xfs: remove old vmap cache
[linux-2.6.git] / fs / direct-io.c
index f11f05d..e82adc2 100644 (file)
@@ -5,11 +5,11 @@
  *
  * O_DIRECT
  *
- * 04Jul2002   akpm@zip.com.au
+ * 04Jul2002   Andrew Morton
  *             Initial version
  * 11Sep2002   janetinc@us.ibm.com
  *             added readv/writev support.
- * 29Oct2002   akpm@zip.com.au
+ * 29Oct2002   Andrew Morton
  *             rewrote bio_add_page() support.
  * 30Oct2002   pbadari@us.ibm.com
  *             added support for non-aligned IO.
  *
  * If blkfactor is zero then the user's request was aligned to the filesystem's
  * blocksize.
- *
- * lock_type is DIO_LOCKING for regular files on direct-IO-naive filesystems.
- * This determines whether we need to do the fancy locking which prevents
- * direct-IO from being able to read uninitialised disk blocks.  If its zero
- * (blockdev) this locking is not done, and if it is DIO_OWN_LOCKING i_mutex is
- * not held for the entire direct write (taken briefly, initially, during a
- * direct read though, but its never held for the duration of a direct-IO).
  */
 
 struct dio {
@@ -68,7 +61,7 @@ struct dio {
        struct inode *inode;
        int rw;
        loff_t i_size;                  /* i_size when submitted */
-       int lock_type;                  /* doesn't change */
+       int flags;                      /* doesn't change */
        unsigned blkbits;               /* doesn't change */
        unsigned blkfactor;             /* When we're using an alignment which
                                           is finer than the filesystem's soft
@@ -104,6 +97,18 @@ struct dio {
        unsigned cur_page_len;          /* Nr of bytes at cur_page_offset */
        sector_t cur_page_block;        /* Where it starts */
 
+       /* BIO completion state */
+       spinlock_t bio_lock;            /* protects BIO fields below */
+       unsigned long refcount;         /* direct_io_worker() and bios */
+       struct bio *bio_list;           /* singly linked via bi_private */
+       struct task_struct *waiter;     /* waiting task (NULL if none) */
+
+       /* AIO related stuff */
+       struct kiocb *iocb;             /* kiocb */
+       int is_async;                   /* is IO async ? */
+       int io_error;                   /* IO error in completion path */
+       ssize_t result;                 /* IO result */
+
        /*
         * Page fetching state. These variables belong to dio_refill_pages().
         */
@@ -115,22 +120,16 @@ struct dio {
         * Page queue.  These variables belong to dio_refill_pages() and
         * dio_get_page().
         */
-       struct page *pages[DIO_PAGES];  /* page buffer */
        unsigned head;                  /* next page to process */
        unsigned tail;                  /* last valid page + 1 */
        int page_errors;                /* errno from get_user_pages() */
 
-       /* BIO completion state */
-       atomic_t refcount;              /* direct_io_worker() and bios */
-       spinlock_t bio_lock;            /* protects BIO fields below */
-       struct bio *bio_list;           /* singly linked via bi_private */
-       struct task_struct *waiter;     /* waiting task (NULL if none) */
-
-       /* AIO related stuff */
-       struct kiocb *iocb;             /* kiocb */
-       int is_async;                   /* is IO async ? */
-       int io_error;                   /* IO error in completion path */
-       ssize_t result;                 /* IO result */
+       /*
+        * pages[] (and any fields placed after it) are not zeroed out at
+        * allocation time.  Don't add new fields after pages[] unless you
+        * wish that they not be zeroed.
+        */
+       struct page *pages[DIO_PAGES];  /* page buffer */
 };
 
 /*
@@ -150,20 +149,14 @@ static int dio_refill_pages(struct dio *dio)
        int nr_pages;
 
        nr_pages = min(dio->total_pages - dio->curr_page, DIO_PAGES);
-       down_read(&current->mm->mmap_sem);
-       ret = get_user_pages(
-               current,                        /* Task for fault acounting */
-               current->mm,                    /* whose pages? */
+       ret = get_user_pages_fast(
                dio->curr_user_address,         /* Where from? */
                nr_pages,                       /* How many pages? */
                dio->rw == READ,                /* Write to memory? */
-               0,                              /* force (?) */
-               &dio->pages[0],
-               NULL);                          /* vmas */
-       up_read(&current->mm->mmap_sem);
+               &dio->pages[0]);                /* Put results here */
 
        if (ret < 0 && dio->blocks_available && (dio->rw & WRITE)) {
-               struct page *page = ZERO_PAGE(dio->curr_user_address);
+               struct page *page = ZERO_PAGE(0);
                /*
                 * A memory fault, but the filesystem has some outstanding
                 * mapped blocks.  We need to use those blocks up to avoid
@@ -226,6 +219,15 @@ static int dio_complete(struct dio *dio, loff_t offset, int ret)
 {
        ssize_t transferred = 0;
 
