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 <asm/atomic.h>
  39
  40/*
  41 * How many user pages to map in one call to get_user_pages().  This determines
  42 * the size of a structure on the stack.
  43 */
  44#define DIO_PAGES       64
  45
  46/*
  47 * This code generally works in units of "dio_blocks".  A dio_block is
  48 * somewhere between the hard sector size and the filesystem block size.  it
  49 * is determined on a per-invocation basis.   When talking to the filesystem
  50 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
  51 * down by dio->blkfactor.  Similarly, fs-blocksize quantities are converted
  52 * to bio_block quantities by shifting left by blkfactor.
  53 *
  54 * If blkfactor is zero then the user's request was aligned to the filesystem's
  55 * blocksize.
  56 *
  57 * lock_type is DIO_LOCKING for regular files on direct-IO-naive filesystems.
  58 * This determines whether we need to do the fancy locking which prevents
  59 * direct-IO from being able to read uninitialised disk blocks.  If its zero
  60 * (blockdev) this locking is not done, and if it is DIO_OWN_LOCKING i_mutex is
  61 * not held for the entire direct write (taken briefly, initially, during a
  62 * direct read though, but its never held for the duration of a direct-IO).
  63 */
  64
  65struct dio {
  66        /* BIO submission state */
  67        struct bio *bio;                /* bio under assembly */
  68        struct inode *inode;
  69        int rw;
  70        loff_t i_size;                  /* i_size when submitted */
  71        int lock_type;                  /* doesn't change */
  72        unsigned blkbits;               /* doesn't change */
  73        unsigned blkfactor;             /* When we're using an alignment which
  74                                           is finer than the filesystem's soft
  75                                           blocksize, this specifies how much
  76                                           finer.  blkfactor=2 means 1/4-block
  77                                           alignment.  Does not change */
  78        unsigned start_zero_done;       /* flag: sub-blocksize zeroing has
  79                                           been performed at the start of a
  80                                           write */
  81        int pages_in_io;                /* approximate total IO pages */
  82        size_t  size;                   /* total request size (doesn't change)*/
  83        sector_t block_in_file;         /* Current offset into the underlying
  84                                           file in dio_block units. */
  85        unsigned blocks_available;      /* At block_in_file.  changes */
  86        sector_t final_block_in_request;/* doesn't change */
  87        unsigned first_block_in_page;   /* doesn't change, Used only once */
  88        int boundary;                   /* prev block is at a boundary */
  89        int reap_counter;               /* rate limit reaping */
  90        get_block_t *get_block;         /* block mapping function */
  91        dio_iodone_t *end_io;           /* IO completion function */
  92        sector_t final_block_in_bio;    /* current final block in bio + 1 */
  93        sector_t next_block_for_io;     /* next block to be put under IO,
  94                                           in dio_blocks units */
  95        struct buffer_head map_bh;      /* last get_block() result */
  96
  97        /*
  98         * Deferred addition of a page to the dio.  These variables are
  99         * private to dio_send_cur_page(), submit_page_section() and
 100         * dio_bio_add_page().
 101         */
 102        struct page *cur_page;          /* The page */
 103        unsigned cur_page_offset;       /* Offset into it, in bytes */
 104        unsigned cur_page_len;          /* Nr of bytes at cur_page_offset */
 105        sector_t cur_page_block;        /* Where it starts */
 106
 107        /*
 108         * Page fetching state. These variables belong to dio_refill_pages().
 109         */
 110        int curr_page;                  /* changes */
 111        int total_pages;                /* doesn't change */
 112        unsigned long curr_user_address;/* changes */
 113
 114        /*
 115         * Page queue.  These variables belong to dio_refill_pages() and
 116         * dio_get_page().
 117         */
 118        struct page *pages[DIO_PAGES];  /* page buffer */
 119        unsigned head;                  /* next page to process */
 120        unsigned tail;                  /* last valid page + 1 */
 121        int page_errors;                /* errno from get_user_pages() */
 122
 123        /* BIO completion state */
 124        spinlock_t bio_lock;            /* protects BIO fields below */
 125        unsigned long refcount;         /* direct_io_worker() and bios */
 126        struct bio *bio_list;           /* singly linked via bi_private */
 127        struct task_struct *waiter;     /* waiting task (NULL if none) */
 128
 129        /* AIO related stuff */
 130        struct kiocb *iocb;             /* kiocb */
 131        int is_async;                   /* is IO async ? */
 132        int io_error;                   /* IO error in completion path */
 133        ssize_t result;                 /* IO result */
 134};
 135
 136/*
 137 * How many pages are in the queue?
 138 */
 139static inline unsigned dio_pages_present(struct dio *dio)
 140{
 141        return dio->tail - dio->head;
 142}
 143
 144/*
 145 * Go grab and pin some userspace pages.   Typically we'll get 64 at a time.
 146 */
 147static int dio_refill_pages(struct dio *dio)
 148{
 149        int ret;
 150        int nr_pages;
 151
 152        nr_pages = min(dio->total_pages - dio->curr_page, DIO_PAGES);
 153        ret = get_user_pages_fast(
 154                dio->curr_user_address,         /* Where from? */
 155                nr_pages,                       /* How many pages? */
 156                dio->rw == READ,                /* Write to memory? */
 157                &dio->pages[0]);                /* Put results here */
 158
 159        if (ret < 0 && dio->blocks_available && (dio->rw & WRITE)) {
 160                struct page *page = ZERO_PAGE(0);
 161                /*
 162                 * A memory fault, but the filesystem has some outstanding
 163                 * mapped blocks.  We need to use those blocks up to avoid
 164                 * leaking stale data in the file.
