags);
506                 dio->waiter = NULL;
507         }
508         if (dio->bio_list) {
509                 bio = dio->bio_list;
510                 dio->bio_list = bio->bi_private;
511         }
512         spin_unlock_irqrestore(&dio->bio_lock, flags);
513         return bio;
514 }
515 
516 /*
517  * Process one completed BIO.  No locks are held.
518  */
519 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
520 {
521         blk_status_t err = bio->bi_status;
522         bool should_dirty = dio->op == REQ_OP_READ && dio->should_dirty;
523 
524         if (err) {
525                 if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
526                         dio->io_error = -EAGAIN;
527                 else
528                         dio->io_error = -EIO;
529         }
530 
531         if (dio->is_async && should_dirty) {
532                 bio_check_pages_dirty(bio);     /* transfers ownership */
533         } else {
534                 bio_release_pages(bio, should_dirty);
535                 bio_put(bio);
536         }
537         return err;
538 }
539 
540 /*
541  * Wait on and process all in-flight BIOs.  This must only be called once
542  * all bios have been issued so that the refcount can only decrease.
543  * This just waits for all bios to make it through dio_bio_complete.  IO
544  * errors are propagated through dio->io_error and should be propagated via
545  * dio_complete().
546  */
547 static void dio_await_completion(struct dio *dio)
548 {
549         struct bio *bio;
550         do {
551                 bio = dio_await_one(dio);
552                 if (bio)
553                         dio_bio_complete(dio, bio);
554         } while (bio);
555 }
556 
557 /*
558  * A really large O_DIRECT read or write can generate a lot of BIOs.  So
559  * to keep the memory consumption sane we periodically reap any completed BIOs
560  * during the BIO generation phase.
561  *
562  * This also helps to limit the peak amount of pinned userspace memory.
563  */
564 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
565 {
566         int ret = 0;
567 
568         if (sdio->reap_counter++ >= 64) {
569                 while (dio->bio_list) {
570                         unsigned long flags;
571                         struct bio *bio;
572                         int ret2;
573 
574                         spin_lock_irqsave(&dio->bio_lock, flags);
575                         bio = dio->bio_list;
576                         dio->bio_list = bio->bi_private;
577                         spin_unlock_irqrestore(&dio->bio_lock, flags);
578                         ret2 = blk_status_to_errno(dio_bio_complete(dio, bio));
579                         if (ret == 0)
580                                 ret = ret2;
581                 }
582                 sdio->reap_counter = 0;
583         }
584         return ret;
585 }
586 
587 /*
588  * Create workqueue for deferred direct IO completions. We allocate the
589  * workqueue when it's first needed. This avoids creating workqueue for
590  * filesystems that don't need it and also allows us to create the workqueue
591  * late enough so the we can include s_id in the name of the workqueue.
592  */
593 int sb_init_dio_done_wq(struct super_block *sb)
594 {
595         struct workqueue_struct *old;
596         struct workqueue_struct *wq = alloc_workqueue("dio/%s",
597                                                       WQ_MEM_RECLAIM, 0,
598                                                       sb->s_id);
599         if (!wq)
600                 return -ENOMEM;
601         /*
602          * This has to be atomic as more DIOs can race to create the workqueue
603          */
604         old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
605         /* Someone created workqueue before us? Free ours... */
606         if (old)
607                 destroy_workqueue(wq);
608         return 0;
609 }
610 
611 static int dio_set_defer_completion(struct dio *dio)
612 {
613         struct super_block *sb = dio->inode->i_sb;
614 
615         if (dio->defer_completion)
616                 return 0;
617         dio->defer_completion = true;
618         if (!sb->s_dio_done_wq)
619                 return sb_init_dio_done_wq(sb);
620         return 0;
621 }
622 
623 /*
624  * Call into the fs to map some more disk blocks.  We record the current number
625  * of available blocks at sdio->blocks_available.  These are in units of the
626  * fs blocksize, i_blocksize(inode).
