linux/fs/direct-io.c
<<
>>
Prefs
   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#include <linux/aio.h>
  41
  42/*
  43 * How many user pages to map in one call to get_user_pages().  This determines
  44 * the size of a structure in the slab cache
  45 */
  46#define DIO_PAGES       64
  47
  48/*
  49 * This code generally works in units of "dio_blocks".  A dio_block is
  50 * somewhere between the hard sector size and the filesystem block size.  it
  51 * is determined on a per-invocation basis.   When talking to the filesystem
  52 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
  53 * down by dio->blkfactor.  Similarly, fs-blocksize quantities are converted
  54 * to bio_block quantities by shifting left by blkfactor.
  55 *
  56 * If blkfactor is zero then the user's request was aligned to the filesystem's
  57 * blocksize.
  58 */
  59
  60/* dio_state only used in the submission path */
  61
  62struct dio_submit {
  63        struct bio *bio;                /* bio under assembly */
  64        unsigned blkbits;               /* doesn't change */
  65        unsigned blkfactor;             /* When we're using an alignment which
  66                                           is finer than the filesystem's soft
  67                                           blocksize, this specifies how much
  68                                           finer.  blkfactor=2 means 1/4-block
  69                                           alignment.  Does not change */
  70        unsigned start_zero_done;       /* flag: sub-blocksize zeroing has
  71                                           been performed at the start of a
  72                                           write */
  73        int pages_in_io;                /* approximate total IO pages */
  74        size_t  size;                   /* total request size (doesn't change)*/
  75        sector_t block_in_file;         /* Current offset into the underlying
  76                                           file in dio_block units. */
  77        unsigned blocks_available;      /* At block_in_file.  changes */
  78        int reap_counter;               /* rate limit reaping */
  79        sector_t final_block_in_request;/* doesn't change */
  80        unsigned first_block_in_page;   /* doesn't change, Used only once */
  81        int boundary;                   /* prev block is at a boundary */
  82        get_block_t *get_block;         /* block mapping function */
  83        dio_submit_t *submit_io;        /* IO submition function */
  84
  85        loff_t logical_offset_in_bio;   /* current first logical block in bio */
  86        sector_t final_block_in_bio;    /* current final block in bio + 1 */
  87        sector_t next_block_for_io;     /* next block to be put under IO,
  88                                           in dio_blocks units */
  89
  90        /*
  91         * Deferred addition of a page to the dio.  These variables are
  92         * private to dio_send_cur_page(), submit_page_section() and
  93         * dio_bio_add_page().
  94         */
  95        struct page *cur_page;          /* The page */
  96        unsigned cur_page_offset;       /* Offset into it, in bytes */
  97        unsigned cur_page_len;          /* Nr of bytes at cur_page_offset */
  98        sector_t cur_page_block;        /* Where it starts */
  99        loff_t cur_page_fs_offset;      /* Offset in file */
 100
 101        /*
 102         * Page fetching state. These variables belong to dio_refill_pages().
 103         */
 104        int curr_page;                  /* changes */
 105        int total_pages;                /* doesn't change */
 106        unsigned long curr_user_address;/* changes */
 107
 108        /*
 109         * Page queue.  These variables belong to dio_refill_pages() and
 110         * dio_get_page().
 111         */
 112        unsigned head;                  /* next page to process */
 113        unsigned tail;                  /* last valid page + 1 */
 114};
 115
 116/* dio_state communicated between submission path and end_io */
 117struct dio {
 118        int flags;                      /* doesn't change */
 119        int rw;
 120        struct inode *inode;
 121        loff_t i_size;                  /* i_size when submitted */
 122        dio_iodone_t *end_io;           /* IO completion function */
 123
 124        void *private;                  /* copy from map_bh.b_private */
 125
 126        /* BIO completion state */
 127        spinlock_t bio_lock;            /* protects BIO fields below */
 128        int page_errors;                /* errno from get_user_pages() */
 129        int is_async;                   /* is IO async ? */
 130        bool defer_completion;          /* defer AIO completion to workqueue? */
 131        int io_error;                   /* IO error in completion path */
 132        unsigned long refcount;         /* direct_io_worker() and bios */
 133        struct bio *bio_list;           /* singly linked via bi_private */
 134        struct task_struct *waiter;     /* waiting task (NULL if none) */
 135
 136        /* AIO related stuff */
 137        struct kiocb *iocb;             /* kiocb */
 138        ssize_t result;                 /* IO result */
 139
 140        /*
 141         * pages[] (and any fields placed after it) are not zeroed out at
 142         * allocation time.  Don't add new fields after pages[] unless you
 143         * wish that they not be zeroed.
 144         */
 145        union {
 146                struct page *pages[DIO_PAGES];  /* page buffer */
 147                struct work_struct complete_work;/* deferred AIO completion */
 148        };
 149} ____cacheline_aligned_in_smp;
 150
 151static struct kmem_cache *dio_cache __read_mostly;
 152
 153/*
 154 * How many pages are in the queue?
 155 */
 156static inline unsigned dio_pages_present(struct dio_submit *sdio)
 157{
 158        return sdio->tail - sdio->head;
 159}
 160
 161/*
 162 * Go grab and pin some userspace pages.   Typically we'll get 64 at a time.
 163 */
 164static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
 165{
 166        int ret;
 167        int nr_pages;
 168
 169        nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES);
 170        ret = get_user_pages_fast(
 171                sdio->curr_user_address,                /* Where from? */
 172                nr_pages,                       /* How many pages? */
 173                dio->rw == READ,                /* Write to memory? */
 174                &dio->pages[0]);                /* Put results here */
 175
 176        if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
 177                struct page *page = ZERO_PAGE(0);
 178                /*
 179                 * A memory fault, but the filesystem has some outstanding
 180                 * mapped blocks.  We need to use those blocks up to avoid
 181                 * leaking stale data in the file.
