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  * needs_locking is set for regular files on direct-IO-naive filesystems.  It
  57  * determines whether we need to do the fancy locking which prevents direct-IO
  58  * from being able to read uninitialised disk blocks.
  59  */
  60
  61 struct dio {
  62         /* BIO submission state */
  63         struct bio *bio;                /* bio under assembly */
  64         struct inode *inode;
  65         int rw;
  66         int needs_locking;              /* doesn't change */
  67         unsigned blkbits;               /* doesn't change */
  68         unsigned blkfactor;             /* When we're using an alignment which
  69                                            is finer than the filesystem's soft
  70                                            blocksize, this specifies how much
  71                                            finer.  blkfactor=2 means 1/4-block
  72                                            alignment.  Does not change */
  73         unsigned start_zero_done;       /* flag: sub-blocksize zeroing has
  74                                            been performed at the start of a
  75                                            write */
  76         int pages_in_io;                /* approximate total IO pages */
  77         size_t  size;                   /* total request size (doesn't change)*/
  78         sector_t block_in_file;         /* Current offset into the underlying
  79                                            file in dio_block units. */
  80         unsigned blocks_available;      /* At block_in_file.  changes */
  81         sector_t final_block_in_request;/* doesn't change */
  82         unsigned first_block_in_page;   /* doesn't change, Used only once */
  83         int boundary;                   /* prev block is at a boundary */
  84         int reap_counter;               /* rate limit reaping */
  85         get_blocks_t *get_blocks;       /* block mapping function */
  86         dio_iodone_t *end_io;           /* IO completion function */
  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_blocks() 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
 102         /*
 103          * Page fetching state. These variables belong to dio_refill_pages().
 104          */
 105         int curr_page;                  /* changes */
 106         int total_pages;                /* doesn't change */
 107         unsigned long curr_user_address;/* changes */
 108
 109         /*
 110          * Page queue.  These variables belong to dio_refill_pages() and
 111          * dio_get_page().
 112          */
 113         struct page *pages[DIO_PAGES];  /* page buffer */
 114         unsigned head;                  /* next page to process */
 115         unsigned tail;                  /* last valid page + 1 */
 116         int page_errors;                /* errno from get_user_pages() */
 117
 118         /* BIO completion state */
 119         spinlock_t bio_lock;            /* protects BIO fields below */
 120         int bio_count;                  /* nr bios to be completed */
 121         int bios_in_flight;             /* nr bios in flight */
 122         struct bio *bio_list;           /* singly linked via bi_private */
 123         struct task_struct *waiter;     /* waiting task (NULL if none) */
 124
 125         /* AIO related stuff */
 126         struct kiocb *iocb;             /* kiocb */
 127         int is_async;                   /* is IO async ? */
 128         ssize_t result;                 /* IO result */
 129 };
 130
 131 /*
 132  * How many pages are in the queue?
 133  */
 134 static inline unsigned dio_pages_present(struct dio *dio)
 135 {
 136         return dio->tail - dio->head;
 137 }
 138
 139 /*
 140  * Go grab and pin some userspace pages.   Typically we'll get 64 at a time.
 141  */
 142 static int dio_refill_pages(struct dio *dio)
 143 {
 144         int ret;
 145         int nr_pages;
 146
 147         nr_pages = min(dio->total_pages - dio->curr_page, DIO_PAGES);
 148         down_read(&current->mm->mmap_sem);
 149         ret = get_user_pages(
 150                 current,                        /* Task for fault acounting */
 151                 current->mm,                    /* whose pages? */
 152                 dio->curr_user_address,         /* Where from? */
 153                 nr_pages,                       /* How many pages? */
 154                 dio->rw == READ,                /* Write to memory? */
 155                 0,                              /* force (?) */
 156                 &dio->pages[0],
 157                 NULL);                          /* vmas */
 158         up_read(&current->mm->mmap_sem);
 159
 160         if (ret < 0 && dio->blocks_available && (dio->rw == WRITE)) {
 161                 /*
 162                  * A memory fault, but the filesystem has some outstanding
 163                  * mapped blocks.  We need to use those blocks up to avoid
 164                  * leaking stale data in the file.
 165                  */
 166                 if (dio->page_errors == 0)
 167                         dio->page_errors = ret;
 168                 dio->pages[0] = ZERO_PAGE(dio->curr_user_address);
 169                 dio->head = 0;
 170                 dio->tail = 1;
 171                 ret = 0;
 172                 goto out;
 173         }
 174
 175         if (ret >= 0) {
 176                 dio->curr_user_address += ret * PAGE_SIZE;
 177                 dio->curr_page += ret;
 178                 dio->head = 0;
 179                 dio->tail = ret;
 180                 ret = 0;
 181         }
 182 out:
 183         return ret;
 184 }
 185
 186 /*
 187  * Get another userspace page.  Returns an ERR_PTR on error.  Pages are
 188  * buffered inside the dio so that we can call get_user_pages() against a
 189  * decent number of pages, less frequently.  To provide nicer use of the
 190  * L1 cache.