 165                 */
 166                if (dio->page_errors == 0)
 167                        dio->page_errors = ret;
 168                page_cache_get(page);
 169                dio->pages[0] = page;
 170                dio->head = 0;
 171                dio->tail = 1;
 172                ret = 0;
 173                goto out;
 174        }
 175
 176        if (ret >= 0) {
 177                dio->curr_user_address += ret * PAGE_SIZE;
 178                dio->curr_page += ret;
 179                dio->head = 0;
 180                dio->tail = ret;
 181                ret = 0;
 182        }
 183out:
 184        return ret;     
 185}
 186
 187/*
 188 * Get another userspace page.  Returns an ERR_PTR on error.  Pages are
 189 * buffered inside the dio so that we can call get_user_pages() against a
 190 * decent number of pages, less frequently.  To provide nicer use of the
 191 * L1 cache.
 192 */
 193static struct page *dio_get_page(struct dio *dio)
 194{
 195        if (dio_pages_present(dio) == 0) {
 196                int ret;
 197
 198                ret = dio_refill_pages(dio);
 199                if (ret)
 200                        return ERR_PTR(ret);
 201                BUG_ON(dio_pages_present(dio) == 0);
 202        }
 203        return dio->pages[dio->head++];
 204}
 205
 206/**
 207 * dio_complete() - called when all DIO BIO I/O has been completed
 208 * @offset: the byte offset in the file of the completed operation
 209 *
 210 * This releases locks as dictated by the locking type, lets interested parties
 211 * know that a DIO operation has completed, and calculates the resulting return
 212 * code for the operation.
 213 *
 214 * It lets the filesystem know if it registered an interest earlier via
 215 * get_block.  Pass the private field of the map buffer_head so that
 216 * filesystems can use it to hold additional state between get_block calls and
 217 * dio_complete.
 218 */
 219static int dio_complete(struct dio *dio, loff_t offset, int ret)
 220{
 221        ssize_t transferred = 0;
 222
 223        /*
 224         * AIO submission can race with bio completion to get here while
 225         * expecting to have the last io completed by bio completion.
 226         * In that case -EIOCBQUEUED is in fact not an error we want
 227         * to preserve through this call.
 228         */
 229        if (ret == -EIOCBQUEUED)
 230                ret = 0;
 231
 232        if (dio->result) {
 233                transferred = dio->result;
 234
 235                /* Check for short read case */
 236                if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
 237                        transferred = dio->i_size - offset;
 238        }
 239
 240        if (dio->end_io && dio->result)
 241                dio->end_io(dio->iocb, offset, transferred,
 242                            dio->map_bh.b_private);
 243        if (dio->lock_type == DIO_LOCKING)
 244                /* lockdep: non-owner release */
 245                up_read_non_owner(&dio->inode->i_alloc_sem);
 246
 247        if (ret == 0)
 248                ret = dio->page_errors;
 249        if (ret == 0)
 250                ret = dio->io_error;
 251        if (ret == 0)
 252                ret = transferred;
 253
 254        return ret;
 255}
 256
 257static int dio_bio_complete(struct dio *dio, struct bio *bio);
 258/*
 259 * Asynchronous IO callback. 
 260 */
 261static void dio_bio_end_aio(struct bio *bio, int error)
 262{
 263        struct dio *dio = bio->bi_private;
 264        unsigned long remaining;
 265        unsigned long flags;
 266
 267        /* cleanup the bio */
 268        dio_bio_complete(dio, bio);
 269
 270        spin_lock_irqsave(&dio->bio_lock, flags);
 271        remaining = --dio->refcount;
 272        if (remaining == 1 && dio->waiter)
 273                wake_up_process(dio->waiter);
 274        spin_unlock_irqrestore(&dio->bio_lock, flags);
 275
 276        if (remaining == 0) {
 277                int ret = dio_complete(dio, dio->iocb->ki_pos, 0);
 278                aio_complete(dio->iocb, ret, 0);
 279                kfree(dio);
 280        }
 281}
 282
 283/*
 284 * The BIO completion handler simply queues the BIO up for the process-context
 285 * handler.
 286 *
 287 * During I/O bi_private points at the dio.  After I/O, bi_private is used to
 288 * implement a singly-linked list of completed BIOs, at dio->bio_list.
 289 */
 290static void dio_bio_end_io(struct bio *bio, int error)
 291{
 292        struct dio *dio = bio->bi_private;
 293        unsigned long flags;
 294
 295        spin_lock_irqsave(&dio->bio_lock, flags);
 296        bio->bi_private = dio->bio_list;
 297        dio->bio_list = bio;
 298        if (--dio->refcount == 1 && dio->waiter)
 299                wake_up_process(dio->waiter);
 300        spin_unlock_irqrestore(&dio->bio_lock, flags);
 301}
 302
 303static int
 304dio_bio_alloc(struct dio *dio, struct block_device *bdev,
 305                sector_t first_sector, int nr_vecs)
 306{
 307        struct bio *bio;
 308
 309        bio = bio_alloc(GFP_KERNEL, nr_vecs);
 310        if (bio == NULL)
 311                return -ENOMEM;
 312
 313        bio->bi_bdev = bdev;
 314        bio->bi_sector = first_sector;
 315        if (dio->is_async)
 316                bio->bi_end_io = dio_bio_end_aio;
 317        else
 318                bio->bi_end_io = dio_bio_end_io;
 319
 320        dio->bio = bio;
 321        return 0;
 322}
 323
 324/*
 325 * In the AIO read case we speculatively dirty the pages before starting IO.
