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

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

Version: ~ [ linux-5.7 ] ~ [ linux-5.6.15 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.43 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.125 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.182 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.225 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.225 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.84 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ 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 /*
  2  * fs/direct-io.c
  3  *
  4  * Copyright (C) 2002, Linus Torvalds.
  5  *
  6  * O_DIRECT
  7  *
  8  * 04Jul2002    Andrew Morton
  9  *              Initial version
 10  * 11Sep2002    janetinc@us.ibm.com
 11  *              added readv/writev support.
 12  * 29Oct2002    Andrew Morton
 13  *              rewrote bio_add_page() support.
 14  * 30Oct2002    pbadari@us.ibm.com
 15  *              added support for non-aligned IO.
 16  * 06Nov2002    pbadari@us.ibm.com
 17  *              added asynchronous IO support.
 18  * 21Jul2003    nathans@sgi.com
 19  *              added IO completion notifier.
 20  */
 21 
 22 #include <linux/kernel.h>
 23 #include <linux/module.h>
 24 #include <linux/types.h>
 25 #include <linux/fs.h>
 26 #include <linux/mm.h>
 27 #include <linux/slab.h>
 28 #include <linux/highmem.h>
 29 #include <linux/pagemap.h>
 30 #include <linux/task_io_accounting_ops.h>
 31 #include <linux/bio.h>
 32 #include <linux/wait.h>
 33 #include <linux/err.h>
 34 #include <linux/blkdev.h>
 35 #include <linux/buffer_head.h>
 36 #include <linux/rwsem.h>
 37 #include <linux/uio.h>
 38 #include <linux/atomic.h>
 39 #include <linux/prefetch.h>
 40 
 41 /*
 42  * How many user pages to map in one call to get_user_pages().  This determines
 43  * the size of a structure in the slab cache
 44  */
 45 #define DIO_PAGES       64
 46 
 47 /*
 48  * Flags for dio_complete()
 49  */
 50 #define DIO_COMPLETE_ASYNC              0x01    /* This is async IO */
 51 #define DIO_COMPLETE_INVALIDATE         0x02    /* Can invalidate pages */
 52 
 53 /*
 54  * This code generally works in units of "dio_blocks".  A dio_block is
 55  * somewhere between the hard sector size and the filesystem block size.  it
 56  * is determined on a per-invocation basis.   When talking to the filesystem
 57  * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
 58  * down by dio->blkfactor.  Similarly, fs-blocksize quantities are converted
 59  * to bio_block quantities by shifting left by blkfactor.
 60  *
 61  * If blkfactor is zero then the user's request was aligned to the filesystem's
 62  * blocksize.
 63  */
 64 
 65 /* dio_state only used in the submission path */
 66 
 67 struct dio_submit {
 68         struct bio *bio;                /* bio under assembly */
 69         unsigned blkbits;               /* doesn't change */
 70         unsigned blkfactor;             /* When we're using an alignment which
 71                                            is finer than the filesystem's soft
 72                                            blocksize, this specifies how much
 73                                            finer.  blkfactor=2 means 1/4-block
 74                                            alignment.  Does not change */
 75         unsigned start_zero_done;       /* flag: sub-blocksize zeroing has
 76                                            been performed at the start of a
 77                                            write */
 78         int pages_in_io;                /* approximate total IO pages */
 79         sector_t block_in_file;         /* Current offset into the underlying
 80                                            file in dio_block units. */
 81         unsigned blocks_available;      /* At block_in_file.  changes */
 82         int reap_counter;               /* rate limit reaping */
 83         sector_t final_block_in_request;/* doesn't change */
 84         int boundary;                   /* prev block is at a boundary */
 85         get_block_t *get_block;         /* block mapping function */
 86         dio_submit_t *submit_io;        /* IO submition function */
 87 
 88         loff_t logical_offset_in_bio;   /* current first logical block in bio */
 89         sector_t final_block_in_bio;    /* current final block in bio + 1 */
 90         sector_t next_block_for_io;     /* next block to be put under IO,
 91                                            in dio_blocks units */
 92 
 93         /*
 94          * Deferred addition of a page to the dio.  These variables are
 95          * private to dio_send_cur_page(), submit_page_section() and
 96          * dio_bio_add_page().
 97          */
 98         struct page *cur_page;          /* The page */
 99         unsigned cur_page_offset;       /* Offset into it, in bytes */
100         unsigned cur_page_len;          /* Nr of bytes at cur_page_offset */
101         sector_t cur_page_block;        /* Where it starts */
102         loff_t cur_page_fs_offset;      /* Offset in file */
103 
104         struct iov_iter *iter;
105         /*
106          * Page queue.  These variables belong to dio_refill_pages() and
107          * dio_get_page().
108          */
109         unsigned head;                  /* next page to process */
110         unsigned tail;                  /* last valid page + 1 */
111         size_t from, to;
112 };
113 
114 /* dio_state communicated between submission path and end_io */
115 struct dio {
116         int flags;                      /* doesn't change */
117         int op;
118         int op_flags;
119         blk_qc_t bio_cookie;
120         struct gendisk *bio_disk;
121         struct inode *inode;
122         loff_t i_size;                  /* i_size when submitted */
123         dio_iodone_t *end_io;           /* IO completion function */
124 
125         void *private;                  /* copy from map_bh.b_private */
126 
127         /* BIO completion state */
128         spinlock_t bio_lock;            /* protects BIO fields below */
129         int page_errors;                /* errno from get_user_pages() */
130         int is_async;                   /* is IO async ? */
131         bool defer_completion;          /* defer AIO completion to workqueue? */
132         bool should_dirty;              /* if pages should be dirtied */
133         int io_error;                   /* IO error in completion path */
134         unsigned long refcount;         /* direct_io_worker() and bios */
135         struct bio *bio_list;           /* singly linked via bi_private */
136         struct task_struct *waiter;     /* waiting task (NULL if none) */
137 
138         /* AIO related stuff */
139         struct kiocb *iocb;             /* kiocb */
140         ssize_t result;                 /* IO result */
141 
142         /*
143          * pages[] (and any fields placed after it) are not zeroed out at
144          * allocation time.  Don't add new fields after pages[] unless you
145          * wish that they not be zeroed.
146          */
147         union {
148                 struct page *pages[DIO_PAGES];  /* page buffer */
149                 struct work_struct complete_work;/* deferred AIO completion */
150         };
151 } ____cacheline_aligned_in_smp;
152 
153 static struct kmem_cache *dio_cache __read_mostly;
154 
155 /*
156  * How many pages are in the queue?