627  *
628  * The fs is allowed to map lots of blocks at once.  If it wants to do that,
629  * it uses the passed inode-relative block number as the file offset, as usual.
630  *
631  * get_block() is passed the number of i_blkbits-sized blocks which direct_io
632  * has remaining to do.  The fs should not map more than this number of blocks.
633  *
634  * If the fs has mapped a lot of blocks, it should populate bh->b_size to
635  * indicate how much contiguous disk space has been made available at
636  * bh->b_blocknr.
637  *
638  * If *any* of the mapped blocks are new, then the fs must set buffer_new().
639  * This isn't very efficient...
640  *
641  * In the case of filesystem holes: the fs may return an arbitrarily-large
642  * hole by returning an appropriate value in b_size and by clearing
643  * buffer_mapped().  However the direct-io code will only process holes one
644  * block at a time - it will repeatedly call get_block() as it walks the hole.
645  */
646 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
647                            struct buffer_head *map_bh)
648 {
649         int ret;
650         sector_t fs_startblk;   /* Into file, in filesystem-sized blocks */
651         sector_t fs_endblk;     /* Into file, in filesystem-sized blocks */
652         unsigned long fs_count; /* Number of filesystem-sized blocks */
653         int create;
654         unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
655         loff_t i_size;
656 
657         /*
658          * If there was a memory error and we've overwritten all the
659          * mapped blocks then we can now return that memory error
660          */
661         ret = dio->page_errors;
662         if (ret == 0) {
663                 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
664                 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
665                 fs_endblk = (sdio->final_block_in_request - 1) >>
666                                         sdio->blkfactor;
667                 fs_count = fs_endblk - fs_startblk + 1;
668 
669                 map_bh->b_state = 0;
670                 map_bh->b_size = fs_count << i_blkbits;
671 
672                 /*
673                  * For writes that could fill holes inside i_size on a
674                  * DIO_SKIP_HOLES filesystem we forbid block creations: only
675                  * overwrites are permitted. We will return early to the caller
676                  * once we see an unmapped buffer head returned, and the caller
677                  * will fall back to buffered I/O.
678                  *
679                  * Otherwise the decision is left to the get_blocks method,
680                  * which may decide to handle it or also return an unmapped
681                  * buffer head.
682                  */
683                 create = dio->op == REQ_OP_WRITE;
684                 if (dio->flags & DIO_SKIP_HOLES) {
685                         i_size = i_size_read(dio->inode);
686                         if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits)
687                                 create = 0;
688                 }
689 
690                 ret = (*sdio->get_block)(dio->inode, fs_startblk,
691                                                 map_bh, create);
692 
693                 /* Store for completion */
694                 dio->private = map_bh->b_private;
695 
696                 if (ret == 0 && buffer_defer_completion(map_bh))
697                         ret = dio_set_defer_completion(dio);
698         }
699         return ret;
700 }
701 
702 /*
703  * There is no bio.  Make one now.
704  */
705 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
706                 sector_t start_sector, struct buffer_head *map_bh)
707 {
708         sector_t sector;
709         int ret, nr_pages;
710 
711         ret = dio_bio_reap(dio, sdio);
712         if (ret)
713                 goto out;
714         sector = start_sector << (sdio->blkbits - 9);
715         nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES);
716         BUG_ON(nr_pages <= 0);
717         dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
718         sdio->boundary = 0;
719 out:
720         return ret;
721 }
722 
723 /*
724  * Attempt to put the current chunk of 'cur_page' into the current BIO.  If
725  * that was successful then update final_block_in_bio and take a ref against
726  * the just-added page.
727  *
728  * Return zero on success.  Non-zero means the caller needs to start a new BIO.
