fs/direct-io.c

   1 /*
   2  * fs/direct-io.c
   3  *
   4  * Copyright (C) 2002, Linus Torvalds.
   5  *
   6  * O_DIRECT
   7  *
   8  * 04Jul2002    akpm@zip.com.au
   9  *              Initial version
  10  * 11Sep2002    janetinc@us.ibm.com
  11  *              added readv/writev support.
  12  * 29Oct2002    akpm@zip.com.au
  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/bio.h>
  31 #include <linux/wait.h>
  32 #include <linux/err.h>
  33 #include <linux/blkdev.h>
  34 #include <linux/buffer_head.h>
  35 #include <linux/rwsem.h>
  36 #include <linux/uio.h>
  37 #include <asm/atomic.h>
  38
  39 /*
  40  * How many user pages to map in one call to get_user_pages().  This determines
  41  * the size of a structure on the stack.
  42  */
  43 #define DIO_PAGES       64
  44
  45 /*
  46  * This code generally works in units of "dio_blocks".  A dio_block is
  47  * somewhere between the hard sector size and the filesystem block size.  it
  48  * is determined on a per-invocation basis.   When talking to the filesystem
  49  * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
  50  * down by dio->blkfactor.  Similarly, fs-blocksize quantities are converted
  51  * to bio_block quantities by shifting left by blkfactor.
  52  *
  53  * If blkfactor is zero then the user's request was aligned to the filesystem's
  54  * blocksize.
  55  *
  56  * lock_type is DIO_LOCKING for regular files on direct-IO-naive filesystems.
  57  * This determines whether we need to do the fancy locking which prevents
  58  * direct-IO from being able to read uninitialised disk blocks.  If its zero
  59  * (blockdev) this locking is not done, and if it is DIO_OWN_LOCKING i_sem is
  60  * not held for the entire direct write (taken briefly, initially, during a
  61  * direct read though, but its never held for the duration of a direct-IO).
  62  */
  63
  64 struct dio {
  65         /* BIO submission state */
  66         struct bio *bio;                /* bio under assembly */
  67         struct inode *inode;
  68         int rw;
  69         int lock_type;                  /* doesn't change */
  70         unsigned blkbits;               /* doesn't change */
  71         unsigned blkfactor;             /* When we're using an alignment which
  72                                            is finer than the filesystem's soft
  73                                            blocksize, this specifies how much
  74                                            finer.  blkfactor=2 means 1/4-block
  75                                            alignment.  Does not change */
  76         unsigned start_zero_done;       /* flag: sub-blocksize zeroing has
  77                                            been performed at the start of a
  78                                            write */
  79         int pages_in_io;                /* approximate total IO pages */
  80         size_t  size;                   /* total request size (doesn't change)*/
  81         sector_t block_in_file;         /* Current offset into the underlying
  82                                            file in dio_block units. */
  83         unsigned blocks_available;      /* At block_in_file.  changes */
  84         sector_t final_block_in_request;/* doesn't change */
  85         unsigned first_block_in_page;   /* doesn't change, Used only once */
  86         int boundary;                   /* prev block is at a boundary */
  87         int reap_counter;               /* rate limit reaping */
  88         get_blocks_t *get_blocks;       /* block mapping function */
  89         dio_iodone_t *end_io;           /* IO completion function */
  90         sector_t final_block_in_bio;    /* current final block in bio + 1 */
  91         sector_t next_block_for_io;     /* next block to be put under IO,
  92                                            in dio_blocks units */
  93         struct buffer_head map_bh;      /* last get_blocks() result */
  94
  95         /*
  96          * Deferred addition of a page to the dio.  These variables are
  97          * private to dio_send_cur_page(), submit_page_section() and
  98          * dio_bio_add_page().
  99          */
 100         struct page *cur_page;          /* The page */
 101         unsigned cur_page_offset;       /* Offset into it, in bytes */
 102         unsigned cur_page_len;          /* Nr of bytes at cur_page_offset */
 103         sector_t cur_page_block;        /* Where it starts */
 104
 105         /*
 106          * Page fetching state. These variables belong to dio_refill_pages().
 107          */
 108         int curr_page;                  /* changes */
 109         int total_pages;                /* doesn't change */
 110         unsigned long curr_user_address;/* changes */
 111
 112         /*
 113          * Page queue.  These variables belong to dio_refill_pages() and
 114          * dio_get_page().
 115          */
 116         struct page *pages[DIO_PAGES];  /* page buffer */
 117         unsigned head;                  /* next page to process */
 118         unsigned tail;                  /* last valid page + 1 */
 119         int page_errors;                /* errno from get_user_pages() */
 120
 121         /* BIO completion state */
 122         spinlock_t bio_lock;            /* protects BIO fields below */
 123         int bio_count;                  /* nr bios to be completed */
 124         int bios_in_flight;             /* nr bios in flight */
 125         struct bio *bio_list;           /* singly linked via bi_private */
 126         struct task_struct *waiter;     /* waiting task (NULL if none) */
 127
 128         /* AIO related stuff */
 129         struct kiocb *iocb;             /* kiocb */
 130         int is_async;                   /* is IO async ? */
 131         ssize_t result;                 /* IO result */
 132 };
 133
 134 /*
 135  * How many pages are in the queue?
 136  */
 137 static inline unsigned dio_pages_present(struct dio *dio)
 138 {
 139         return dio->tail - dio->head;
 140 }
 141
 142 /*
 143  * Go grab and pin some userspace pages.   Typically we'll get 64 at a time.