157  */
158 static inline unsigned dio_pages_present(struct dio_submit *sdio)
159 {
160         return sdio->tail - sdio->head;
161 }
162 
163 /*
164  * Go grab and pin some userspace pages.   Typically we'll get 64 at a time.
165  */
166 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
167 {
168         ssize_t ret;
169 
170         ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
171                                 &sdio->from);
172 
173         if (ret < 0 && sdio->blocks_available && (dio->op == REQ_OP_WRITE)) {
174                 struct page *page = ZERO_PAGE(0);
175                 /*
176                  * A memory fault, but the filesystem has some outstanding
177                  * mapped blocks.  We need to use those blocks up to avoid
178                  * leaking stale data in the file.
179                  */
180                 if (dio->page_errors == 0)
181                         dio->page_errors = ret;
182                 get_page(page);
183                 dio->pages[0] = page;
184                 sdio->head = 0;
185                 sdio->tail = 1;
186                 sdio->from = 0;
187                 sdio->to = PAGE_SIZE;
188                 return 0;
189         }
190 
191         if (ret >= 0) {
192                 iov_iter_advance(sdio->iter, ret);
193                 ret += sdio->from;
194                 sdio->head = 0;
195                 sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
196                 sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
197                 return 0;
198         }
199         return ret;     
200 }
201 
202 /*
203  * Get another userspace page.  Returns an ERR_PTR on error.  Pages are
204  * buffered inside the dio so that we can call get_user_pages() against a
205  * decent number of pages, less frequently.  To provide nicer use of the
206  * L1 cache.
207  */
208 static inline struct page *dio_get_page(struct dio *dio,
209                                         struct dio_submit *sdio)
210 {
211         if (dio_pages_present(sdio) == 0) {
212                 int ret;
213 
214                 ret = dio_refill_pages(dio, sdio);
215                 if (ret)
216                         return ERR_PTR(ret);
217                 BUG_ON(dio_pages_present(sdio) == 0);
218         }
219         return dio->pages[sdio->head];
220 }
221 
222 /*
223  * Warn about a page cache invalidation failure during a direct io write.
224  */
225 void dio_warn_stale_pagecache(struct file *filp)
226 {
227         static DEFINE_RATELIMIT_STATE(_rs, 86400 * HZ, DEFAULT_RATELIMIT_BURST);
228         char pathname[128];
229         struct inode *inode = file_inode(filp);
230         char *path;
231 
232         errseq_set(&inode->i_mapping->wb_err, -EIO);
233         if (__ratelimit(&_rs)) {
234                 path = file_path(filp, pathname, sizeof(pathname));
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         }
241 }
242 
243 /**
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
246  *
247  * This drops i_dio_count, lets interested parties know that a DIO operation
248  * has completed, and calculates the resulting return code for the operation.
249  *
250  * It lets the filesystem know if it registered an interest earlier via
251  * get_block.  Pass the private field of the map buffer_head so that
252  * filesystems can use it to hold additional state between get_block calls and
253  * dio_complete.
254  */
255 static ssize_t dio_complete(struct dio *dio, ssize_t ret, unsigned int flags)
256 {
257         loff_t offset = dio->iocb->ki_pos;
258         ssize_t transferred = 0;
259         int err;
260 
261         /*
262          * AIO submission can race with bio completion to get here while
263          * expecting to have the last io completed by bio completion.
264          * In that case -EIOCBQUEUED is in fact not an error we want
265          * to preserve through this call.
266          */
267         if (ret == -EIOCBQUEUED)
268                 ret = 0;
269 
270         if (dio->result) {
271                 transferred = dio->result;
272 
273                 /* Check for short read case */
274                 if ((dio->op == REQ_OP_READ) &&
275                     ((offset + transferred) > dio->i_size))
276                         transferred = dio->i_size - offset;
277                 /* ignore EFAULT if some IO has been done */
278                 if (unlikely(ret == -EFAULT) && transferred)
279                         ret = 0;
280         }
281 
282         if (ret == 0)
283                 ret = dio->page_errors;
284         if (ret == 0)
285                 ret = dio->io_error;
286         if (ret == 0)
287                 ret = transferred;
288 
289         if (dio->end_io) {
290                 // XXX: ki_pos??
291                 err = dio->end_io(dio->iocb, offset, ret, dio->private);
292                 if (err)
293                         ret = err;
294         }
295 
296         /*
297          * Try again to invalidate clean pages which might have been cached by
298          * non-direct readahead, or faulted in by get_user_pages() if the source
299          * of the write was an mmap'ed region of the file we're writing.  Either
300          * one is a pretty crazy thing to do, so we don't support it 100%.  If
301          * this invalidation fails, tough, the write still worked...
302          *
303          * And this page cache invalidation has to be after dio->end_io(), as
304          * some filesystems convert unwritten extents to real allocations in
305          * end_io() when necessary, otherwise a racing buffer read would cache
306          * zeros from unwritten extents.
307          */
308         if (flags & DIO_COMPLETE_INVALIDATE &&
309             ret > 0 && dio->op == REQ_OP_WRITE &&
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         }
317 
318         inode_dio_end(dio->inode);
319 
320         if (flags & DIO_COMPLETE_ASYNC) {
321                 /*
322                  * generic_write_sync expects ki_pos to have been updated
323                  * already, but the submission path only does this for
324                  * synchronous I/O.
325                  */
326                 dio->iocb->ki_pos += transferred;
327 
328                 if (ret > 0 && dio->op == REQ_OP_WRITE)
329                         ret = generic_write_sync(dio->iocb, ret);
330                 dio->iocb->ki_complete(dio->iocb, ret, 0);
331         }
332 
333         kmem_cache_free(dio_cache, dio);
334         return ret;
335 }
336 
337 static void dio_aio_complete_work(struct work_struct *work)
338 {
339         struct dio *dio = container_of(work, struct dio, complete_work);
340 
341         dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
342 }
343 
344 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio);
345 
346 /*
347  * Asynchronous IO callback. 