 182                 */
 183                if (dio->page_errors == 0)
 184                        dio->page_errors = ret;
 185                page_cache_get(page);
 186                dio->pages[0] = page;
 187                sdio->head = 0;
 188                sdio->tail = 1;
 189                ret = 0;
 190                goto out;
 191        }
 192
 193        if (ret >= 0) {
 194                sdio->curr_user_address += ret * PAGE_SIZE;
 195                sdio->curr_page += ret;
 196                sdio->head = 0;
 197                sdio->tail = ret;
 198                ret = 0;
 199        }
 200out:
 201        return ret;     
 202}
 203
 204/*
 205 * Get another userspace page.  Returns an ERR_PTR on error.  Pages are
 206 * buffered inside the dio so that we can call get_user_pages() against a
 207 * decent number of pages, less frequently.  To provide nicer use of the
 208 * L1 cache.
 209 */
 210static inline struct page *dio_get_page(struct dio *dio,
 211                struct dio_submit *sdio)
 212{
 213        if (dio_pages_present(sdio) == 0) {
 214                int ret;
 215
 216                ret = dio_refill_pages(dio, sdio);
 217                if (ret)
 218                        return ERR_PTR(ret);
 219                BUG_ON(dio_pages_present(sdio) == 0);
 220        }
 221        return dio->pages[sdio->head++];
 222}
 223
 224/**
 225 * dio_complete() - called when all DIO BIO I/O has been completed
 226 * @offset: the byte offset in the file of the completed operation
 227 *
 228 * This drops i_dio_count, lets interested parties know that a DIO operation
 229 * has completed, and calculates the resulting return code for the operation.
 230 *
 231 * It lets the filesystem know if it registered an interest earlier via
 232 * get_block.  Pass the private field of the map buffer_head so that
 233 * filesystems can use it to hold additional state between get_block calls and
 234 * dio_complete.
 235 */
 236static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret,
 237                bool is_async)
 238{
 239        ssize_t transferred = 0;
 240
 241        /*
 242         * AIO submission can race with bio completion to get here while
 243         * expecting to have the last io completed by bio completion.
 244         * In that case -EIOCBQUEUED is in fact not an error we want
 245         * to preserve through this call.
 246         */
 247        if (ret == -EIOCBQUEUED)
 248                ret = 0;
 249
 250        if (dio->result) {
 251                transferred = dio->result;
 252
 253                /* Check for short read case */
 254                if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
 255                        transferred = dio->i_size - offset;
 256        }
 257
 258        if (ret == 0)
 259                ret = dio->page_errors;
 260        if (ret == 0)
 261                ret = dio->io_error;
 262        if (ret == 0)
 263                ret = transferred;
 264
 265        if (dio->end_io && dio->result)
 266                dio->end_io(dio->iocb, offset, transferred, dio->private);
 267
 268        inode_dio_done(dio->inode);
 269        if (is_async) {
 270                if (dio->rw & WRITE) {
 271                        int err;
 272
 273                        err = generic_write_sync(dio->iocb->ki_filp, offset,
 274                                                 transferred);
 275                        if (err < 0 && ret > 0)
 276                                ret = err;
 277                }
 278
 279                aio_complete(dio->iocb, ret, 0);
 280        }
 281
 282        kmem_cache_free(dio_cache, dio);
 283        return ret;
 284}
 285
 286static void dio_aio_complete_work(struct work_struct *work)
 287{
 288        struct dio *dio = container_of(work, struct dio, complete_work);
 289
 290        dio_complete(dio, dio->iocb->ki_pos, 0, true);
 291}
 292
 293static int dio_bio_complete(struct dio *dio, struct bio *bio);
 294
 295/*
 296 * Asynchronous IO callback. 
 297 */
 298static void dio_bio_end_aio(struct bio *bio, int error)
 299{
 300        struct dio *dio = bio->bi_private;
 301        unsigned long remaining;
 302        unsigned long flags;
 303
 304        /* cleanup the bio */
 305        dio_bio_complete(dio, bio);
 306
 307        spin_lock_irqsave(&dio->bio_lock, flags);
 308        remaining = --dio->refcount;
 309        if (remaining == 1 && dio->waiter)
 310                wake_up_process(dio->waiter);
 311        spin_unlock_irqrestore(&dio->bio_lock, flags);
 312
 313        if (remaining == 0) {
 314                if (dio->result && dio->defer_completion) {
 315                        INIT_WORK(&dio->complete_work, dio_aio_complete_work);
 316                        queue_work(dio->inode->i_sb->s_dio_done_wq,
 317                                   &dio->complete_work);
 318                } else {
 319                        dio_complete(dio, dio->iocb->ki_pos, 0, true);
 320                }
 321        }
 322}
 323
 324/*
 325 * The BIO completion handler simply queues the BIO up for the process-context
 326 * handler.
 327 *
 328 * During I/O bi_private points at the dio.  After I/O, bi_private is used to
 329 * implement a singly-linked list of completed BIOs, at dio->bio_list.
 330 */
 331static void dio_bio_end_io(struct bio *bio, int error)
 332{
 333        struct dio *dio = bio->bi_private;
 334        unsigned long flags;
 335
 336        spin_lock_irqsave(&dio->bio_lock, flags);
 337        bio->bi_private = dio->bio_list;
 338        dio->bio_list = bio;
 339        if (--dio->refcount == 1 && dio->waiter)
 340                wake_up_process(dio->waiter);
 341        spin_unlock_irqrestore(&dio->bio_lock, flags);
 342}
 343
 344/**
 345 * dio_end_io - handle the end io action for the given bio
 346 * @bio: The direct io bio thats being completed
 347 * @error: Error if there was one
 348 *
 349 * This is meant to be called by any filesystem that uses their own dio_submit_t
 350 * so that the DIO specific endio actions are dealt with after the filesystem
 351 * has done it's completion work.
 352 */
 353void dio_end_io(struct bio *bio, int error)
 354{
 355        struct dio *dio = bio->bi_private;
 356
 357        if (dio->is_async)
 358                dio_bio_end_aio(bio, error);
 359        else
 360                dio_bio_end_io(bio, error);
 361}
 362EXPORT_SYMBOL_GPL(dio_end_io);
 363
 364static inline void
 365dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
 366              struct block_device *bdev,
 367              sector_t first_sector, int nr_vecs)
 368{
 369        struct bio *bio;
 370
 371        /*
 372         * bio_alloc() is guaranteed to return a bio when called with
 373         * __GFP_WAIT and we request a valid number of vectors.
