xfs: remove nr_to_write writeback windup.
[safe/jmp/linux-2.6] / fs / xfs / linux-2.6 / xfs_aops.c
index a79b84f..34640d6 100644 (file)
@@ -21,7 +21,6 @@
 #include "xfs_inum.h"
 #include "xfs_sb.h"
 #include "xfs_ag.h"
-#include "xfs_dir.h"
 #include "xfs_dir2.h"
 #include "xfs_trans.h"
 #include "xfs_dmapi.h"
@@ -29,7 +28,6 @@
 #include "xfs_bmap_btree.h"
 #include "xfs_alloc_btree.h"
 #include "xfs_ialloc_btree.h"
-#include "xfs_dir_sf.h"
 #include "xfs_dir2_sf.h"
 #include "xfs_attr_sf.h"
 #include "xfs_dinode.h"
 #include "xfs_error.h"
 #include "xfs_rw.h"
 #include "xfs_iomap.h"
+#include "xfs_vnodeops.h"
+#include "xfs_trace.h"
+#include "xfs_bmap.h"
+#include <linux/gfp.h>
 #include <linux/mpage.h>
 #include <linux/pagevec.h>
 #include <linux/writeback.h>
 
+/*
+ * Types of I/O for bmap clustering and I/O completion tracking.
+ */
+enum {
+       IO_READ,        /* mapping for a read */
+       IO_DELAY,       /* mapping covers delalloc region */
+       IO_UNWRITTEN,   /* mapping covers allocated but uninitialized data */
+       IO_NEW          /* just allocated */
+};
+
+/*
+ * Prime number of hash buckets since address is used as the key.
+ */
+#define NVSYNC         37
+#define to_ioend_wq(v) (&xfs_ioend_wq[((unsigned long)v) % NVSYNC])
+static wait_queue_head_t xfs_ioend_wq[NVSYNC];
+
+void __init
+xfs_ioend_init(void)
+{
+       int i;
+
+       for (i = 0; i < NVSYNC; i++)
+               init_waitqueue_head(&xfs_ioend_wq[i]);
+}
+
+void
+xfs_ioend_wait(
+       xfs_inode_t     *ip)
+{
+       wait_queue_head_t *wq = to_ioend_wq(ip);
+
+       wait_event(*wq, (atomic_read(&ip->i_iocount) == 0));
+}
+
 STATIC void
+xfs_ioend_wake(
+       xfs_inode_t     *ip)
+{
+       if (atomic_dec_and_test(&ip->i_iocount))
+               wake_up(to_ioend_wq(ip));
+}
+
+void
 xfs_count_page_state(
        struct page             *page,
        int                     *delalloc,
@@ -58,8 +103,6 @@ xfs_count_page_state(
        do {
                if (buffer_uptodate(bh) && !buffer_mapped(bh))
                        (*unmapped) = 1;
-               else if (buffer_unwritten(bh) && !buffer_delay(bh))
-                       clear_buffer_unwritten(bh);
                else if (buffer_unwritten(bh))
                        (*unwritten) = 1;
                else if (buffer_delay(bh))
@@ -67,59 +110,17 @@ xfs_count_page_state(
        } while ((bh = bh->b_this_page) != head);
 }
 
-#if defined(XFS_RW_TRACE)
-void
-xfs_page_trace(
-       int             tag,
-       struct inode    *inode,
-       struct page     *page,
-       int             mask)
+STATIC struct block_device *
+xfs_find_bdev_for_inode(
+       struct inode            *inode)
 {
-       xfs_inode_t     *ip;
-       vnode_t         *vp = vn_from_inode(inode);
-       loff_t          isize = i_size_read(inode);
-       loff_t          offset = page_offset(page);
-       int             delalloc = -1, unmapped = -1, unwritten = -1;
-
-       if (page_has_buffers(page))
-               xfs_count_page_state(page, &delalloc, &unmapped, &unwritten);
-
-       ip = xfs_vtoi(vp);
-       if (!ip->i_rwtrace)
-               return;
-
-       ktrace_enter(ip->i_rwtrace,
-               (void *)((unsigned long)tag),
-               (void *)ip,
-               (void *)inode,
-               (void *)page,
-               (void *)((unsigned long)mask),
-               (void *)((unsigned long)((ip->i_d.di_size >> 32) & 0xffffffff)),
-               (void *)((unsigned long)(ip->i_d.di_size & 0xffffffff)),
-               (void *)((unsigned long)((isize >> 32) & 0xffffffff)),
-               (void *)((unsigned long)(isize & 0xffffffff)),
-               (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
-               (void *)((unsigned long)(offset & 0xffffffff)),
-               (void *)((unsigned long)delalloc),
-               (void *)((unsigned long)unmapped),
-               (void *)((unsigned long)unwritten),
-               (void *)((unsigned long)current_pid()),
-               (void *)NULL);
-}
-#else
-#define xfs_page_trace(tag, inode, page, mask)
-#endif
+       struct xfs_inode        *ip = XFS_I(inode);
+       struct xfs_mount        *mp = ip->i_mount;
 