348  */
349 static void dio_bio_end_aio(struct bio *bio)
350 {
351         struct dio *dio = bio->bi_private;
352         unsigned long remaining;
353         unsigned long flags;
354         bool defer_completion = false;
355 
356         /* cleanup the bio */
357         dio_bio_complete(dio, bio);
358 
359         spin_lock_irqsave(&dio->bio_lock, flags);
360         remaining = --dio->refcount;
361         if (remaining == 1 && dio->waiter)
362                 wake_up_process(dio->waiter);
363         spin_unlock_irqrestore(&dio->bio_lock, flags);
364 
365         if (remaining == 0) {
366                 /*
367                  * Defer completion when defer_completion is set or
368                  * when the inode has pages mapped and this is AIO write.
369                  * We need to invalidate those pages because there is a
370                  * chance they contain stale data in the case buffered IO
371                  * went in between AIO submission and completion into the
372                  * same region.
373                  */
374                 if (dio->result)
375                         defer_completion = dio->defer_completion ||
376                                            (dio->op == REQ_OP_WRITE &&
377                                             dio->inode->i_mapping->nrpages);
378                 if (defer_completion) {
379                         INIT_WORK(&dio->complete_work, dio_aio_complete_work);
380                         queue_work(dio->inode->i_sb->s_dio_done_wq,
381                                    &dio->complete_work);
382                 } else {
383                         dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
384                 }
385         }
386 }
387 
388 /*
389  * The BIO completion handler simply queues the BIO up for the process-context
390  * handler.
391  *
392  * During I/O bi_private points at the dio.  After I/O, bi_private is used to
393  * implement a singly-linked list of completed BIOs, at dio->bio_list.
394  */
395 static void dio_bio_end_io(struct bio *bio)
396 {
397         struct dio *dio = bio->bi_private;
398         unsigned long flags;
399 
400         spin_lock_irqsave(&dio->bio_lock, flags);
401         bio->bi_private = dio->bio_list;
402         dio->bio_list = bio;
403         if (--dio->refcount == 1 && dio->waiter)
404                 wake_up_process(dio->waiter);
405         spin_unlock_irqrestore(&dio->bio_lock, flags);
406 }
407 
408 /**
409  * dio_end_io - handle the end io action for the given bio
410  * @bio: The direct io bio thats being completed
411  *
412  * This is meant to be called by any filesystem that uses their own dio_submit_t
413  * so that the DIO specific endio actions are dealt with after the filesystem
414  * has done it's completion work.
415  */
416 void dio_end_io(struct bio *bio)
417 {
418         struct dio *dio = bio->bi_private;
419 
420         if (dio->is_async)
421                 dio_bio_end_aio(bio);
422         else
423                 dio_bio_end_io(bio);
424 }
425 EXPORT_SYMBOL_GPL(dio_end_io);
426 
427 static inline void
428 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
429               struct block_device *bdev,
430               sector_t first_sector, int nr_vecs)
431 {
432         struct bio *bio;
433 
434         /*
435          * bio_alloc() is guaranteed to return a bio when allowed to sleep and
436          * we request a valid number of vectors.
437          */
438         bio = bio_alloc(GFP_KERNEL, nr_vecs);
439 
440         bio_set_dev(bio, bdev);
441         bio->bi_iter.bi_sector = first_sector;
442         bio_set_op_attrs(bio, dio->op, dio->op_flags);
443         if (dio->is_async)
444                 bio->bi_end_io = dio_bio_end_aio;
445         else
446                 bio->bi_end_io = dio_bio_end_io;
447 
448         bio->bi_write_hint = dio->iocb->ki_hint;
449 
450         sdio->bio = bio;
451         sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
452 }
453 
454 /*
455  * In the AIO read case we speculatively dirty the pages before starting IO.
456  * During IO completion, any of these pages which happen to have been written
457  * back will be redirtied by bio_check_pages_dirty().
458  *
459  * bios hold a dio reference between submit_bio and ->end_io.
460  */
461 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
462 {
463         struct bio *bio = sdio->bio;
464         unsigned long flags;
465 
466         bio->bi_private = dio;
467 
468         spin_lock_irqsave(&dio->bio_lock, flags);
469         dio->refcount++;
470         spin_unlock_irqrestore(&dio->bio_lock, flags);
471 
472         if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty)
473                 bio_set_pages_dirty(bio);
474 
475         dio->bio_disk = bio->bi_disk;
476 
477         if (sdio->submit_io) {
478                 sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio);
479                 dio->bio_cookie = BLK_QC_T_NONE;
480         } else
481                 dio->bio_cookie = submit_bio(bio);
482 
483         sdio->bio = NULL;
484         sdio->boundary = 0;
485         sdio->logical_offset_in_bio = 0;
486 }
487 
488 /*
489  * Release any resources in case of a failure
490  */
491 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
492 {
493         while (sdio->head < sdio->tail)
494                 put_page(dio->pages[sdio->head++]);
495 }
496 
497 /*
498  * Wait for the next BIO to complete.  Remove it and return it.  NULL is
499  * returned once all BIOs have been completed.  This must only be called once
500  * all bios have been issued so that dio->refcount can only decrease.  This
501  * requires that that the caller hold a reference on the dio.
502  */
503 static struct bio *dio_await_one(struct dio *dio)
504 {
505         unsigned long flags;
506         struct bio *bio = NULL;
507 
508         spin_lock_irqsave(&dio->bio_lock, flags);
509 
510         /*
511          * Wait as long as the list is empty and there are bios in flight.  bio
512          * completion drops the count, maybe adds to the list, and wakes while
513          * holding the bio_lock so we don't need set_current_state()'s barrier
514          * and can call it after testing our condition.
515          */
516         while (dio->refcount > 1 && dio->bio_list == NULL) {
517                 __set_current_state(TASK_UNINTERRUPTIBLE);
518                 dio->waiter = current;
519                 spin_unlock_irqrestore(&dio->bio_lock, flags);
520                 if (!(dio->iocb->ki_flags & IOCB_HIPRI) ||
521                     !blk_poll(dio->bio_disk->queue, dio->bio_cookie, true))
522                         io_schedule();
523                 /* wake up sets us TASK_RUNNING */
524                 spin_lock_irqsave(&dio->bio_lock, flags);
525                 dio->waiter = NULL;
526         }
527         if (dio->bio_list) {
528                 bio = dio->bio_list;
529                 dio->bio_list = bio->bi_private;
530         }
531         spin_unlock_irqrestore(&dio->bio_lock, flags);
532         return bio;
533 }
534 
535 /*
536  * Process one completed BIO.  No locks are held.