 374         */
 375        bio = bio_alloc(GFP_KERNEL, nr_vecs);
 376
 377        bio->bi_bdev = bdev;
 378        bio->bi_sector = first_sector;
 379        if (dio->is_async)
 380                bio->bi_end_io = dio_bio_end_aio;
 381        else
 382                bio->bi_end_io = dio_bio_end_io;
 383
 384        sdio->bio = bio;
 385        sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
 386}
 387
 388/*
 389 * In the AIO read case we speculatively dirty the pages before starting IO.
 390 * During IO completion, any of these pages which happen to have been written
 391 * back will be redirtied by bio_check_pages_dirty().
 392 *
 393 * bios hold a dio reference between submit_bio and ->end_io.
 394 */
 395static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
 396{
 397        struct bio *bio = sdio->bio;
 398        unsigned long flags;
 399
 400        bio->bi_private = dio;
 401
 402        spin_lock_irqsave(&dio->bio_lock, flags);
 403        dio->refcount++;
 404        spin_unlock_irqrestore(&dio->bio_lock, flags);
 405
 406        if (dio->is_async && dio->rw == READ)
 407                bio_set_pages_dirty(bio);
 408
 409        if (sdio->submit_io)
 410                sdio->submit_io(dio->rw, bio, dio->inode,
 411                               sdio->logical_offset_in_bio);
 412        else
 413                submit_bio(dio->rw, bio);
 414
 415        sdio->bio = NULL;
 416        sdio->boundary = 0;
 417        sdio->logical_offset_in_bio = 0;
 418}
 419
 420/*
 421 * Release any resources in case of a failure
 422 */
 423static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
 424{
 425        while (dio_pages_present(sdio))
 426                page_cache_release(dio_get_page(dio, sdio));
 427}
 428
 429/*
 430 * Wait for the next BIO to complete.  Remove it and return it.  NULL is
 431 * returned once all BIOs have been completed.  This must only be called once
 432 * all bios have been issued so that dio->refcount can only decrease.  This
 433 * requires that that the caller hold a reference on the dio.
 434 */
 435static struct bio *dio_await_one(struct dio *dio)
 436{
 437        unsigned long flags;
 438        struct bio *bio = NULL;
 439
 440        spin_lock_irqsave(&dio->bio_lock, flags);
 441
 442        /*
 443         * Wait as long as the list is empty and there are bios in flight.  bio
 444         * completion drops the count, maybe adds to the list, and wakes while
 445         * holding the bio_lock so we don't need set_current_state()'s barrier
 446         * and can call it after testing our condition.
 447         */
 448        while (dio->refcount > 1 && dio->bio_list == NULL) {
 449                __set_current_state(TASK_UNINTERRUPTIBLE);
 450                dio->waiter = current;
 451                spin_unlock_irqrestore(&dio->bio_lock, flags);
 452                io_schedule();
 453                /* wake up sets us TASK_RUNNING */
 454                spin_lock_irqsave(&dio->bio_lock, flags);
 455                dio->waiter = NULL;
 456        }
 457        if (dio->bio_list) {
 458                bio = dio->bio_list;
 459                dio->bio_list = bio->bi_private;
 460        }
 461        spin_unlock_irqrestore(&dio->bio_lock, flags);
 462        return bio;
 463}
 464
 465/*
 466 * Process one completed BIO.  No locks are held.
 467 */
 468static int dio_bio_complete(struct dio *dio, struct bio *bio)
 469{
 470        const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 471        struct bio_vec *bvec;
 472        unsigned i;
 473
 474        if (!uptodate)
 475                dio->io_error = -EIO;
 476
 477        if (dio->is_async && dio->rw == READ) {
 478                bio_check_pages_dirty(bio);     /* transfers ownership */
 479        } else {
 480                bio_for_each_segment_all(bvec, bio, i) {
 481                        struct page *page = bvec->bv_page;
 482
 483                        if (dio->rw == READ && !PageCompound(page))
 484                                set_page_dirty_lock(page);
 485                        page_cache_release(page);
 486                }
 487                bio_put(bio);
 488        }
 489        return uptodate ? 0 : -EIO;
 490}
 491
 492/*
 493 * Wait on and process all in-flight BIOs.  This must only be called once
 494 * all bios have been issued so that the refcount can only decrease.
 495 * This just waits for all bios to make it through dio_bio_complete.  IO
 496 * errors are propagated through dio->io_error and should be propagated via
 497 * dio_complete().
 498 */
 499static void dio_await_completion(struct dio *dio)
 500{
 501        struct bio *bio;
 502        do {
 503                bio = dio_await_one(dio);
 504                if (bio)
 505                        dio_bio_complete(dio, bio);
 506        } while (bio);
 507}
 508
 509/*
 510 * A really large O_DIRECT read or write can generate a lot of BIOs.  So
 511 * to keep the memory consumption sane we periodically reap any completed BIOs
 512 * during the BIO generation phase.
 513 *
 514 * This also helps to limit the peak amount of pinned userspace memory.
 515 */
 516static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
 517{
 518        int ret = 0;
 519
 520        if (sdio->reap_counter++ >= 64) {
 521                while (dio->bio_list) {
 522                        unsigned long flags;
 523                        struct bio *bio;
 524                        int ret2;
 525
 526                        spin_lock_irqsave(&dio->bio_lock, flags);
 527                        bio = dio->bio_list;
 528                        dio->bio_list = bio->bi_private;
 529                        spin_unlock_irqrestore(&dio->bio_lock, flags);
 530                        ret2 = dio_bio_complete(dio, bio);
 531                        if (ret == 0)
 532                                ret = ret2;
 533                }
 534                sdio->reap_counter = 0;
 535        }
 536        return ret;
 537}
 538
 539/*
 540 * Create workqueue for deferred direct IO completions. We allocate the
 541 * workqueue when it's first needed. This avoids creating workqueue for
 542 * filesystems that don't need it and also allows us to create the workqueue
 543 * late enough so the we can include s_id in the name of the workqueue.
