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