537  */
538 static blk_status_t dio_bio_complete(struct dio *dio, struct bio *bio)
539 {
540         struct bio_vec *bvec;
541         unsigned i;
542         blk_status_t err = bio->bi_status;
543 
544         if (err) {
545                 if (err == BLK_STS_AGAIN && (bio->bi_opf & REQ_NOWAIT))
546                         dio->io_error = -EAGAIN;
547                 else
548                         dio->io_error = -EIO;
549         }
550 
551         if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty) {
552                 bio_check_pages_dirty(bio);     /* transfers ownership */
553         } else {
554                 bio_for_each_segment_all(bvec, bio, i) {
555                         struct page *page = bvec->bv_page;
556 
557                         if (dio->op == REQ_OP_READ && !PageCompound(page) &&
558                                         dio->should_dirty)
559                                 set_page_dirty_lock(page);
560                         put_page(page);
561                 }
562                 bio_put(bio);
563         }
564         return err;
565 }
566 
567 /*
568  * Wait on and process all in-flight BIOs.  This must only be called once
569  * all bios have been issued so that the refcount can only decrease.
570  * This just waits for all bios to make it through dio_bio_complete.  IO
571  * errors are propagated through dio->io_error and should be propagated via
572  * dio_complete().
573  */
574 static void dio_await_completion(struct dio *dio)
575 {
576         struct bio *bio;
577         do {
578                 bio = dio_await_one(dio);
579                 if (bio)
580                         dio_bio_complete(dio, bio);
581         } while (bio);
582 }
583 
584 /*
585  * A really large O_DIRECT read or write can generate a lot of BIOs.  So
586  * to keep the memory consumption sane we periodically reap any completed BIOs
587  * during the BIO generation phase.
588  *
589  * This also helps to limit the peak amount of pinned userspace memory.
590  */
591 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
592 {
593         int ret = 0;
594 
595         if (sdio->reap_counter++ >= 64) {
596                 while (dio->bio_list) {
597                         unsigned long flags;
598                         struct bio *bio;
599                         int ret2;
600 
601                         spin_lock_irqsave(&dio->bio_lock, flags);
602                         bio = dio->bio_list;
603                         dio->bio_list = bio->bi_private;
604                         spin_unlock_irqrestore(&dio->bio_lock, flags);
605                         ret2 = blk_status_to_errno(dio_bio_complete(dio, bio));
606                         if (ret == 0)
607                                 ret = ret2;
608                 }
609                 sdio->reap_counter = 0;
610         }
611         return ret;
612 }
613 
614 /*
615  * Create workqueue for deferred direct IO completions. We allocate the
616  * workqueue when it's first needed. This avoids creating workqueue for
617  * filesystems that don't need it and also allows us to create the workqueue
618  * late enough so the we can include s_id in the name of the workqueue.
619  */
620 int sb_init_dio_done_wq(struct super_block *sb)
621 {
622         struct workqueue_struct *old;
623         struct workqueue_struct *wq = alloc_workqueue("dio/%s",
624                                                       WQ_MEM_RECLAIM, 0,
625                                                       sb->s_id);
626         if (!wq)
627                 return -ENOMEM;
628         /*
629          * This has to be atomic as more DIOs can race to create the workqueue
630          */
631         old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
632         /* Someone created workqueue before us? Free ours... */
633         if (old)
634                 destroy_workqueue(wq);
635         return 0;
636 }
637 
638 static int dio_set_defer_completion(struct dio *dio)
639 {
640         struct super_block *sb = dio->inode->i_sb;
641 
642         if (dio->defer_completion)
643                 return 0;
644         dio->defer_completion = true;
645         if (!sb->s_dio_done_wq)
646                 return sb_init_dio_done_wq(sb);
647         return 0;
648 }
649 
650 /*
651  * Call into the fs to map some more disk blocks.  We record the current number
652  * of available blocks at sdio->blocks_available.  These are in units of the
653  * fs blocksize, i_blocksize(inode).
654  *
655  * The fs is allowed to map lots of blocks at once.  If it wants to do that,
656  * it uses the passed inode-relative block number as the file offset, as usual.
657  *
658  * get_block() is passed the number of i_blkbits-sized blocks which direct_io
659  * has remaining to do.  The fs should not map more than this number of blocks.
660  *
661  * If the fs has mapped a lot of blocks, it should populate bh->b_size to
662  * indicate how much contiguous disk space has been made available at
663  * bh->b_blocknr.
664  *
665  * If *any* of the mapped blocks are new, then the fs must set buffer_new().
666  * This isn't very efficient...
667  *
668  * In the case of filesystem holes: the fs may return an arbitrarily-large
669  * hole by returning an appropriate value in b_size and by clearing
670  * buffer_mapped().  However the direct-io code will only process holes one
671  * block at a time - it will repeatedly call get_block() as it walks the hole.
672  */
673 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
674                            struct buffer_head *map_bh)
675 {
676         int ret;
677         sector_t fs_startblk;   /* Into file, in filesystem-sized blocks */
678         sector_t fs_endblk;     /* Into file, in filesystem-sized blocks */
679         unsigned long fs_count; /* Number of filesystem-sized blocks */
680         int create;
681         unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
682         loff_t i_size;
683 
684         /*
685          * If there was a memory error and we've overwritten all the
686          * mapped blocks then we can now return that memory error
687          */
688         ret = dio->page_errors;
689         if (ret == 0) {
690                 BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
691                 fs_startblk = sdio->block_in_file >> sdio->blkfactor;
692                 fs_endblk = (sdio->final_block_in_request - 1) >>
693                                         sdio->blkfactor;
694                 fs_count = fs_endblk - fs_startblk + 1;
695 
696                 map_bh->b_state = 0;
697                 map_bh->b_size = fs_count << i_blkbits;
698 
699                 /*
700                  * For writes that could fill holes inside i_size on a
701                  * DIO_SKIP_HOLES filesystem we forbid block creations: only
702                  * overwrites are permitted. We will return early to the caller
703                  * once we see an unmapped buffer head returned, and the caller
704                  * will fall back to buffered I/O.
705                  *
706                  * Otherwise the decision is left to the get_blocks method,
707                  * which may decide to handle it or also return an unmapped
708                  * buffer head.