 326 * During IO completion, any of these pages which happen to have been written
 327 * back will be redirtied by bio_check_pages_dirty().
 328 *
 329 * bios hold a dio reference between submit_bio and ->end_io.
 330 */
 331static void dio_bio_submit(struct dio *dio)
 332{
 333        struct bio *bio = dio->bio;
 334        unsigned long flags;
 335
 336        bio->bi_private = dio;
 337
 338        spin_lock_irqsave(&dio->bio_lock, flags);
 339        dio->refcount++;
 340        spin_unlock_irqrestore(&dio->bio_lock, flags);
 341
 342        if (dio->is_async && dio->rw == READ)
 343                bio_set_pages_dirty(bio);
 344
 345        submit_bio(dio->rw, bio);
 346
 347        dio->bio = NULL;
 348        dio->boundary = 0;
 349}
 350
 351/*
 352 * Release any resources in case of a failure
 353 */
 354static void dio_cleanup(struct dio *dio)
 355{
 356        while (dio_pages_present(dio))
 357                page_cache_release(dio_get_page(dio));
 358}
 359
 360/*
 361 * Wait for the next BIO to complete.  Remove it and return it.  NULL is
 362 * returned once all BIOs have been completed.  This must only be called once
 363 * all bios have been issued so that dio->refcount can only decrease.  This
 364 * requires that that the caller hold a reference on the dio.
 365 */
 366static struct bio *dio_await_one(struct dio *dio)
 367{
 368        unsigned long flags;
 369        struct bio *bio = NULL;
 370
 371        spin_lock_irqsave(&dio->bio_lock, flags);
 372
 373        /*
 374         * Wait as long as the list is empty and there are bios in flight.  bio
 375         * completion drops the count, maybe adds to the list, and wakes while
 376         * holding the bio_lock so we don't need set_current_state()'s barrier
 377         * and can call it after testing our condition.
 378         */
 379        while (dio->refcount > 1 && dio->bio_list == NULL) {
 380                __set_current_state(TASK_UNINTERRUPTIBLE);
 381                dio->waiter = current;
 382                spin_unlock_irqrestore(&dio->bio_lock, flags);
 383                io_schedule();
 384                /* wake up sets us TASK_RUNNING */
 385                spin_lock_irqsave(&dio->bio_lock, flags);
 386                dio->waiter = NULL;
 387        }
 388        if (dio->bio_list) {
 389                bio = dio->bio_list;
 390                dio->bio_list = bio->bi_private;
 391        }
 392        spin_unlock_irqrestore(&dio->bio_lock, flags);
 393        return bio;
 394}
 395
 396/*
 397 * Process one completed BIO.  No locks are held.
 398 */
 399static int dio_bio_complete(struct dio *dio, struct bio *bio)
 400{
 401        const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 402        struct bio_vec *bvec = bio->bi_io_vec;
 403        int page_no;
 404
 405        if (!uptodate)
 406                dio->io_error = -EIO;
 407
 408        if (dio->is_async && dio->rw == READ) {
 409                bio_check_pages_dirty(bio);     /* transfers ownership */
 410        } else {
 411                for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
 412                        struct page *page = bvec[page_no].bv_page;
 413
 414                        if (dio->rw == READ && !PageCompound(page))
 415                                set_page_dirty_lock(page);
 416                        page_cache_release(page);
 417                }
 418                bio_put(bio);
 419        }
 420        return uptodate ? 0 : -EIO;
 421}
 422
 423/*
 424 * Wait on and process all in-flight BIOs.  This must only be called once
 425 * all bios have been issued so that the refcount can only decrease.
 426 * This just waits for all bios to make it through dio_bio_complete.  IO
 427 * errors are propagated through dio->io_error and should be propagated via
 428 * dio_complete().
 429 */
 430static void dio_await_completion(struct dio *dio)
 431{
 432        struct bio *bio;
 433        do {
 434                bio = dio_await_one(dio);
 435                if (bio)
 436                        dio_bio_complete(dio, bio);
 437        } while (bio);
 438}
 439
 440/*
 441 * A really large O_DIRECT read or write can generate a lot of BIOs.  So
 442 * to keep the memory consumption sane we periodically reap any completed BIOs
 443 * during the BIO generation phase.
 444 *
 445 * This also helps to limit the peak amount of pinned userspace memory.
 446 */
 447static int dio_bio_reap(struct dio *dio)
 448{
 449        int ret = 0;
 450
 451        if (dio->reap_counter++ >= 64) {
 452                while (dio->bio_list) {
 453                        unsigned long flags;
 454                        struct bio *bio;
 455                        int ret2;
 456
 457                        spin_lock_irqsave(&dio->bio_lock, flags);
 458                        bio = dio->bio_list;
 459                        dio->bio_list = bio->bi_private;
 460                        spin_unlock_irqrestore(&dio->bio_lock, flags);
 461                        ret2 = dio_bio_complete(dio, bio);
 462                        if (ret == 0)
 463                                ret = ret2;
 464                }
 465                dio->reap_counter = 0;
 466        }
 467        return ret;
 468}
 469
 470/*
 471 * Call into the fs to map some more disk blocks.  We record the current number
 472 * of available blocks at dio->blocks_available.  These are in units of the
 473 * fs blocksize, (1 << inode->i_blkbits).
 474 *
 475 * The fs is allowed to map lots of blocks at once.  If it wants to do that,
 476 * it uses the passed inode-relative block number as the file offset, as usual.
 477 *
 478 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
 479 * has remaining to do.  The fs should not map more than this number of blocks.