 144  */
 145 static int dio_refill_pages(struct dio *dio)
 146 {
 147         int ret;
 148         int nr_pages;
 149
 150         nr_pages = min(dio->total_pages - dio->curr_page, DIO_PAGES);
 151         down_read(&current->mm->mmap_sem);
 152         ret = get_user_pages(
 153                 current,                        /* Task for fault acounting */
 154                 current->mm,                    /* whose pages? */
 155                 dio->curr_user_address,         /* Where from? */
 156                 nr_pages,                       /* How many pages? */
 157                 dio->rw == READ,                /* Write to memory? */
 158                 0,                              /* force (?) */
 159                 &dio->pages[0],
 160                 NULL);                          /* vmas */
 161         up_read(&current->mm->mmap_sem);
 162
 163         if (ret < 0 && dio->blocks_available && (dio->rw == WRITE)) {
 164                 /*
 165                  * A memory fault, but the filesystem has some outstanding
 166                  * mapped blocks.  We need to use those blocks up to avoid
 167                  * leaking stale data in the file.
 168                  */
 169                 if (dio->page_errors == 0)
 170                         dio->page_errors = ret;
 171                 dio->pages[0] = ZERO_PAGE(dio->curr_user_address);
 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         }
 185 out:
 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  */
 195 static 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  * Called when all DIO BIO I/O has been completed - let the filesystem
 210  * know, if it registered an interest earlier via get_blocks.  Pass the
 211  * private field of the map buffer_head so that filesystems can use it
 212  * to hold additional state between get_blocks calls and dio_complete.
 213  */
 214 static void dio_complete(struct dio *dio, loff_t offset, ssize_t bytes)
 215 {
 216         if (dio->end_io && dio->result)
 217                 dio->end_io(dio->inode, offset, bytes, dio->map_bh.b_private);
 218         if (dio->lock_type != DIO_NO_LOCKING)
 219                 up_read(&dio->inode->i_alloc_sem);
 220 }
 221
 222 /*
 223  * Called when a BIO has been processed.  If the count goes to zero then IO is
 224  * complete and we can signal this to the AIO layer.
 225  */
 226 static void finished_one_bio(struct dio *dio)
 227 {
 228         unsigned long flags;
 229
 230         spin_lock_irqsave(&dio->bio_lock, flags);
 231         if (dio->bio_count == 1) {
 232                 if (dio->is_async) {
 233                         /*
 234                          * Last reference to the dio is going away.
 235                          * Drop spinlock and complete the DIO.
 236                          */
 237                         spin_unlock_irqrestore(&dio->bio_lock, flags);
 238                         dio_complete(dio, dio->block_in_file << dio->blkbits,
 239                                         dio->result);
 240                         /* Complete AIO later if falling back to buffered i/o */
 241                         if (dio->result == dio->size ||
 242                                 ((dio->rw == READ) && dio->result)) {
 243                                 aio_complete(dio->iocb, dio->result, 0);
 244                                 kfree(dio);
 245                                 return;
 246                         } else {
 247                                 /*
 248                                  * Falling back to buffered
 249                                  */
 250                                 spin_lock_irqsave(&dio->bio_lock, flags);
 251                                 dio->bio_count--;
 252                                 if (dio->waiter)
 253                                         wake_up_process(dio->waiter);
 254                                 spin_unlock_irqrestore(&dio->bio_lock, flags);
 255                                 return;
 256                         }
 257                 }
 258         }
 259         dio->bio_count--;
 260         spin_unlock_irqrestore(&dio->bio_lock, flags);
 261 }
 262
 263 static int dio_bio_complete(struct dio *dio, struct bio *bio);
 264 /*
 265  * Asynchronous IO callback.
 266  */
 267 static int dio_bio_end_aio(struct bio *bio, unsigned int bytes_done, int error)
 268 {
 269         struct dio *dio = bio->bi_private;
 270
 271         if (bio->bi_size)
 272                 return 1;
 273
 274         /* cleanup the bio */
 275         dio_bio_complete(dio, bio);
 276         return 0;
 277 }
 278
 279 /*
 280  * The BIO completion handler simply queues the BIO up for the process-context
 281  * handler.
 282  *
 283  * During I/O bi_private points at the dio.  After I/O, bi_private is used to
 284  * implement a singly-linked list of completed BIOs, at dio->bio_list.
 285  */
 286 static int dio_bio_end_io(struct bio *bio, unsigned int bytes_done, int error)
 287 {
 288         struct dio *dio = bio->bi_private;
 289         unsigned long flags;
 290
 291         if (bio->bi_size)
 292                 return 1;
 293
 294         spin_lock_irqsave(&dio->bio_lock, flags);
 295         bio->bi_private = dio->bio_list;
 296         dio->bio_list = bio;
 297         dio->bios_in_flight--;
 298         if (dio->waiter && dio->bios_in_flight == 0)
 299                 wake_up_process(dio->waiter);
 300         spin_unlock_irqrestore(&dio->bio_lock, flags);
 301         return 0;
 302 }
 303
 304 static int
 305 dio_bio_alloc(struct dio *dio, struct block_device *bdev,
 306                 sector_t first_sector, int nr_vecs)
 307 {
 308         struct bio *bio;
 309
 310         bio = bio_alloc(GFP_KERNEL, nr_vecs);
 311         if (bio == NULL)
 312                 return -ENOMEM;
 313
 314         bio->bi_bdev = bdev;
 315         bio->bi_sector = first_sector;
 316         if (dio->is_async)
 317                 bio->bi_end_io = dio_bio_end_aio;
 318         else
 319                 bio->bi_end_io = dio_bio_end_io;
 320
 321         dio->bio = bio;
 322         return 0;
 323 }
 324
 325 /*
 326  * In the AIO read case we speculatively dirty the pages before starting IO.