709                  */
710                 create = dio->op == REQ_OP_WRITE;
711                 if (dio->flags & DIO_SKIP_HOLES) {
712                         i_size = i_size_read(dio->inode);
713                         if (i_size && fs_startblk <= (i_size - 1) >> i_blkbits)
714                                 create = 0;
715                 }
716 
717                 ret = (*sdio->get_block)(dio->inode, fs_startblk,
718                                                 map_bh, create);
719 
720                 /* Store for completion */
721                 dio->private = map_bh->b_private;
722 
723                 if (ret == 0 && buffer_defer_completion(map_bh))
724                         ret = dio_set_defer_completion(dio);
725         }
726         return ret;
727 }
728 
729 /*
730  * There is no bio.  Make one now.
731  */
732 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
733                 sector_t start_sector, struct buffer_head *map_bh)
734 {
735         sector_t sector;
736         int ret, nr_pages;
737 
738         ret = dio_bio_reap(dio, sdio);
739         if (ret)
740                 goto out;
741         sector = start_sector << (sdio->blkbits - 9);
742         nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES);
743         BUG_ON(nr_pages <= 0);
744         dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
745         sdio->boundary = 0;
746 out:
747         return ret;
748 }
749 
750 /*
751  * Attempt to put the current chunk of 'cur_page' into the current BIO.  If
752  * that was successful then update final_block_in_bio and take a ref against
753  * the just-added page.
754  *
755  * Return zero on success.  Non-zero means the caller needs to start a new BIO.
756  */
757 static inline int dio_bio_add_page(struct dio_submit *sdio)
758 {
759         int ret;
760 
761         ret = bio_add_page(sdio->bio, sdio->cur_page,
762                         sdio->cur_page_len, sdio->cur_page_offset);
763         if (ret == sdio->cur_page_len) {
764                 /*
765                  * Decrement count only, if we are done with this page
766                  */
767                 if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
768                         sdio->pages_in_io--;
769                 get_page(sdio->cur_page);
770                 sdio->final_block_in_bio = sdio->cur_page_block +
771                         (sdio->cur_page_len >> sdio->blkbits);
772                 ret = 0;
773         } else {
774                 ret = 1;
775         }
776         return ret;
777 }
778                 
779 /*
780  * Put cur_page under IO.  The section of cur_page which is described by
781  * cur_page_offset,cur_page_len is put into a BIO.  The section of cur_page
782  * starts on-disk at cur_page_block.
783  *
784  * We take a ref against the page here (on behalf of its presence in the bio).
785  *
786  * The caller of this function is responsible for removing cur_page from the
787  * dio, and for dropping the refcount which came from that presence.
788  */
789 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
790                 struct buffer_head *map_bh)
791 {
792         int ret = 0;
793 
794         if (sdio->bio) {
795                 loff_t cur_offset = sdio->cur_page_fs_offset;
796                 loff_t bio_next_offset = sdio->logical_offset_in_bio +
797                         sdio->bio->bi_iter.bi_size;
798 
799                 /*
800                  * See whether this new request is contiguous with the old.
801                  *
802                  * Btrfs cannot handle having logically non-contiguous requests
803                  * submitted.  For example if you have
804                  *
805                  * Logical:  [0-4095][HOLE][8192-12287]
806                  * Physical: [0-4095]      [4096-8191]
807                  *
808                  * We cannot submit those pages together as one BIO.  So if our
809                  * current logical offset in the file does not equal what would
810                  * be the next logical offset in the bio, submit the bio we
811                  * have.
812                  */
813                 if (sdio->final_block_in_bio != sdio->cur_page_block ||
814                     cur_offset != bio_next_offset)
815                         dio_bio_submit(dio, sdio);
816         }
817 
818         if (sdio->bio == NULL) {
819                 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
820                 if (ret)
821                         goto out;
822         }
823 
824         if (dio_bio_add_page(sdio) != 0) {
825                 dio_bio_submit(dio, sdio);
826                 ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
827                 if (ret == 0) {
828                         ret = dio_bio_add_page(sdio);
829                         BUG_ON(ret != 0);
830                 }
831         }
832 out:
833         return ret;
834 }
835 
836 /*
837  * An autonomous function to put a chunk of a page under deferred IO.
838  *
839  * The caller doesn't actually know (or care) whether this piece of page is in
840  * a BIO, or is under IO or whatever.  We just take care of all possible 
841  * situations here.  The separation between the logic of do_direct_IO() and
842  * that of submit_page_section() is important for clarity.  Please don't break.
843  *
844  * The chunk of page starts on-disk at blocknr.
845  *
846  * We perform deferred IO, by recording the last-submitted page inside our
847  * private part of the dio structure.  If possible, we just expand the IO
848  * across that page here.
849  *
850  * If that doesn't work out then we put the old page into the bio and add this
851  * page to the dio instead.
852  */
853 static inline int
854 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
855                     unsigned offset, unsigned len, sector_t blocknr,
856                     struct buffer_head *map_bh)
857 {
858         int ret = 0;
859 
860         if (dio->op == REQ_OP_WRITE) {
861                 /*
862                  * Read accounting is performed in submit_bio()
863                  */
864                 task_io_account_write(len);
865         }
866 
867         /*
868          * Can we just grow the current page's presence in the dio?
869          */
870         if (sdio->cur_page == page &&
871             sdio->cur_page_offset + sdio->cur_page_len == offset &&
872             sdio->cur_page_block +
873             (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
874                 sdio->cur_page_len += len;
875                 goto out;
876         }
877 
878         /*
879          * If there's a deferred page already there then send it.
880          */
881         if (sdio->cur_page) {
882                 ret = dio_send_cur_page(dio, sdio, map_bh);
883                 put_page(sdio->cur_page);
884                 sdio->cur_page = NULL;
885                 if (ret)
886                         return ret;
887         }
888 
889         get_page(page);         /* It is in dio */
890         sdio->cur_page = page;
891         sdio->cur_page_offset = offset;
892         sdio->cur_page_len = len;
893         sdio->cur_page_block = blocknr;
894         sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
895 out:
896         /*
897          * If sdio->boundary then we want to schedule the IO now to
898          * avoid metadata seeks.
899          */
900         if (sdio->boundary) {
901                 ret = dio_send_cur_page(dio, sdio, map_bh);
902                 if (sdio->bio)
903                         dio_bio_submit(dio, sdio);
904                 put_page(sdio->cur_page);
905                 sdio->cur_page = NULL;
906         }
907         return ret;
908 }
909 
910 /*
911  * If we are not writing the entire block and get_block() allocated
912  * the block for us, we need to fill-in the unused portion of the
913  * block with zeros. This happens only if user-buffer, fileoffset or
914  * io length is not filesystem block-size multiple.