729  */
730 static inline int dio_bio_add_page(struct dio_submit *sdio)
731 {
732         int ret;
733 
734         ret = bio_add_page(sdio->bio, sdio->cur_page,
735                         sdio->cur_page_len, sdio->cur_page_offset);
736         if (ret == sdio->cur_page_len) {
737                 /*
738                  * Decrement count only, if we are done with this page
739                  */
740                 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
741                         sdio->pages_in_io--;
742                 get_page(sdio->cur_page);
743                 sdio->final_block_in_bio = sdio->cur_page_block +
744                         (sdio->cur_page_len >> sdio->blkbits);
745                 ret = 0;
746         } else {
747                 ret = 1;
748         }
749         return ret;
750 }
751                 
752 /*
753  * Put cur_page under IO.  The section of cur_page which is described by
754  * cur_page_offset,cur_page_len is put into a BIO.  The section of cur_page
755  * starts on-disk at cur_page_block.
756  *
757  * We take a ref against the page here (on behalf of its presence in the bio).
758  *
759  * The caller of this function is responsible for removing cur_page from the
760  * dio, and for dropping the refcount which came from that presence.
761  */
762 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
763                 struct buffer_head *map_bh)
764 {
765         int ret = 0;
766 
767         if (sdio->bio) {
768                 loff_t cur_offset = sdio->cur_page_fs_offset;
769                 loff_t bio_next_offset = sdio->logical_offset_in_bio +
770                         sdio->bio->bi_iter.bi_size;
771 
772                 /*
773                  * See whether this new request is contiguous with the old.
774                  *
775                  * Btrfs cannot handle having logically non-contiguous requests
776                  * submitted.  For example if you have
777                  *
778                  * Logical:  [0-4095][HOLE][8192-12287]
779                  * Physical: [0-4095]      [4096-8191]
780                  *
781                  * We cannot submit those pages together as one BIO.  So if our
782                  * current logical offset in the file does not equal what would
783                  * be the next logical offset in the bio, submit the bio we
784                  * have.
785                  */
786                 if (sdio->final_block_in_bio != sdio->cur_page_block ||
787                     cur_offset != bio_next_offset)
788                         dio_bio_submit(dio, sdio);
789         }
790 
791         if (sdio->bio == NULL) {
792                 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
793                 if (ret)
794                         goto out;
795         }
796 
797         if (dio_bio_add_page(sdio) != 0) {
798                 dio_bio_submit(dio, sdio);
799                 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
800                 if (ret == 0) {
801                         ret = dio_bio_add_page(sdio);
802                         BUG_ON(ret != 0);
803                 }
804         }
805 out:
806         return ret;
807 }
808 
809 /*
810  * An autonomous function to put a chunk of a page under deferred IO.
811  *
812  * The caller doesn't actually know (or care) whether this piece of page is in
813  * a BIO, or is under IO or whatever.  We just take care of all possible 
814  * situations here.  The separation between the logic of do_direct_IO() and
815  * that of submit_page_section() is important for clarity.  Please don't break.
816  *
817  * The chunk of page starts on-disk at blocknr.
818  *
819  * We perform deferred IO, by recording the last-submitted page inside our
820  * private part of the dio structure.  If possible, we just expand the IO
821  * across that page here.
822  *
823  * If that doesn't work out then we put the old page into the bio and add this
824  * page to the dio instead.
825  */
826 static inline int
827 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
828                     unsigned offset, unsigned len, sector_t blocknr,
829                     struct buffer_head *map_bh)
830 {
831         int ret = 0;
832 
833         if (dio->op == REQ_OP_WRITE) {
834                 /*
835                  * Read accounting is performed in submit_bio()
836                  */
837                 task_io_account_write(len);
838         }
839 
840         /*
841          * Can we just grow the current page's presence in the dio?
842          */
843         if (sdio->cur_page == page &&
844             sdio->cur_page_offset + sdio->cur_page_len == offset &&
845             sdio->cur_page_block +
846             (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
847                 sdio->cur_page_len += len;
848                 goto out;
849         }
850 
851         /*
852          * If there's a deferred page already there then send it.
853          */
854         if (sdio->cur_page) {
855                 ret = dio_send_cur_page(dio, sdio, map_bh);
856                 put_page(sdio->cur_page);
857                 sdio->cur_page = NULL;
858                 if (ret)
859                         return ret;
860         }
861 
862         get_page(page);         /* It is in dio */
863         sdio->cur_page = page;
864         sdio->cur_page_offset = offset;
865         sdio->cur_page_len = len;
866         sdio->cur_page_block = blocknr;
867         sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
868 out:
869         /*
870          * If sdio->boundary then we want to schedule the IO now to
871          * avoid metadata seeks.