 480 *
 481 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
 482 * indicate how much contiguous disk space has been made available at
 483 * bh->b_blocknr.
 484 *
 485 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
 486 * This isn't very efficient...
 487 *
 488 * In the case of filesystem holes: the fs may return an arbitrarily-large
 489 * hole by returning an appropriate value in b_size and by clearing
 490 * buffer_mapped().  However the direct-io code will only process holes one
 491 * block at a time - it will repeatedly call get_block() as it walks the hole.
 492 */
 493static int get_more_blocks(struct dio *dio)
 494{
 495        int ret;
 496        struct buffer_head *map_bh = &dio->map_bh;
 497        sector_t fs_startblk;   /* Into file, in filesystem-sized blocks */
 498        unsigned long fs_count; /* Number of filesystem-sized blocks */
 499        unsigned long dio_count;/* Number of dio_block-sized blocks */
 500        unsigned long blkmask;
 501        int create;
 502
 503        /*
 504         * If there was a memory error and we've overwritten all the
 505         * mapped blocks then we can now return that memory error
 506         */
 507        ret = dio->page_errors;
 508        if (ret == 0) {
 509                BUG_ON(dio->block_in_file >= dio->final_block_in_request);
 510                fs_startblk = dio->block_in_file >> dio->blkfactor;
 511                dio_count = dio->final_block_in_request - dio->block_in_file;
 512                fs_count = dio_count >> dio->blkfactor;
 513                blkmask = (1 << dio->blkfactor) - 1;
 514                if (dio_count & blkmask)        
 515                        fs_count++;
 516
 517                map_bh->b_state = 0;
 518                map_bh->b_size = fs_count << dio->inode->i_blkbits;
 519
 520                create = dio->rw & WRITE;
 521                if (dio->lock_type == DIO_LOCKING) {
 522                        if (dio->block_in_file < (i_size_read(dio->inode) >>
 523                                                        dio->blkbits))
 524                                create = 0;
 525                } else if (dio->lock_type == DIO_NO_LOCKING) {
 526                        create = 0;
 527                }
 528
 529                /*
 530                 * For writes inside i_size we forbid block creations: only
 531                 * overwrites are permitted.  We fall back to buffered writes
 532                 * at a higher level for inside-i_size block-instantiating
 533                 * writes.
 534                 */
 535                ret = (*dio->get_block)(dio->inode, fs_startblk,
 536                                                map_bh, create);
 537        }
 538        return ret;
 539}
 540
 541/*
 542 * There is no bio.  Make one now.
 543 */
 544static int dio_new_bio(struct dio *dio, sector_t start_sector)
 545{
 546        sector_t sector;
 547        int ret, nr_pages;
 548
 549        ret = dio_bio_reap(dio);
 550        if (ret)
 551                goto out;
 552        sector = start_sector << (dio->blkbits - 9);
 553        nr_pages = min(dio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev));
 554        BUG_ON(nr_pages <= 0);
 555        ret = dio_bio_alloc(dio, dio->map_bh.b_bdev, sector, nr_pages);
 556        dio->boundary = 0;
 557out:
 558        return ret;
 559}
 560
 561/*
 562 * Attempt to put the current chunk of 'cur_page' into the current BIO.  If
 563 * that was successful then update final_block_in_bio and take a ref against
 564 * the just-added page.
 565 *
 566 * Return zero on success.  Non-zero means the caller needs to start a new BIO.
 567 */
 568static int dio_bio_add_page(struct dio *dio)
 569{
 570        int ret;
 571
 572        ret = bio_add_page(dio->bio, dio->cur_page,
 573                        dio->cur_page_len, dio->cur_page_offset);
 574        if (ret == dio->cur_page_len) {
 575                /*
 576                 * Decrement count only, if we are done with this page
 577                 */
 578                if ((dio->cur_page_len + dio->cur_page_offset) == PAGE_SIZE)
 579                        dio->pages_in_io--;
 580                page_cache_get(dio->cur_page);
 581                dio->final_block_in_bio = dio->cur_page_block +
 582                        (dio->cur_page_len >> dio->blkbits);
 583                ret = 0;
 584        } else {
 585                ret = 1;
 586        }
 587        return ret;
 588}
 589                
 590/*
 591 * Put cur_page under IO.  The section of cur_page which is described by
 592 * cur_page_offset,cur_page_len is put into a BIO.  The section of cur_page
 593 * starts on-disk at cur_page_block.
 594 *
 595 * We take a ref against the page here (on behalf of its presence in the bio).
 596 *
 597 * The caller of this function is responsible for removing cur_page from the
 598 * dio, and for dropping the refcount which came from that presence.
 599 */
 600static int dio_send_cur_page(struct dio *dio)
 601{
 602        int ret = 0;
 603
 604        if (dio->bio) {
 605                /*
 606                 * See whether this new request is contiguous with the old
 607                 */
 608                if (dio->final_block_in_bio != dio->cur_page_block)
 609                        dio_bio_submit(dio);
 610                /*
 611                 * Submit now if the underlying fs is about to perform a
 612                 * metadata read
 613                 */
 614                if (dio->boundary)
 615                        dio_bio_submit(dio);
 616        }
 617
 618        if (dio->bio == NULL) {
 619                ret = dio_new_bio(dio, dio->cur_page_block);
 620                if (ret)
 621                        goto out;
 622        }
 623
 624        if (dio_bio_add_page(dio) != 0) {
 625                dio_bio_submit(dio);
 626                ret = dio_new_bio(dio, dio->cur_page_block);
 627                if (ret == 0) {
 628                        ret = dio_bio_add_page(dio);
 629                        BUG_ON(ret != 0);
 630                }
 631        }
 632out:
 633        return ret;
 634}
 635
 636/*
 637 * An autonomous function to put a chunk of a page under deferred IO.