 544 */
 545static int sb_init_dio_done_wq(struct super_block *sb)
 546{
 547        struct workqueue_struct *old;
 548        struct workqueue_struct *wq = alloc_workqueue("dio/%s",
 549                                                      WQ_MEM_RECLAIM, 0,
 550                                                      sb->s_id);
 551        if (!wq)
 552                return -ENOMEM;
 553        /*
 554         * This has to be atomic as more DIOs can race to create the workqueue
 555         */
 556        old = cmpxchg(&sb->s_dio_done_wq, NULL, wq);
 557        /* Someone created workqueue before us? Free ours... */
 558        if (old)
 559                destroy_workqueue(wq);
 560        return 0;
 561}
 562
 563static int dio_set_defer_completion(struct dio *dio)
 564{
 565        struct super_block *sb = dio->inode->i_sb;
 566
 567        if (dio->defer_completion)
 568                return 0;
 569        dio->defer_completion = true;
 570        if (!sb->s_dio_done_wq)
 571                return sb_init_dio_done_wq(sb);
 572        return 0;
 573}
 574
 575/*
 576 * Call into the fs to map some more disk blocks.  We record the current number
 577 * of available blocks at sdio->blocks_available.  These are in units of the
 578 * fs blocksize, (1 << inode->i_blkbits).
 579 *
 580 * The fs is allowed to map lots of blocks at once.  If it wants to do that,
 581 * it uses the passed inode-relative block number as the file offset, as usual.
 582 *
 583 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
 584 * has remaining to do.  The fs should not map more than this number of blocks.
 585 *
 586 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
 587 * indicate how much contiguous disk space has been made available at
 588 * bh->b_blocknr.
 589 *
 590 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
 591 * This isn't very efficient...
 592 *
 593 * In the case of filesystem holes: the fs may return an arbitrarily-large
 594 * hole by returning an appropriate value in b_size and by clearing
 595 * buffer_mapped().  However the direct-io code will only process holes one
 596 * block at a time - it will repeatedly call get_block() as it walks the hole.
 597 */
 598static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
 599                           struct buffer_head *map_bh)
 600{
 601        int ret;
 602        sector_t fs_startblk;   /* Into file, in filesystem-sized blocks */
 603        sector_t fs_endblk;     /* Into file, in filesystem-sized blocks */
 604        unsigned long fs_count; /* Number of filesystem-sized blocks */
 605        int create;
 606        unsigned int i_blkbits = sdio->blkbits + sdio->blkfactor;
 607
 608        /*
 609         * If there was a memory error and we've overwritten all the
 610         * mapped blocks then we can now return that memory error
 611         */
 612        ret = dio->page_errors;
 613        if (ret == 0) {
 614                BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
 615                fs_startblk = sdio->block_in_file >> sdio->blkfactor;
 616                fs_endblk = (sdio->final_block_in_request - 1) >>
 617                                        sdio->blkfactor;
 618                fs_count = fs_endblk - fs_startblk + 1;
 619
 620                map_bh->b_state = 0;
 621                map_bh->b_size = fs_count << i_blkbits;
 622
 623                /*
 624                 * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
 625                 * forbid block creations: only overwrites are permitted.
 626                 * We will return early to the caller once we see an
 627                 * unmapped buffer head returned, and the caller will fall
 628                 * back to buffered I/O.
 629                 *
 630                 * Otherwise the decision is left to the get_blocks method,
 631                 * which may decide to handle it or also return an unmapped
 632                 * buffer head.
 633                 */
 634                create = dio->rw & WRITE;
 635                if (dio->flags & DIO_SKIP_HOLES) {
 636                        if (sdio->block_in_file < (i_size_read(dio->inode) >>
 637                                                        sdio->blkbits))
 638                                create = 0;
 639                }
 640
 641                ret = (*sdio->get_block)(dio->inode, fs_startblk,
 642                                                map_bh, create);
 643
 644                /* Store for completion */
 645                dio->private = map_bh->b_private;
 646
 647                if (ret == 0 && buffer_defer_completion(map_bh))
 648                        ret = dio_set_defer_completion(dio);
 649        }
 650        return ret;
 651}
 652
 653/*
 654 * There is no bio.  Make one now.
 655 */
 656static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
 657                sector_t start_sector, struct buffer_head *map_bh)
 658{
 659        sector_t sector;
 660        int ret, nr_pages;
 661
 662        ret = dio_bio_reap(dio, sdio);
 663        if (ret)
 664                goto out;
 665        sector = start_sector << (sdio->blkbits - 9);
 666        nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
 667        nr_pages = min(nr_pages, BIO_MAX_PAGES);
 668        BUG_ON(nr_pages <= 0);
 669        dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
 670        sdio->boundary = 0;
 671out:
 672        return ret;
 673}
 674
 675/*
 676 * Attempt to put the current chunk of 'cur_page' into the current BIO.  If
 677 * that was successful then update final_block_in_bio and take a ref against
 678 * the just-added page.
 679 *
 680 * Return zero on success.  Non-zero means the caller needs to start a new BIO.
 681 */
 682static inline int dio_bio_add_page(struct dio_submit *sdio)
 683{
 684        int ret;
 685
 686        ret = bio_add_page(sdio->bio, sdio->cur_page,
 687                        sdio->cur_page_len, sdio->cur_page_offset);
 688        if (ret == sdio->cur_page_len) {
 689                /*
 690                 * Decrement count only, if we are done with this page
 691                 */
 692                if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
 693                        sdio->pages_in_io--;
 694                page_cache_get(sdio->cur_page);
 695                sdio->final_block_in_bio = sdio->cur_page_block +
 696                        (sdio->cur_page_len >> sdio->blkbits);
 697                ret = 0;
 698        } else {
 699                ret = 1;
 700        }
 701        return ret;
 702}
 703                
 704/*
 705 * Put cur_page under IO.  The section of cur_page which is described by
 706 * cur_page_offset,cur_page_len is put into a BIO.  The section of cur_page
 707 * starts on-disk at cur_page_block.
 708 *
 709 * We take a ref against the page here (on behalf of its presence in the bio).
 710 *
 711 * The caller of this function is responsible for removing cur_page from the
 712 * dio, and for dropping the refcount which came from that presence.