872          */
873         if (sdio->boundary) {
874                 ret = dio_send_cur_page(dio, sdio, map_bh);
875                 if (sdio->bio)
876                         dio_bio_submit(dio, sdio);
877                 put_page(sdio->cur_page);
878                 sdio->cur_page = NULL;
879         }
880         return ret;
881 }
882 
883 /*
884  * If we are not writing the entire block and get_block() allocated
885  * the block for us, we need to fill-in the unused portion of the
886  * block with zeros. This happens only if user-buffer, fileoffset or
887  * io length is not filesystem block-size multiple.
888  *
889  * `end' is zero if we're doing the start of the IO, 1 at the end of the
890  * IO.
891  */
892 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
893                 int end, struct buffer_head *map_bh)
894 {
895         unsigned dio_blocks_per_fs_block;
896         unsigned this_chunk_blocks;     /* In dio_blocks */
897         unsigned this_chunk_bytes;
898         struct page *page;
899 
900         sdio->start_zero_done = 1;
901         if (!sdio->blkfactor || !buffer_new(map_bh))
902                 return;
903 
904         dio_blocks_per_fs_block = 1 << sdio->blkfactor;
905         this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
906 
907         if (!this_chunk_blocks)
908                 return;
909 
910         /*
911          * We need to zero out part of an fs block.  It is either at the
912          * beginning or the end of the fs block.
913          */
914         if (end) 
915                 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
916 
917         this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
918 
919         page = ZERO_PAGE(0);
920         if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
921                                 sdio->next_block_for_io, map_bh))
922                 return;
923 
924         sdio->next_block_for_io += this_chunk_blocks;
925 }
926 
927 /*
928  * Walk the user pages, and the file, mapping blocks to disk and generating
929  * a sequence of (page,offset,len,block) mappings.  These mappings are injected
930  * into submit_page_section(), which takes care of the next stage of submission
931  *
932  * Direct IO against a blockdev is different from a file.  Because we can
933  * happily perform page-sized but 512-byte aligned IOs.  It is important that
934  * blockdev IO be able to have fine alignment and large sizes.
935  *
936  * So what we do is to permit the ->get_block function to populate bh.b_size
937  * with the size of IO which is permitted at this offset and this i_blkbits.
938  *
939  * For best results, the blockdev should be set up with 512-byte i_blkbits and
940  * it should set b_size to PAGE_SIZE or more inside get_block().  This gives
941  * fine alignment but still allows this function to work in PAGE_SIZE units.
942  */
943 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
944                         struct buffer_head *map_bh)
945 {
946         const unsigned blkbits = sdio->blkbits;
947         const unsigned i_blkbits = blkbits + sdio->blkfactor;
948         int ret = 0;
949 
950         while (sdio->block_in_file < sdio->final_block_in_request) {
951                 struct page *page;
952                 size_t from, to;
953 
954                 page = dio_get_page(dio, sdio);
955                 if (IS_ERR(page)) {
956                         ret = PTR_ERR(page);
957                         goto out;
958                 }
959                 from = sdio->head ? 0 : sdio->from;
960                 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
961                 sdio->head++;
962 
963                 while (from < to) {
964                         unsigned this_chunk_bytes;      /* # of bytes mapped */
965                         unsigned this_chunk_blocks;     /* # of blocks */
966                         unsigned u;
967 
968                         if (sdio->blocks_available == 0) {
969                                 /*
970                                  * Need to go and map some more disk
971                                  */
972                                 unsigned long blkmask;
973                                 unsigned long dio_remainder;
974 
975                                 ret = get_more_blocks(dio, sdio, map_bh);
976                                 if (ret) {
977                                         put_page(page);
978                                         goto out;
979                                 }
980                                 if (!buffer_mapped(map_bh))
981                                         goto do_holes;
982 
983                                 sdio->blocks_available =
984                                                 map_bh->b_size >> blkbits;
985                                 sdio->next_block_for_io =
986                                         map_bh->b_blocknr << sdio->blkfactor;
987                                 if (buffer_new(map_bh)) {
988                                         clean_bdev_aliases(
989                                                 map_bh->b_bdev,
990                                                 map_bh->b_blocknr,
991                                                 map_bh->b_size >> i_blkbits);
992                                 }
993 
994                                 if (!sdio->blkfactor)
995                                         goto do_holes;
996 
997                                 blkmask = (1 << sdio->blkfactor) - 1;
998                                 dio_remainder = (sdio->block_in_file & blkmask);
999 
1000                                 /*
1001                                  * If we are at the start of IO and that IO