-/*
- * Schedule IO completion handling on a xfsdatad if this was
- * the final hold on this ioend.
- */
-STATIC void
-xfs_finish_ioend(
-       xfs_ioend_t             *ioend)
-{
-       if (atomic_dec_and_test(&ioend->io_remaining))
-               queue_work(xfsdatad_workqueue, &ioend->io_work);
+       if (XFS_IS_REALTIME_INODE(ip))
+               return mp->m_rtdev_targp->bt_bdev;
+       else
+               return mp->m_ddev_targp->bt_bdev;
 }
 
 /*
@@ -133,61 +134,149 @@ xfs_destroy_ioend(
        xfs_ioend_t             *ioend)
 {
        struct buffer_head      *bh, *next;
+       struct xfs_inode        *ip = XFS_I(ioend->io_inode);
 
        for (bh = ioend->io_buffer_head; bh; bh = next) {
                next = bh->b_private;
-               bh->b_end_io(bh, ioend->io_uptodate);
+               bh->b_end_io(bh, !ioend->io_error);
+       }
+
+       /*
+        * Volume managers supporting multiple paths can send back ENODEV
+        * when the final path disappears.  In this case continuing to fill
+        * the page cache with dirty data which cannot be written out is
+        * evil, so prevent that.
+        */
+       if (unlikely(ioend->io_error == -ENODEV)) {
+               xfs_do_force_shutdown(ip->i_mount, SHUTDOWN_DEVICE_REQ,
+                                     __FILE__, __LINE__);
        }
 
-       vn_iowake(ioend->io_vnode);
+       xfs_ioend_wake(ip);
        mempool_free(ioend, xfs_ioend_pool);
 }
 
 /*
- * Buffered IO write completion for delayed allocate extents.
- * TODO: Update ondisk isize now that we know the file data
- * has been flushed (i.e. the notorious "NULL file" problem).
+ * If the end of the current ioend is beyond the current EOF,
+ * return the new EOF value, otherwise zero.
  */
-STATIC void
-xfs_end_bio_delalloc(
-       void                    *data)
+STATIC xfs_fsize_t
+xfs_ioend_new_eof(
+       xfs_ioend_t             *ioend)
 {
-       xfs_ioend_t             *ioend = data;
+       xfs_inode_t             *ip = XFS_I(ioend->io_inode);
+       xfs_fsize_t             isize;
+       xfs_fsize_t             bsize;
+
+       bsize = ioend->io_offset + ioend->io_size;
+       isize = MAX(ip->i_size, ip->i_new_size);
+       isize = MIN(isize, bsize);
+       return isize > ip->i_d.di_size ? isize : 0;
+}
+
+/*
+ * Update on-disk file size now that data has been written to disk.  The
+ * current in-memory file size is i_size.  If a write is beyond eof i_new_size
+ * will be the intended file size until i_size is updated.  If this write does
+ * not extend all the way to the valid file size then restrict this update to
+ * the end of the write.
+ *
+ * This function does not block as blocking on the inode lock in IO completion
+ * can lead to IO completion order dependency deadlocks.. If it can't get the
+ * inode ilock it will return EAGAIN. Callers must handle this.
+ */
+STATIC int
+xfs_setfilesize(
+       xfs_ioend_t             *ioend)
+{
+       xfs_inode_t             *ip = XFS_I(ioend->io_inode);
+       xfs_fsize_t             isize;
+
+       ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFREG);
+       ASSERT(ioend->io_type != IO_READ);
+
+       if (unlikely(ioend->io_error))
+               return 0;
 