+       /*
+        * AIO submission can race with bio completion to get here while
+        * expecting to have the last io completed by bio completion.
+        * In that case -EIOCBQUEUED is in fact not an error we want
+        * to preserve through this call.
+        */
+       if (ret == -EIOCBQUEUED)
+               ret = 0;
+
        if (dio->result) {
                transferred = dio->result;
 
@@ -237,7 +239,8 @@ static int dio_complete(struct dio *dio, loff_t offset, int ret)
        if (dio->end_io && dio->result)
                dio->end_io(dio->iocb, offset, transferred,
                            dio->map_bh.b_private);
-       if (dio->lock_type == DIO_LOCKING)
+
+       if (dio->flags & DIO_LOCKING)
                /* lockdep: non-owner release */
                up_read_non_owner(&dio->inode->i_alloc_sem);
 
@@ -251,49 +254,30 @@ static int dio_complete(struct dio *dio, loff_t offset, int ret)
        return ret;
 }
 
-/*
- * Called when a BIO has been processed.  If the count goes to zero then IO is
- * complete and we can signal this to the AIO layer.
- */
-static void dio_complete_aio(struct dio *dio)
-{
-       int ret;
-
-       ret = dio_complete(dio, dio->iocb->ki_pos, 0);
-
-       /* Complete AIO later if falling back to buffered i/o */
-       if (dio->result == dio->size ||
-               ((dio->rw == READ) && dio->result)) {
-               aio_complete(dio->iocb, ret, 0);
-               kfree(dio);
-       }
-}
-
 static int dio_bio_complete(struct dio *dio, struct bio *bio);
 /*
  * Asynchronous IO callback. 
  */
-static int dio_bio_end_aio(struct bio *bio, unsigned int bytes_done, int error)
+static void dio_bio_end_aio(struct bio *bio, int error)
 {
        struct dio *dio = bio->bi_private;
-       int waiter_holds_ref = 0;
-       int remaining;
-
-       if (bio->bi_size)
-               return 1;
+       unsigned long remaining;
+       unsigned long flags;
 
        /* cleanup the bio */
        dio_bio_complete(dio, bio);
 
-       waiter_holds_ref = !!dio->waiter;
-       remaining = atomic_sub_return(1, (&dio->refcount));
-       if (remaining == 1 && waiter_holds_ref)
+       spin_lock_irqsave(&dio->bio_lock, flags);
+       remaining = --dio->refcount;
+       if (remaining == 1 && dio->waiter)
                wake_up_process(dio->waiter);
+       spin_unlock_irqrestore(&dio->bio_lock, flags);
 
-       if (remaining == 0)
-               dio_complete_aio(dio);
-
-       return 0;
+       if (remaining == 0) {
+               int ret = dio_complete(dio, dio->iocb->ki_pos, 0);
+               aio_complete(dio->iocb, ret, 0);
+               kfree(dio);
+       }
 }
 
 /*
@@ -303,21 +287,17 @@ static int dio_bio_end_aio(struct bio *bio, unsigned int bytes_done, int error)
  * During I/O bi_private points at the dio.  After I/O, bi_private is used to
  * implement a singly-linked list of completed BIOs, at dio->bio_list.
  */
-static int dio_bio_end_io(struct bio *bio, unsigned int bytes_done, int error)
+static void dio_bio_end_io(struct bio *bio, int error)
 {
        struct dio *dio = bio->bi_private;
        unsigned long flags;
 