1002                                  * starts partway into a fs-block,
1003                                  * dio_remainder will be non-zero.  If the IO
1004                                  * is a read then we can simply advance the IO
1005                                  * cursor to the first block which is to be
1006                                  * read.  But if the IO is a write and the
1007                                  * block was newly allocated we cannot do that;
1008                                  * the start of the fs block must be zeroed out
1009                                  * on-disk
1010                                  */
1011                                 if (!buffer_new(map_bh))
1012                                         sdio->next_block_for_io += dio_remainder;
1013                                 sdio->blocks_available -= dio_remainder;
1014                         }
1015 do_holes:
1016                         /* Handle holes */
1017                         if (!buffer_mapped(map_bh)) {
1018                                 loff_t i_size_aligned;
1019 
1020                                 /* AKPM: eargh, -ENOTBLK is a hack */
1021                                 if (dio->op == REQ_OP_WRITE) {
1022                                         put_page(page);
1023                                         return -ENOTBLK;
1024                                 }
1025 
1026                                 /*
1027                                  * Be sure to account for a partial block as the
1028                                  * last block in the file
1029                                  */
1030                                 i_size_aligned = ALIGN(i_size_read(dio->inode),
1031                                                         1 << blkbits);
1032                                 if (sdio->block_in_file >=
1033                                                 i_size_aligned >> blkbits) {
1034                                         /* We hit eof */
1035                                         put_page(page);
1036                                         goto out;
1037                                 }
1038                                 zero_user(page, from, 1 << blkbits);
1039                                 sdio->block_in_file++;
1040                                 from += 1 << blkbits;
1041                                 dio->result += 1 << blkbits;
1042                                 goto next_block;
1043                         }
1044 
1045                         /*
1046                          * If we're performing IO which has an alignment which
1047                          * is finer than the underlying fs, go check to see if
1048                          * we must zero out the start of this block.
1049                          */
1050                         if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1051                                 dio_zero_block(dio, sdio, 0, map_bh);
1052 
1053                         /*
1054                          * Work out, in this_chunk_blocks, how much disk we
1055                          * can add to this page
1056                          */
1057                         this_chunk_blocks = sdio->blocks_available;
1058                         u = (to - from) >> blkbits;
1059                         if (this_chunk_blocks > u)
1060                                 this_chunk_blocks = u;
1061                         u = sdio->final_block_in_request - sdio->block_in_file;
1062                         if (this_chunk_blocks > u)
1063                                 this_chunk_blocks = u;
1064                         this_chunk_bytes = this_chunk_blocks << blkbits;
1065                         BUG_ON(this_chunk_bytes == 0);
1066 
1067                         if (this_chunk_blocks == sdio->blocks_available)
1068                                 sdio->boundary = buffer_boundary(map_bh);
1069                         ret = submit_page_section(dio, sdio, page,
1070                                                   from,
1071                                                   this_chunk_bytes,
1072                                                   sdio->next_block_for_io,
1073                                                   map_bh);
1074                         if (ret) {
1075                                 put_page(page);
1076                                 goto out;
1077                         }
1078                         sdio->next_block_for_io += this_chunk_blocks;
1079 
1080                         sdio->block_in_file += this_chunk_blocks;
1081                         from += this_chunk_bytes;
1082                         dio->result += this_chunk_bytes;
1083                         sdio->blocks_available -= this_chunk_blocks;
1084 next_block:
1085                         BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1086                         if (sdio->block_in_file == sdio->final_block_in_request)
1087                                 break;
1088                 }
1089 
1090                 /* Drop the ref which was taken in get_user_pages() */
1091                 put_page(page);
1092         }
1093 out:
1094         return ret;
1095 }
1096 
1097 static inline int drop_refcount(struct dio *dio)
1098 {
1099         int ret2;
1100         unsigned long flags;
1101 
1102         /*
1103          * Sync will always be dropping the final ref and completing the
1104          * operation.  AIO can if it was a broken operation described above or
1105          * in fact if all the bios race to complete before we get here.  In
1106          * that case dio_complete() translates the EIOCBQUEUED into the proper
1107          * return code that the caller will hand to ->complete().