915  *
916  * `end' is zero if we're doing the start of the IO, 1 at the end of the
917  * IO.
918  */
919 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
920                 int end, struct buffer_head *map_bh)
921 {
922         unsigned dio_blocks_per_fs_block;
923         unsigned this_chunk_blocks;     /* In dio_blocks */
924         unsigned this_chunk_bytes;
925         struct page *page;
926 
927         sdio->start_zero_done = 1;
928         if (!sdio->blkfactor || !buffer_new(map_bh))
929                 return;
930 
931         dio_blocks_per_fs_block = 1 << sdio->blkfactor;
932         this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
933 
934         if (!this_chunk_blocks)
935                 return;
936 
937         /*
938          * We need to zero out part of an fs block.  It is either at the
939          * beginning or the end of the fs block.
940          */
941         if (end) 
942                 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
943 
944         this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
945 
946         page = ZERO_PAGE(0);
947         if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
948                                 sdio->next_block_for_io, map_bh))
949                 return;
950 
951         sdio->next_block_for_io += this_chunk_blocks;
952 }
953 
954 /*
955  * Walk the user pages, and the file, mapping blocks to disk and generating
956  * a sequence of (page,offset,len,block) mappings.  These mappings are injected
957  * into submit_page_section(), which takes care of the next stage of submission
958  *
959  * Direct IO against a blockdev is different from a file.  Because we can
960  * happily perform page-sized but 512-byte aligned IOs.  It is important that
961  * blockdev IO be able to have fine alignment and large sizes.
962  *
963  * So what we do is to permit the ->get_block function to populate bh.b_size
964  * with the size of IO which is permitted at this offset and this i_blkbits.
965  *
966  * For best results, the blockdev should be set up with 512-byte i_blkbits and
967  * it should set b_size to PAGE_SIZE or more inside get_block().  This gives
968  * fine alignment but still allows this function to work in PAGE_SIZE units.
969  */
970 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
971                         struct buffer_head *map_bh)
972 {
973         const unsigned blkbits = sdio->blkbits;
974         const unsigned i_blkbits = blkbits + sdio->blkfactor;
975         int ret = 0;
976 
977         while (sdio->block_in_file < sdio->final_block_in_request) {
978                 struct page *page;
979                 size_t from, to;
980 
981                 page = dio_get_page(dio, sdio);
982                 if (IS_ERR(page)) {
983                         ret = PTR_ERR(page);
984                         goto out;
985                 }
986                 from = sdio->head ? 0 : sdio->from;
987                 to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
988                 sdio->head++;
989 
990                 while (from < to) {
991                         unsigned this_chunk_bytes;      /* # of bytes mapped */
992                         unsigned this_chunk_blocks;     /* # of blocks */
993                         unsigned u;
994 
995                         if (sdio->blocks_available == 0) {
996                                 /*
997                                  * Need to go and map some more disk
998                                  */
999                                 unsigned long blkmask;
1000                                 unsigned long dio_remainder;
1001 
1002                                 ret = get_more_blocks(dio, sdio, map_bh);
1003                                 if (ret) {
1004                                         put_page(page);
1005                                         goto out;
1006                                 }
1007                                 if (!buffer_mapped(map_bh))
1008                                         goto do_holes;
1009 
1010                                 sdio->blocks_available =
1011                                                 map_bh->b_size >> blkbits;
1012                                 sdio->next_block_for_io =
1013                                         map_bh->b_blocknr << sdio->blkfactor;
1014                                 if (buffer_new(map_bh)) {
1015                                         clean_bdev_aliases(
1016                                                 map_bh->b_bdev,
1017                                                 map_bh->b_blocknr,
1018                                                 map_bh->b_size >> i_blkbits);
1019                                 }
1020 
1021                                 if (!sdio->blkfactor)
1022                                         goto do_holes;
1023 
1024                                 blkmask = (1 << sdio->blkfactor) - 1;
1025                                 dio_remainder = (sdio->block_in_file & blkmask);
1026 
1027                                 /*
1028                                  * If we are at the start of IO and that IO
1029                                  * starts partway into a fs-block,
1030                                  * dio_remainder will be non-zero.  If the IO
1031                                  * is a read then we can simply advance the IO
1032                                  * cursor to the first block which is to be
1033                                  * read.  But if the IO is a write and the
1034                                  * block was newly allocated we cannot do that;
1035                                  * the start of the fs block must be zeroed out
1036                                  * on-disk
1037                                  */
1038                                 if (!buffer_new(map_bh))
1039                                         sdio->next_block_for_io += dio_remainder;
1040                                 sdio->blocks_available -= dio_remainder;
1041                         }
1042 do_holes:
1043                         /* Handle holes */
1044                         if (!buffer_mapped(map_bh)) {
1045                                 loff_t i_size_aligned;
1046 
1047                                 /* AKPM: eargh, -ENOTBLK is a hack */
1048                                 if (dio->op == REQ_OP_WRITE) {
1049                                         put_page(page);
1050                                         return -ENOTBLK;
1051                                 }
1052 
1053                                 /*
1054                                  * Be sure to account for a partial block as the
1055                                  * last block in the file
1056                                  */
1057                                 i_size_aligned = ALIGN(i_size_read(dio->inode),
1058                                                         1 << blkbits);
1059                                 if (sdio->block_in_file >=
1060                                                 i_size_aligned >> blkbits) {
1061                                         /* We hit eof */
1062                                         put_page(page);
1063                                         goto out;
1064                                 }
1065                                 zero_user(page, from, 1 << blkbits);
1066                                 sdio->block_in_file++;
1067                                 from += 1 << blkbits;
1068                                 dio->result += 1 << blkbits;
1069                                 goto next_block;
1070                         }
1071 
1072                         /*
1073                          * If we're performing IO which has an alignment which
1074                          * is finer than the underlying fs, go check to see if
1075                          * we must zero out the start of this block.
1076                          */
1077                         if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1078                                 dio_zero_block(dio, sdio, 0, map_bh);
1079 
1080                         /*
1081                          * Work out, in this_chunk_blocks, how much disk we
1082                          * can add to this page
1083                          */
1084                         this_chunk_blocks = sdio->blocks_available;
1085                         u = (to - from) >> blkbits;
1086                         if (this_chunk_blocks > u)
1087                                 this_chunk_blocks = u;
1088                         u = sdio->final_block_in_request - sdio->block_in_file;
1089                         if (this_chunk_blocks > u)
1090                                 this_chunk_blocks = u;
1091                         this_chunk_bytes = this_chunk_blocks << blkbits;
1092                         BUG_ON(this_chunk_bytes == 0);
1093 
1094                         if (this_chunk_blocks == sdio->blocks_available)
1095                                 sdio->boundary = buffer_boundary(map_bh);
1096                         ret = submit_page_section(dio, sdio, page,
1097                                                   from,
1098                                                   this_chunk_bytes,
1099                                                   sdio->next_block_for_io,
1100                                                   map_bh);
1101                         if (ret) {
1102                                 put_page(page);
1103                                 goto out;
1104                         }
1105                         sdio->next_block_for_io += this_chunk_blocks;
1106 
1107                         sdio->block_in_file += this_chunk_blocks;
1108                         from += this_chunk_bytes;
1109                         dio->result += this_chunk_bytes;
1110                         sdio->blocks_available -= this_chunk_blocks;
1111 next_block:
1112                         BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1113                         if (sdio->block_in_file == sdio->final_block_in_request)
1114                                 break;
1115                 }
1116 
1117                 /* Drop the ref which was taken in get_user_pages() */
1118                 put_page(page);
1119         }
1120 out:
1121         return ret;
1122 }
1123 
1124 static inline int drop_refcount(struct dio *dio)
1125 {
1126         int ret2;
1127         unsigned long flags;
1128 
1129         /*
1130          * Sync will always be dropping the final ref and completing the
1131          * operation.  AIO can if it was a broken operation described above or
1132          * in fact if all the bios race to complete before we get here.  In
1133          * that case dio_complete() translates the EIOCBQUEUED into the proper
1134          * return code that the caller will hand to ->complete().
1135          *
1136          * This is managed by the bio_lock instead of being an atomic_t so that
1137          * completion paths can drop their ref and use the remaining count to
1138          * decide to wake the submission path atomically.
1139          */
1140         spin_lock_irqsave(&dio->bio_lock, flags);
1141         ret2 = --dio->refcount;
1142         spin_unlock_irqrestore(&dio->bio_lock, flags);
1143         return ret2;
1144 }
1145 
1146 /*
1147  * This is a library function for use by filesystem drivers.
1148  *
1149  * The locking rules are governed by the flags parameter:
1150  *  - if the flags value contains DIO_LOCKING we use a fancy locking
1151  *    scheme for dumb filesystems.
1152  *    For writes this function is called under i_mutex and returns with
1153  *    i_mutex held, for reads, i_mutex is not held on entry, but it is
1154  *    taken and dropped again before returning.
1155  *  - if the flags value does NOT contain DIO_LOCKING we don't use any
1156  *    internal locking but rather rely on the filesystem to synchronize
1157  *    direct I/O reads/writes versus each other and truncate.
1158  *
1159  * To help with locking against truncate we incremented the i_dio_count
1160  * counter before starting direct I/O, and decrement it once we are done.
1161  * Truncate can wait for it to reach zero to provide exclusion.  It is
1162  * expected that filesystem provide exclusion between new direct I/O
1163  * and truncates.  For DIO_LOCKING filesystems this is done by i_mutex,
1164  * but other filesystems need to take care of this on their own.
1165  *
1166  * NOTE: if you pass "sdio" to anything by pointer make sure that function
1167  * is always inlined. Otherwise gcc is unable to split the structure into
1168  * individual fields and will generate much worse code. This is important
1169  * for the whole file.
1170  */
1171 static inline ssize_t
1172 do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1173                       struct block_device *bdev, struct iov_iter *iter,
1174                       get_block_t get_block, dio_iodone_t end_io,
1175                       dio_submit_t submit_io, int flags)
1176 {
1177         unsigned i_blkbits = READ_ONCE(inode->i_blkbits);
1178         unsigned blkbits = i_blkbits;
1179         unsigned blocksize_mask = (1 << blkbits) - 1;
1180         ssize_t retval = -EINVAL;
1181         const size_t count = iov_iter_count(iter);
1182         loff_t offset = iocb->ki_pos;
1183         const loff_t end = offset + count;
1184         struct dio *dio;
1185         struct dio_submit sdio = { 0, };
1186         struct buffer_head map_bh = { 0, };
1187         struct blk_plug plug;
1188         unsigned long align = offset | iov_iter_alignment(iter);
1189 
1190         /*
1191          * Avoid references to bdev if not absolutely needed to give
1192          * the early prefetch in the caller enough time.
1193          */
1194 
1195         if (align & blocksize_mask) {
1196                 if (bdev)
1197                         blkbits = blksize_bits(bdev_logical_block_size(bdev));
1198                 blocksize_mask = (1 << blkbits) - 1;
1199                 if (align & blocksize_mask)
1200                         goto out;
1201         }
1202 
1203         /* watch out for a 0 len io from a tricksy fs */
1204         if (iov_iter_rw(iter) == READ && !count)
1205                 return 0;
1206 
1207         dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1208         retval = -ENOMEM;
1209         if (!dio)
1210                 goto out;
1211         /*
1212          * Believe it or not, zeroing out the page array caused a .5%
1213          * performance regression in a database benchmark.  So, we take
1214          * care to only zero out what's needed.
1215          */
1216         memset(dio, 0, offsetof(struct dio, pages));
1217 
1218         dio->flags = flags;
1219         if (dio->flags & DIO_LOCKING) {
1220                 if (iov_iter_rw(iter) == READ) {
1221                         struct address_space *mapping =
1222                                         iocb->ki_filp->f_mapping;
1223 
1224                         /* will be released by direct_io_worker */
1225                         inode_lock(inode);
1226 
1227                         retval = filemap_write_and_wait_range(mapping, offset,
1228                                                               end - 1);
1229                         if (retval) {
1230                                 inode_unlock(inode);
1231                                 kmem_cache_free(dio_cache, dio);
1232                                 goto out;
1233                         }
1234                 }
1235         }
1236 
1237         /* Once we sampled i_size check for reads beyond EOF */
1238         dio->i_size = i_size_read(inode);
1239         if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1240                 if (dio->flags & DIO_LOCKING)
1241                         inode_unlock(inode);
1242                 kmem_cache_free(dio_cache, dio);
1243                 retval = 0;
1244                 goto out;
1245         }
1246 
1247         /*
1248          * For file extending writes updating i_size before data writeouts
1249          * complete can expose uninitialized blocks in dumb filesystems.