 713 */
 714static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
 715                struct buffer_head *map_bh)
 716{
 717        int ret = 0;
 718
 719        if (sdio->bio) {
 720                loff_t cur_offset = sdio->cur_page_fs_offset;
 721                loff_t bio_next_offset = sdio->logical_offset_in_bio +
 722                        sdio->bio->bi_size;
 723
 724                /*
 725                 * See whether this new request is contiguous with the old.
 726                 *
 727                 * Btrfs cannot handle having logically non-contiguous requests
 728                 * submitted.  For example if you have
 729                 *
 730                 * Logical:  [0-4095][HOLE][8192-12287]
 731                 * Physical: [0-4095]      [4096-8191]
 732                 *
 733                 * We cannot submit those pages together as one BIO.  So if our
 734                 * current logical offset in the file does not equal what would
 735                 * be the next logical offset in the bio, submit the bio we
 736                 * have.
 737                 */
 738                if (sdio->final_block_in_bio != sdio->cur_page_block ||
 739                    cur_offset != bio_next_offset)
 740                        dio_bio_submit(dio, sdio);
 741        }
 742
 743        if (sdio->bio == NULL) {
 744                ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
 745                if (ret)
 746                        goto out;
 747        }
 748
 749        if (dio_bio_add_page(sdio) != 0) {
 750                dio_bio_submit(dio, sdio);
 751                ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
 752                if (ret == 0) {
 753                        ret = dio_bio_add_page(sdio);
 754                        BUG_ON(ret != 0);
 755                }
 756        }
 757out:
 758        return ret;
 759}
 760
 761/*
 762 * An autonomous function to put a chunk of a page under deferred IO.
 763 *
 764 * The caller doesn't actually know (or care) whether this piece of page is in
 765 * a BIO, or is under IO or whatever.  We just take care of all possible 
 766 * situations here.  The separation between the logic of do_direct_IO() and
 767 * that of submit_page_section() is important for clarity.  Please don't break.
 768 *
 769 * The chunk of page starts on-disk at blocknr.
 770 *
 771 * We perform deferred IO, by recording the last-submitted page inside our
 772 * private part of the dio structure.  If possible, we just expand the IO
 773 * across that page here.
 774 *
 775 * If that doesn't work out then we put the old page into the bio and add this
 776 * page to the dio instead.
 777 */
 778static inline int
 779submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
 780                    unsigned offset, unsigned len, sector_t blocknr,
 781                    struct buffer_head *map_bh)
 782{
 783        int ret = 0;
 784
 785        if (dio->rw & WRITE) {
 786                /*
 787                 * Read accounting is performed in submit_bio()
 788                 */
 789                task_io_account_write(len);
 790        }
 791
 792        /*
 793         * Can we just grow the current page's presence in the dio?
 794         */
 795        if (sdio->cur_page == page &&
 796            sdio->cur_page_offset + sdio->cur_page_len == offset &&
 797            sdio->cur_page_block +
 798            (sdio->cur_page_len >> sdio->blkbits) == blocknr) {
 799                sdio->cur_page_len += len;
 800                goto out;
 801        }
 802
 803        /*
 804         * If there's a deferred page already there then send it.
 805         */
 806        if (sdio->cur_page) {
 807                ret = dio_send_cur_page(dio, sdio, map_bh);
 808                page_cache_release(sdio->cur_page);
 809                sdio->cur_page = NULL;
 810                if (ret)
 811                        return ret;
 812        }
 813
 814        page_cache_get(page);           /* It is in dio */
 815        sdio->cur_page = page;
 816        sdio->cur_page_offset = offset;
 817        sdio->cur_page_len = len;
 818        sdio->cur_page_block = blocknr;
 819        sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
 820out:
 821        /*
 822         * If sdio->boundary then we want to schedule the IO now to
 823         * avoid metadata seeks.
 824         */
 825        if (sdio->boundary) {
 826                ret = dio_send_cur_page(dio, sdio, map_bh);
 827                dio_bio_submit(dio, sdio);
 828                page_cache_release(sdio->cur_page);
 829                sdio->cur_page = NULL;
 830        }
 831        return ret;
 832}
 833
 834/*
 835 * Clean any dirty buffers in the blockdev mapping which alias newly-created
 836 * file blocks.  Only called for S_ISREG files - blockdevs do not set
 837 * buffer_new
 838 */
 839static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
 840{
 841        unsigned i;
 842        unsigned nblocks;
 843
 844        nblocks = map_bh->b_size >> dio->inode->i_blkbits;
 845
 846        for (i = 0; i < nblocks; i++) {
 847                unmap_underlying_metadata(map_bh->b_bdev,
 848                                          map_bh->b_blocknr + i);
 849        }
 850}
 851
 852/*
 853 * If we are not writing the entire block and get_block() allocated
 854 * the block for us, we need to fill-in the unused portion of the
 855 * block with zeros. This happens only if user-buffer, fileoffset or
 856 * io length is not filesystem block-size multiple.
 857 *
 858 * `end' is zero if we're doing the start of the IO, 1 at the end of the
 859 * IO.
 860 */
 861static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio,
 862                int end, struct buffer_head *map_bh)
 863{
 864        unsigned dio_blocks_per_fs_block;
 865        unsigned this_chunk_blocks;     /* In dio_blocks */
 866        unsigned this_chunk_bytes;
 867        struct page *page;
 868
 869        sdio->start_zero_done = 1;
 870        if (!sdio->blkfactor || !buffer_new(map_bh))
 871                return;
 872
 873        dio_blocks_per_fs_block = 1 << sdio->blkfactor;
 874        this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
 875
 876        if (!this_chunk_blocks)
 877                return;
 878
 879        /*
 880         * We need to zero out part of an fs block.  It is either at the
 881         * beginning or the end of the fs block.
 882         */
 883        if (end) 
 884                this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
 885
 886        this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
 887
 888        page = ZERO_PAGE(0);
 889        if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
 890                                sdio->next_block_for_io, map_bh))
 891                return;
 892
 893        sdio->next_block_for_io += this_chunk_blocks;
 894}
 895
 896/*
 897 * Walk the user pages, and the file, mapping blocks to disk and generating
 898 * a sequence of (page,offset,len,block) mappings.  These mappings are injected
 899 * into submit_page_section(), which takes care of the next stage of submission
 900 *
 901 * Direct IO against a blockdev is different from a file.  Because we can
 902 * happily perform page-sized but 512-byte aligned IOs.  It is important that
 903 * blockdev IO be able to have fine alignment and large sizes.