 638 *
 639 * The caller doesn't actually know (or care) whether this piece of page is in
 640 * a BIO, or is under IO or whatever.  We just take care of all possible 
 641 * situations here.  The separation between the logic of do_direct_IO() and
 642 * that of submit_page_section() is important for clarity.  Please don't break.
 643 *
 644 * The chunk of page starts on-disk at blocknr.
 645 *
 646 * We perform deferred IO, by recording the last-submitted page inside our
 647 * private part of the dio structure.  If possible, we just expand the IO
 648 * across that page here.
 649 *
 650 * If that doesn't work out then we put the old page into the bio and add this
 651 * page to the dio instead.
 652 */
 653static int
 654submit_page_section(struct dio *dio, struct page *page,
 655                unsigned offset, unsigned len, sector_t blocknr)
 656{
 657        int ret = 0;
 658
 659        if (dio->rw & WRITE) {
 660                /*
 661                 * Read accounting is performed in submit_bio()
 662                 */
 663                task_io_account_write(len);
 664        }
 665
 666        /*
 667         * Can we just grow the current page's presence in the dio?
 668         */
 669        if (    (dio->cur_page == page) &&
 670                (dio->cur_page_offset + dio->cur_page_len == offset) &&
 671                (dio->cur_page_block +
 672                        (dio->cur_page_len >> dio->blkbits) == blocknr)) {
 673                dio->cur_page_len += len;
 674
 675                /*
 676                 * If dio->boundary then we want to schedule the IO now to
 677                 * avoid metadata seeks.
 678                 */
 679                if (dio->boundary) {
 680                        ret = dio_send_cur_page(dio);
 681                        page_cache_release(dio->cur_page);
 682                        dio->cur_page = NULL;
 683                }
 684                goto out;
 685        }
 686
 687        /*
 688         * If there's a deferred page already there then send it.
 689         */
 690        if (dio->cur_page) {
 691                ret = dio_send_cur_page(dio);
 692                page_cache_release(dio->cur_page);
 693                dio->cur_page = NULL;
 694                if (ret)
 695                        goto out;
 696        }
 697
 698        page_cache_get(page);           /* It is in dio */
 699        dio->cur_page = page;
 700        dio->cur_page_offset = offset;
 701        dio->cur_page_len = len;
 702        dio->cur_page_block = blocknr;
 703out:
 704        return ret;
 705}
 706
 707/*
 708 * Clean any dirty buffers in the blockdev mapping which alias newly-created
 709 * file blocks.  Only called for S_ISREG files - blockdevs do not set
 710 * buffer_new
 711 */
 712static void clean_blockdev_aliases(struct dio *dio)
 713{
 714        unsigned i;
 715        unsigned nblocks;
 716
 717        nblocks = dio->map_bh.b_size >> dio->inode->i_blkbits;
 718
 719        for (i = 0; i < nblocks; i++) {
 720                unmap_underlying_metadata(dio->map_bh.b_bdev,
 721                                        dio->map_bh.b_blocknr + i);
 722        }
 723}
 724
 725/*
 726 * If we are not writing the entire block and get_block() allocated
 727 * the block for us, we need to fill-in the unused portion of the
 728 * block with zeros. This happens only if user-buffer, fileoffset or
 729 * io length is not filesystem block-size multiple.
 730 *
 731 * `end' is zero if we're doing the start of the IO, 1 at the end of the
 732 * IO.
 733 */
 734static void dio_zero_block(struct dio *dio, int end)
 735{
 736        unsigned dio_blocks_per_fs_block;
 737        unsigned this_chunk_blocks;     /* In dio_blocks */
 738        unsigned this_chunk_bytes;
 739        struct page *page;
 740
 741        dio->start_zero_done = 1;
 742        if (!dio->blkfactor || !buffer_new(&dio->map_bh))
 743                return;
 744
 745        dio_blocks_per_fs_block = 1 << dio->blkfactor;
 746        this_chunk_blocks = dio->block_in_file & (dio_blocks_per_fs_block - 1);
 747
 748        if (!this_chunk_blocks)
 749                return;
 750
 751        /*
 752         * We need to zero out part of an fs block.  It is either at the
 753         * beginning or the end of the fs block.
 754         */
 755        if (end) 
 756                this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
 757
 758        this_chunk_bytes = this_chunk_blocks << dio->blkbits;
 759
 760        page = ZERO_PAGE(0);
 761        if (submit_page_section(dio, page, 0, this_chunk_bytes, 
 762                                dio->next_block_for_io))
 763                return;
 764
 765        dio->next_block_for_io += this_chunk_blocks;
 766}
 767
 768/*
 769 * Walk the user pages, and the file, mapping blocks to disk and generating
 770 * a sequence of (page,offset,len,block) mappings.  These mappings are injected
 771 * into submit_page_section(), which takes care of the next stage of submission
 772 *
 773 * Direct IO against a blockdev is different from a file.  Because we can
 774 * happily perform page-sized but 512-byte aligned IOs.  It is important that
 775 * blockdev IO be able to have fine alignment and large sizes.
 776 *
 777 * So what we do is to permit the ->get_block function to populate bh.b_size
 778 * with the size of IO which is permitted at this offset and this i_blkbits.
 779 *
 780 * For best results, the blockdev should be set up with 512-byte i_blkbits and
 781 * it should set b_size to PAGE_SIZE or more inside get_block().  This gives
 782 * fine alignment but still allows this function to work in PAGE_SIZE units.