 191  */
 192 static struct page *dio_get_page(struct dio *dio)
 193 {
 194         if (dio_pages_present(dio) == 0) {
 195                 int ret;
 196
 197                 ret = dio_refill_pages(dio);
 198                 if (ret)
 199                         return ERR_PTR(ret);
 200                 BUG_ON(dio_pages_present(dio) == 0);
 201         }
 202         return dio->pages[dio->head++];
 203 }
 204
 205 /*
 206  * Called when all DIO BIO I/O has been completed - let the filesystem
 207  * know, if it registered an interest earlier via get_blocks.  Pass the
 208  * private field of the map buffer_head so that filesystems can use it
 209  * to hold additional state between get_blocks calls and dio_complete.
 210  */
 211 static void dio_complete(struct dio *dio, loff_t offset, ssize_t bytes)
 212 {
 213         if (dio->end_io && dio->result)
 214                 dio->end_io(dio->inode, offset, bytes, dio->map_bh.b_private);
 215         if (dio->needs_locking)
 216                 up_read(&dio->inode->i_alloc_sem);
 217 }
 218
 219 /*
 220  * Called when a BIO has been processed.  If the count goes to zero then IO is
 221  * complete and we can signal this to the AIO layer.
 222  */
 223 static void finished_one_bio(struct dio *dio)
 224 {
 225         unsigned long flags;
 226
 227         spin_lock_irqsave(&dio->bio_lock, flags);
 228         if (dio->bio_count == 1) {
 229                 if (dio->is_async) {
 230                         /*
 231                          * Last reference to the dio is going away.
 232                          * Drop spinlock and complete the DIO.
 233                          */
 234                         spin_unlock_irqrestore(&dio->bio_lock, flags);
 235                         dio_complete(dio, dio->block_in_file << dio->blkbits,
 236                                         dio->result);
 237                         /* Complete AIO later if falling back to buffered i/o */
 238                         if (dio->result == dio->size || dio->rw == READ) {
 239                                 aio_complete(dio->iocb, dio->result, 0);
 240                                 kfree(dio);
 241                                 return;
 242                         } else {
 243                                 /*
 244                                  * Falling back to buffered
 245                                  */
 246                                 spin_lock_irqsave(&dio->bio_lock, flags);
 247                                 dio->bio_count--;
 248                                 if (dio->waiter)
 249                                         wake_up_process(dio->waiter);
 250                                 spin_unlock_irqrestore(&dio->bio_lock, flags);
 251                                 return;
 252                         }
 253                 }
 254         }
 255         dio->bio_count--;
 256         spin_unlock_irqrestore(&dio->bio_lock, flags);
 257 }
 258
 259 static int dio_bio_complete(struct dio *dio, struct bio *bio);
 260 /*
 261  * Asynchronous IO callback.
 262  */
 263 static int dio_bio_end_aio(struct bio *bio, unsigned int bytes_done, int error)
 264 {
 265         struct dio *dio = bio->bi_private;
 266
 267         if (bio->bi_size)
 268                 return 1;
 269
 270         /* cleanup the bio */
 271         dio_bio_complete(dio, bio);
 272         return 0;
 273 }
 274
 275 /*
 276  * The BIO completion handler simply queues the BIO up for the process-context
 277  * handler.
 278  *
 279  * During I/O bi_private points at the dio.  After I/O, bi_private is used to
 280  * implement a singly-linked list of completed BIOs, at dio->bio_list.
 281  */
 282 static int dio_bio_end_io(struct bio *bio, unsigned int bytes_done, int error)
 283 {
 284         struct dio *dio = bio->bi_private;
 285         unsigned long flags;
 286
 287         if (bio->bi_size)
 288                 return 1;
 289
 290         spin_lock_irqsave(&dio->bio_lock, flags);
 291         bio->bi_private = dio->bio_list;
 292         dio->bio_list = bio;
 293         dio->bios_in_flight--;
 294         if (dio->waiter && dio->bios_in_flight == 0)
 295                 wake_up_process(dio->waiter);
 296         spin_unlock_irqrestore(&dio->bio_lock, flags);
 297         return 0;
 298 }
 299
 300 static int
 301 dio_bio_alloc(struct dio *dio, struct block_device *bdev,
 302                 sector_t first_sector, int nr_vecs)
 303 {
 304         struct bio *bio;
 305
 306         bio = bio_alloc(GFP_KERNEL, nr_vecs);
 307         if (bio == NULL)
 308                 return -ENOMEM;
 309
 310         bio->bi_bdev = bdev;
 311         bio->bi_sector = first_sector;
 312         if (dio->is_async)
 313                 bio->bi_end_io = dio_bio_end_aio;
 314         else
 315                 bio->bi_end_io = dio_bio_end_io;
 316
 317         dio->bio = bio;
 318         return 0;
 319 }
 320
 321 /*
 322  * In the AIO read case we speculatively dirty the pages before starting IO.
 323  * During IO completion, any of these pages which happen to have been written
 324  * back will be redirtied by bio_check_pages_dirty().
 325  */
 326 static void dio_bio_submit(struct dio *dio)
 327 {
 328         struct bio *bio = dio->bio;
 329         unsigned long flags;
 330
 331         bio->bi_private = dio;
 332         spin_lock_irqsave(&dio->bio_lock, flags);
 333         dio->bio_count++;
 334         dio->bios_in_flight++;
 335         spin_unlock_irqrestore(&dio->bio_lock, flags);
 336         if (dio->is_async && dio->rw == READ)
 337                 bio_set_pages_dirty(bio);
 338         submit_bio(dio->rw, bio);
 339
 340         dio->bio = NULL;
 341         dio->boundary = 0;
 342 }
 343
 344 /*
 345  * Release any resources in case of a failure
 346  */
 347 static void dio_cleanup(struct dio *dio)
 348 {
 349         while (dio_pages_present(dio))
 350                 page_cache_release(dio_get_page(dio));
 351 }
 352
 353 /*
 354  * Wait for the next BIO to complete.  Remove it and return it.