 904 *
 905 * So what we do is to permit the ->get_block function to populate bh.b_size
 906 * with the size of IO which is permitted at this offset and this i_blkbits.
 907 *
 908 * For best results, the blockdev should be set up with 512-byte i_blkbits and
 909 * it should set b_size to PAGE_SIZE or more inside get_block().  This gives
 910 * fine alignment but still allows this function to work in PAGE_SIZE units.
 911 */
 912static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
 913                        struct buffer_head *map_bh)
 914{
 915        const unsigned blkbits = sdio->blkbits;
 916        const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
 917        struct page *page;
 918        unsigned block_in_page;
 919        int ret = 0;
 920
 921        /* The I/O can start at any block offset within the first page */
 922        block_in_page = sdio->first_block_in_page;
 923
 924        while (sdio->block_in_file < sdio->final_block_in_request) {
 925                page = dio_get_page(dio, sdio);
 926                if (IS_ERR(page)) {
 927                        ret = PTR_ERR(page);
 928                        goto out;
 929                }
 930
 931                while (block_in_page < blocks_per_page) {
 932                        unsigned offset_in_page = block_in_page << blkbits;
 933                        unsigned this_chunk_bytes;      /* # of bytes mapped */
 934                        unsigned this_chunk_blocks;     /* # of blocks */
 935                        unsigned u;
 936
 937                        if (sdio->blocks_available == 0) {
 938                                /*
 939                                 * Need to go and map some more disk
 940                                 */
 941                                unsigned long blkmask;
 942                                unsigned long dio_remainder;
 943
 944                                ret = get_more_blocks(dio, sdio, map_bh);
 945                                if (ret) {
 946                                        page_cache_release(page);
 947                                        goto out;
 948                                }
 949                                if (!buffer_mapped(map_bh))
 950                                        goto do_holes;
 951
 952                                sdio->blocks_available =
 953                                                map_bh->b_size >> sdio->blkbits;
 954                                sdio->next_block_for_io =
 955                                        map_bh->b_blocknr << sdio->blkfactor;
 956                                if (buffer_new(map_bh))
 957                                        clean_blockdev_aliases(dio, map_bh);
 958
 959                                if (!sdio->blkfactor)
 960                                        goto do_holes;
 961
 962                                blkmask = (1 << sdio->blkfactor) - 1;
 963                                dio_remainder = (sdio->block_in_file & blkmask);
 964
 965                                /*
 966                                 * If we are at the start of IO and that IO
 967                                 * starts partway into a fs-block,
 968                                 * dio_remainder will be non-zero.  If the IO
 969                                 * is a read then we can simply advance the IO
 970                                 * cursor to the first block which is to be
 971                                 * read.  But if the IO is a write and the
 972                                 * block was newly allocated we cannot do that;
 973                                 * the start of the fs block must be zeroed out
 974                                 * on-disk
 975                                 */
 976                                if (!buffer_new(map_bh))
 977                                        sdio->next_block_for_io += dio_remainder;
 978                                sdio->blocks_available -= dio_remainder;
 979                        }
 980do_holes:
 981                        /* Handle holes */
 982                        if (!buffer_mapped(map_bh)) {
 983                                loff_t i_size_aligned;
 984
 985                                /* AKPM: eargh, -ENOTBLK is a hack */
 986                                if (dio->rw & WRITE) {
 987                                        page_cache_release(page);
 988                                        return -ENOTBLK;
 989                                }
 990
 991                                /*
 992                                 * Be sure to account for a partial block as the
 993                                 * last block in the file
 994                                 */
 995                                i_size_aligned = ALIGN(i_size_read(dio->inode),
 996                                                        1 << blkbits);
 997                                if (sdio->block_in_file >=
 998                                                i_size_aligned >> blkbits) {
 999                                        /* We hit eof */
1000                                        page_cache_release(page);
1001                                        goto out;
1002                                }
1003                                zero_user(page, block_in_page << blkbits,
1004                                                1 << blkbits);
1005                                sdio->block_in_file++;
1006                                block_in_page++;
1007                                goto next_block;
1008                        }
1009
1010                        /*
1011                         * If we're performing IO which has an alignment which
1012                         * is finer than the underlying fs, go check to see if
1013                         * we must zero out the start of this block.
1014                         */
1015                        if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
1016                                dio_zero_block(dio, sdio, 0, map_bh);
1017
1018                        /*
1019                         * Work out, in this_chunk_blocks, how much disk we
1020                         * can add to this page
1021                         */
1022                        this_chunk_blocks = sdio->blocks_available;
1023                        u = (PAGE_SIZE - offset_in_page) >> blkbits;
1024                        if (this_chunk_blocks > u)
1025                                this_chunk_blocks = u;
1026                        u = sdio->final_block_in_request - sdio->block_in_file;
1027                        if (this_chunk_blocks > u)
1028                                this_chunk_blocks = u;
1029                        this_chunk_bytes = this_chunk_blocks << blkbits;
1030                        BUG_ON(this_chunk_bytes == 0);
1031
1032                        if (this_chunk_blocks == sdio->blocks_available)
1033                                sdio->boundary = buffer_boundary(map_bh);
1034                        ret = submit_page_section(dio, sdio, page,
1035                                                  offset_in_page,
1036                                                  this_chunk_bytes,
1037                                                  sdio->next_block_for_io,
1038                                                  map_bh);
1039                        if (ret) {
1040                                page_cache_release(page);
1041                                goto out;
1042                        }
1043                        sdio->next_block_for_io += this_chunk_blocks;
1044
1045                        sdio->block_in_file += this_chunk_blocks;
1046                        block_in_page += this_chunk_blocks;
1047                        sdio->blocks_available -= this_chunk_blocks;
1048next_block:
1049                        BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
1050                        if (sdio->block_in_file == sdio->final_block_in_request)
1051                                break;
1052                }
1053
1054                /* Drop the ref which was taken in get_user_pages() */
1055                page_cache_release(page);
1056                block_in_page = 0;
1057        }
1058out:
1059        return ret;
1060}
1061
1062static inline int drop_refcount(struct dio *dio)
1063{
1064        int ret2;
1065        unsigned long flags;
1066
1067        /*
1068         * Sync will always be dropping the final ref and completing the
1069         * operation.  AIO can if it was a broken operation described above or
1070         * in fact if all the bios race to complete before we get here.  In
1071         * that case dio_complete() translates the EIOCBQUEUED into the proper
1072         * return code that the caller will hand to aio_complete().
1073         *
1074         * This is managed by the bio_lock instead of being an atomic_t so that
1075         * completion paths can drop their ref and use the remaining count to
1076         * decide to wake the submission path atomically.
1077         */
1078        spin_lock_irqsave(&dio->bio_lock, flags);
1079        ret2 = --dio->refcount;
1080        spin_unlock_irqrestore(&dio->bio_lock, flags);
1081        return ret2;
1082}
1083
1084/*
1085 * This is a library function for use by filesystem drivers.
1086 *
1087 * The locking rules are governed by the flags parameter:
1088 *  - if the flags value contains DIO_LOCKING we use a fancy locking
1089 *    scheme for dumb filesystems.
1090 *    For writes this function is called under i_mutex and returns with
1091 *    i_mutex held, for reads, i_mutex is not held on entry, but it is
1092 *    taken and dropped again before returning.
1093 *  - if the flags value does NOT contain DIO_LOCKING we don't use any
1094 *    internal locking but rather rely on the filesystem to synchronize
1095 *    direct I/O reads/writes versus each other and truncate.
1096 *
1097 * To help with locking against truncate we incremented the i_dio_count
1098 * counter before starting direct I/O, and decrement it once we are done.
1099 * Truncate can wait for it to reach zero to provide exclusion.  It is
1100 * expected that filesystem provide exclusion between new direct I/O
1101 * and truncates.  For DIO_LOCKING filesystems this is done by i_mutex,
1102 * but other filesystems need to take care of this on their own.
1103 *
1104 * NOTE: if you pass "sdio" to anything by pointer make sure that function
1105 * is always inlined. Otherwise gcc is unable to split the structure into
1106 * individual fields and will generate much worse code. This is important
1107 * for the whole file.
1108 */
1109static inline ssize_t
1110do_blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1111        struct block_device *bdev, const struct iovec *iov, loff_t offset, 
1112        unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1113        dio_submit_t submit_io, int flags)
1114{
1115        int seg;
1116        size_t size;
1117        unsigned long addr;
1118        unsigned i_blkbits = ACCESS_ONCE(inode->i_blkbits);
1119        unsigned blkbits = i_blkbits;
1120        unsigned blocksize_mask = (1 << blkbits) - 1;
1121        ssize_t retval = -EINVAL;
1122        loff_t end = offset;
1123        struct dio *dio;
1124        struct dio_submit sdio = { 0, };
1125        unsigned long user_addr;
1126        size_t bytes;
1127        struct buffer_head map_bh = { 0, };
1128        struct blk_plug plug;
1129
1130        if (rw & WRITE)
1131                rw = WRITE_ODIRECT;
1132
1133        /*
1134         * Avoid references to bdev if not absolutely needed to give
1135         * the early prefetch in the caller enough time.
1136         */
1137
1138        if (offset & blocksize_mask) {
1139                if (bdev)
1140                        blkbits = blksize_bits(bdev_logical_block_size(bdev));
1141                blocksize_mask = (1 << blkbits) - 1;
1142                if (offset & blocksize_mask)
1143                        goto out;
1144        }
1145
1146        /* Check the memory alignment.  Blocks cannot straddle pages */
1147        for (seg = 0; seg < nr_segs; seg++) {
1148                addr = (unsigned long)iov[seg].iov_base;
1149                size = iov[seg].iov_len;
1150                end += size;
1151                if (unlikely((addr & blocksize_mask) ||
1152                             (size & blocksize_mask))) {
1153                        if (bdev)
1154                                blkbits = blksize_bits(
1155                                         bdev_logical_block_size(bdev));
1156                        blocksize_mask = (1 << blkbits) - 1;
1157                        if ((addr & blocksize_mask) || (size & blocksize_mask))
1158                                goto out;
1159                }
1160        }
1161
1162        /* watch out for a 0 len io from a tricksy fs */
1163        if (rw == READ && end == offset)
1164                return 0;
1165
1166        dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
1167        retval = -ENOMEM;
1168        if (!dio)
1169                goto out;
1170        /*
1171         * Believe it or not, zeroing out the page array caused a .5%
1172         * performance regression in a database benchmark.  So, we take
1173         * care to only zero out what's needed.
1174         */
1175        memset(dio, 0, offsetof(struct dio, pages));
1176
1177        dio->flags = flags;
1178        if (dio->flags & DIO_LOCKING) {
1179                if (rw == READ) {
1180                        struct address_space *mapping =
1181                                        iocb->ki_filp->f_mapping;
1182
1183                        /* will be released by direct_io_worker */
1184                        mutex_lock(&inode->i_mutex);
1185
1186                        retval = filemap_write_and_wait_range(mapping, offset,
1187                                                              end - 1);
1188                        if (retval) {
1189                                mutex_unlock(&inode->i_mutex);
1190                                kmem_cache_free(dio_cache, dio);
1191                                goto out;
1192                        }
1193                }
1194        }
1195
1196        /*
1197         * For file extending writes updating i_size before data
1198         * writeouts complete can expose uninitialized blocks. So
1199         * even for AIO, we need to wait for i/o to complete before
1200         * returning in this case.
1201         */
1202        dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
1203                (end > i_size_read(inode)));
1204        dio->inode = inode;
1205        dio->rw = rw;
1206
1207        /*
1208         * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
1209         * so that we can call ->fsync.
1210         */
1211        if (dio->is_async && (rw & WRITE) &&
1212            ((iocb->ki_filp->f_flags & O_DSYNC) ||
1213             IS_SYNC(iocb->ki_filp->f_mapping->host))) {
1214                retval = dio_set_defer_completion(dio);
1215                if (retval) {
1216                        /*
1217                         * We grab i_mutex only for reads so we don't have
1218                         * to release it here
1219                         */
1220                        kmem_cache_free(dio_cache, dio);
1221                        goto out;
1222                }
1223        }
1224
1225        /*
1226         * Will be decremented at I/O completion time.