 327  * During IO completion, any of these pages which happen to have been written
 328  * back will be redirtied by bio_check_pages_dirty().
 329  */
 330 static void dio_bio_submit(struct dio *dio)
 331 {
 332         struct bio *bio = dio->bio;
 333         unsigned long flags;
 334
 335         bio->bi_private = dio;
 336         spin_lock_irqsave(&dio->bio_lock, flags);
 337         dio->bio_count++;
 338         dio->bios_in_flight++;
 339         spin_unlock_irqrestore(&dio->bio_lock, flags);
 340         if (dio->is_async && dio->rw == READ)
 341                 bio_set_pages_dirty(bio);
 342         submit_bio(dio->rw, bio);
 343
 344         dio->bio = NULL;
 345         dio->boundary = 0;
 346 }
 347
 348 /*
 349  * Release any resources in case of a failure
 350  */
 351 static void dio_cleanup(struct dio *dio)
 352 {
 353         while (dio_pages_present(dio))
 354                 page_cache_release(dio_get_page(dio));
 355 }
 356
 357 /*
 358  * Wait for the next BIO to complete.  Remove it and return it.
 359  */
 360 static struct bio *dio_await_one(struct dio *dio)
 361 {
 362         unsigned long flags;
 363         struct bio *bio;
 364
 365         spin_lock_irqsave(&dio->bio_lock, flags);
 366         while (dio->bio_list == NULL) {
 367                 set_current_state(TASK_UNINTERRUPTIBLE);
 368                 if (dio->bio_list == NULL) {
 369                         dio->waiter = current;
 370                         spin_unlock_irqrestore(&dio->bio_lock, flags);
 371                         blk_run_address_space(dio->inode->i_mapping);
 372                         io_schedule();
 373                         spin_lock_irqsave(&dio->bio_lock, flags);
 374                         dio->waiter = NULL;
 375                 }
 376                 set_current_state(TASK_RUNNING);
 377         }
 378         bio = dio->bio_list;
 379         dio->bio_list = bio->bi_private;
 380         spin_unlock_irqrestore(&dio->bio_lock, flags);
 381         return bio;
 382 }
 383
 384 /*
 385  * Process one completed BIO.  No locks are held.
 386  */
 387 static int dio_bio_complete(struct dio *dio, struct bio *bio)
 388 {
 389         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 390         struct bio_vec *bvec = bio->bi_io_vec;
 391         int page_no;
 392
 393         if (!uptodate)
 394                 dio->result = -EIO;
 395
 396         if (dio->is_async && dio->rw == READ) {
 397                 bio_check_pages_dirty(bio);     /* transfers ownership */
 398         } else {
 399                 for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
 400                         struct page *page = bvec[page_no].bv_page;
 401
 402                         if (dio->rw == READ && !PageCompound(page))
 403                                 set_page_dirty_lock(page);
 404                         page_cache_release(page);
 405                 }
 406                 bio_put(bio);
 407         }
 408         finished_one_bio(dio);
 409         return uptodate ? 0 : -EIO;
 410 }
 411
 412 /*
 413  * Wait on and process all in-flight BIOs.
 414  */
 415 static int dio_await_completion(struct dio *dio)
 416 {
 417         int ret = 0;
 418
 419         if (dio->bio)
 420                 dio_bio_submit(dio);
 421
 422         /*
 423          * The bio_lock is not held for the read of bio_count.
 424          * This is ok since it is the dio_bio_complete() that changes
 425          * bio_count.
 426          */
 427         while (dio->bio_count) {
 428                 struct bio *bio = dio_await_one(dio);
 429                 int ret2;
 430
 431                 ret2 = dio_bio_complete(dio, bio);
 432                 if (ret == 0)
 433                         ret = ret2;
 434         }
 435         return ret;
 436 }
 437
 438 /*
 439  * A really large O_DIRECT read or write can generate a lot of BIOs.  So
 440  * to keep the memory consumption sane we periodically reap any completed BIOs
 441  * during the BIO generation phase.
 442  *
 443  * This also helps to limit the peak amount of pinned userspace memory.
 444  */
 445 static int dio_bio_reap(struct dio *dio)
 446 {
 447         int ret = 0;
 448
 449         if (dio->reap_counter++ >= 64) {
 450                 while (dio->bio_list) {
 451                         unsigned long flags;
 452                         struct bio *bio;
 453                         int ret2;
 454
 455                         spin_lock_irqsave(&dio->bio_lock, flags);
 456                         bio = dio->bio_list;
 457                         dio->bio_list = bio->bi_private;
 458                         spin_unlock_irqrestore(&dio->bio_lock, flags);
 459                         ret2 = dio_bio_complete(dio, bio);
 460                         if (ret == 0)
 461                                 ret = ret2;
 462                 }
 463                 dio->reap_counter = 0;
 464         }
 465         return ret;
 466 }
 467
 468 /*
 469  * Call into the fs to map some more disk blocks.  We record the current number
 470  * of available blocks at dio->blocks_available.  These are in units of the
 471  * fs blocksize, (1 << inode->i_blkbits).
 472  *
 473  * The fs is allowed to map lots of blocks at once.  If it wants to do that,
 474  * it uses the passed inode-relative block number as the file offset, as usual.
 475  *
 476  * get_blocks() is passed the number of i_blkbits-sized blocks which direct_io
 477  * has remaining to do.  The fs should not map more than this number of blocks.
 478  *
 479  * If the fs has mapped a lot of blocks, it should populate bh->b_size to
 480  * indicate how much contiguous disk space has been made available at
 481  * bh->b_blocknr.