 783 */
 784static int do_direct_IO(struct dio *dio)
 785{
 786        const unsigned blkbits = dio->blkbits;
 787        const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
 788        struct page *page;
 789        unsigned block_in_page;
 790        struct buffer_head *map_bh = &dio->map_bh;
 791        int ret = 0;
 792
 793        /* The I/O can start at any block offset within the first page */
 794        block_in_page = dio->first_block_in_page;
 795
 796        while (dio->block_in_file < dio->final_block_in_request) {
 797                page = dio_get_page(dio);
 798                if (IS_ERR(page)) {
 799                        ret = PTR_ERR(page);
 800                        goto out;
 801                }
 802
 803                while (block_in_page < blocks_per_page) {
 804                        unsigned offset_in_page = block_in_page << blkbits;
 805                        unsigned this_chunk_bytes;      /* # of bytes mapped */
 806                        unsigned this_chunk_blocks;     /* # of blocks */
 807                        unsigned u;
 808
 809                        if (dio->blocks_available == 0) {
 810                                /*
 811                                 * Need to go and map some more disk
 812                                 */
 813                                unsigned long blkmask;
 814                                unsigned long dio_remainder;
 815
 816                                ret = get_more_blocks(dio);
 817                                if (ret) {
 818                                        page_cache_release(page);
 819                                        goto out;
 820                                }
 821                                if (!buffer_mapped(map_bh))
 822                                        goto do_holes;
 823
 824                                dio->blocks_available =
 825                                                map_bh->b_size >> dio->blkbits;
 826                                dio->next_block_for_io =
 827                                        map_bh->b_blocknr << dio->blkfactor;
 828                                if (buffer_new(map_bh))
 829                                        clean_blockdev_aliases(dio);
 830
 831                                if (!dio->blkfactor)
 832                                        goto do_holes;
 833
 834                                blkmask = (1 << dio->blkfactor) - 1;
 835                                dio_remainder = (dio->block_in_file & blkmask);
 836
 837                                /*
 838                                 * If we are at the start of IO and that IO
 839                                 * starts partway into a fs-block,
 840                                 * dio_remainder will be non-zero.  If the IO
 841                                 * is a read then we can simply advance the IO
 842                                 * cursor to the first block which is to be
 843                                 * read.  But if the IO is a write and the
 844                                 * block was newly allocated we cannot do that;
 845                                 * the start of the fs block must be zeroed out
 846                                 * on-disk
 847                                 */
 848                                if (!buffer_new(map_bh))
 849                                        dio->next_block_for_io += dio_remainder;
 850                                dio->blocks_available -= dio_remainder;
 851                        }
 852do_holes:
 853                        /* Handle holes */
 854                        if (!buffer_mapped(map_bh)) {
 855                                loff_t i_size_aligned;
 856
 857                                /* AKPM: eargh, -ENOTBLK is a hack */
 858                                if (dio->rw & WRITE) {
 859                                        page_cache_release(page);
 860                                        return -ENOTBLK;
 861                                }
 862
 863                                /*
 864                                 * Be sure to account for a partial block as the
 865                                 * last block in the file
 866                                 */
 867                                i_size_aligned = ALIGN(i_size_read(dio->inode),
 868                                                        1 << blkbits);
 869                                if (dio->block_in_file >=
 870                                                i_size_aligned >> blkbits) {
 871                                        /* We hit eof */
 872                                        page_cache_release(page);
 873                                        goto out;
 874                                }
 875                                zero_user(page, block_in_page << blkbits,
 876                                                1 << blkbits);
 877                                dio->block_in_file++;
 878                                block_in_page++;
 879                                goto next_block;
 880                        }
 881
 882                        /*
 883                         * If we're performing IO which has an alignment which
 884                         * is finer than the underlying fs, go check to see if
 885                         * we must zero out the start of this block.
 886                         */
 887                        if (unlikely(dio->blkfactor && !dio->start_zero_done))
 888                                dio_zero_block(dio, 0);
 889
 890                        /*
 891                         * Work out, in this_chunk_blocks, how much disk we
 892                         * can add to this page
 893                         */
 894                        this_chunk_blocks = dio->blocks_available;
 895                        u = (PAGE_SIZE - offset_in_page) >> blkbits;
 896                        if (this_chunk_blocks > u)
 897                                this_chunk_blocks = u;
 898                        u = dio->final_block_in_request - dio->block_in_file;
 899                        if (this_chunk_blocks > u)
 900                                this_chunk_blocks = u;
 901                        this_chunk_bytes = this_chunk_blocks << blkbits;
 902                        BUG_ON(this_chunk_bytes == 0);
 903
 904                        dio->boundary = buffer_boundary(map_bh);
 905                        ret = submit_page_section(dio, page, offset_in_page,
 906                                this_chunk_bytes, dio->next_block_for_io);
 907                        if (ret) {
 908                                page_cache_release(page);
 909                                goto out;
 910                        }
 911                        dio->next_block_for_io += this_chunk_blocks;
 912
 913                        dio->block_in_file += this_chunk_blocks;
 914                        block_in_page += this_chunk_blocks;
 915                        dio->blocks_available -= this_chunk_blocks;
 916next_block:
 917                        BUG_ON(dio->block_in_file > dio->final_block_in_request);
 918                        if (dio->block_in_file == dio->final_block_in_request)
 919                                break;
 920                }
 921
 922                /* Drop the ref which was taken in get_user_pages() */
 923                page_cache_release(page);
 924                block_in_page = 0;
 925        }
 926out:
 927        return ret;
 928}
 929
 930/*
 931 * Releases both i_mutex and i_alloc_sem
 932 */
 933static ssize_t
 934direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode, 
 935        const struct iovec *iov, loff_t offset, unsigned long nr_segs, 
 936        unsigned blkbits, get_block_t get_block, dio_iodone_t end_io,
 937        struct dio *dio)
 938{
 939        unsigned long user_addr; 
 940        unsigned long flags;
 941        int seg;
 942        ssize_t ret = 0;
 943        ssize_t ret2;
 944        size_t bytes;
 945
 946        dio->inode = inode;
 947        dio->rw = rw;
 948        dio->blkbits = blkbits;
 949        dio->blkfactor = inode->i_blkbits - blkbits;
 950        dio->block_in_file = offset >> blkbits;
 951
 952        dio->get_block = get_block;
 953        dio->end_io = end_io;
 954        dio->final_block_in_bio = -1;
 955        dio->next_block_for_io = -1;
 956
 957        dio->iocb = iocb;
 958        dio->i_size = i_size_read(inode);
 959
 960        spin_lock_init(&dio->bio_lock);
 961        dio->refcount = 1;
 962
 963        /*
 964         * In case of non-aligned buffers, we may need 2 more
 965         * pages since we need to zero out first and last block.