 355  */
 356 static struct bio *dio_await_one(struct dio *dio)
 357 {
 358         unsigned long flags;
 359         struct bio *bio;
 360
 361         spin_lock_irqsave(&dio->bio_lock, flags);
 362         while (dio->bio_list == NULL) {
 363                 set_current_state(TASK_UNINTERRUPTIBLE);
 364                 if (dio->bio_list == NULL) {
 365                         dio->waiter = current;
 366                         spin_unlock_irqrestore(&dio->bio_lock, flags);
 367                         blk_run_address_space(dio->inode->i_mapping);
 368                         io_schedule();
 369                         spin_lock_irqsave(&dio->bio_lock, flags);
 370                         dio->waiter = NULL;
 371                 }
 372                 set_current_state(TASK_RUNNING);
 373         }
 374         bio = dio->bio_list;
 375         dio->bio_list = bio->bi_private;
 376         spin_unlock_irqrestore(&dio->bio_lock, flags);
 377         return bio;
 378 }
 379
 380 /*
 381  * Process one completed BIO.  No locks are held.
 382  */
 383 static int dio_bio_complete(struct dio *dio, struct bio *bio)
 384 {
 385         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 386         struct bio_vec *bvec = bio->bi_io_vec;
 387         int page_no;
 388
 389         if (!uptodate)
 390                 dio->result = -EIO;
 391
 392         if (dio->is_async && dio->rw == READ) {
 393                 bio_check_pages_dirty(bio);     /* transfers ownership */
 394         } else {
 395                 for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
 396                         struct page *page = bvec[page_no].bv_page;
 397
 398                         if (dio->rw == READ)
 399                                 set_page_dirty_lock(page);
 400                         page_cache_release(page);
 401                 }
 402                 bio_put(bio);
 403         }
 404         finished_one_bio(dio);
 405         return uptodate ? 0 : -EIO;
 406 }
 407
 408 /*
 409  * Wait on and process all in-flight BIOs.
 410  */
 411 static int dio_await_completion(struct dio *dio)
 412 {
 413         int ret = 0;
 414
 415         if (dio->bio)
 416                 dio_bio_submit(dio);
 417
 418         /*
 419          * The bio_lock is not held for the read of bio_count.
 420          * This is ok since it is the dio_bio_complete() that changes
 421          * bio_count.
 422          */
 423         while (dio->bio_count) {
 424                 struct bio *bio = dio_await_one(dio);
 425                 int ret2;
 426
 427                 ret2 = dio_bio_complete(dio, bio);
 428                 if (ret == 0)
 429                         ret = ret2;
 430         }
 431         return ret;
 432 }
 433
 434 /*
 435  * A really large O_DIRECT read or write can generate a lot of BIOs.  So
 436  * to keep the memory consumption sane we periodically reap any completed BIOs
 437  * during the BIO generation phase.
 438  *
 439  * This also helps to limit the peak amount of pinned userspace memory.
 440  */
 441 static int dio_bio_reap(struct dio *dio)
 442 {
 443         int ret = 0;
 444
 445         if (dio->reap_counter++ >= 64) {
 446                 while (dio->bio_list) {
 447                         unsigned long flags;
 448                         struct bio *bio;
 449                         int ret2;
 450
 451                         spin_lock_irqsave(&dio->bio_lock, flags);
 452                         bio = dio->bio_list;
 453                         dio->bio_list = bio->bi_private;
 454                         spin_unlock_irqrestore(&dio->bio_lock, flags);
 455                         ret2 = dio_bio_complete(dio, bio);
 456                         if (ret == 0)
 457                                 ret = ret2;
 458                 }
 459                 dio->reap_counter = 0;
 460         }
 461         return ret;
 462 }
 463
 464 /*
 465  * Call into the fs to map some more disk blocks.  We record the current number
 466  * of available blocks at dio->blocks_available.  These are in units of the
 467  * fs blocksize, (1 << inode->i_blkbits).
 468  *
 469  * The fs is allowed to map lots of blocks at once.  If it wants to do that,
 470  * it uses the passed inode-relative block number as the file offset, as usual.
 471  *
 472  * get_blocks() is passed the number of i_blkbits-sized blocks which direct_io
 473  * has remaining to do.  The fs should not map more than this number of blocks.
 474  *
 475  * If the fs has mapped a lot of blocks, it should populate bh->b_size to
 476  * indicate how much contiguous disk space has been made available at
 477  * bh->b_blocknr.
 478  *
 479  * If *any* of the mapped blocks are new, then the fs must set buffer_new().
 480  * This isn't very efficient...
 481  *
 482  * In the case of filesystem holes: the fs may return an arbitrarily-large
 483  * hole by returning an appropriate value in b_size and by clearing
 484  * buffer_mapped().  However the direct-io code will only process holes one
 485  * block at a time - it will repeatedly call get_blocks() as it walks the hole.