-       if (bio->bi_size)
-               return 1;
-
        spin_lock_irqsave(&dio->bio_lock, flags);
        bio->bi_private = dio->bio_list;
        dio->bio_list = bio;
-       if ((atomic_sub_return(1, &dio->refcount) == 1) && dio->waiter)
+       if (--dio->refcount == 1 && dio->waiter)
                wake_up_process(dio->waiter);
        spin_unlock_irqrestore(&dio->bio_lock, flags);
-       return 0;
 }
 
 static int
@@ -327,8 +307,6 @@ dio_bio_alloc(struct dio *dio, struct block_device *bdev,
        struct bio *bio;
 
        bio = bio_alloc(GFP_KERNEL, nr_vecs);
-       if (bio == NULL)
-               return -ENOMEM;
 
        bio->bi_bdev = bdev;
        bio->bi_sector = first_sector;
@@ -351,11 +329,17 @@ dio_bio_alloc(struct dio *dio, struct block_device *bdev,
 static void dio_bio_submit(struct dio *dio)
 {
        struct bio *bio = dio->bio;
+       unsigned long flags;
 
        bio->bi_private = dio;
-       atomic_inc(&dio->refcount);
+
+       spin_lock_irqsave(&dio->bio_lock, flags);
+       dio->refcount++;
+       spin_unlock_irqrestore(&dio->bio_lock, flags);
+
        if (dio->is_async && dio->rw == READ)
                bio_set_pages_dirty(bio);
+
        submit_bio(dio->rw, bio);
 
        dio->bio = NULL;
@@ -371,13 +355,6 @@ static void dio_cleanup(struct dio *dio)
                page_cache_release(dio_get_page(dio));
 }
 
-static int wait_for_more_bios(struct dio *dio)
-{
-       assert_spin_locked(&dio->bio_lock);
-
-       return (atomic_read(&dio->refcount) > 1) && (dio->bio_list == NULL);
-}
-
 /*
  * Wait for the next BIO to complete.  Remove it and return it.  NULL is
  * returned once all BIOs have been completed.  This must only be called once
@@ -390,16 +367,21 @@ static struct bio *dio_await_one(struct dio *dio)
        struct bio *bio = NULL;
 
        spin_lock_irqsave(&dio->bio_lock, flags);
-       while (wait_for_more_bios(dio)) {
-               set_current_state(TASK_UNINTERRUPTIBLE);
-               if (wait_for_more_bios(dio)) {
-                       dio->waiter = current;
-                       spin_unlock_irqrestore(&dio->bio_lock, flags);
-                       io_schedule();
-                       spin_lock_irqsave(&dio->bio_lock, flags);
-                       dio->waiter = NULL;
-               }
-               set_current_state(TASK_RUNNING);
+
+       /*
+        * Wait as long as the list is empty and there are bios in flight.  bio
+        * completion drops the count, maybe adds to the list, and wakes while
+        * holding the bio_lock so we don't need set_current_state()'s barrier
+        * and can call it after testing our condition.
+        */
+       while (dio->refcount > 1 && dio->bio_list == NULL) {
+               __set_current_state(TASK_UNINTERRUPTIBLE);
+               dio->waiter = current;
+               spin_unlock_irqrestore(&dio->bio_lock, flags);
+               io_schedule();
+               /* wake up sets us TASK_RUNNING */
+               spin_lock_irqsave(&dio->bio_lock, flags);
+               dio->waiter = NULL;
        }
        if (dio->bio_list) {
                bio = dio->bio_list;
@@ -440,7 +422,7 @@ static int dio_bio_complete(struct dio *dio, struct bio *bio)
  * Wait on and process all in-flight BIOs.  This must only be called once
  * all bios have been issued so that the refcount can only decrease.
  * This just waits for all bios to make it through dio_bio_complete.  IO
- * errors are propogated through dio->io_error and should be propogated via
+ * errors are propagated through dio->io_error and should be propagated via
  * dio_complete().
  */
 static void dio_await_completion(struct dio *dio)
@@ -533,21 +515,24 @@ static int get_more_blocks(struct dio *dio)
                map_bh->b_state = 0;
                map_bh->b_size = fs_count << dio->inode->i_blkbits;
 
+               /*
+                * For writes inside i_size on a DIO_SKIP_HOLES filesystem we
+                * forbid block creations: only overwrites are permitted.
+                * We will return early to the caller once we see an
+                * unmapped buffer head returned, and the caller will fall
+                * back to buffered I/O.
+                *
+                * Otherwise the decision is left to the get_blocks method,
+                * which may decide to handle it or also return an unmapped
+                * buffer head.
+                */
                create = dio->rw & WRITE;
-               if (dio->lock_type == DIO_LOCKING) {
+               if (dio->flags & DIO_SKIP_HOLES) {
                        if (dio->block_in_file < (i_size_read(dio->inode) >>
                                                        dio->blkbits))
                                create = 0;
-               } else if (dio->lock_type == DIO_NO_LOCKING) {
-                       create = 0;
                }
 