1227         */
1228        atomic_inc(&inode->i_dio_count);
1229
1230        retval = 0;
1231        sdio.blkbits = blkbits;
1232        sdio.blkfactor = i_blkbits - blkbits;
1233        sdio.block_in_file = offset >> blkbits;
1234
1235        sdio.get_block = get_block;
1236        dio->end_io = end_io;
1237        sdio.submit_io = submit_io;
1238        sdio.final_block_in_bio = -1;
1239        sdio.next_block_for_io = -1;
1240
1241        dio->iocb = iocb;
1242        dio->i_size = i_size_read(inode);
1243
1244        spin_lock_init(&dio->bio_lock);
1245        dio->refcount = 1;
1246
1247        /*
1248         * In case of non-aligned buffers, we may need 2 more
1249         * pages since we need to zero out first and last block.
1250         */
1251        if (unlikely(sdio.blkfactor))
1252                sdio.pages_in_io = 2;
1253
1254        for (seg = 0; seg < nr_segs; seg++) {
1255                user_addr = (unsigned long)iov[seg].iov_base;
1256                sdio.pages_in_io +=
1257                        ((user_addr + iov[seg].iov_len + PAGE_SIZE-1) /
1258                                PAGE_SIZE - user_addr / PAGE_SIZE);
1259        }
1260
1261        blk_start_plug(&plug);
1262
1263        for (seg = 0; seg < nr_segs; seg++) {
1264                user_addr = (unsigned long)iov[seg].iov_base;
1265                sdio.size += bytes = iov[seg].iov_len;
1266
1267                /* Index into the first page of the first block */
1268                sdio.first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1269                sdio.final_block_in_request = sdio.block_in_file +
1270                                                (bytes >> blkbits);
1271                /* Page fetching state */
1272                sdio.head = 0;
1273                sdio.tail = 0;
1274                sdio.curr_page = 0;
1275
1276                sdio.total_pages = 0;
1277                if (user_addr & (PAGE_SIZE-1)) {
1278                        sdio.total_pages++;
1279                        bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1280                }
1281                sdio.total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1282                sdio.curr_user_address = user_addr;
1283
1284                retval = do_direct_IO(dio, &sdio, &map_bh);
1285
1286                dio->result += iov[seg].iov_len -
1287                        ((sdio.final_block_in_request - sdio.block_in_file) <<
1288                                        blkbits);
1289
1290                if (retval) {
1291                        dio_cleanup(dio, &sdio);
1292                        break;
1293                }
1294        } /* end iovec loop */
1295
1296        if (retval == -ENOTBLK) {
1297                /*
1298                 * The remaining part of the request will be
1299                 * be handled by buffered I/O when we return
1300                 */
1301                retval = 0;
1302        }
1303        /*
1304         * There may be some unwritten disk at the end of a part-written
1305         * fs-block-sized block.  Go zero that now.
1306         */
1307        dio_zero_block(dio, &sdio, 1, &map_bh);
1308
1309        if (sdio.cur_page) {
1310                ssize_t ret2;
1311
1312                ret2 = dio_send_cur_page(dio, &sdio, &map_bh);
1313                if (retval == 0)
1314                        retval = ret2;
1315                page_cache_release(sdio.cur_page);
1316                sdio.cur_page = NULL;
1317        }
1318        if (sdio.bio)
1319                dio_bio_submit(dio, &sdio);
1320
1321        blk_finish_plug(&plug);
1322
1323        /*
1324         * It is possible that, we return short IO due to end of file.
1325         * In that case, we need to release all the pages we got hold on.
1326         */
1327        dio_cleanup(dio, &sdio);
1328
1329        /*
1330         * All block lookups have been performed. For READ requests
1331         * we can let i_mutex go now that its achieved its purpose
1332         * of protecting us from looking up uninitialized blocks.
1333         */
1334        if (rw == READ && (dio->flags & DIO_LOCKING))
1335                mutex_unlock(&dio->inode->i_mutex);
1336
1337        /*
1338         * The only time we want to leave bios in flight is when a successful
1339         * partial aio read or full aio write have been setup.  In that case
1340         * bio completion will call aio_complete.  The only time it's safe to
1341         * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1342         * This had *better* be the only place that raises -EIOCBQUEUED.
1343         */
1344        BUG_ON(retval == -EIOCBQUEUED);
1345        if (dio->is_async && retval == 0 && dio->result &&
1346            ((rw == READ) || (dio->result == sdio.size)))
1347                retval = -EIOCBQUEUED;
1348
1349        if (retval != -EIOCBQUEUED)
1350                dio_await_completion(dio);
1351
1352        if (drop_refcount(dio) == 0) {
1353                retval = dio_complete(dio, offset, retval, false);
1354        } else
1355                BUG_ON(retval != -EIOCBQUEUED);
1356
1357out:
1358        return retval;
1359}
1360
1361ssize_t
1362__blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1363        struct block_device *bdev, const struct iovec *iov, loff_t offset,
1364        unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1365        dio_submit_t submit_io, int flags)
1366{
1367        /*
1368         * The block device state is needed in the end to finally
1369         * submit everything.  Since it's likely to be cache cold
1370         * prefetch it here as first thing to hide some of the
1371         * latency.
1372         *
1373         * Attempt to prefetch the pieces we likely need later.
1374         */
1375        prefetch(&bdev->bd_disk->part_tbl);
1376        prefetch(bdev->bd_queue);
1377        prefetch((char *)bdev->bd_queue + SMP_CACHE_BYTES);
1378
1379        return do_blockdev_direct_IO(rw, iocb, inode, bdev, iov, offset,
1380                                     nr_segs, get_block, end_io,
1381                                     submit_io, flags);
1382}
1383
1384EXPORT_SYMBOL(__blockdev_direct_IO);
1385
1386static __init int dio_init(void)
1387{
1388        dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
1389        return 0;
1390}
1391module_init(dio_init)
1392
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.