 966         */
 967        if (unlikely(dio->blkfactor))
 968                dio->pages_in_io = 2;
 969
 970        for (seg = 0; seg < nr_segs; seg++) {
 971                user_addr = (unsigned long)iov[seg].iov_base;
 972                dio->pages_in_io +=
 973                        ((user_addr+iov[seg].iov_len +PAGE_SIZE-1)/PAGE_SIZE
 974                                - user_addr/PAGE_SIZE);
 975        }
 976
 977        for (seg = 0; seg < nr_segs; seg++) {
 978                user_addr = (unsigned long)iov[seg].iov_base;
 979                dio->size += bytes = iov[seg].iov_len;
 980
 981                /* Index into the first page of the first block */
 982                dio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
 983                dio->final_block_in_request = dio->block_in_file +
 984                                                (bytes >> blkbits);
 985                /* Page fetching state */
 986                dio->head = 0;
 987                dio->tail = 0;
 988                dio->curr_page = 0;
 989
 990                dio->total_pages = 0;
 991                if (user_addr & (PAGE_SIZE-1)) {
 992                        dio->total_pages++;
 993                        bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
 994                }
 995                dio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
 996                dio->curr_user_address = user_addr;
 997        
 998                ret = do_direct_IO(dio);
 999
1000                dio->result += iov[seg].iov_len -
1001                        ((dio->final_block_in_request - dio->block_in_file) <<
1002                                        blkbits);
1003
1004                if (ret) {
1005                        dio_cleanup(dio);
1006                        break;
1007                }
1008        } /* end iovec loop */
1009
1010        if (ret == -ENOTBLK && (rw & WRITE)) {
1011                /*
1012                 * The remaining part of the request will be
1013                 * be handled by buffered I/O when we return
1014                 */
1015                ret = 0;
1016        }
1017        /*
1018         * There may be some unwritten disk at the end of a part-written
1019         * fs-block-sized block.  Go zero that now.
1020         */
1021        dio_zero_block(dio, 1);
1022
1023        if (dio->cur_page) {
1024                ret2 = dio_send_cur_page(dio);
1025                if (ret == 0)
1026                        ret = ret2;
1027                page_cache_release(dio->cur_page);
1028                dio->cur_page = NULL;
1029        }
1030        if (dio->bio)
1031                dio_bio_submit(dio);
1032
1033        /* All IO is now issued, send it on its way */
1034        blk_run_address_space(inode->i_mapping);
1035
1036        /*
1037         * It is possible that, we return short IO due to end of file.
1038         * In that case, we need to release all the pages we got hold on.
1039         */
1040        dio_cleanup(dio);
1041
1042        /*
1043         * All block lookups have been performed. For READ requests
1044         * we can let i_mutex go now that its achieved its purpose
1045         * of protecting us from looking up uninitialized blocks.
1046         */
1047        if ((rw == READ) && (dio->lock_type == DIO_LOCKING))
1048                mutex_unlock(&dio->inode->i_mutex);
1049
1050        /*
1051         * The only time we want to leave bios in flight is when a successful
1052         * partial aio read or full aio write have been setup.  In that case
1053         * bio completion will call aio_complete.  The only time it's safe to
1054         * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
1055         * This had *better* be the only place that raises -EIOCBQUEUED.
1056         */
1057        BUG_ON(ret == -EIOCBQUEUED);
1058        if (dio->is_async && ret == 0 && dio->result &&
1059            ((rw & READ) || (dio->result == dio->size)))
1060                ret = -EIOCBQUEUED;
1061
1062        if (ret != -EIOCBQUEUED)
1063                dio_await_completion(dio);
1064
1065        /*
1066         * Sync will always be dropping the final ref and completing the
1067         * operation.  AIO can if it was a broken operation described above or
1068         * in fact if all the bios race to complete before we get here.  In
1069         * that case dio_complete() translates the EIOCBQUEUED into the proper
1070         * return code that the caller will hand to aio_complete().
1071         *
1072         * This is managed by the bio_lock instead of being an atomic_t so that
1073         * completion paths can drop their ref and use the remaining count to
1074         * decide to wake the submission path atomically.