1108          *
1109          * This is managed by the bio_lock instead of being an atomic_t so that
1110          * completion paths can drop their ref and use the remaining count to
1111          * decide to wake the submission path atomically.
1112          */
1113         spin_lock_irqsave(&dio->bio_lock, flags);
1114         ret2 = --dio->refcount;
1115         spin_unlock_irqrestore(&dio->bio_lock, flags);
1116         return ret2;
1117 }
1118 
1119 /*
1120  * This is a library function for use by filesystem drivers.
1121  *
1122  * The locking rules are governed by the flags parameter:
1123  *  - if the flags value contains DIO_LOCKING we use a fancy locking
1124  *    scheme for dumb filesystems.
1125  *    For writes this function is called under i_mutex and returns with
1126  *    i_mutex held, for reads, i_mutex is not held on entry, but it is
1127  *    taken and dropped again before returning.
1128  *  - if the flags value does NOT contain DIO_LOCKING we don't use any
1129  *    internal locking but rather rely on the filesystem to synchronize
1130  *    direct I/O reads/writes versus each other and truncate.
1131  *
1132  * To help with locking against truncate we incremented the i_dio_count
1133  * counter before starting direct I/O, and decrement it once we are done.
1134  * Truncate can wait for it to reach zero to provide exclusion.  It is
1135  * expected that filesystem provide exclusion between new direct I/O
1136  * and truncates.  For DIO_LOCKING filesystems this is done by i_mutex,
1137  * but other filesystems need to take care of this on their own.
1138  *
1139  * NOTE: if you pass "sdio" to anything by pointer make sure that function
1140  * is always inlined. Otherwise gcc is unable to split the structure into
1141  * individual fields and will generate much worse code. This is important
1142  * for the whole file.
1143  */
1144 static inline ssize_t
1145 do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1146                       struct block_device *bdev, struct iov_iter *iter,
1147                       get_block_t get_block, dio_iodone_t end_io,
1148                       dio_submit_t submit_io, int flags)
1149 {
1150         unsigned i_blkbits = READ_ONCE(inode->i_blkbits);
1151         unsigned blkbits = i_blkbits;
1152         unsigned blocksize_mask = (1 << blkbits) - 1;
1153         ssize_t retval = -EINVAL;
1154         const size_t count = iov_iter_count(iter);
1155         loff_t offset = iocb->ki_pos;
1156         const loff_t end = offset + count;
1157         struct dio *dio;
1158         struct dio_submit sdio = { 0, };
1159         struct buffer_head map_bh = { 0, };
1160         struct blk_plug plug;
1161         unsigned long align = offset | iov_iter_alignment(iter);
1162 
1163         /*
1164          * Avoid references to bdev if not absolutely needed to give
1165          * the early prefetch in the caller enough time.