-               /*
-                * For writes inside i_size we forbid block creations: only
-                * overwrites are permitted.  We fall back to buffered writes
-                * at a higher level for inside-i_size block-instantiating
-                * writes.
-                */
                ret = (*dio->get_block)(dio->inode, fs_startblk,
                                                map_bh, create);
        }
@@ -773,7 +758,7 @@ static void dio_zero_block(struct dio *dio, int end)
 
        this_chunk_bytes = this_chunk_blocks << dio->blkbits;
 
-       page = ZERO_PAGE(dio->curr_user_address);
+       page = ZERO_PAGE(0);
        if (submit_page_section(dio, page, 0, this_chunk_bytes, 
                                dio->next_block_for_io))
                return;
@@ -868,7 +853,6 @@ static int do_direct_IO(struct dio *dio)
 do_holes:
                        /* Handle holes */
                        if (!buffer_mapped(map_bh)) {
-                               char *kaddr;
                                loff_t i_size_aligned;
 
                                /* AKPM: eargh, -ENOTBLK is a hack */
@@ -889,11 +873,8 @@ do_holes:
                                        page_cache_release(page);
                                        goto out;
                                }
-                               kaddr = kmap_atomic(page, KM_USER0);
-                               memset(kaddr + (block_in_page << blkbits),
-                                               0, 1 << blkbits);
-                               flush_dcache_page(page);
-                               kunmap_atomic(kaddr, KM_USER0);
+                               zero_user(page, block_in_page << blkbits,
+                                               1 << blkbits);
                                dio->block_in_file++;
                                block_in_page++;
                                goto next_block;
@@ -957,40 +938,28 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
        struct dio *dio)
 {
        unsigned long user_addr; 
+       unsigned long flags;
        int seg;
        ssize_t ret = 0;
        ssize_t ret2;
        size_t bytes;
 
-       dio->bio = NULL;
        dio->inode = inode;
        dio->rw = rw;
        dio->blkbits = blkbits;
        dio->blkfactor = inode->i_blkbits - blkbits;
-       dio->start_zero_done = 0;
-       dio->size = 0;
        dio->block_in_file = offset >> blkbits;
-       dio->blocks_available = 0;
-       dio->cur_page = NULL;
 
-       dio->boundary = 0;
-       dio->reap_counter = 0;
        dio->get_block = get_block;
        dio->end_io = end_io;
-       dio->map_bh.b_private = NULL;
        dio->final_block_in_bio = -1;
        dio->next_block_for_io = -1;
 
-       dio->page_errors = 0;
-       dio->io_error = 0;
-       dio->result = 0;
        dio->iocb = iocb;
        dio->i_size = i_size_read(inode);
 
-       atomic_set(&dio->refcount, 1);
        spin_lock_init(&dio->bio_lock);
-       dio->bio_list = NULL;
-       dio->waiter = NULL;
+       dio->refcount = 1;
 
        /*
         * In case of non-aligned buffers, we may need 2 more
@@ -998,8 +967,6 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
         */
        if (unlikely(dio->blkfactor))
                dio->pages_in_io = 2;
-       else
-               dio->pages_in_io = 0;
 
        for (seg = 0; seg < nr_segs; seg++) {
                user_addr = (unsigned long)iov[seg].iov_base;
@@ -1064,9 +1031,6 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
        if (dio->bio)
                dio_bio_submit(dio);
 
-       /* All IO is now issued, send it on its way */
-       blk_run_address_space(inode->i_mapping);
-
        /*
         * It is possible that, we return short IO due to end of file.
         * In that case, we need to release all the pages we got hold on.
@@ -1078,80 +1042,75 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
         * we can let i_mutex go now that its achieved its purpose
         * of protecting us from looking up uninitialized blocks.
         */
-       if ((rw == READ) && (dio->lock_type == DIO_LOCKING))
+       if (rw == READ && (dio->flags & DIO_LOCKING))
                mutex_unlock(&dio->inode->i_mutex);
 
        /*
-        * OK, all BIOs are submitted, so we can decrement bio_count to truly
-        * reflect the number of to-be-processed BIOs.
+        * The only time we want to leave bios in flight is when a successful
+        * partial aio read or full aio write have been setup.  In that case
+        * bio completion will call aio_complete.  The only time it's safe to
+        * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
+        * This had *better* be the only place that raises -EIOCBQUEUED.
         */
-       if (dio->is_async) {
-               int should_wait = 0;
-
-               if (dio->result < dio->size && (rw & WRITE)) {
-                       dio->waiter = current;
-                       should_wait = 1;
-               }
-               if (ret == 0)
-                       ret = dio->result;
-
-               if (should_wait)
-                       dio_await_completion(dio);
-
-               /* this can free the dio */
-               if (atomic_dec_and_test(&dio->refcount))
-                       dio_complete_aio(dio);
-
-               if (should_wait)
-                       kfree(dio);
-       } else {
+       BUG_ON(ret == -EIOCBQUEUED);
+       if (dio->is_async && ret == 0 && dio->result &&
+           ((rw & READ) || (dio->result == dio->size)))
+               ret = -EIOCBQUEUED;
+
+       if (ret != -EIOCBQUEUED) {
+               /* All IO is now issued, send it on its way */
+               blk_run_address_space(inode->i_mapping);
                dio_await_completion(dio);
+       }
 
-               ret = dio_complete(dio, offset, ret);
+       /*
+        * Sync will always be dropping the final ref and completing the
+        * operation.  AIO can if it was a broken operation described above or
+        * in fact if all the bios race to complete before we get here.  In
+        * that case dio_complete() translates the EIOCBQUEUED into the proper
+        * return code that the caller will hand to aio_complete().
+        *
+        * This is managed by the bio_lock instead of being an atomic_t so that
+        * completion paths can drop their ref and use the remaining count to
+        * decide to wake the submission path atomically.
+        */
+       spin_lock_irqsave(&dio->bio_lock, flags);
+       ret2 = --dio->refcount;
+       spin_unlock_irqrestore(&dio->bio_lock, flags);
 
-               /* We could have also come here on an AIO file extend */
-               if (!is_sync_kiocb(iocb) && (rw & WRITE) &&
-                   ret >= 0 && dio->result == dio->size)
-                       /*
-                        * For AIO writes where we have completed the
-                        * i/o, we have to mark the the aio complete.
-                        */
-                       aio_complete(iocb, ret, 0);
+       if (ret2 == 0) {
+               ret = dio_complete(dio, offset, ret);
+               kfree(dio);
+       } else
+               BUG_ON(ret != -EIOCBQUEUED);
 
-               if (atomic_dec_and_test(&dio->refcount))
-                       kfree(dio);
-               else
-                       BUG();
-       }
        return ret;
 }
 
 /*
  * This is a library function for use by filesystem drivers.
- * The locking rules are governed by the dio_lock_type parameter.
- *
- * DIO_NO_LOCKING (no locking, for raw block device access)
- * For writes, i_mutex is not held on entry; it is never taken.
- *
- * DIO_LOCKING (simple locking for regular files)
- * For writes we are called under i_mutex and return with i_mutex held, even
- * though it is internally dropped.
- * For reads, i_mutex is not held on entry, but it is taken and dropped before
- * returning.
  *
- * DIO_OWN_LOCKING (filesystem provides synchronisation and handling of
- *     uninitialised data, allowing parallel direct readers and writers)
- * For writes we are called without i_mutex, return without it, never touch it.
- * For reads we are called under i_mutex and return with i_mutex held, even
- * though it may be internally dropped.
+ * The locking rules are governed by the flags parameter:
+ *  - if the flags value contains DIO_LOCKING we use a fancy locking
+ *    scheme for dumb filesystems.
+ *    For writes this function is called under i_mutex and returns with
+ *    i_mutex held, for reads, i_mutex is not held on entry, but it is
+ *    taken and dropped again before returning.
+ *    For reads and writes i_alloc_sem is taken in shared mode and released
+ *    on I/O completion (which may happen asynchronously after returning to
+ *    the caller).
  *
- * Additional i_alloc_sem locking requirements described inline below.
+ *  - if the flags value does NOT contain DIO_LOCKING we don't use any
+ *    internal locking but rather rely on the filesystem to synchronize
+ *    direct I/O reads/writes versus each other and truncate.
+ *    For reads and writes both i_mutex and i_alloc_sem are not held on
+ *    entry and are never taken.
  */
 ssize_t
 __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
        struct block_device *bdev, const struct iovec *iov, loff_t offset, 
        unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
-       int dio_lock_type)
+       int flags)
 {
        int seg;
        size_t size;
@@ -1162,14 +1121,12 @@ __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
        ssize_t retval = -EINVAL;
        loff_t end = offset;
        struct dio *dio;
-       int release_i_mutex = 0;
-       int acquire_i_mutex = 0;
 
        if (rw & WRITE)
-               rw = WRITE_SYNC;
+               rw = WRITE_ODIRECT_PLUG;
 
        if (bdev)
-               bdev_blkbits = blksize_bits(bdev_hardsect_size(bdev));
+               bdev_blkbits = blksize_bits(bdev_logical_block_size(bdev));
 
        if (offset & blocksize_mask) {
                if (bdev)
@@ -1197,44 +1154,37 @@ __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
        retval = -ENOMEM;
        if (!dio)
                goto out;
-
        /*
-        * For block device access DIO_NO_LOCKING is used,
-        *      neither readers nor writers do any locking at all
-        * For regular files using DIO_LOCKING,
-        *      readers need to grab i_mutex and i_alloc_sem
-        *      writers need to grab i_alloc_sem only (i_mutex is already held)
-        * For regular files using DIO_OWN_LOCKING,
-        *      neither readers nor writers take any locks here
+        * Believe it or not, zeroing out the page array caused a .5%
+        * performance regression in a database benchmark.  So, we take
+        * care to only zero out what's needed.
         */
-       dio->lock_type = dio_lock_type;
-       if (dio_lock_type != DIO_NO_LOCKING) {
+       memset(dio, 0, offsetof(struct dio, pages));
+
+       dio->flags = flags;
+       if (dio->flags & DIO_LOCKING) {
                /* watch out for a 0 len io from a tricksy fs */
                if (rw == READ && end > offset) {
-                       struct address_space *mapping;
+                       struct address_space *mapping =
+                                       iocb->ki_filp->f_mapping;
 
-                       mapping = iocb->ki_filp->f_mapping;
-                       if (dio_lock_type != DIO_OWN_LOCKING) {
-                               mutex_lock(&inode->i_mutex);
-                               release_i_mutex = 1;
-                       }
+                       /* will be released by direct_io_worker */
+                       mutex_lock(&inode->i_mutex);
 
                        retval = filemap_write_and_wait_range(mapping, offset,
                                                              end - 1);
                        if (retval) {
+                               mutex_unlock(&inode->i_mutex);
                                kfree(dio);
                                goto out;
                        }
-
-                       if (dio_lock_type == DIO_OWN_LOCKING) {
-                               mutex_unlock(&inode->i_mutex);
-                               acquire_i_mutex = 1;
-                       }
                }
 
-               if (dio_lock_type == DIO_LOCKING)
-                       /* lockdep: not the owner will release it */
-                       down_read_non_owner(&inode->i_alloc_sem);
+               /*
+                * Will be released at I/O completion, possibly in a
+                * different thread.
+                */
+               down_read_non_owner(&inode->i_alloc_sem);
        }
 
        /*
@@ -1249,14 +1199,22 @@ __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
        retval = direct_io_worker(rw, iocb, inode, iov, offset,
                                nr_segs, blkbits, get_block, end_io, dio);
 
-       if (rw == READ && dio_lock_type == DIO_LOCKING)
-               release_i_mutex = 0;
+       /*
+        * In case of error extending write may have instantiated a few
+        * blocks outside i_size. Trim these off again for DIO_LOCKING.
+        *
+        * NOTE: filesystems with their own locking have to handle this
+        * on their own.
+        */
+       if (flags & DIO_LOCKING) {
+               if (unlikely((rw & WRITE) && retval < 0)) {
+                       loff_t isize = i_size_read(inode);
+                       if (end > isize)
+                               vmtruncate(inode, isize);
+               }
+       }
 
 out:
-       if (release_i_mutex)
-               mutex_unlock(&inode->i_mutex);
-       else if (acquire_i_mutex)
-               mutex_lock(&inode->i_mutex);
        return retval;
 }
 EXPORT_SYMBOL(__blockdev_direct_IO);