 482  *
 483  * If *any* of the mapped blocks are new, then the fs must set buffer_new().
 484  * This isn't very efficient...
 485  *
 486  * In the case of filesystem holes: the fs may return an arbitrarily-large
 487  * hole by returning an appropriate value in b_size and by clearing
 488  * buffer_mapped().  However the direct-io code will only process holes one
 489  * block at a time - it will repeatedly call get_blocks() as it walks the hole.
 490  */
 491 static int get_more_blocks(struct dio *dio)
 492 {
 493         int ret;
 494         struct buffer_head *map_bh = &dio->map_bh;
 495         sector_t fs_startblk;   /* Into file, in filesystem-sized blocks */
 496         unsigned long fs_count; /* Number of filesystem-sized blocks */
 497         unsigned long dio_count;/* Number of dio_block-sized blocks */
 498         unsigned long blkmask;
 499         int create;
 500
 501         /*
 502          * If there was a memory error and we've overwritten all the
 503          * mapped blocks then we can now return that memory error
 504          */
 505         ret = dio->page_errors;
 506         if (ret == 0) {
 507                 map_bh->b_state = 0;
 508                 map_bh->b_size = 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                 create = dio->rw == WRITE;
 518                 if (dio->lock_type == DIO_LOCKING) {
 519                         if (dio->block_in_file < (i_size_read(dio->inode) >>
 520                                                         dio->blkbits))
 521                                 create = 0;
 522                 } else if (dio->lock_type == DIO_NO_LOCKING) {
 523                         create = 0;
 524                 }
 525                 /*
 526                  * For writes inside i_size we forbid block creations: only
 527                  * overwrites are permitted.  We fall back to buffered writes
 528                  * at a higher level for inside-i_size block-instantiating
 529                  * writes.
 530                  */
 531                 ret = (*dio->get_blocks)(dio->inode, fs_startblk, fs_count,
 532                                                 map_bh, create);
 533         }
 534         return ret;
 535 }
 536
 537 /*
 538  * There is no bio.  Make one now.
 539  */
 540 static int dio_new_bio(struct dio *dio, sector_t start_sector)
 541 {
 542         sector_t sector;
 543         int ret, nr_pages;
 544
 545         ret = dio_bio_reap(dio);
 546         if (ret)
 547                 goto out;
 548         sector = start_sector << (dio->blkbits - 9);
 549         nr_pages = min(dio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev));
 550         BUG_ON(nr_pages <= 0);
 551         ret = dio_bio_alloc(dio, dio->map_bh.b_bdev, sector, nr_pages);
 552         dio->boundary = 0;
 553 out:
 554         return ret;
 555 }
 556
 557 /*
 558  * Attempt to put the current chunk of 'cur_page' into the current BIO.  If
 559  * that was successful then update final_block_in_bio and take a ref against
 560  * the just-added page.
 561  *
 562  * Return zero on success.  Non-zero means the caller needs to start a new BIO.
 563  */
 564 static int dio_bio_add_page(struct dio *dio)
 565 {
 566         int ret;
 567
 568         ret = bio_add_page(dio->bio, dio->cur_page,
 569                         dio->cur_page_len, dio->cur_page_offset);
 570         if (ret == dio->cur_page_len) {
 571                 /*
 572                  * Decrement count only, if we are done with this page
 573                  */
 574                 if ((dio->cur_page_len + dio->cur_page_offset) == PAGE_SIZE)
 575                         dio->pages_in_io--;
 576                 page_cache_get(dio->cur_page);
 577                 dio->final_block_in_bio = dio->cur_page_block +
 578                         (dio->cur_page_len >> dio->blkbits);
 579                 ret = 0;
 580         } else {
 581                 ret = 1;
 582         }
 583         return ret;
 584 }
 585
 586 /*
 587  * Put cur_page under IO.  The section of cur_page which is described by
 588  * cur_page_offset,cur_page_len is put into a BIO.  The section of cur_page
 589  * starts on-disk at cur_page_block.
 590  *
 591  * We take a ref against the page here (on behalf of its presence in the bio).
 592  *
 593  * The caller of this function is responsible for removing cur_page from the
 594  * dio, and for dropping the refcount which came from that presence.
 595  */
 596 static int dio_send_cur_page(struct dio *dio)
 597 {
 598         int ret = 0;
 599
 600         if (dio->bio) {
 601                 /*
 602                  * See whether this new request is contiguous with the old
 603                  */
 604                 if (dio->final_block_in_bio != dio->cur_page_block)
 605                         dio_bio_submit(dio);
 606                 /*
 607                  * Submit now if the underlying fs is about to perform a
 608                  * metadata read
 609                  */
 610                 if (dio->boundary)
 611                         dio_bio_submit(dio);
 612         }
 613
 614         if (dio->bio == NULL) {
 615                 ret = dio_new_bio(dio, dio->cur_page_block);
 616                 if (ret)
 617                         goto out;
 618         }
 619
 620         if (dio_bio_add_page(dio) != 0) {
 621                 dio_bio_submit(dio);
 622                 ret = dio_new_bio(dio, dio->cur_page_block);
 623                 if (ret == 0) {
 624                         ret = dio_bio_add_page(dio);
 625                         BUG_ON(ret != 0);
 626                 }
 627         }
 628 out:
 629         return ret;
 630 }
 631
 632 /*
 633  * An autonomous function to put a chunk of a page under deferred IO.
 634  *
 635  * The caller doesn't actually know (or care) whether this piece of page is in
 636  * a BIO, or is under IO or whatever.  We just take care of all possible
 637  * situations here.  The separation between the logic of do_direct_IO() and
 638  * that of submit_page_section() is important for clarity.  Please don't break.