1166          */
1167 
1168         if (align & blocksize_mask) {
1169                 if (bdev)
1170                         blkbits = blksize_bits(bdev_logical_block_size(bdev));
1171                 blocksize_mask = (1 << blkbits) - 1;
1172                 if (align & blocksize_mask)
1173                         goto out;
1174         }
1175 
1176         /* watch out for a 0 len io from a tricksy fs */
1177         if (iov_iter_rw(iter) == READ && !count)
1178                 return 0;
1179 
1180         dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1181         retval = -ENOMEM;
1182         if (!dio)
1183                 goto out;
1184         /*
1185          * Believe it or not, zeroing out the page array caused a .5%
1186          * performance regression in a database benchmark.  So, we take
1187          * care to only zero out what's needed.
1188          */
1189         memset(dio, 0, offsetof(struct dio, pages));
1190 
1191         dio->flags = flags;
1192         if (dio->flags & DIO_LOCKING) {
1193                 if (iov_iter_rw(iter) == READ) {
1194                         struct address_space *mapping =
1195                                         iocb->ki_filp->f_mapping;
1196 
1197                         /* will be released by direct_io_worker */
1198                         inode_lock(inode);
1199 
1200                         retval = filemap_write_and_wait_range(mapping, offset,
1201                                                               end - 1);
1202                         if (retval) {
1203                                 inode_unlock(inode);
1204                                 kmem_cache_free(dio_cache, dio);
1205                                 goto out;
1206                         }
1207                 }
1208         }
1209 
1210         /* Once we sampled i_size check for reads beyond EOF */
1211         dio->i_size = i_size_read(inode);
1212         if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1213                 if (dio->flags & DIO_LOCKING)
1214                         inode_unlock(inode);
1215                 kmem_cache_free(dio_cache, dio);
1216                 retval = 0;
1217                 goto out;
1218         }
1219 
1220         /*
1221          * For file extending writes updating i_size before data writeouts
1222          * complete can expose uninitialized blocks in dumb filesystems.
1223          * In that case we need to wait for I/O completion even if asked
1224          * for an asynchronous write.
1225          */
1226         if (is_sync_kiocb(iocb))
1227                 dio->is_async = false;
1228         else if (iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
1229                 dio->is_async = false;
1230         else
1231                 dio->is_async = true;
1232 
1233         dio->inode = inode;
1234         if (iov_iter_rw(iter) == WRITE) {
1235                 dio->op = REQ_OP_WRITE;
1236                 dio->op_flags = REQ_SYNC | REQ_IDLE;
1237                 if (iocb->ki_flags & IOCB_NOWAIT)
1238                         dio->op_flags |= REQ_NOWAIT;
1239         } else {
1240                 dio->op = REQ_OP_READ;
1241         }
1242         if (iocb->ki_flags & IOCB_HIPRI)
1243                 dio->op_flags |= REQ_HIPRI;
1244 
1245         /*
1246          * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1247          * so that we can call ->fsync.
1248          */
1249         if (dio->is_async && iov_iter_rw(iter) == WRITE) {
1250                 retval = 0;
1251                 if (iocb->ki_flags & IOCB_DSYNC)
1252                         retval = dio_set_defer_completion(dio);
1253                 else if (!dio->inode->i_sb->s_dio_done_wq) {
1254                         /*
1255                          * In case of AIO write racing with buffered read we
1256                          * need to defer completion. We can't decide this now,
1257                          * however the workqueue needs to be initialized here.
1258                          */
1259                         retval = sb_init_dio_done_wq(dio->inode->i_sb);
1260                 }
1261                 if (retval) {
1262                         /*
1263                          * We grab i_mutex only for reads so we don't have
1264                          * to release it here
1265                          */
1266                         kmem_cache_free(dio_cache, dio);
1267                         goto out;
1268                 }
1269         }
1270 
1271         /*
1272          * Will be decremented at I/O completion time.