 486  */
 487 static int get_more_blocks(struct dio *dio)
 488 {
 489         int ret;
 490         struct buffer_head *map_bh = &dio->map_bh;
 491         sector_t fs_startblk;   /* Into file, in filesystem-sized blocks */
 492         unsigned long fs_count; /* Number of filesystem-sized blocks */
 493         unsigned long dio_count;/* Number of dio_block-sized blocks */
 494         unsigned long blkmask;
 495         int beyond_eof = 0;
 496
 497         /*
 498          * If there was a memory error and we've overwritten all the
 499          * mapped blocks then we can now return that memory error
 500          */
 501         ret = dio->page_errors;
 502         if (ret == 0) {
 503                 map_bh->b_state = 0;
 504                 map_bh->b_size = 0;
 505                 BUG_ON(dio->block_in_file >= dio->final_block_in_request);
 506                 fs_startblk = dio->block_in_file >> dio->blkfactor;
 507                 dio_count = dio->final_block_in_request - dio->block_in_file;
 508                 fs_count = dio_count >> dio->blkfactor;
 509                 blkmask = (1 << dio->blkfactor) - 1;
 510                 if (dio_count & blkmask)
 511                         fs_count++;
 512
 513                 if (dio->needs_locking) {
 514                         if (dio->block_in_file >= (i_size_read(dio->inode) >>
 515                                                         dio->blkbits))
 516                                 beyond_eof = 1;
 517                 }
 518                 /*
 519                  * For writes inside i_size we forbid block creations: only
 520                  * overwrites are permitted.  We fall back to buffered writes
 521                  * at a higher level for inside-i_size block-instantiating
 522                  * writes.
 523                  */
 524                 ret = (*dio->get_blocks)(dio->inode, fs_startblk, fs_count,
 525                                 map_bh, (dio->rw == WRITE) && beyond_eof);
 526         }
 527         return ret;
 528 }
 529
 530 /*
 531  * There is no bio.  Make one now.
 532  */
 533 static int dio_new_bio(struct dio *dio, sector_t start_sector)
 534 {
 535         sector_t sector;
 536         int ret, nr_pages;
 537
 538         ret = dio_bio_reap(dio);
 539         if (ret)
 540                 goto out;
 541         sector = start_sector << (dio->blkbits - 9);
 542         nr_pages = min(dio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev));
 543         BUG_ON(nr_pages <= 0);
 544         ret = dio_bio_alloc(dio, dio->map_bh.b_bdev, sector, nr_pages);
 545         dio->boundary = 0;
 546 out:
 547         return ret;
 548 }
 549
 550 /*
 551  * Attempt to put the current chunk of 'cur_page' into the current BIO.  If
 552  * that was successful then update final_block_in_bio and take a ref against
 553  * the just-added page.
 554  *
 555  * Return zero on success.  Non-zero means the caller needs to start a new BIO.
 556  */
 557 static int dio_bio_add_page(struct dio *dio)
 558 {
 559         int ret;
 560
 561         ret = bio_add_page(dio->bio, dio->cur_page,
 562                         dio->cur_page_len, dio->cur_page_offset);
 563         if (ret == dio->cur_page_len) {
 564                 dio->pages_in_io--;
 565                 page_cache_get(dio->cur_page);
 566                 dio->final_block_in_bio = dio->cur_page_block +
 567                         (dio->cur_page_len >> dio->blkbits);
 568                 ret = 0;
 569         } else {
 570                 ret = 1;
 571         }
 572         return ret;
 573 }
 574
 575 /*
 576  * Put cur_page under IO.  The section of cur_page which is described by
 577  * cur_page_offset,cur_page_len is put into a BIO.  The section of cur_page
 578  * starts on-disk at cur_page_block.
 579  *
 580  * We take a ref against the page here (on behalf of its presence in the bio).
 581  *
 582  * The caller of this function is responsible for removing cur_page from the
 583  * dio, and for dropping the refcount which came from that presence.
 584  */
 585 static int dio_send_cur_page(struct dio *dio)
 586 {
 587         int ret = 0;
 588
 589         if (dio->bio) {
 590                 /*
 591                  * See whether this new request is contiguous with the old
 592                  */
 593                 if (dio->final_block_in_bio != dio->cur_page_block)
 594                         dio_bio_submit(dio);
 595                 /*
 596                  * Submit now if the underlying fs is about to perform a
 597                  * metadata read
 598                  */
 599                 if (dio->boundary)
 600                         dio_bio_submit(dio);
 601         }
 602
 603         if (dio->bio == NULL) {
 604                 ret = dio_new_bio(dio, dio->cur_page_block);
 605                 if (ret)
 606                         goto out;
 607         }
 608
 609         if (dio_bio_add_page(dio) != 0) {
 610                 dio_bio_submit(dio);
 611                 ret = dio_new_bio(dio, dio->cur_page_block);
 612                 if (ret == 0) {
 613                         ret = dio_bio_add_page(dio);
 614                         BUG_ON(ret != 0);
 615                 }
 616         }
 617 out:
 618         return ret;
 619 }
 620
 621 /*
 622  * An autonomous function to put a chunk of a page under deferred IO.
 623  *
 624  * The caller doesn't actually know (or care) whether this piece of page is in
 625  * a BIO, or is under IO or whatever.  We just take care of all possible
 626  * situations here.  The separation between the logic of do_direct_IO() and
 627  * that of submit_page_section() is important for clarity.  Please don't break.
 628  *
 629  * The chunk of page starts on-disk at blocknr.
 630  *
 631  * We perform deferred IO, by recording the last-submitted page inside our
 632  * private part of the dio structure.  If possible, we just expand the IO
 633  * across that page here.
 634  *
 635  * If that doesn't work out then we put the old page into the bio and add this
 636  * page to the dio instead.
 637  */
 638 static int
 639 submit_page_section(struct dio *dio, struct page *page,
 640                 unsigned offset, unsigned len, sector_t blocknr)
 641 {
 642         int ret = 0;
 643
 644         /*
 645          * Can we just grow the current page's presence in the dio?
 646          */
 647         if (    (dio->cur_page == page) &&
 648                 (dio->cur_page_offset + dio->cur_page_len == offset) &&
 649                 (dio->cur_page_block +
 650                         (dio->cur_page_len >> dio->blkbits) == blocknr)) {
 651                 dio->cur_page_len += len;
 652
 653                 /*
 654                  * If dio->boundary then we want to schedule the IO now to
 655                  * avoid metadata seeks.
 656                  */
 657                 if (dio->boundary) {
 658                         ret = dio_send_cur_page(dio);
 659                         page_cache_release(dio->cur_page);
 660                         dio->cur_page = NULL;
 661                 }
 662                 goto out;
 663         }
 664
 665         /*
 666          * If there's a deferred page already there then send it.
 667          */
 668         if (dio->cur_page) {
 669                 ret = dio_send_cur_page(dio);
 670                 page_cache_release(dio->cur_page);
 671                 dio->cur_page = NULL;
 672                 if (ret)
 673                         goto out;
 674         }
 675
 676         page_cache_get(page);           /* It is in dio */
 677         dio->cur_page = page;
 678         dio->cur_page_offset = offset;
 679         dio->cur_page_len = len;
 680         dio->cur_page_block = blocknr;
 681 out:
 682         return ret;
 683 }
 684
 685 /*
 686  * Clean any dirty buffers in the blockdev mapping which alias newly-created
 687  * file blocks.  Only called for S_ISREG files - blockdevs do not set
 688  * buffer_new
 689  */
 690 static void clean_blockdev_aliases(struct dio *dio)
 691 {
 692         unsigned i;
 693         unsigned nblocks;
 694
 695         nblocks = dio->map_bh.b_size >> dio->inode->i_blkbits;
 696
 697         for (i = 0; i < nblocks; i++) {
 698                 unmap_underlying_metadata(dio->map_bh.b_bdev,
 699                                         dio->map_bh.b_blocknr + i);
 700         }
 701 }
 702
 703 /*
 704  * If we are not writing the entire block and get_block() allocated
 705  * the block for us, we need to fill-in the unused portion of the
 706  * block with zeros. This happens only if user-buffer, fileoffset or
 707  * io length is not filesystem block-size multiple.
 708  *
 709  * `end' is zero if we're doing the start of the IO, 1 at the end of the
 710  * IO.
 711  */
 712 static void dio_zero_block(struct dio *dio, int end)
 713 {
 714         unsigned dio_blocks_per_fs_block;
 715         unsigned this_chunk_blocks;     /* In dio_blocks */
 716         unsigned this_chunk_bytes;
 717         struct page *page;
 718
 719         dio->start_zero_done = 1;
 720         if (!dio->blkfactor || !buffer_new(&dio->map_bh))
 721                 return;
 722
 723         dio_blocks_per_fs_block = 1 << dio->blkfactor;
 724         this_chunk_blocks = dio->block_in_file & (dio_blocks_per_fs_block - 1);
 725
 726         if (!this_chunk_blocks)
 727                 return;
 728
 729         /*
 730          * We need to zero out part of an fs block.  It is either at the
 731          * beginning or the end of the fs block.
 732          */
 733         if (end)
 734                 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
 735
 736         this_chunk_bytes = this_chunk_blocks << dio->blkbits;
 737
 738         page = ZERO_PAGE(dio->curr_user_address);
 739         if (submit_page_section(dio, page, 0, this_chunk_bytes,
 740                                 dio->next_block_for_io))
 741                 return;
 742
 743         dio->next_block_for_io += this_chunk_blocks;
 744 }
 745
 746 /*
 747  * Walk the user pages, and the file, mapping blocks to disk and generating
 748  * a sequence of (page,offset,len,block) mappings.  These mappings are injected
 749  * into submit_page_section(), which takes care of the next stage of submission
 750  *
 751  * Direct IO against a blockdev is different from a file.  Because we can
 752  * happily perform page-sized but 512-byte aligned IOs.  It is important that
 753  * blockdev IO be able to have fine alignment and large sizes.
 754  *
 755  * So what we do is to permit the ->get_blocks function to populate bh.b_size
 756  * with the size of IO which is permitted at this offset and this i_blkbits.
 757  *
 758  * For best results, the blockdev should be set up with 512-byte i_blkbits and
 759  * it should set b_size to PAGE_SIZE or more inside get_blocks().  This gives
 760  * fine alignment but still allows this function to work in PAGE_SIZE units.
 761  */
 762 static int do_direct_IO(struct dio *dio)
 763 {
 764         const unsigned blkbits = dio->blkbits;
 765         const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
 766         struct page *page;
 767         unsigned block_in_page;
 768         struct buffer_head *map_bh = &dio->map_bh;
 769         int ret = 0;
 770
 771         /* The I/O can start at any block offset within the first page */
 772         block_in_page = dio->first_block_in_page;
 773
 774         while (dio->block_in_file < dio->final_block_in_request) {
 775                 page = dio_get_page(dio);
 776                 if (IS_ERR(page)) {
 777                         ret = PTR_ERR(page);
 778                         goto out;
 779                 }
 780
 781                 while (block_in_page < blocks_per_page) {
 782                         unsigned offset_in_page = block_in_page << blkbits;
 783                         unsigned this_chunk_bytes;      /* # of bytes mapped */
 784                         unsigned this_chunk_blocks;     /* # of blocks */
 785                         unsigned u;
 786
 787                         if (dio->blocks_available == 0) {
 788                                 /*
 789                                  * Need to go and map some more disk
 790                                  */
 791                                 unsigned long blkmask;
 792                                 unsigned long dio_remainder;
 793
 794                                 ret = get_more_blocks(dio);
 795                                 if (ret) {
 796                                         page_cache_release(page);
 797                                         goto out;
 798                                 }
 799                                 if (!buffer_mapped(map_bh))
 800                                         goto do_holes;
 801
 802                                 dio->blocks_available =
 803                                                 map_bh->b_size >> dio->blkbits;
 804                                 dio->next_block_for_io =
 805                                         map_bh->b_blocknr << dio->blkfactor;
 806                                 if (buffer_new(map_bh))
 807                                         clean_blockdev_aliases(dio);
 808
 809                                 if (!dio->blkfactor)
 810                                         goto do_holes;
 811
 812                                 blkmask = (1 << dio->blkfactor) - 1;
 813                                 dio_remainder = (dio->block_in_file & blkmask);
 814
 815                                 /*
 816                                  * If we are at the start of IO and that IO
 817                                  * starts partway into a fs-block,
 818                                  * dio_remainder will be non-zero.  If the IO
 819                                  * is a read then we can simply advance the IO
 820                                  * cursor to the first block which is to be
 821                                  * read.  But if the IO is a write and the
 822                                  * block was newly allocated we cannot do that;
 823                                  * the start of the fs block must be zeroed out
 824                                  * on-disk
 825                                  */
 826                                 if (!buffer_new(map_bh))
 827                                         dio->next_block_for_io += dio_remainder;
 828                                 dio->blocks_available -= dio_remainder;
 829                         }
 830 do_holes:
 831                         /* Handle holes */
 832                         if (!buffer_mapped(map_bh)) {
 833                                 char *kaddr;
 834
 835                                 /* AKPM: eargh, -ENOTBLK is a hack */
 836                                 if (dio->rw == WRITE)
 837                                         return -ENOTBLK;
 838
 839                                 if (dio->block_in_file >=
 840                                         i_size_read(dio->inode)>>blkbits) {
 841                                         /* We hit eof */
 842                                         page_cache_release(page);
 843                                         goto out;
 844                                 }
 845                                 kaddr = kmap_atomic(page, KM_USER0);
 846                                 memset(kaddr + (block_in_page << blkbits),
 847                                                 0, 1 << blkbits);
 848                                 flush_dcache_page(page);
 849                                 kunmap_atomic(kaddr, KM_USER0);
 850                                 dio->block_in_file++;
 851                                 block_in_page++;
 852                                 goto next_block;
 853                         }
 854
 855                         /*
 856                          * If we're performing IO which has an alignment which
 857                          * is finer than the underlying fs, go check to see if
 858                          * we must zero out the start of this block.
 859                          */
 860                         if (unlikely(dio->blkfactor && !dio->start_zero_done))
 861                                 dio_zero_block(dio, 0);
 862
 863                         /*
 864                          * Work out, in this_chunk_blocks, how much disk we
 865                          * can add to this page
 866                          */
 867                         this_chunk_blocks = dio->blocks_available;
 868                         u = (PAGE_SIZE - offset_in_page) >> blkbits;
 869                         if (this_chunk_blocks > u)
 870                                 this_chunk_blocks = u;
 871                         u = dio->final_block_in_request - dio->block_in_file;
 872                         if (this_chunk_blocks > u)
 873                                 this_chunk_blocks = u;
 874                         this_chunk_bytes = this_chunk_blocks << blkbits;
 875                         BUG_ON(this_chunk_bytes == 0);
 876
 877                         dio->boundary = buffer_boundary(map_bh);
 878                         ret = submit_page_section(dio, page, offset_in_page,
 879                                 this_chunk_bytes, dio->next_block_for_io);
 880                         if (ret) {
 881                                 page_cache_release(page);
 882                                 goto out;
 883                         }
 884                         dio->next_block_for_io += this_chunk_blocks;
 885
 886                         dio->block_in_file += this_chunk_blocks;
 887                         block_in_page += this_chunk_blocks;
 888                         dio->blocks_available -= this_chunk_blocks;
 889 next_block:
 890                         if (dio->block_in_file > dio->final_block_in_request)
 891                                 BUG();
 892                         if (dio->block_in_file == dio->final_block_in_request)
 893                                 break;
 894                 }
 895
 896                 /* Drop the ref which was taken in get_user_pages() */
 897                 page_cache_release(page);
 898                 block_in_page = 0;
 899         }
 900 out:
 901         return ret;
 902 }
 903
 904 /*
 905  * Releases both i_sem and i_alloc_sem
 906  */
 907 static ssize_t
 908 direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
 909         const struct iovec *iov, loff_t offset, unsigned long nr_segs,
 910         unsigned blkbits, get_blocks_t get_blocks, dio_iodone_t end_io,
 911         struct dio *dio)
 912 {
 913         unsigned long user_addr;
 914         int seg;
 915         ssize_t ret = 0;
 916         ssize_t ret2;
 917         size_t bytes;
 918
 919         dio->bio = NULL;
 920         dio->inode = inode;
 921         dio->rw = rw;
 922         dio->blkbits = blkbits;
 923         dio->blkfactor = inode->i_blkbits - blkbits;
 924         dio->start_zero_done = 0;
 925         dio->size = 0;
 926         dio->block_in_file = offset >> blkbits;
 927         dio->blocks_available = 0;
 928         dio->cur_page = NULL;
 929
 930         dio->boundary = 0;
 931         dio->reap_counter = 0;
 932         dio->get_blocks = get_blocks;
 933         dio->end_io = end_io;
 934         dio->map_bh.b_private = NULL;
 935         dio->final_block_in_bio = -1;
 936         dio->next_block_for_io = -1;
 937
 938         dio->page_errors = 0;
 939         dio->result = 0;
 940         dio->iocb = iocb;
 941
 942         /*
 943          * BIO completion state.
 944          *
 945          * ->bio_count starts out at one, and we decrement it to zero after all
 946          * BIOs are submitted.  This to avoid the situation where a really fast
 947          * (or synchronous) device could take the count to zero while we're
 948          * still submitting BIOs.
 949          */
 950         dio->bio_count = 1;
 951         dio->bios_in_flight = 0;
 952         spin_lock_init(&dio->bio_lock);
 953         dio->bio_list = NULL;
 954         dio->waiter = NULL;
 955
 956         dio->pages_in_io = 0;
 957         for (seg = 0; seg < nr_segs; seg++)
 958                 dio->pages_in_io += (iov[seg].iov_len >> blkbits) + 2;
 959
 960         for (seg = 0; seg < nr_segs; seg++) {
 961                 user_addr = (unsigned long)iov[seg].iov_base;
 962                 dio->size += bytes = iov[seg].iov_len;
 963
 964                 /* Index into the first page of the first block */
 965                 dio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
 966                 dio->final_block_in_request = dio->block_in_file +
 967                                                 (bytes >> blkbits);
 968                 /* Page fetching state */
 969                 dio->head = 0;
 970                 dio->tail = 0;
 971                 dio->curr_page = 0;
 972
 973                 dio->total_pages = 0;
 974                 if (user_addr & (PAGE_SIZE-1)) {
 975                         dio->total_pages++;
 976                         bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
 977                 }
 978                 dio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
 979                 dio->curr_user_address = user_addr;
 980
 981                 ret = do_direct_IO(dio);
 982
 983                 dio->result += iov[seg].iov_len -
 984                         ((dio->final_block_in_request - dio->block_in_file) <<
 985                                         blkbits);
 986
 987                 if (ret) {
 988                         dio_cleanup(dio);
 989                         break;
 990                 }
 991         } /* end iovec loop */
 992
 993         /*
 994          * There may be some unwritten disk at the end of a part-written
 995          * fs-block-sized block.  Go zero that now.
 996          */
 997         dio_zero_block(dio, 1);
 998
 999         if (dio->cur_page) {
1000                 ret2 = dio_send_cur_page(dio);
1001                 if (ret == 0)
1002                         ret = ret2;
1003                 page_cache_release(dio->cur_page);
1004                 dio->cur_page = NULL;
1005         }
1006         if (dio->bio)
1007                 dio_bio_submit(dio);
1008
1009         /*
1010          * It is possible that, we return short IO due to end of file.
1011          * In that case, we need to release all the pages we got hold on.
1012          */
1013         dio_cleanup(dio);
1014
1015         /*
1016          * All block lookups have been performed. For READ requests
1017          * we can let i_sem go now that its achieved its purpose
1018          * of protecting us from looking up uninitialized blocks.
1019          */
1020         if ((rw == READ) && dio->needs_locking)
1021                 up(&dio->inode->i_sem);
1022
1023         /*
1024          * OK, all BIOs are submitted, so we can decrement bio_count to truly
1025          * reflect the number of to-be-processed BIOs.
1026          */
1027         if (dio->is_async) {
1028                 int should_wait = 0;
1029
1030                 if (dio->result < dio->size && rw == WRITE) {
1031                         dio->waiter = current;
1032                         should_wait = 1;
1033                 }
1034                 if (ret == 0)
1035                         ret = dio->result;
1036                 finished_one_bio(dio);          /* This can free the dio */
1037                 blk_run_address_space(inode->i_mapping);
1038                 if (should_wait) {
1039                         unsigned long flags;
1040                         /*
1041                          * Wait for already issued I/O to drain out and
1042                          * release its references to user-space pages
1043                          * before returning to fallback on buffered I/O
1044                          */
1045
1046                         spin_lock_irqsave(&dio->bio_lock, flags);
1047                         set_current_state(TASK_UNINTERRUPTIBLE);
1048                         while (dio->bio_count) {
1049                                 spin_unlock_irqrestore(&dio->bio_lock, flags);
1050                                 io_schedule();
1051                                 spin_lock_irqsave(&dio->bio_lock, flags);
1052                                 set_current_state(TASK_UNINTERRUPTIBLE);
1053                         }
1054                         spin_unlock_irqrestore(&dio->bio_lock, flags);
1055                         set_current_state(TASK_RUNNING);
1056                         kfree(dio);
1057                 }
1058         } else {
1059                 ssize_t transferred = 0;
1060
1061                 finished_one_bio(dio);
1062                 ret2 = dio_await_completion(dio);
1063                 if (ret == 0)
1064                         ret = ret2;
1065                 if (ret == 0)
1066                         ret = dio->page_errors;
1067                 if (dio->result) {
1068                         loff_t i_size = i_size_read(inode);
1069
1070                         transferred = dio->result;
1071                         /*
1072                          * Adjust the return value if the read crossed a
1073                          * non-block-aligned EOF.