1075         */
1076        spin_lock_irqsave(&dio->bio_lock, flags);
1077        ret2 = --dio->refcount;
1078        spin_unlock_irqrestore(&dio->bio_lock, flags);
1079
1080        if (ret2 == 0) {
1081                ret = dio_complete(dio, offset, ret);
1082                kfree(dio);
1083        } else
1084                BUG_ON(ret != -EIOCBQUEUED);
1085
1086        return ret;
1087}
1088
1089/*
1090 * This is a library function for use by filesystem drivers.
1091 * The locking rules are governed by the dio_lock_type parameter.
1092 *
1093 * DIO_NO_LOCKING (no locking, for raw block device access)
1094 * For writes, i_mutex is not held on entry; it is never taken.
1095 *
1096 * DIO_LOCKING (simple locking for regular files)
1097 * For writes we are called under i_mutex and return with i_mutex held, even
1098 * though it is internally dropped.
1099 * For reads, i_mutex is not held on entry, but it is taken and dropped before
1100 * returning.
1101 *
1102 * DIO_OWN_LOCKING (filesystem provides synchronisation and handling of
1103 *      uninitialised data, allowing parallel direct readers and writers)
1104 * For writes we are called without i_mutex, return without it, never touch it.
1105 * For reads we are called under i_mutex and return with i_mutex held, even
1106 * though it may be internally dropped.
1107 *
1108 * Additional i_alloc_sem locking requirements described inline below.
1109 */
1110ssize_t
1111__blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1112        struct block_device *bdev, const struct iovec *iov, loff_t offset, 
1113        unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
1114        int dio_lock_type)
1115{
1116        int seg;
1117        size_t size;
1118        unsigned long addr;
1119        unsigned blkbits = inode->i_blkbits;
1120        unsigned bdev_blkbits = 0;
1121        unsigned blocksize_mask = (1 << blkbits) - 1;
1122        ssize_t retval = -EINVAL;
1123        loff_t end = offset;
1124        struct dio *dio;
1125        int release_i_mutex = 0;
1126        int acquire_i_mutex = 0;
1127
1128        if (rw & WRITE)
1129                rw = WRITE_SYNC;
1130
1131        if (bdev)
1132                bdev_blkbits = blksize_bits(bdev_hardsect_size(bdev));
1133
1134        if (offset & blocksize_mask) {
1135                if (bdev)
1136                         blkbits = bdev_blkbits;
1137                blocksize_mask = (1 << blkbits) - 1;
1138                if (offset & blocksize_mask)
1139                        goto out;
1140        }
1141
1142        /* Check the memory alignment.  Blocks cannot straddle pages */
1143        for (seg = 0; seg < nr_segs; seg++) {
1144                addr = (unsigned long)iov[seg].iov_base;
1145                size = iov[seg].iov_len;
1146                end += size;
1147                if ((addr & blocksize_mask) || (size & blocksize_mask))  {
1148                        if (bdev)
1149                                 blkbits = bdev_blkbits;
1150                        blocksize_mask = (1 << blkbits) - 1;
1151                        if ((addr & blocksize_mask) || (size & blocksize_mask))  
1152                                goto out;
1153                }
1154        }
1155
1156        dio = kzalloc(sizeof(*dio), GFP_KERNEL);
1157        retval = -ENOMEM;
1158        if (!dio)
1159                goto out;
1160
1161        /*
1162         * For block device access DIO_NO_LOCKING is used,
1163         *      neither readers nor writers do any locking at all
1164         * For regular files using DIO_LOCKING,
1165         *      readers need to grab i_mutex and i_alloc_sem
1166         *      writers need to grab i_alloc_sem only (i_mutex is already held)
1167         * For regular files using DIO_OWN_LOCKING,
1168         *      neither readers nor writers take any locks here
1169         */
1170        dio->lock_type = dio_lock_type;
1171        if (dio_lock_type != DIO_NO_LOCKING) {
1172                /* watch out for a 0 len io from a tricksy fs */
1173                if (rw == READ && end > offset) {
1174                        struct address_space *mapping;
1175
1176                        mapping = iocb->ki_filp->f_mapping;
1177                        if (dio_lock_type != DIO_OWN_LOCKING) {
1178                                mutex_lock(&inode->i_mutex);
1179                                release_i_mutex = 1;
1180                        }
1181
1182                        retval = filemap_write_and_wait_range(mapping, offset,
1183                                                              end - 1);
1184                        if (retval) {
1185                                kfree(dio);
1186                                goto out;
1187                        }
1188
1189                        if (dio_lock_type == DIO_OWN_LOCKING) {
1190                                mutex_unlock(&inode->i_mutex);
1191                                acquire_i_mutex = 1;
1192                        }
1193                }
1194
1195                if (dio_lock_type == DIO_LOCKING)
1196                        /* lockdep: not the owner will release it */
1197                        down_read_non_owner(&inode->i_alloc_sem);
1198        }
1199
1200        /*
1201         * For file extending writes updating i_size before data
1202         * writeouts complete can expose uninitialized blocks. So
1203         * even for AIO, we need to wait for i/o to complete before
1204         * returning in this case.
1205         */
1206        dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
1207                (end > i_size_read(inode)));
1208
1209        retval = direct_io_worker(rw, iocb, inode, iov, offset,
1210                                nr_segs, blkbits, get_block, end_io, dio);
1211
1212        if (rw == READ && dio_lock_type == DIO_LOCKING)
1213                release_i_mutex = 0;
1214
1215out:
1216        if (release_i_mutex)
1217                mutex_unlock(&inode->i_mutex);
1218        else if (acquire_i_mutex)
1219                mutex_lock(&inode->i_mutex);
1220        return retval;
1221}
1222EXPORT_SYMBOL(__blockdev_direct_IO);
1223