 639  *
 640  * The chunk of page starts on-disk at blocknr.
 641  *
 642  * We perform deferred IO, by recording the last-submitted page inside our
 643  * private part of the dio structure.  If possible, we just expand the IO
 644  * across that page here.
 645  *
 646  * If that doesn't work out then we put the old page into the bio and add this
 647  * page to the dio instead.
 648  */
 649 static int
 650 submit_page_section(struct dio *dio, struct page *page,
 651                 unsigned offset, unsigned len, sector_t blocknr)
 652 {
 653         int ret = 0;
 654
 655         /*
 656          * Can we just grow the current page's presence in the dio?
 657          */
 658         if (    (dio->cur_page == page) &&
 659                 (dio->cur_page_offset + dio->cur_page_len == offset) &&
 660                 (dio->cur_page_block +
 661                         (dio->cur_page_len >> dio->blkbits) == blocknr)) {
 662                 dio->cur_page_len += len;
 663
 664                 /*
 665                  * If dio->boundary then we want to schedule the IO now to
 666                  * avoid metadata seeks.
 667                  */
 668                 if (dio->boundary) {
 669                         ret = dio_send_cur_page(dio);
 670                         page_cache_release(dio->cur_page);
 671                         dio->cur_page = NULL;
 672                 }
 673                 goto out;
 674         }
 675
 676         /*
 677          * If there's a deferred page already there then send it.
 678          */
 679         if (dio->cur_page) {
 680                 ret = dio_send_cur_page(dio);
 681                 page_cache_release(dio->cur_page);
 682                 dio->cur_page = NULL;
 683                 if (ret)
 684                         goto out;
 685         }
 686
 687         page_cache_get(page);           /* It is in dio */
 688         dio->cur_page = page;
 689         dio->cur_page_offset = offset;
 690         dio->cur_page_len = len;
 691         dio->cur_page_block = blocknr;
 692 out:
 693         return ret;
 694 }
 695
 696 /*
 697  * Clean any dirty buffers in the blockdev mapping which alias newly-created
 698  * file blocks.  Only called for S_ISREG files - blockdevs do not set
 699  * buffer_new
 700  */
 701 static void clean_blockdev_aliases(struct dio *dio)
 702 {
 703         unsigned i;
 704         unsigned nblocks;
 705
 706         nblocks = dio->map_bh.b_size >> dio->inode->i_blkbits;
 707
 708         for (i = 0; i < nblocks; i++) {
 709                 unmap_underlying_metadata(dio->map_bh.b_bdev,
 710                                         dio->map_bh.b_blocknr + i);
 711         }
 712 }
 713
 714 /*
 715  * If we are not writing the entire block and get_block() allocated
 716  * the block for us, we need to fill-in the unused portion of the
 717  * block with zeros. This happens only if user-buffer, fileoffset or
 718  * io length is not filesystem block-size multiple.
 719  *
 720  * `end' is zero if we're doing the start of the IO, 1 at the end of the
 721  * IO.
 722  */
 723 static void dio_zero_block(struct dio *dio, int end)
 724 {
 725         unsigned dio_blocks_per_fs_block;
 726         unsigned this_chunk_blocks;     /* In dio_blocks */
 727         unsigned this_chunk_bytes;
 728         struct page *page;
 729
 730         dio->start_zero_done = 1;
 731         if (!dio->blkfactor || !buffer_new(&dio->map_bh))
 732                 return;
 733
 734         dio_blocks_per_fs_block = 1 << dio->blkfactor;
 735         this_chunk_blocks = dio->block_in_file & (dio_blocks_per_fs_block - 1);
 736
 737         if (!this_chunk_blocks)
 738                 return;
 739
 740         /*
 741          * We need to zero out part of an fs block.  It is either at the
 742          * beginning or the end of the fs block.
 743          */
 744         if (end)
 745                 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
 746
 747         this_chunk_bytes = this_chunk_blocks << dio->blkbits;
 748
 749         page = ZERO_PAGE(dio->curr_user_address);
 750         if (submit_page_section(dio, page, 0, this_chunk_bytes,
 751                                 dio->next_block_for_io))
 752                 return;
 753
 754         dio->next_block_for_io += this_chunk_blocks;
 755 }
 756
 757 /*
 758  * Walk the user pages, and the file, mapping blocks to disk and generating
 759  * a sequence of (page,offset,len,block) mappings.  These mappings are injected
 760  * into submit_page_section(), which takes care of the next stage of submission
 761  *
 762  * Direct IO against a blockdev is different from a file.  Because we can
 763  * happily perform page-sized but 512-byte aligned IOs.  It is important that
 764  * blockdev IO be able to have fine alignment and large sizes.
 765  *
 766  * So what we do is to permit the ->get_blocks function to populate bh.b_size
 767  * with the size of IO which is permitted at this offset and this i_blkbits.
 768  *
 769  * For best results, the blockdev should be set up with 512-byte i_blkbits and
 770  * it should set b_size to PAGE_SIZE or more inside get_blocks().  This gives
 771  * fine alignment but still allows this function to work in PAGE_SIZE units.