1074                          */
1075                         if (rw == READ && (offset + transferred > i_size))
1076                                 transferred = i_size - offset;
1077                 }
1078                 dio_complete(dio, offset, transferred);
1079                 if (ret == 0)
1080                         ret = transferred;
1081
1082                 /* We could have also come here on an AIO file extend */
1083                 if (!is_sync_kiocb(iocb) && rw == WRITE &&
1084                     ret >= 0 && dio->result == dio->size)
1085                         /*
1086                          * For AIO writes where we have completed the
1087                          * i/o, we have to mark the the aio complete.
1088                          */
1089                         aio_complete(iocb, ret, 0);
1090                 kfree(dio);
1091         }
1092         if (ret == -ENOTBLK && rw == WRITE) {
1093                 /*
1094                  * The entire request will be be handled by buffered I/O
1095                  * when we return
1096                  */
1097                 ret = 0;
1098         }
1099         return ret;
1100 }
1101
1102 /*
1103  * This is a library function for use by filesystem drivers.
1104  *
1105  * For writes to S_ISREG files, we are called under i_sem and return with i_sem
1106  * held, even though it is internally dropped.
1107  *
1108  * For writes to S_ISBLK files, i_sem is not held on entry; it is never taken.
1109  */
1110 ssize_t
1111 __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1112         struct block_device *bdev, const struct iovec *iov, loff_t offset,
1113         unsigned long nr_segs, get_blocks_t get_blocks, dio_iodone_t end_io,
1114         int needs_special_locking)
1115 {
1116         int seg;
1117         size_t size;
1118         unsigned long addr;
1119         unsigned blkbits = inode->i_blkbits;
1120         unsigned bdev_blkbits = 0;
1121         unsigned blocksize_mask = (1 << blkbits) - 1;
1122         ssize_t retval = -EINVAL;
1123         loff_t end = offset;
1124         struct dio *dio;
1125         int needs_locking;
1126
1127         if (bdev)
1128                 bdev_blkbits = blksize_bits(bdev_hardsect_size(bdev));
1129
1130         if (offset & blocksize_mask) {
1131                 if (bdev)
1132                          blkbits = bdev_blkbits;
1133                 blocksize_mask = (1 << blkbits) - 1;
1134                 if (offset & blocksize_mask)
1135                         goto out;
1136         }
1137
1138         /* Check the memory alignment.  Blocks cannot straddle pages */
1139         for (seg = 0; seg < nr_segs; seg++) {
1140                 addr = (unsigned long)iov[seg].iov_base;
1141                 size = iov[seg].iov_len;
1142                 end += size;
1143                 if ((addr & blocksize_mask) || (size & blocksize_mask))  {
1144                         if (bdev)
1145                                  blkbits = bdev_blkbits;
1146                         blocksize_mask = (1 << blkbits) - 1;
1147                         if ((addr & blocksize_mask) || (size & blocksize_mask))
1148                                 goto out;
1149                 }
1150         }
1151
1152         dio = kmalloc(sizeof(*dio), GFP_KERNEL);
1153         retval = -ENOMEM;
1154         if (!dio)
1155                 goto out;
1156
1157         /*
1158          * For regular files,
1159          *      readers need to grab i_sem and i_alloc_sem
1160          *      writers need to grab i_alloc_sem only (i_sem is already held)
1161          */
1162         needs_locking = 0;
1163         if (S_ISREG(inode->i_mode) && needs_special_locking) {
1164                 needs_locking = 1;
1165                 if (rw == READ) {
1166                         struct address_space *mapping;
1167
1168                         mapping = iocb->ki_filp->f_mapping;
1169                         down(&inode->i_sem);
1170                         retval = filemap_write_and_wait(mapping);
1171                         if (retval) {
1172                                 up(&inode->i_sem);
1173                                 kfree(dio);
1174                                 goto out;
1175                         }
1176                 }
1177                 down_read(&inode->i_alloc_sem);
1178         }
1179         dio->needs_locking = needs_locking;
1180         /*
1181          * For file extending writes updating i_size before data
1182          * writeouts complete can expose uninitialized blocks. So
1183          * even for AIO, we need to wait for i/o to complete before
1184          * returning in this case.
1185          */
1186         dio->is_async = !is_sync_kiocb(iocb) && !((rw == WRITE) &&
1187                 (end > i_size_read(inode)));
1188
1189         retval = direct_io_worker(rw, iocb, inode, iov, offset,
1190                                 nr_segs, blkbits, get_blocks, end_io, dio);
1191 out:
1192         return retval;
1193 }
1194 EXPORT_SYMBOL(__blockdev_direct_IO);