-       xfs_destroy_ioend(ioend);
+       if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
+               return EAGAIN;
+
+       isize = xfs_ioend_new_eof(ioend);
+       if (isize) {
+               ip->i_d.di_size = isize;
+               xfs_mark_inode_dirty(ip);
+       }
+
+       xfs_iunlock(ip, XFS_ILOCK_EXCL);
+       return 0;
 }
 
 /*
- * Buffered IO write completion for regular, written extents.
+ * Schedule IO completion handling on a xfsdatad if this was
+ * the final hold on this ioend. If we are asked to wait,
+ * flush the workqueue.
  */
 STATIC void
-xfs_end_bio_written(
-       void                    *data)
+xfs_finish_ioend(
+       xfs_ioend_t     *ioend,
+       int             wait)
 {
-       xfs_ioend_t             *ioend = data;
-
-       xfs_destroy_ioend(ioend);
+       if (atomic_dec_and_test(&ioend->io_remaining)) {
+               struct workqueue_struct *wq;
+
+               wq = (ioend->io_type == IO_UNWRITTEN) ?
+                       xfsconvertd_workqueue : xfsdatad_workqueue;
+               queue_work(wq, &ioend->io_work);
+               if (wait)
+                       flush_workqueue(wq);
+       }
 }
 
 /*
- * IO write completion for unwritten extents.
- *
- * Issue transactions to convert a buffer range from unwritten
- * to written extents.
+ * IO write completion.
  */
 STATIC void
-xfs_end_bio_unwritten(
-       void                    *data)
+xfs_end_io(
+       struct work_struct *work)
 {
-       xfs_ioend_t             *ioend = data;
-       vnode_t                 *vp = ioend->io_vnode;
-       xfs_off_t               offset = ioend->io_offset;
-       size_t                  size = ioend->io_size;
-       int                     error;
+       xfs_ioend_t     *ioend = container_of(work, xfs_ioend_t, io_work);
+       struct xfs_inode *ip = XFS_I(ioend->io_inode);
+       int             error = 0;
+
+       /*
+        * For unwritten extents we need to issue transactions to convert a
+        * range to normal written extens after the data I/O has finished.
+        */
+       if (ioend->io_type == IO_UNWRITTEN &&
+           likely(!ioend->io_error && !XFS_FORCED_SHUTDOWN(ip->i_mount))) {
+
+               error = xfs_iomap_write_unwritten(ip, ioend->io_offset,
+                                                ioend->io_size);
+               if (error)
+                       ioend->io_error = error;
+       }
+
+       /*
+        * We might have to update the on-disk file size after extending
+        * writes.
+        */
+       if (ioend->io_type != IO_READ) {
+               error = xfs_setfilesize(ioend);
+               ASSERT(!error || error == EAGAIN);
+       }
 
-       if (ioend->io_uptodate)
-               VOP_BMAP(vp, offset, size, BMAPI_UNWRITTEN, NULL, NULL, error);
-       xfs_destroy_ioend(ioend);
+       /*
+        * If we didn't complete processing of the ioend, requeue it to the
+        * tail of the workqueue for another attempt later. Otherwise destroy
+        * it.
+        */
+       if (error == EAGAIN) {
+               atomic_inc(&ioend->io_remaining);
+               xfs_finish_ioend(ioend, 0);
+               /* ensure we don't spin on blocked ioends */
+               delay(1);
+       } else
+               xfs_destroy_ioend(ioend);
 }
 
 /*
@@ -211,23 +300,17 @@ xfs_alloc_ioend(
         * all the I/O from calling the completion routine too early.
         */
        atomic_set(&ioend->io_remaining, 1);
-       ioend->io_uptodate = 1; /* cleared if any I/O fails */
+       ioend->io_error = 0;
        ioend->io_list = NULL;
        ioend->io_type = type;
-       ioend->io_vnode = vn_from_inode(inode);
+       ioend->io_inode = inode;
        ioend->io_buffer_head = NULL;
        ioend->io_buffer_tail = NULL;
-       atomic_inc(&ioend->io_vnode->v_iocount);
+       atomic_inc(&XFS_I(ioend->io_inode)->i_iocount);
        ioend->io_offset = 0;
        ioend->io_size = 0;
 
-       if (type == IOMAP_UNWRITTEN)
-               INIT_WORK(&ioend->io_work, xfs_end_bio_unwritten, ioend);
-       else if (type == IOMAP_DELAY)
-               INIT_WORK(&ioend->io_work, xfs_end_bio_delalloc, ioend);
-       else
-               INIT_WORK(&ioend->io_work, xfs_end_bio_written, ioend);
-
+       INIT_WORK(&ioend->io_work, xfs_end_io);
        return ioend;
 }
 