 772  */
 773 static int do_direct_IO(struct dio *dio)
 774 {
 775         const unsigned blkbits = dio->blkbits;
 776         const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
 777         struct page *page;
 778         unsigned block_in_page;
 779         struct buffer_head *map_bh = &dio->map_bh;
 780         int ret = 0;
 781
 782         /* The I/O can start at any block offset within the first page */
 783         block_in_page = dio->first_block_in_page;
 784
 785         while (dio->block_in_file < dio->final_block_in_request) {
 786                 page = dio_get_page(dio);
 787                 if (IS_ERR(page)) {
 788                         ret = PTR_ERR(page);
 789                         goto out;
 790                 }
 791
 792                 while (block_in_page < blocks_per_page) {
 793                         unsigned offset_in_page = block_in_page << blkbits;
 794                         unsigned this_chunk_bytes;      /* # of bytes mapped */
 795                         unsigned this_chunk_blocks;     /* # of blocks */
 796                         unsigned u;
 797
 798                         if (dio->blocks_available == 0) {
 799                                 /*
 800                                  * Need to go and map some more disk
 801                                  */
 802                                 unsigned long blkmask;
 803                                 unsigned long dio_remainder;
 804
 805                                 ret = get_more_blocks(dio);
 806                                 if (ret) {
 807                                         page_cache_release(page);
 808                                         goto out;
 809                                 }
 810                                 if (!buffer_mapped(map_bh))
 811                                         goto do_holes;
 812
 813                                 dio->blocks_available =
 814                                                 map_bh->b_size >> dio->blkbits;
 815                                 dio->next_block_for_io =
 816                                         map_bh->b_blocknr << dio->blkfactor;
 817                                 if (buffer_new(map_bh))
 818                                         clean_blockdev_aliases(dio);
 819
 820                                 if (!dio->blkfactor)
 821                                         goto do_holes;
 822
 823                                 blkmask = (1 << dio->blkfactor) - 1;
 824                                 dio_remainder = (dio->block_in_file & blkmask);
 825
 826                                 /*
 827                                  * If we are at the start of IO and that IO
 828                                  * starts partway into a fs-block,
 829                                  * dio_remainder will be non-zero.  If the IO
 830                                  * is a read then we can simply advance the IO
 831                                  * cursor to the first block which is to be
 832                                  * read.  But if the IO is a write and the
 833                                  * block was newly allocated we cannot do that;
 834                                  * the start of the fs block must be zeroed out
 835                                  * on-disk
 836                                  */
 837                                 if (!buffer_new(map_bh))
 838                                         dio->next_block_for_io += dio_remainder;
 839                                 dio->blocks_available -= dio_remainder;
 840                         }
 841 do_holes:
 842                         /* Handle holes */
 843                         if (!buffer_mapped(map_bh)) {
 844                                 char *kaddr;
 845
 846                                 /* AKPM: eargh, -ENOTBLK is a hack */
 847                                 if (dio->rw == WRITE)
 848                                         return -ENOTBLK;
 849
 850                                 if (dio->block_in_file >=
 851                                         i_size_read(dio->inode)>>blkbits) {
 852                                         /* We hit eof */
 853                                         page_cache_release(page);
 854                                         goto out;
 855                                 }
 856                                 kaddr = kmap_atomic(page, KM_USER0);
 857                                 memset(kaddr + (block_in_page << blkbits),
 858                                                 0, 1 << blkbits);
 859                                 flush_dcache_page(page);
 860                                 kunmap_atomic(kaddr, KM_USER0);
 861                                 dio->block_in_file++;
 862                                 block_in_page++;
 863                                 goto next_block;
 864                         }
 865
 866                         /*
 867                          * If we're performing IO which has an alignment which
 868                          * is finer than the underlying fs, go check to see if
 869                          * we must zero out the start of this block.
 870                          */
 871                         if (unlikely(dio->blkfactor && !dio->start_zero_done))
 872                                 dio_zero_block(dio, 0);
 873
 874                         /*
 875                          * Work out, in this_chunk_blocks, how much disk we
 876                          * can add to this page
 877                          */
 878                         this_chunk_blocks = dio->blocks_available;
 879                         u = (PAGE_SIZE - offset_in_page) >> blkbits;
 880                         if (this_chunk_blocks > u)
 881                                 this_chunk_blocks = u;
 882                         u = dio->final_block_in_request - dio->block_in_file;
 883                         if (this_chunk_blocks > u)
 884                                 this_chunk_blocks = u;
 885                         this_chunk_bytes = this_chunk_blocks << blkbits;
 886                         BUG_ON(this_chunk_bytes == 0);
 887
 888                         dio->boundary = buffer_boundary(map_bh);
 889                         ret = submit_page_section(dio, page, offset_in_page,
 890                                 this_chunk_bytes, dio->next_block_for_io);
 891                         if (ret) {
 892                                 page_cache_release(page);
 893                                 goto out;
 894                         }
 895                         dio->next_block_for_io += this_chunk_blocks;
 896
 897                         dio->block_in_file += this_chunk_blocks;
 898                         block_in_page += this_chunk_blocks;
 899                         dio->blocks_available -= this_chunk_blocks;
 900 next_block:
 901                         if (dio->block_in_file > dio->final_block_in_request)
 902                                 BUG();
 903                         if (dio->block_in_file == dio->final_block_in_request)
 904                                 break;
 905                 }
 906
 907                 /* Drop the ref which was taken in get_user_pages() */
 908                 page_cache_release(page);
 909                 block_in_page = 0;
 910         }
 911 out:
 912         return ret;
 913 }
 914
 915 /*
 916  * Releases both i_sem and i_alloc_sem
 917  */
 918 static ssize_t
 919 direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
 920         const struct iovec *iov, loff_t offset, unsigned long nr_segs,
 921         unsigned blkbits, get_blocks_t get_blocks, dio_iodone_t end_io,
 922         struct dio *dio)
 923 {
 924         unsigned long user_addr;
 925         int seg;
 926         ssize_t ret = 0;
 927         ssize_t ret2;
 928         size_t bytes;
 929
 930         dio->bio = NULL;
 931         dio->inode = inode;
 932         dio->rw = rw;
 933         dio->blkbits = blkbits;
 934         dio->blkfactor = inode->i_blkbits - blkbits;
 935         dio->start_zero_done = 0;
 936         dio->size = 0;
 937         dio->block_in_file = offset >> blkbits;
 938         dio->blocks_available = 0;
 939         dio->cur_page = NULL;
 940
 941         dio->boundary = 0;
 942         dio->reap_counter = 0;
 943         dio->get_blocks = get_blocks;
 944         dio->end_io = end_io;
 945         dio->map_bh.b_private = NULL;
 946         dio->final_block_in_bio = -1;
 947         dio->next_block_for_io = -1;
 948
 949         dio->page_errors = 0;
 950         dio->result = 0;
 951         dio->iocb = iocb;
 952
 953         /*
 954          * BIO completion state.
 955          *
 956          * ->bio_count starts out at one, and we decrement it to zero after all
 957          * BIOs are submitted.  This to avoid the situation where a really fast
 958          * (or synchronous) device could take the count to zero while we're
 959          * still submitting BIOs.
 960          */
 961         dio->bio_count = 1;
 962         dio->bios_in_flight = 0;
 963         spin_lock_init(&dio->bio_lock);
 964         dio->bio_list = NULL;
 965         dio->waiter = NULL;
 966
 967         /*
 968          * In case of non-aligned buffers, we may need 2 more
 969          * pages since we need to zero out first and last block.
 970          */
 971         if (unlikely(dio->blkfactor))
 972                 dio->pages_in_io = 2;
 973         else
 974                 dio->pages_in_io = 0;
 975
 976         for (seg = 0; seg < nr_segs; seg++) {
 977                 user_addr = (unsigned long)iov[seg].iov_base;
 978                 dio->pages_in_io +=
 979                         ((user_addr+iov[seg].iov_len +PAGE_SIZE-1)/PAGE_SIZE
 980                                 - user_addr/PAGE_SIZE);
 981         }
 982
 983         for (seg = 0; seg < nr_segs; seg++) {
 984                 user_addr = (unsigned long)iov[seg].iov_base;
 985                 dio->size += bytes = iov[seg].iov_len;
 986
 987                 /* Index into the first page of the first block */
 988                 dio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
 989                 dio->final_block_in_request = dio->block_in_file +
 990                                                 (bytes >> blkbits);
 991                 /* Page fetching state */
 992                 dio->head = 0;
 993                 dio->tail = 0;
 994                 dio->curr_page = 0;
 995
 996                 dio->total_pages = 0;
 997                 if (user_addr & (PAGE_SIZE-1)) {
 998                         dio->total_pages++;
 999                         bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1000                 }
1001                 dio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1002                 dio->curr_user_address = user_addr;
1003
1004                 ret = do_direct_IO(dio);
1005
1006                 dio->result += iov[seg].iov_len -
1007                         ((dio->final_block_in_request - dio->block_in_file) <<
1008                                         blkbits);
1009
1010                 if (ret) {
1011                         dio_cleanup(dio);
1012                         break;
1013                 }
1014         } /* end iovec loop */
1015
1016         if (ret == -ENOTBLK && rw == WRITE) {
1017                 /*
1018                  * The remaining part of the request will be
1019                  * be handled by buffered I/O when we return
1020                  */
1021                 ret = 0;
1022         }
1023         /*
1024          * There may be some unwritten disk at the end of a part-written
1025          * fs-block-sized block.  Go zero that now.
1026          */
1027         dio_zero_block(dio, 1);
1028
1029         if (dio->cur_page) {
1030                 ret2 = dio_send_cur_page(dio);
1031                 if (ret == 0)
1032                         ret = ret2;
1033                 page_cache_release(dio->cur_page);
1034                 dio->cur_page = NULL;
1035         }
1036         if (dio->bio)
1037                 dio_bio_submit(dio);
1038
1039         /*
1040          * It is possible that, we return short IO due to end of file.
1041          * In that case, we need to release all the pages we got hold on.
1042          */
1043         dio_cleanup(dio);
1044
1045         /*
1046          * All block lookups have been performed. For READ requests
1047          * we can let i_sem go now that its achieved its purpose
1048          * of protecting us from looking up uninitialized blocks.
1049          */
1050         if ((rw == READ) && (dio->lock_type == DIO_LOCKING))
1051                 up(&dio->inode->i_sem);
1052
1053         /*
1054          * OK, all BIOs are submitted, so we can decrement bio_count to truly
1055          * reflect the number of to-be-processed BIOs.
1056          */
1057         if (dio->is_async) {
1058                 int should_wait = 0;
1059
1060                 if (dio->result < dio->size && rw == WRITE) {
1061                         dio->waiter = current;
1062                         should_wait = 1;
1063                 }
1064                 if (ret == 0)
1065                         ret = dio->result;
1066                 finished_one_bio(dio);          /* This can free the dio */
1067                 blk_run_address_space(inode->i_mapping);
1068                 if (should_wait) {
1069                         unsigned long flags;
1070                         /*
1071                          * Wait for already issued I/O to drain out and
1072                          * release its references to user-space pages
1073                          * before returning to fallback on buffered I/O
1074                          */
1075
1076                         spin_lock_irqsave(&dio->bio_lock, flags);
1077                         set_current_state(TASK_UNINTERRUPTIBLE);
1078                         while (dio->bio_count) {
1079                                 spin_unlock_irqrestore(&dio->bio_lock, flags);
1080                                 io_schedule();
1081                                 spin_lock_irqsave(&dio->bio_lock, flags);
1082                                 set_current_state(TASK_UNINTERRUPTIBLE);
1083                         }
1084                         spin_unlock_irqrestore(&dio->bio_lock, flags);
1085                         set_current_state(TASK_RUNNING);
1086                         kfree(dio);
1087                 }
1088         } else {
1089                 ssize_t transferred = 0;
1090
1091                 finished_one_bio(dio);
1092                 ret2 = dio_await_completion(dio);
1093                 if (ret == 0)
1094                         ret = ret2;
1095                 if (ret == 0)
1096                         ret = dio->page_errors;
1097                 if (dio->result) {
1098                         loff_t i_size = i_size_read(inode);
1099
1100                         transferred = dio->result;
1101                         /*
1102                          * Adjust the return value if the read crossed a
1103                          * non-block-aligned EOF.
1104                          */
1105                         if (rw == READ && (offset + transferred > i_size))
1106                                 transferred = i_size - offset;
1107                 }
1108                 dio_complete(dio, offset, transferred);
1109                 if (ret == 0)
1110                         ret = transferred;
1111
1112                 /* We could have also come here on an AIO file extend */
1113                 if (!is_sync_kiocb(iocb) && rw == WRITE &&
1114                     ret >= 0 && dio->result == dio->size)
1115                         /*
1116                          * For AIO writes where we have completed the
1117                          * i/o, we have to mark the the aio complete.
1118                          */
1119                         aio_complete(iocb, ret, 0);
1120                 kfree(dio);
1121         }
1122         return ret;
1123 }
1124
1125 /*
1126  * This is a library function for use by filesystem drivers.
1127  *
1128  * For writes to S_ISREG files, we are called under i_sem and return with i_sem
1129  * held, even though it is internally dropped.
1130  *
1131  * For writes to S_ISBLK files, i_sem is not held on entry; it is never taken.
1132  */
1133 ssize_t
1134 __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1135         struct block_device *bdev, const struct iovec *iov, loff_t offset,
1136         unsigned long nr_segs, get_blocks_t get_blocks, dio_iodone_t end_io,
1137         int dio_lock_type)
1138 {
1139         int seg;
1140         size_t size;
1141         unsigned long addr;
1142         unsigned blkbits = inode->i_blkbits;
1143         unsigned bdev_blkbits = 0;
1144         unsigned blocksize_mask = (1 << blkbits) - 1;
1145         ssize_t retval = -EINVAL;
1146         loff_t end = offset;
1147         struct dio *dio;
1148
1149         if (bdev)
1150                 bdev_blkbits = blksize_bits(bdev_hardsect_size(bdev));
1151
1152         if (offset & blocksize_mask) {
1153                 if (bdev)
1154                          blkbits = bdev_blkbits;
1155                 blocksize_mask = (1 << blkbits) - 1;
1156                 if (offset & blocksize_mask)
1157                         goto out;
1158         }
1159
1160         /* Check the memory alignment.  Blocks cannot straddle pages */
1161         for (seg = 0; seg < nr_segs; seg++) {
1162                 addr = (unsigned long)iov[seg].iov_base;
1163                 size = iov[seg].iov_len;
1164                 end += size;
1165                 if ((addr & blocksize_mask) || (size & blocksize_mask))  {
1166                         if (bdev)
1167                                  blkbits = bdev_blkbits;
1168                         blocksize_mask = (1 << blkbits) - 1;
1169                         if ((addr & blocksize_mask) || (size & blocksize_mask))
1170                                 goto out;
1171                 }
1172         }
1173
1174         dio = kmalloc(sizeof(*dio), GFP_KERNEL);
1175         retval = -ENOMEM;
1176         if (!dio)
1177                 goto out;
1178
1179         /*
1180          * For regular files using DIO_LOCKING,
1181          *      readers need to grab i_sem and i_alloc_sem
1182          *      writers need to grab i_alloc_sem only (i_sem is already held)
1183          * For regular files using DIO_OWN_LOCKING,
1184          *      both readers and writers need to grab i_alloc_sem
1185          *      neither readers nor writers hold i_sem on entry (nor exit)
1186          */
1187         dio->lock_type = dio_lock_type;
1188         if (dio_lock_type != DIO_NO_LOCKING) {
1189                 if (rw == READ) {
1190                         struct address_space *mapping;
1191
1192                         mapping = iocb->ki_filp->f_mapping;
1193                         down(&inode->i_sem);
1194                         retval = filemap_write_and_wait(mapping);
1195                         if (retval) {
1196                                 up(&inode->i_sem);
1197                                 kfree(dio);
1198                                 goto out;
1199                         }
1200                         down_read(&inode->i_alloc_sem);
1201                         if (dio_lock_type == DIO_OWN_LOCKING)
1202                                 up(&inode->i_sem);
1203                 } else {
1204                         down_read(&inode->i_alloc_sem);
1205                 }
1206         }
1207         /*
1208          * For file extending writes updating i_size before data
1209          * writeouts complete can expose uninitialized blocks. So
1210          * even for AIO, we need to wait for i/o to complete before
1211          * returning in this case.
1212          */
1213         dio->is_async = !is_sync_kiocb(iocb) && !((rw == WRITE) &&
1214                 (end > i_size_read(inode)));
1215
1216         retval = direct_io_worker(rw, iocb, inode, iov, offset,
1217                                 nr_segs, blkbits, get_blocks, end_io, dio);
1218 out:
1219         return retval;
1220 }
1221 EXPORT_SYMBOL(__blockdev_direct_IO);