1273          */
1274         inode_dio_begin(inode);
1275 
1276         retval = 0;
1277         sdio.blkbits = blkbits;
1278         sdio.blkfactor = i_blkbits - blkbits;
1279         sdio.block_in_file = offset >> blkbits;
1280 
1281         sdio.get_block = get_block;
1282         dio->end_io = end_io;
1283         sdio.submit_io = submit_io;
1284         sdio.final_block_in_bio = -1;
1285         sdio.next_block_for_io = -1;
1286 
1287         dio->iocb = iocb;
1288 
1289         spin_lock_init(&dio->bio_lock);
1290         dio->refcount = 1;
1291 
1292         dio->should_dirty = iter_is_iovec(iter) && iov_iter_rw(iter) == READ;
1293         sdio.iter = iter;
1294         sdio.final_block_in_request = end >> blkbits;
1295 
1296         /*
1297          * In case of non-aligned buffers, we may need 2 more
1298          * pages since we need to zero out first and last block.
1299          */
1300         if (unlikely(sdio.blkfactor))
1301                 sdio.pages_in_io = 2;
1302 
1303         sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
1304 
1305         blk_start_plug(&plug);
1306 
1307         retval = do_direct_IO(dio, &sdio, &map_bh);
1308         if (retval)
1309                 dio_cleanup(dio, &sdio);
1310 
1311         if (retval == -ENOTBLK) {
1312                 /*
1313                  * The remaining part of the request will be
1314                  * be handled by buffered I/O when we return
1315                  */
1316                 retval = 0;
1317         }
1318         /*
1319          * There may be some unwritten disk at the end of a part-written
1320          * fs-block-sized block.  Go zero that now.
1321          */
1322         dio_zero_block(dio, &sdio, 1, &map_bh);
1323 
1324         if (sdio.cur_page) {
1325                 ssize_t ret2;
1326 
1327                 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1328                 if (retval == 0)
1329                         retval = ret2;
1330                 put_page(sdio.cur_page);
1331                 sdio.cur_page = NULL;
1332         }
1333         if (sdio.bio)
1334                 dio_bio_submit(dio, &sdio);
1335 
1336         blk_finish_plug(&plug);
1337 
1338         /*
1339          * It is possible that, we return short IO due to end of file.
1340          * In that case, we need to release all the pages we got hold on.
1341          */
1342         dio_cleanup(dio, &sdio);
1343 
1344         /*
1345          * All block lookups have been performed. For READ requests
1346          * we can let i_mutex go now that its achieved its purpose
1347          * of protecting us from looking up uninitialized blocks.
1348          */
1349         if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
1350                 inode_unlock(dio->inode);
1351 
1352         /*
1353          * The only time we want to leave bios in flight is when a successful
1354          * partial aio read or full aio write have been setup.  In that case
1355          * bio completion will call aio_complete.  The only time it's safe to
1356          * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1357          * This had *better* be the only place that raises -EIOCBQUEUED.
1358          */
1359         BUG_ON(retval == -EIOCBQUEUED);
1360         if (dio->is_async && retval == 0 && dio->result &&
1361             (iov_iter_rw(iter) == READ || dio->result == count))
1362                 retval = -EIOCBQUEUED;
1363         else
1364                 dio_await_completion(dio);
1365 
1366         if (drop_refcount(dio) == 0) {
1367                 retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE);
1368         } else
1369                 BUG_ON(retval != -EIOCBQUEUED);
1370 
1371 out:
1372         return retval;
1373 }
1374 
1375 ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1376                              struct block_device *bdev, struct iov_iter *iter,
1377                              get_block_t get_block,
1378                              dio_iodone_t end_io, dio_submit_t submit_io,
1379                              int flags)
1380 {
1381         /*
1382          * The block device state is needed in the end to finally
1383          * submit everything.  Since it's likely to be cache cold
1384          * prefetch it here as first thing to hide some of the
1385          * latency.
1386          *
1387          * Attempt to prefetch the pieces we likely need later.
1388          */
1389         prefetch(&bdev->bd_disk->part_tbl);
1390         prefetch(bdev->bd_queue);
1391         prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1392 
1393         return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block,
1394                                      end_io, submit_io, flags);
1395 }
1396 
1397 EXPORT_SYMBOL(__blockdev_direct_IO);
1398 
1399 static __init int dio_init(void)
1400 {
1401         dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1402         return 0;
1403 }
1404 module_init(dio_init)
1405 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp