Back to home page

LXR

 
 

    


0001 /*
0002  * fs/direct-io.c
0003  *
0004  * Copyright (C) 2002, Linus Torvalds.
0005  *
0006  * O_DIRECT
0007  *
0008  * 04Jul2002    Andrew Morton
0009  *      Initial version
0010  * 11Sep2002    janetinc@us.ibm.com
0011  *      added readv/writev support.
0012  * 29Oct2002    Andrew Morton
0013  *      rewrote bio_add_page() support.
0014  * 30Oct2002    pbadari@us.ibm.com
0015  *      added support for non-aligned IO.
0016  * 06Nov2002    pbadari@us.ibm.com
0017  *      added asynchronous IO support.
0018  * 21Jul2003    nathans@sgi.com
0019  *      added IO completion notifier.
0020  */
0021 
0022 #include <linux/kernel.h>
0023 #include <linux/module.h>
0024 #include <linux/types.h>
0025 #include <linux/fs.h>
0026 #include <linux/mm.h>
0027 #include <linux/slab.h>
0028 #include <linux/highmem.h>
0029 #include <linux/pagemap.h>
0030 #include <linux/task_io_accounting_ops.h>
0031 #include <linux/bio.h>
0032 #include <linux/wait.h>
0033 #include <linux/err.h>
0034 #include <linux/blkdev.h>
0035 #include <linux/buffer_head.h>
0036 #include <linux/rwsem.h>
0037 #include <linux/uio.h>
0038 #include <linux/atomic.h>
0039 #include <linux/prefetch.h>
0040 
0041 /*
0042  * How many user pages to map in one call to get_user_pages().  This determines
0043  * the size of a structure in the slab cache
0044  */
0045 #define DIO_PAGES   64
0046 
0047 /*
0048  * This code generally works in units of "dio_blocks".  A dio_block is
0049  * somewhere between the hard sector size and the filesystem block size.  it
0050  * is determined on a per-invocation basis.   When talking to the filesystem
0051  * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
0052  * down by dio->blkfactor.  Similarly, fs-blocksize quantities are converted
0053  * to bio_block quantities by shifting left by blkfactor.
0054  *
0055  * If blkfactor is zero then the user's request was aligned to the filesystem's
0056  * blocksize.
0057  */
0058 
0059 /* dio_state only used in the submission path */
0060 
0061 struct dio_submit {
0062     struct bio *bio;        /* bio under assembly */
0063     unsigned blkbits;       /* doesn't change */
0064     unsigned blkfactor;     /* When we're using an alignment which
0065                        is finer than the filesystem's soft
0066                        blocksize, this specifies how much
0067                        finer.  blkfactor=2 means 1/4-block
0068                        alignment.  Does not change */
0069     unsigned start_zero_done;   /* flag: sub-blocksize zeroing has
0070                        been performed at the start of a
0071                        write */
0072     int pages_in_io;        /* approximate total IO pages */
0073     sector_t block_in_file;     /* Current offset into the underlying
0074                        file in dio_block units. */
0075     unsigned blocks_available;  /* At block_in_file.  changes */
0076     int reap_counter;       /* rate limit reaping */
0077     sector_t final_block_in_request;/* doesn't change */
0078     int boundary;           /* prev block is at a boundary */
0079     get_block_t *get_block;     /* block mapping function */
0080     dio_submit_t *submit_io;    /* IO submition function */
0081 
0082     loff_t logical_offset_in_bio;   /* current first logical block in bio */
0083     sector_t final_block_in_bio;    /* current final block in bio + 1 */
0084     sector_t next_block_for_io; /* next block to be put under IO,
0085                        in dio_blocks units */
0086 
0087     /*
0088      * Deferred addition of a page to the dio.  These variables are
0089      * private to dio_send_cur_page(), submit_page_section() and
0090      * dio_bio_add_page().
0091      */
0092     struct page *cur_page;      /* The page */
0093     unsigned cur_page_offset;   /* Offset into it, in bytes */
0094     unsigned cur_page_len;      /* Nr of bytes at cur_page_offset */
0095     sector_t cur_page_block;    /* Where it starts */
0096     loff_t cur_page_fs_offset;  /* Offset in file */
0097 
0098     struct iov_iter *iter;
0099     /*
0100      * Page queue.  These variables belong to dio_refill_pages() and
0101      * dio_get_page().
0102      */
0103     unsigned head;          /* next page to process */
0104     unsigned tail;          /* last valid page + 1 */
0105     size_t from, to;
0106 };
0107 
0108 /* dio_state communicated between submission path and end_io */
0109 struct dio {
0110     int flags;          /* doesn't change */
0111     int op;
0112     int op_flags;
0113     blk_qc_t bio_cookie;
0114     struct block_device *bio_bdev;
0115     struct inode *inode;
0116     loff_t i_size;          /* i_size when submitted */
0117     dio_iodone_t *end_io;       /* IO completion function */
0118 
0119     void *private;          /* copy from map_bh.b_private */
0120 
0121     /* BIO completion state */
0122     spinlock_t bio_lock;        /* protects BIO fields below */
0123     int page_errors;        /* errno from get_user_pages() */
0124     int is_async;           /* is IO async ? */
0125     bool defer_completion;      /* defer AIO completion to workqueue? */
0126     bool should_dirty;      /* if pages should be dirtied */
0127     int io_error;           /* IO error in completion path */
0128     unsigned long refcount;     /* direct_io_worker() and bios */
0129     struct bio *bio_list;       /* singly linked via bi_private */
0130     struct task_struct *waiter; /* waiting task (NULL if none) */
0131 
0132     /* AIO related stuff */
0133     struct kiocb *iocb;     /* kiocb */
0134     ssize_t result;                 /* IO result */
0135 
0136     /*
0137      * pages[] (and any fields placed after it) are not zeroed out at
0138      * allocation time.  Don't add new fields after pages[] unless you
0139      * wish that they not be zeroed.
0140      */
0141     union {
0142         struct page *pages[DIO_PAGES];  /* page buffer */
0143         struct work_struct complete_work;/* deferred AIO completion */
0144     };
0145 } ____cacheline_aligned_in_smp;
0146 
0147 static struct kmem_cache *dio_cache __read_mostly;
0148 
0149 /*
0150  * How many pages are in the queue?
0151  */
0152 static inline unsigned dio_pages_present(struct dio_submit *sdio)
0153 {
0154     return sdio->tail - sdio->head;
0155 }
0156 
0157 /*
0158  * Go grab and pin some userspace pages.   Typically we'll get 64 at a time.
0159  */
0160 static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
0161 {
0162     ssize_t ret;
0163 
0164     ret = iov_iter_get_pages(sdio->iter, dio->pages, LONG_MAX, DIO_PAGES,
0165                 &sdio->from);
0166 
0167     if (ret < 0 && sdio->blocks_available && (dio->op == REQ_OP_WRITE)) {
0168         struct page *page = ZERO_PAGE(0);
0169         /*
0170          * A memory fault, but the filesystem has some outstanding
0171          * mapped blocks.  We need to use those blocks up to avoid
0172          * leaking stale data in the file.
0173          */
0174         if (dio->page_errors == 0)
0175             dio->page_errors = ret;
0176         get_page(page);
0177         dio->pages[0] = page;
0178         sdio->head = 0;
0179         sdio->tail = 1;
0180         sdio->from = 0;
0181         sdio->to = PAGE_SIZE;
0182         return 0;
0183     }
0184 
0185     if (ret >= 0) {
0186         iov_iter_advance(sdio->iter, ret);
0187         ret += sdio->from;
0188         sdio->head = 0;
0189         sdio->tail = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
0190         sdio->to = ((ret - 1) & (PAGE_SIZE - 1)) + 1;
0191         return 0;
0192     }
0193     return ret; 
0194 }
0195 
0196 /*
0197  * Get another userspace page.  Returns an ERR_PTR on error.  Pages are
0198  * buffered inside the dio so that we can call get_user_pages() against a
0199  * decent number of pages, less frequently.  To provide nicer use of the
0200  * L1 cache.
0201  */
0202 static inline struct page *dio_get_page(struct dio *dio,
0203                     struct dio_submit *sdio)
0204 {
0205     if (dio_pages_present(sdio) == 0) {
0206         int ret;
0207 
0208         ret = dio_refill_pages(dio, sdio);
0209         if (ret)
0210             return ERR_PTR(ret);
0211         BUG_ON(dio_pages_present(sdio) == 0);
0212     }
0213     return dio->pages[sdio->head];
0214 }
0215 
0216 /**
0217  * dio_complete() - called when all DIO BIO I/O has been completed
0218  * @offset: the byte offset in the file of the completed operation
0219  *
0220  * This drops i_dio_count, lets interested parties know that a DIO operation
0221  * has completed, and calculates the resulting return code for the operation.
0222  *
0223  * It lets the filesystem know if it registered an interest earlier via
0224  * get_block.  Pass the private field of the map buffer_head so that
0225  * filesystems can use it to hold additional state between get_block calls and
0226  * dio_complete.
0227  */
0228 static ssize_t dio_complete(struct dio *dio, ssize_t ret, bool is_async)
0229 {
0230     loff_t offset = dio->iocb->ki_pos;
0231     ssize_t transferred = 0;
0232 
0233     /*
0234      * AIO submission can race with bio completion to get here while
0235      * expecting to have the last io completed by bio completion.
0236      * In that case -EIOCBQUEUED is in fact not an error we want
0237      * to preserve through this call.
0238      */
0239     if (ret == -EIOCBQUEUED)
0240         ret = 0;
0241 
0242     if (dio->result) {
0243         transferred = dio->result;
0244 
0245         /* Check for short read case */
0246         if ((dio->op == REQ_OP_READ) &&
0247             ((offset + transferred) > dio->i_size))
0248             transferred = dio->i_size - offset;
0249         /* ignore EFAULT if some IO has been done */
0250         if (unlikely(ret == -EFAULT) && transferred)
0251             ret = 0;
0252     }
0253 
0254     if (ret == 0)
0255         ret = dio->page_errors;
0256     if (ret == 0)
0257         ret = dio->io_error;
0258     if (ret == 0)
0259         ret = transferred;
0260 
0261     if (dio->end_io) {
0262         int err;
0263 
0264         // XXX: ki_pos??
0265         err = dio->end_io(dio->iocb, offset, ret, dio->private);
0266         if (err)
0267             ret = err;
0268     }
0269 
0270     if (!(dio->flags & DIO_SKIP_DIO_COUNT))
0271         inode_dio_end(dio->inode);
0272 
0273     if (is_async) {
0274         /*
0275          * generic_write_sync expects ki_pos to have been updated
0276          * already, but the submission path only does this for
0277          * synchronous I/O.
0278          */
0279         dio->iocb->ki_pos += transferred;
0280 
0281         if (dio->op == REQ_OP_WRITE)
0282             ret = generic_write_sync(dio->iocb,  transferred);
0283         dio->iocb->ki_complete(dio->iocb, ret, 0);
0284     }
0285 
0286     kmem_cache_free(dio_cache, dio);
0287     return ret;
0288 }
0289 
0290 static void dio_aio_complete_work(struct work_struct *work)
0291 {
0292     struct dio *dio = container_of(work, struct dio, complete_work);
0293 
0294     dio_complete(dio, 0, true);
0295 }
0296 
0297 static int dio_bio_complete(struct dio *dio, struct bio *bio);
0298 
0299 /*
0300  * Asynchronous IO callback. 
0301  */
0302 static void dio_bio_end_aio(struct bio *bio)
0303 {
0304     struct dio *dio = bio->bi_private;
0305     unsigned long remaining;
0306     unsigned long flags;
0307 
0308     /* cleanup the bio */
0309     dio_bio_complete(dio, bio);
0310 
0311     spin_lock_irqsave(&dio->bio_lock, flags);
0312     remaining = --dio->refcount;
0313     if (remaining == 1 && dio->waiter)
0314         wake_up_process(dio->waiter);
0315     spin_unlock_irqrestore(&dio->bio_lock, flags);
0316 
0317     if (remaining == 0) {
0318         if (dio->result && dio->defer_completion) {
0319             INIT_WORK(&dio->complete_work, dio_aio_complete_work);
0320             queue_work(dio->inode->i_sb->s_dio_done_wq,
0321                    &dio->complete_work);
0322         } else {
0323             dio_complete(dio, 0, true);
0324         }
0325     }
0326 }
0327 
0328 /*
0329  * The BIO completion handler simply queues the BIO up for the process-context
0330  * handler.
0331  *
0332  * During I/O bi_private points at the dio.  After I/O, bi_private is used to
0333  * implement a singly-linked list of completed BIOs, at dio->bio_list.
0334  */
0335 static void dio_bio_end_io(struct bio *bio)
0336 {
0337     struct dio *dio = bio->bi_private;
0338     unsigned long flags;
0339 
0340     spin_lock_irqsave(&dio->bio_lock, flags);
0341     bio->bi_private = dio->bio_list;
0342     dio->bio_list = bio;
0343     if (--dio->refcount == 1 && dio->waiter)
0344         wake_up_process(dio->waiter);
0345     spin_unlock_irqrestore(&dio->bio_lock, flags);
0346 }
0347 
0348 /**
0349  * dio_end_io - handle the end io action for the given bio
0350  * @bio: The direct io bio thats being completed
0351  * @error: Error if there was one
0352  *
0353  * This is meant to be called by any filesystem that uses their own dio_submit_t
0354  * so that the DIO specific endio actions are dealt with after the filesystem
0355  * has done it's completion work.
0356  */
0357 void dio_end_io(struct bio *bio, int error)
0358 {
0359     struct dio *dio = bio->bi_private;
0360 
0361     if (dio->is_async)
0362         dio_bio_end_aio(bio);
0363     else
0364         dio_bio_end_io(bio);
0365 }
0366 EXPORT_SYMBOL_GPL(dio_end_io);
0367 
0368 static inline void
0369 dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
0370           struct block_device *bdev,
0371           sector_t first_sector, int nr_vecs)
0372 {
0373     struct bio *bio;
0374 
0375     /*
0376      * bio_alloc() is guaranteed to return a bio when called with
0377      * __GFP_RECLAIM and we request a valid number of vectors.
0378      */
0379     bio = bio_alloc(GFP_KERNEL, nr_vecs);
0380 
0381     bio->bi_bdev = bdev;
0382     bio->bi_iter.bi_sector = first_sector;
0383     bio_set_op_attrs(bio, dio->op, dio->op_flags);
0384     if (dio->is_async)
0385         bio->bi_end_io = dio_bio_end_aio;
0386     else
0387         bio->bi_end_io = dio_bio_end_io;
0388 
0389     sdio->bio = bio;
0390     sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
0391 }
0392 
0393 /*
0394  * In the AIO read case we speculatively dirty the pages before starting IO.
0395  * During IO completion, any of these pages which happen to have been written
0396  * back will be redirtied by bio_check_pages_dirty().
0397  *
0398  * bios hold a dio reference between submit_bio and ->end_io.
0399  */
0400 static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
0401 {
0402     struct bio *bio = sdio->bio;
0403     unsigned long flags;
0404 
0405     bio->bi_private = dio;
0406 
0407     spin_lock_irqsave(&dio->bio_lock, flags);
0408     dio->refcount++;
0409     spin_unlock_irqrestore(&dio->bio_lock, flags);
0410 
0411     if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty)
0412         bio_set_pages_dirty(bio);
0413 
0414     dio->bio_bdev = bio->bi_bdev;
0415 
0416     if (sdio->submit_io) {
0417         sdio->submit_io(bio, dio->inode, sdio->logical_offset_in_bio);
0418         dio->bio_cookie = BLK_QC_T_NONE;
0419     } else
0420         dio->bio_cookie = submit_bio(bio);
0421 
0422     sdio->bio = NULL;
0423     sdio->boundary = 0;
0424     sdio->logical_offset_in_bio = 0;
0425 }
0426 
0427 /*
0428  * Release any resources in case of a failure
0429  */
0430 static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
0431 {
0432     while (sdio->head < sdio->tail)
0433         put_page(dio->pages[sdio->head++]);
0434 }
0435 
0436 /*
0437  * Wait for the next BIO to complete.  Remove it and return it.  NULL is
0438  * returned once all BIOs have been completed.  This must only be called once
0439  * all bios have been issued so that dio->refcount can only decrease.  This
0440  * requires that that the caller hold a reference on the dio.
0441  */
0442 static struct bio *dio_await_one(struct dio *dio)
0443 {
0444     unsigned long flags;
0445     struct bio *bio = NULL;
0446 
0447     spin_lock_irqsave(&dio->bio_lock, flags);
0448 
0449     /*
0450      * Wait as long as the list is empty and there are bios in flight.  bio
0451      * completion drops the count, maybe adds to the list, and wakes while
0452      * holding the bio_lock so we don't need set_current_state()'s barrier
0453      * and can call it after testing our condition.
0454      */
0455     while (dio->refcount > 1 && dio->bio_list == NULL) {
0456         __set_current_state(TASK_UNINTERRUPTIBLE);
0457         dio->waiter = current;
0458         spin_unlock_irqrestore(&dio->bio_lock, flags);
0459         if (!(dio->iocb->ki_flags & IOCB_HIPRI) ||
0460             !blk_mq_poll(bdev_get_queue(dio->bio_bdev), dio->bio_cookie))
0461             io_schedule();
0462         /* wake up sets us TASK_RUNNING */
0463         spin_lock_irqsave(&dio->bio_lock, flags);
0464         dio->waiter = NULL;
0465     }
0466     if (dio->bio_list) {
0467         bio = dio->bio_list;
0468         dio->bio_list = bio->bi_private;
0469     }
0470     spin_unlock_irqrestore(&dio->bio_lock, flags);
0471     return bio;
0472 }
0473 
0474 /*
0475  * Process one completed BIO.  No locks are held.
0476  */
0477 static int dio_bio_complete(struct dio *dio, struct bio *bio)
0478 {
0479     struct bio_vec *bvec;
0480     unsigned i;
0481     int err;
0482 
0483     if (bio->bi_error)
0484         dio->io_error = -EIO;
0485 
0486     if (dio->is_async && dio->op == REQ_OP_READ && dio->should_dirty) {
0487         err = bio->bi_error;
0488         bio_check_pages_dirty(bio); /* transfers ownership */
0489     } else {
0490         bio_for_each_segment_all(bvec, bio, i) {
0491             struct page *page = bvec->bv_page;
0492 
0493             if (dio->op == REQ_OP_READ && !PageCompound(page) &&
0494                     dio->should_dirty)
0495                 set_page_dirty_lock(page);
0496             put_page(page);
0497         }
0498         err = bio->bi_error;
0499         bio_put(bio);
0500     }
0501     return err;
0502 }
0503 
0504 /*
0505  * Wait on and process all in-flight BIOs.  This must only be called once
0506  * all bios have been issued so that the refcount can only decrease.
0507  * This just waits for all bios to make it through dio_bio_complete.  IO
0508  * errors are propagated through dio->io_error and should be propagated via
0509  * dio_complete().
0510  */
0511 static void dio_await_completion(struct dio *dio)
0512 {
0513     struct bio *bio;
0514     do {
0515         bio = dio_await_one(dio);
0516         if (bio)
0517             dio_bio_complete(dio, bio);
0518     } while (bio);
0519 }
0520 
0521 /*
0522  * A really large O_DIRECT read or write can generate a lot of BIOs.  So
0523  * to keep the memory consumption sane we periodically reap any completed BIOs
0524  * during the BIO generation phase.
0525  *
0526  * This also helps to limit the peak amount of pinned userspace memory.
0527  */
0528 static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
0529 {
0530     int ret = 0;
0531 
0532     if (sdio->reap_counter++ >= 64) {
0533         while (dio->bio_list) {
0534             unsigned long flags;
0535             struct bio *bio;
0536             int ret2;
0537 
0538             spin_lock_irqsave(&dio->bio_lock, flags);
0539             bio = dio->bio_list;
0540             dio->bio_list = bio->bi_private;
0541             spin_unlock_irqrestore(&dio->bio_lock, flags);
0542             ret2 = dio_bio_complete(dio, bio);
0543             if (ret == 0)
0544                 ret = ret2;
0545         }
0546         sdio->reap_counter = 0;
0547     }
0548     return ret;
0549 }
0550 
0551 /*
0552  * Create workqueue for deferred direct IO completions. We allocate the
0553  * workqueue when it's first needed. This avoids creating workqueue for
0554  * filesystems that don't need it and also allows us to create the workqueue
0555  * late enough so the we can include s_id in the name of the workqueue.
0556  */
0557 int sb_init_dio_done_wq(struct super_block *sb)
0558 {
0559     struct workqueue_struct *old;
0560     struct workqueue_struct *wq = alloc_workqueue("dio/%s",
0561                               WQ_MEM_RECLAIM, 0,
0562                               sb->s_id);
0563     if (!wq)
0564         return -ENOMEM;
0565     /*
0566      * This has to be atomic as more DIOs can race to create the workqueue
0567      */
0568     old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
0569     /* Someone created workqueue before us? Free ours... */
0570     if (old)
0571         destroy_workqueue(wq);
0572     return 0;
0573 }
0574 
0575 static int dio_set_defer_completion(struct dio *dio)
0576 {
0577     struct super_block *sb = dio->inode->i_sb;
0578 
0579     if (dio->defer_completion)
0580         return 0;
0581     dio->defer_completion = true;
0582     if (!sb->s_dio_done_wq)
0583         return sb_init_dio_done_wq(sb);
0584     return 0;
0585 }
0586 
0587 /*
0588  * Call into the fs to map some more disk blocks.  We record the current number
0589  * of available blocks at sdio->blocks_available.  These are in units of the
0590  * fs blocksize, (1 << inode->i_blkbits).
0591  *
0592  * The fs is allowed to map lots of blocks at once.  If it wants to do that,
0593  * it uses the passed inode-relative block number as the file offset, as usual.
0594  *
0595  * get_block() is passed the number of i_blkbits-sized blocks which direct_io
0596  * has remaining to do.  The fs should not map more than this number of blocks.
0597  *
0598  * If the fs has mapped a lot of blocks, it should populate bh->b_size to
0599  * indicate how much contiguous disk space has been made available at
0600  * bh->b_blocknr.
0601  *
0602  * If *any* of the mapped blocks are new, then the fs must set buffer_new().
0603  * This isn't very efficient...
0604  *
0605  * In the case of filesystem holes: the fs may return an arbitrarily-large
0606  * hole by returning an appropriate value in b_size and by clearing
0607  * buffer_mapped().  However the direct-io code will only process holes one
0608  * block at a time - it will repeatedly call get_block() as it walks the hole.
0609  */
0610 static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
0611                struct buffer_head *map_bh)
0612 {
0613     int ret;
0614     sector_t fs_startblk;   /* Into file, in filesystem-sized blocks */
0615     sector_t fs_endblk; /* Into file, in filesystem-sized blocks */
0616     unsigned long fs_count; /* Number of filesystem-sized blocks */
0617     int create;
0618     unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
0619 
0620     /*
0621      * If there was a memory error and we've overwritten all the
0622      * mapped blocks then we can now return that memory error
0623      */
0624     ret = dio->page_errors;
0625     if (ret == 0) {
0626         BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
0627         fs_startblk = sdio->block_in_file >> sdio->blkfactor;
0628         fs_endblk = (sdio->final_block_in_request - 1) >>
0629                     sdio->blkfactor;
0630         fs_count = fs_endblk - fs_startblk + 1;
0631 
0632         map_bh->b_state = 0;
0633         map_bh->b_size = fs_count << i_blkbits;
0634 
0635         /*
0636          * For writes that could fill holes inside i_size on a
0637          * DIO_SKIP_HOLES filesystem we forbid block creations: only
0638          * overwrites are permitted. We will return early to the caller
0639          * once we see an unmapped buffer head returned, and the caller
0640          * will fall back to buffered I/O.
0641          *
0642          * Otherwise the decision is left to the get_blocks method,
0643          * which may decide to handle it or also return an unmapped
0644          * buffer head.
0645          */
0646         create = dio->op == REQ_OP_WRITE;
0647         if (dio->flags & DIO_SKIP_HOLES) {
0648             if (fs_startblk <= ((i_size_read(dio->inode) - 1) >>
0649                             i_blkbits))
0650                 create = 0;
0651         }
0652 
0653         ret = (*sdio->get_block)(dio->inode, fs_startblk,
0654                         map_bh, create);
0655 
0656         /* Store for completion */
0657         dio->private = map_bh->b_private;
0658 
0659         if (ret == 0 && buffer_defer_completion(map_bh))
0660             ret = dio_set_defer_completion(dio);
0661     }
0662     return ret;
0663 }
0664 
0665 /*
0666  * There is no bio.  Make one now.
0667  */
0668 static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
0669         sector_t start_sector, struct buffer_head *map_bh)
0670 {
0671     sector_t sector;
0672     int ret, nr_pages;
0673 
0674     ret = dio_bio_reap(dio, sdio);
0675     if (ret)
0676         goto out;
0677     sector = start_sector << (sdio->blkbits - 9);
0678     nr_pages = min(sdio->pages_in_io, BIO_MAX_PAGES);
0679     BUG_ON(nr_pages <= 0);
0680     dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
0681     sdio->boundary = 0;
0682 out:
0683     return ret;
0684 }
0685 
0686 /*
0687  * Attempt to put the current chunk of 'cur_page' into the current BIO.  If
0688  * that was successful then update final_block_in_bio and take a ref against
0689  * the just-added page.
0690  *
0691  * Return zero on success.  Non-zero means the caller needs to start a new BIO.
0692  */
0693 static inline int dio_bio_add_page(struct dio_submit *sdio)
0694 {
0695     int ret;
0696 
0697     ret = bio_add_page(sdio->bio, sdio->cur_page,
0698             sdio->cur_page_len, sdio->cur_page_offset);
0699     if (ret == sdio->cur_page_len) {
0700         /*
0701          * Decrement count only, if we are done with this page
0702          */
0703         if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
0704             sdio->pages_in_io--;
0705         get_page(sdio->cur_page);
0706         sdio->final_block_in_bio = sdio->cur_page_block +
0707             (sdio->cur_page_len >> sdio->blkbits);
0708         ret = 0;
0709     } else {
0710         ret = 1;
0711     }
0712     return ret;
0713 }
0714         
0715 /*
0716  * Put cur_page under IO.  The section of cur_page which is described by
0717  * cur_page_offset,cur_page_len is put into a BIO.  The section of cur_page
0718  * starts on-disk at cur_page_block.
0719  *
0720  * We take a ref against the page here (on behalf of its presence in the bio).
0721  *
0722  * The caller of this function is responsible for removing cur_page from the
0723  * dio, and for dropping the refcount which came from that presence.
0724  */
0725 static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
0726         struct buffer_head *map_bh)
0727 {
0728     int ret = 0;
0729 
0730     if (sdio->bio) {
0731         loff_t cur_offset = sdio->cur_page_fs_offset;
0732         loff_t bio_next_offset = sdio->logical_offset_in_bio +
0733             sdio->bio->bi_iter.bi_size;
0734 
0735         /*
0736          * See whether this new request is contiguous with the old.
0737          *
0738          * Btrfs cannot handle having logically non-contiguous requests
0739          * submitted.  For example if you have
0740          *
0741          * Logical:  [0-4095][HOLE][8192-12287]
0742          * Physical: [0-4095]      [4096-8191]
0743          *
0744          * We cannot submit those pages together as one BIO.  So if our
0745          * current logical offset in the file does not equal what would
0746          * be the next logical offset in the bio, submit the bio we
0747          * have.
0748          */
0749         if (sdio->final_block_in_bio != sdio->cur_page_block ||
0750             cur_offset != bio_next_offset)
0751             dio_bio_submit(dio, sdio);
0752     }
0753 
0754     if (sdio->bio == NULL) {
0755         ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
0756         if (ret)
0757             goto out;
0758     }
0759 
0760     if (dio_bio_add_page(sdio) != 0) {
0761         dio_bio_submit(dio, sdio);
0762         ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
0763         if (ret == 0) {
0764             ret = dio_bio_add_page(sdio);
0765             BUG_ON(ret != 0);
0766         }
0767     }
0768 out:
0769     return ret;
0770 }
0771 
0772 /*
0773  * An autonomous function to put a chunk of a page under deferred IO.
0774  *
0775  * The caller doesn't actually know (or care) whether this piece of page is in
0776  * a BIO, or is under IO or whatever.  We just take care of all possible 
0777  * situations here.  The separation between the logic of do_direct_IO() and
0778  * that of submit_page_section() is important for clarity.  Please don't break.
0779  *
0780  * The chunk of page starts on-disk at blocknr.
0781  *
0782  * We perform deferred IO, by recording the last-submitted page inside our
0783  * private part of the dio structure.  If possible, we just expand the IO
0784  * across that page here.
0785  *
0786  * If that doesn't work out then we put the old page into the bio and add this
0787  * page to the dio instead.
0788  */
0789 static inline int
0790 submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
0791             unsigned offset, unsigned len, sector_t blocknr,
0792             struct buffer_head *map_bh)
0793 {
0794     int ret = 0;
0795 
0796     if (dio->op == REQ_OP_WRITE) {
0797         /*
0798          * Read accounting is performed in submit_bio()
0799          */
0800         task_io_account_write(len);
0801     }
0802 
0803     /*
0804      * Can we just grow the current page's presence in the dio?
0805      */
0806     if (sdio->cur_page == page &&
0807         sdio->cur_page_offset + sdio->cur_page_len == offset &&
0808         sdio->cur_page_block +
0809         (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
0810         sdio->cur_page_len += len;
0811         goto out;
0812     }
0813 
0814     /*
0815      * If there's a deferred page already there then send it.
0816      */
0817     if (sdio->cur_page) {
0818         ret = dio_send_cur_page(dio, sdio, map_bh);
0819         put_page(sdio->cur_page);
0820         sdio->cur_page = NULL;
0821         if (ret)
0822             return ret;
0823     }
0824 
0825     get_page(page);     /* It is in dio */
0826     sdio->cur_page = page;
0827     sdio->cur_page_offset = offset;
0828     sdio->cur_page_len = len;
0829     sdio->cur_page_block = blocknr;
0830     sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
0831 out:
0832     /*
0833      * If sdio->boundary then we want to schedule the IO now to
0834      * avoid metadata seeks.
0835      */
0836     if (sdio->boundary) {
0837         ret = dio_send_cur_page(dio, sdio, map_bh);
0838         dio_bio_submit(dio, sdio);
0839         put_page(sdio->cur_page);
0840         sdio->cur_page = NULL;
0841     }
0842     return ret;
0843 }
0844 
0845 /*
0846  * If we are not writing the entire block and get_block() allocated
0847  * the block for us, we need to fill-in the unused portion of the
0848  * block with zeros. This happens only if user-buffer, fileoffset or
0849  * io length is not filesystem block-size multiple.
0850  *
0851  * `end' is zero if we're doing the start of the IO, 1 at the end of the
0852  * IO.
0853  */
0854 static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
0855         int end, struct buffer_head *map_bh)
0856 {
0857     unsigned dio_blocks_per_fs_block;
0858     unsigned this_chunk_blocks; /* In dio_blocks */
0859     unsigned this_chunk_bytes;
0860     struct page *page;
0861 
0862     sdio->start_zero_done = 1;
0863     if (!sdio->blkfactor || !buffer_new(map_bh))
0864         return;
0865 
0866     dio_blocks_per_fs_block = 1 << sdio->blkfactor;
0867     this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
0868 
0869     if (!this_chunk_blocks)
0870         return;
0871 
0872     /*
0873      * We need to zero out part of an fs block.  It is either at the
0874      * beginning or the end of the fs block.
0875      */
0876     if (end) 
0877         this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
0878 
0879     this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
0880 
0881     page = ZERO_PAGE(0);
0882     if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
0883                 sdio->next_block_for_io, map_bh))
0884         return;
0885 
0886     sdio->next_block_for_io += this_chunk_blocks;
0887 }
0888 
0889 /*
0890  * Walk the user pages, and the file, mapping blocks to disk and generating
0891  * a sequence of (page,offset,len,block) mappings.  These mappings are injected
0892  * into submit_page_section(), which takes care of the next stage of submission
0893  *
0894  * Direct IO against a blockdev is different from a file.  Because we can
0895  * happily perform page-sized but 512-byte aligned IOs.  It is important that
0896  * blockdev IO be able to have fine alignment and large sizes.
0897  *
0898  * So what we do is to permit the ->get_block function to populate bh.b_size
0899  * with the size of IO which is permitted at this offset and this i_blkbits.
0900  *
0901  * For best results, the blockdev should be set up with 512-byte i_blkbits and
0902  * it should set b_size to PAGE_SIZE or more inside get_block().  This gives
0903  * fine alignment but still allows this function to work in PAGE_SIZE units.
0904  */
0905 static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
0906             struct buffer_head *map_bh)
0907 {
0908     const unsigned blkbits = sdio->blkbits;
0909     const unsigned i_blkbits = blkbits + sdio->blkfactor;
0910     int ret = 0;
0911 
0912     while (sdio->block_in_file < sdio->final_block_in_request) {
0913         struct page *page;
0914         size_t from, to;
0915 
0916         page = dio_get_page(dio, sdio);
0917         if (IS_ERR(page)) {
0918             ret = PTR_ERR(page);
0919             goto out;
0920         }
0921         from = sdio->head ? 0 : sdio->from;
0922         to = (sdio->head == sdio->tail - 1) ? sdio->to : PAGE_SIZE;
0923         sdio->head++;
0924 
0925         while (from < to) {
0926             unsigned this_chunk_bytes;  /* # of bytes mapped */
0927             unsigned this_chunk_blocks; /* # of blocks */
0928             unsigned u;
0929 
0930             if (sdio->blocks_available == 0) {
0931                 /*
0932                  * Need to go and map some more disk
0933                  */
0934                 unsigned long blkmask;
0935                 unsigned long dio_remainder;
0936 
0937                 ret = get_more_blocks(dio, sdio, map_bh);
0938                 if (ret) {
0939                     put_page(page);
0940                     goto out;
0941                 }
0942                 if (!buffer_mapped(map_bh))
0943                     goto do_holes;
0944 
0945                 sdio->blocks_available =
0946                         map_bh->b_size >> blkbits;
0947                 sdio->next_block_for_io =
0948                     map_bh->b_blocknr << sdio->blkfactor;
0949                 if (buffer_new(map_bh)) {
0950                     clean_bdev_aliases(
0951                         map_bh->b_bdev,
0952                         map_bh->b_blocknr,
0953                         map_bh->b_size >> i_blkbits);
0954                 }
0955 
0956                 if (!sdio->blkfactor)
0957                     goto do_holes;
0958 
0959                 blkmask = (1 << sdio->blkfactor) - 1;
0960                 dio_remainder = (sdio->block_in_file & blkmask);
0961 
0962                 /*
0963                  * If we are at the start of IO and that IO
0964                  * starts partway into a fs-block,
0965                  * dio_remainder will be non-zero.  If the IO
0966                  * is a read then we can simply advance the IO
0967                  * cursor to the first block which is to be
0968                  * read.  But if the IO is a write and the
0969                  * block was newly allocated we cannot do that;
0970                  * the start of the fs block must be zeroed out
0971                  * on-disk
0972                  */
0973                 if (!buffer_new(map_bh))
0974                     sdio->next_block_for_io += dio_remainder;
0975                 sdio->blocks_available -= dio_remainder;
0976             }
0977 do_holes:
0978             /* Handle holes */
0979             if (!buffer_mapped(map_bh)) {
0980                 loff_t i_size_aligned;
0981 
0982                 /* AKPM: eargh, -ENOTBLK is a hack */
0983                 if (dio->op == REQ_OP_WRITE) {
0984                     put_page(page);
0985                     return -ENOTBLK;
0986                 }
0987 
0988                 /*
0989                  * Be sure to account for a partial block as the
0990                  * last block in the file
0991                  */
0992                 i_size_aligned = ALIGN(i_size_read(dio->inode),
0993                             1 << blkbits);
0994                 if (sdio->block_in_file >=
0995                         i_size_aligned >> blkbits) {
0996                     /* We hit eof */
0997                     put_page(page);
0998                     goto out;
0999                 }
1000                 zero_user(page, from, 1 << blkbits);
1001                 sdio->block_in_file++;
1002                 from += 1 << blkbits;
1003                 dio->result += 1 << blkbits;
1004                 goto next_block;
1005             }
1006 
1007             /*
1008              * If we're performing IO which has an alignment which
1009              * is finer than the underlying fs, go check to see if
1010              * we must zero out the start of this block.
1011              */
1012             if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1013                 dio_zero_block(dio, sdio, 0, map_bh);
1014 
1015             /*
1016              * Work out, in this_chunk_blocks, how much disk we
1017              * can add to this page
1018              */
1019             this_chunk_blocks = sdio->blocks_available;
1020             u = (to - from) >> blkbits;
1021             if (this_chunk_blocks > u)
1022                 this_chunk_blocks = u;
1023             u = sdio->final_block_in_request - sdio->block_in_file;
1024             if (this_chunk_blocks > u)
1025                 this_chunk_blocks = u;
1026             this_chunk_bytes = this_chunk_blocks << blkbits;
1027             BUG_ON(this_chunk_bytes == 0);
1028 
1029             if (this_chunk_blocks == sdio->blocks_available)
1030                 sdio->boundary = buffer_boundary(map_bh);
1031             ret = submit_page_section(dio, sdio, page,
1032                           from,
1033                           this_chunk_bytes,
1034                           sdio->next_block_for_io,
1035                           map_bh);
1036             if (ret) {
1037                 put_page(page);
1038                 goto out;
1039             }
1040             sdio->next_block_for_io += this_chunk_blocks;
1041 
1042             sdio->block_in_file += this_chunk_blocks;
1043             from += this_chunk_bytes;
1044             dio->result += this_chunk_bytes;
1045             sdio->blocks_available -= this_chunk_blocks;
1046 next_block:
1047             BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1048             if (sdio->block_in_file == sdio->final_block_in_request)
1049                 break;
1050         }
1051 
1052         /* Drop the ref which was taken in get_user_pages() */
1053         put_page(page);
1054     }
1055 out:
1056     return ret;
1057 }
1058 
1059 static inline int drop_refcount(struct dio *dio)
1060 {
1061     int ret2;
1062     unsigned long flags;
1063 
1064     /*
1065      * Sync will always be dropping the final ref and completing the
1066      * operation.  AIO can if it was a broken operation described above or
1067      * in fact if all the bios race to complete before we get here.  In
1068      * that case dio_complete() translates the EIOCBQUEUED into the proper
1069      * return code that the caller will hand to ->complete().
1070      *
1071      * This is managed by the bio_lock instead of being an atomic_t so that
1072      * completion paths can drop their ref and use the remaining count to
1073      * decide to wake the submission path atomically.
1074      */
1075     spin_lock_irqsave(&dio->bio_lock, flags);
1076     ret2 = --dio->refcount;
1077     spin_unlock_irqrestore(&dio->bio_lock, flags);
1078     return ret2;
1079 }
1080 
1081 /*
1082  * This is a library function for use by filesystem drivers.
1083  *
1084  * The locking rules are governed by the flags parameter:
1085  *  - if the flags value contains DIO_LOCKING we use a fancy locking
1086  *    scheme for dumb filesystems.
1087  *    For writes this function is called under i_mutex and returns with
1088  *    i_mutex held, for reads, i_mutex is not held on entry, but it is
1089  *    taken and dropped again before returning.
1090  *  - if the flags value does NOT contain DIO_LOCKING we don't use any
1091  *    internal locking but rather rely on the filesystem to synchronize
1092  *    direct I/O reads/writes versus each other and truncate.
1093  *
1094  * To help with locking against truncate we incremented the i_dio_count
1095  * counter before starting direct I/O, and decrement it once we are done.
1096  * Truncate can wait for it to reach zero to provide exclusion.  It is
1097  * expected that filesystem provide exclusion between new direct I/O
1098  * and truncates.  For DIO_LOCKING filesystems this is done by i_mutex,
1099  * but other filesystems need to take care of this on their own.
1100  *
1101  * NOTE: if you pass "sdio" to anything by pointer make sure that function
1102  * is always inlined. Otherwise gcc is unable to split the structure into
1103  * individual fields and will generate much worse code. This is important
1104  * for the whole file.
1105  */
1106 static inline ssize_t
1107 do_blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1108               struct block_device *bdev, struct iov_iter *iter,
1109               get_block_t get_block, dio_iodone_t end_io,
1110               dio_submit_t submit_io, int flags)
1111 {
1112     unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1113     unsigned blkbits = i_blkbits;
1114     unsigned blocksize_mask = (1 << blkbits) - 1;
1115     ssize_t retval = -EINVAL;
1116     size_t count = iov_iter_count(iter);
1117     loff_t offset = iocb->ki_pos;
1118     loff_t end = offset + count;
1119     struct dio *dio;
1120     struct dio_submit sdio = { 0, };
1121     struct buffer_head map_bh = { 0, };
1122     struct blk_plug plug;
1123     unsigned long align = offset | iov_iter_alignment(iter);
1124 
1125     /*
1126      * Avoid references to bdev if not absolutely needed to give
1127      * the early prefetch in the caller enough time.
1128      */
1129 
1130     if (align & blocksize_mask) {
1131         if (bdev)
1132             blkbits = blksize_bits(bdev_logical_block_size(bdev));
1133         blocksize_mask = (1 << blkbits) - 1;
1134         if (align & blocksize_mask)
1135             goto out;
1136     }
1137 
1138     /* watch out for a 0 len io from a tricksy fs */
1139     if (iov_iter_rw(iter) == READ && !iov_iter_count(iter))
1140         return 0;
1141 
1142     dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1143     retval = -ENOMEM;
1144     if (!dio)
1145         goto out;
1146     /*
1147      * Believe it or not, zeroing out the page array caused a .5%
1148      * performance regression in a database benchmark.  So, we take
1149      * care to only zero out what's needed.
1150      */
1151     memset(dio, 0, offsetof(struct dio, pages));
1152 
1153     dio->flags = flags;
1154     if (dio->flags & DIO_LOCKING) {
1155         if (iov_iter_rw(iter) == READ) {
1156             struct address_space *mapping =
1157                     iocb->ki_filp->f_mapping;
1158 
1159             /* will be released by direct_io_worker */
1160             inode_lock(inode);
1161 
1162             retval = filemap_write_and_wait_range(mapping, offset,
1163                                   end - 1);
1164             if (retval) {
1165                 inode_unlock(inode);
1166                 kmem_cache_free(dio_cache, dio);
1167                 goto out;
1168             }
1169         }
1170     }
1171 
1172     /* Once we sampled i_size check for reads beyond EOF */
1173     dio->i_size = i_size_read(inode);
1174     if (iov_iter_rw(iter) == READ && offset >= dio->i_size) {
1175         if (dio->flags & DIO_LOCKING)
1176             inode_unlock(inode);
1177         kmem_cache_free(dio_cache, dio);
1178         retval = 0;
1179         goto out;
1180     }
1181 
1182     /*
1183      * For file extending writes updating i_size before data writeouts
1184      * complete can expose uninitialized blocks in dumb filesystems.
1185      * In that case we need to wait for I/O completion even if asked
1186      * for an asynchronous write.
1187      */
1188     if (is_sync_kiocb(iocb))
1189         dio->is_async = false;
1190     else if (!(dio->flags & DIO_ASYNC_EXTEND) &&
1191          iov_iter_rw(iter) == WRITE && end > i_size_read(inode))
1192         dio->is_async = false;
1193     else
1194         dio->is_async = true;
1195 
1196     dio->inode = inode;
1197     if (iov_iter_rw(iter) == WRITE) {
1198         dio->op = REQ_OP_WRITE;
1199         dio->op_flags = REQ_SYNC | REQ_IDLE;
1200     } else {
1201         dio->op = REQ_OP_READ;
1202     }
1203 
1204     /*
1205      * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1206      * so that we can call ->fsync.
1207      */
1208     if (dio->is_async && iov_iter_rw(iter) == WRITE &&
1209         ((iocb->ki_filp->f_flags & O_DSYNC) ||
1210          IS_SYNC(iocb->ki_filp->f_mapping->host))) {
1211         retval = dio_set_defer_completion(dio);
1212         if (retval) {
1213             /*
1214              * We grab i_mutex only for reads so we don't have
1215              * to release it here
1216              */
1217             kmem_cache_free(dio_cache, dio);
1218             goto out;
1219         }
1220     }
1221 
1222     /*
1223      * Will be decremented at I/O completion time.
1224      */
1225     if (!(dio->flags & DIO_SKIP_DIO_COUNT))
1226         inode_dio_begin(inode);
1227 
1228     retval = 0;
1229     sdio.blkbits = blkbits;
1230     sdio.blkfactor = i_blkbits - blkbits;
1231     sdio.block_in_file = offset >> blkbits;
1232 
1233     sdio.get_block = get_block;
1234     dio->end_io = end_io;
1235     sdio.submit_io = submit_io;
1236     sdio.final_block_in_bio = -1;
1237     sdio.next_block_for_io = -1;
1238 
1239     dio->iocb = iocb;
1240 
1241     spin_lock_init(&dio->bio_lock);
1242     dio->refcount = 1;
1243 
1244     dio->should_dirty = (iter->type == ITER_IOVEC);
1245     sdio.iter = iter;
1246     sdio.final_block_in_request =
1247         (offset + iov_iter_count(iter)) >> blkbits;
1248 
1249     /*
1250      * In case of non-aligned buffers, we may need 2 more
1251      * pages since we need to zero out first and last block.
1252      */
1253     if (unlikely(sdio.blkfactor))
1254         sdio.pages_in_io = 2;
1255 
1256     sdio.pages_in_io += iov_iter_npages(iter, INT_MAX);
1257 
1258     blk_start_plug(&plug);
1259 
1260     retval = do_direct_IO(dio, &sdio, &map_bh);
1261     if (retval)
1262         dio_cleanup(dio, &sdio);
1263 
1264     if (retval == -ENOTBLK) {
1265         /*
1266          * The remaining part of the request will be
1267          * be handled by buffered I/O when we return
1268          */
1269         retval = 0;
1270     }
1271     /*
1272      * There may be some unwritten disk at the end of a part-written
1273      * fs-block-sized block.  Go zero that now.
1274      */
1275     dio_zero_block(dio, &sdio, 1, &map_bh);
1276 
1277     if (sdio.cur_page) {
1278         ssize_t ret2;
1279 
1280         ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1281         if (retval == 0)
1282             retval = ret2;
1283         put_page(sdio.cur_page);
1284         sdio.cur_page = NULL;
1285     }
1286     if (sdio.bio)
1287         dio_bio_submit(dio, &sdio);
1288 
1289     blk_finish_plug(&plug);
1290 
1291     /*
1292      * It is possible that, we return short IO due to end of file.
1293      * In that case, we need to release all the pages we got hold on.
1294      */
1295     dio_cleanup(dio, &sdio);
1296 
1297     /*
1298      * All block lookups have been performed. For READ requests
1299      * we can let i_mutex go now that its achieved its purpose
1300      * of protecting us from looking up uninitialized blocks.
1301      */
1302     if (iov_iter_rw(iter) == READ && (dio->flags & DIO_LOCKING))
1303         inode_unlock(dio->inode);
1304 
1305     /*
1306      * The only time we want to leave bios in flight is when a successful
1307      * partial aio read or full aio write have been setup.  In that case
1308      * bio completion will call aio_complete.  The only time it's safe to
1309      * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1310      * This had *better* be the only place that raises -EIOCBQUEUED.
1311      */
1312     BUG_ON(retval == -EIOCBQUEUED);
1313     if (dio->is_async && retval == 0 && dio->result &&
1314         (iov_iter_rw(iter) == READ || dio->result == count))
1315         retval = -EIOCBQUEUED;
1316     else
1317         dio_await_completion(dio);
1318 
1319     if (drop_refcount(dio) == 0) {
1320         retval = dio_complete(dio, retval, false);
1321     } else
1322         BUG_ON(retval != -EIOCBQUEUED);
1323 
1324 out:
1325     return retval;
1326 }
1327 
1328 ssize_t __blockdev_direct_IO(struct kiocb *iocb, struct inode *inode,
1329                  struct block_device *bdev, struct iov_iter *iter,
1330                  get_block_t get_block,
1331                  dio_iodone_t end_io, dio_submit_t submit_io,
1332                  int flags)
1333 {
1334     /*
1335      * The block device state is needed in the end to finally
1336      * submit everything.  Since it's likely to be cache cold
1337      * prefetch it here as first thing to hide some of the
1338      * latency.
1339      *
1340      * Attempt to prefetch the pieces we likely need later.
1341      */
1342     prefetch(&bdev->bd_disk->part_tbl);
1343     prefetch(bdev->bd_queue);
1344     prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1345 
1346     return do_blockdev_direct_IO(iocb, inode, bdev, iter, get_block,
1347                      end_io, submit_io, flags);
1348 }
1349 
1350 EXPORT_SYMBOL(__blockdev_direct_IO);
1351 
1352 static __init int dio_init(void)
1353 {
1354     dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1355     return 0;
1356 }
1357 module_init(dio_init)