1250          * In that case we need to wait for I/O completion even if asked
1251          * for an asynchronous write.
1252          */
1253         if (is_sync_kiocb(iocb))
1254                 dio->is_async = false;
1255         else if (iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
1256                 dio->is_async = false;
1257         else
1258                 dio->is_async = true;
1259 
1260         dio->inode = inode;
1261         if (iov_iter_rw(iter) == WRITE) {
1262                 dio->op = REQ_OP_WRITE;
1263                 dio->op_flags = REQ_SYNC | REQ_IDLE;
1264                 if (iocb->ki_flags & IOCB_NOWAIT)
1265                         dio->op_flags |= REQ_NOWAIT;
1266         } else {
1267                 dio->op = REQ_OP_READ;
1268         }
1269         if (iocb->ki_flags & IOCB_HIPRI)
1270                 dio->op_flags |= REQ_HIPRI;
1271 
1272         /*
1273          * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1274          * so that we can call ->fsync.
1275          */
1276         if (dio->is_async && iov_iter_rw(iter) == WRITE) {
1277                 retval = 0;
1278                 if (iocb->ki_flags & IOCB_DSYNC)
1279                         retval = dio_set_defer_completion(dio);
1280                 else if (!dio->inode->i_sb->s_dio_done_wq) {
1281                         /*
1282                          * In case of AIO write racing with buffered read we
1283                          * need to defer completion. We can't decide this now,
1284                          * however the workqueue needs to be initialized here.
1285                          */
1286                         retval = sb_init_dio_done_wq(dio->inode->i_sb);
1287                 }
1288                 if (retval) {
1289                         /*
1290                          * We grab i_mutex only for reads so we don't have
1291                          * to release it here
1292                          */
1293                         kmem_cache_free(dio_cache, dio);
1294                         goto out;
1295                 }
1296         }
1297 
1298         /*
1299          * Will be decremented at I/O completion time.
1300          */
1301         inode_dio_begin(inode);
1302 
1303         retval = 0;
1304         sdio.blkbits = blkbits;
1305         sdio.blkfactor = i_blkbits - blkbits;
1306         sdio.block_in_file = offset >> blkbits;
1307 
1308         sdio.get_block = get_block;
1309         dio->end_io = end_io;
1310         sdio.submit_io = submit_io;
1311         sdio.final_block_in_bio = -1;
1312         sdio.next_block_for_io = -1;
1313 
1314         dio->iocb = iocb;
1315 
1316         spin_lock_init(&dio->bio_lock);
1317         dio->refcount = 1;
1318 
1319         dio->should_dirty = iter_is_iovec(iter) && iov_iter_rw(iter) == READ;
1320         sdio.iter = iter;
1321         sdio.final_block_in_request = end >> blkbits;
1322 
1323         /*
1324          * In case of non-aligned buffers, we may need 2 more
1325          * pages since we need to zero out first and last block.
1326          */
1327         if (unlikely(sdio.blkfactor))
1328                 sdio.pages_in_io = 2;
1329 
1330         sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
1331 
1332         blk_start_plug(&plug);
1333 
1334         retval = do_direct_IO(dio, &sdio, &map_bh);
1335         if (retval)
1336                 dio_cleanup(dio, &sdio);
1337 
1338         if (retval == -ENOTBLK) {
1339                 /*
1340                  * The remaining part of the request will be
1341                  * be handled by buffered I/O when we return
1342                  */
1343                 retval = 0;
1344         }
1345         /*
1346          * There may be some unwritten disk at the end of a part-written
1347          * fs-block-sized block.  Go zero that now.
1348          */
1349         dio_zero_block(dio, &sdio, 1, &map_bh);
1350 
1351         if (sdio.cur_page) {
1352                 ssize_t ret2;
1353 
1354                 ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1355                 if (retval == 0)
1356                         retval = ret2;
1357                 put_page(sdio.cur_page);
1358                 sdio.cur_page = NULL;
1359         }
1360         if (sdio.bio)
1361                 dio_bio_submit(dio, &sdio);
1362 
1363         blk_finish_plug(&plug);
1364 
1365         /*
1366          * It is possible that, we return short IO due to end of file.
1367          * In that case, we need to release all the pages we got hold on.
1368          */
1369         dio_cleanup(dio, &sdio);
1370 
1371         /*
1372          * All block lookups have been performed. For READ requests
1373          * we can let i_mutex go now that its achieved its purpose
1374          * of protecting us from looking up uninitialized blocks.
1375          */
1376         if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
1377                 inode_unlock(dio->inode);
1378 
1379         /*
1380          * The only time we want to leave bios in flight is when a successful
1381          * partial aio read or full aio write have been setup.  In that case
1382          * bio completion will call aio_complete.  The only time it's safe to
1383          * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1384          * This had *better* be the only place that raises -EIOCBQUEUED.
1385          */
1386         BUG_ON(retval == -EIOCBQUEUED);
1387         if (dio->is_async && retval == 0 && dio->result &&
1388             (iov_iter_rw(iter) == READ || dio->result == count))
1389                 retval = -EIOCBQUEUED;
1390         else
1391                 dio_await_completion(dio);
1392 
1393         if (drop_refcount(dio) == 0) {
1394                 retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE);
1395         } else
1396                 BUG_ON(retval != -EIOCBQUEUED);
1397 
1398 out:
1399         return retval;
1400 }
1401 
1402 ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1403                              struct block_device *bdev, struct iov_iter *iter,
1404                              get_block_t get_block,
1405                              dio_iodone_t end_io, dio_submit_t submit_io,
1406                              int flags)
1407 {
1408         /*
1409          * The block device state is needed in the end to finally
1410          * submit everything.  Since it's likely to be cache cold
1411          * prefetch it here as first thing to hide some of the
1412          * latency.
1413          *
1414          * Attempt to prefetch the pieces we likely need later.
1415          */
1416         prefetch(&bdev->bd_disk->part_tbl);
1417         prefetch(bdev->bd_queue);
1418         prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1419 
1420         return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block,
1421                                      end_io, submit_io, flags);
1422 }
1423 
1424 EXPORT_SYMBOL(__blockdev_direct_IO);
1425 
1426 static __init int dio_init(void)
1427 {
1428         dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1429         return 0;
1430 }
1431 module_init(dio_init)
1432 

~ [ 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