@@ -236,66 +319,67 @@ xfs_map_blocks(
        struct inode            *inode,
        loff_t                  offset,
        ssize_t                 count,
-       xfs_iomap_t             *mapp,
+       struct xfs_bmbt_irec    *imap,
        int                     flags)
 {
-       vnode_t                 *vp = vn_from_inode(inode);
-       int                     error, nmaps = 1;
+       int                     nmaps = 1;
+       int                     new = 0;
 
-       VOP_BMAP(vp, offset, count, flags, mapp, &nmaps, error);
-       if (!error && (flags & (BMAPI_WRITE|BMAPI_ALLOCATE)))
-               VMODIFY(vp);
-       return -error;
+       return -xfs_iomap(XFS_I(inode), offset, count, flags, imap, &nmaps, &new);
 }
 
-STATIC inline int
-xfs_iomap_valid(
-       xfs_iomap_t             *iomapp,
-       loff_t                  offset)
+STATIC int
+xfs_imap_valid(
+       struct inode            *inode,
+       struct xfs_bmbt_irec    *imap,
+       xfs_off_t               offset)
 {
-       return offset >= iomapp->iomap_offset &&
-               offset < iomapp->iomap_offset + iomapp->iomap_bsize;
+       offset >>= inode->i_blkbits;
+
+       return offset >= imap->br_startoff &&
+               offset < imap->br_startoff + imap->br_blockcount;
 }
 
 /*
  * BIO completion handler for buffered IO.
  */
-STATIC int
+STATIC void
 xfs_end_bio(
        struct bio              *bio,
-       unsigned int            bytes_done,
        int                     error)
 {
        xfs_ioend_t             *ioend = bio->bi_private;
 
-       if (bio->bi_size)
-               return 1;
-
-       ASSERT(ioend);
        ASSERT(atomic_read(&bio->bi_cnt) >= 1);
+       ioend->io_error = test_bit(BIO_UPTODATE, &bio->bi_flags) ? 0 : error;
 
        /* Toss bio and pass work off to an xfsdatad thread */
-       if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
-               ioend->io_uptodate = 0;
        bio->bi_private = NULL;
        bio->bi_end_io = NULL;
-
        bio_put(bio);
-       xfs_finish_ioend(ioend);
-       return 0;
+
+       xfs_finish_ioend(ioend, 0);
 }
 
 STATIC void
 xfs_submit_ioend_bio(
-       xfs_ioend_t     *ioend,
-       struct bio      *bio)
+       struct writeback_control *wbc,
+       xfs_ioend_t             *ioend,
+       struct bio              *bio)
 {
        atomic_inc(&ioend->io_remaining);
-
        bio->bi_private = ioend;
        bio->bi_end_io = xfs_end_bio;
 
-       submit_bio(WRITE, bio);
+       /*
+        * If the I/O is beyond EOF we mark the inode dirty immediately
+        * but don't update the inode size until I/O completion.
+        */
+       if (xfs_ioend_new_eof(ioend))
+               xfs_mark_inode_dirty(XFS_I(ioend->io_inode));
+
+       submit_bio(wbc->sync_mode == WB_SYNC_ALL ?
+                  WRITE_SYNC_PLUG : WRITE, bio);
        ASSERT(!bio_flagged(bio, BIO_EOPNOTSUPP));
        bio_put(bio);
 }
@@ -336,20 +420,18 @@ xfs_start_buffer_writeback(
 STATIC void
 xfs_start_page_writeback(
        struct page             *page,
-       struct writeback_control *wbc,
        int                     clear_dirty,
        int                     buffers)
 {
        ASSERT(PageLocked(page));
        ASSERT(!PageWriteback(page));
-       set_page_writeback(page);
        if (clear_dirty)
-               clear_page_dirty(page);
+               clear_page_dirty_for_io(page);
+       set_page_writeback(page);
        unlock_page(page);
-       if (!buffers) {
+       /* If no buffers on the page are to be written, finish it here */
+       if (!buffers)
                end_page_writeback(page);
-               wbc->pages_skipped++;   /* We didn't write this page */
-       }
 }
 
 static inline int bio_add_buffer(struct bio *bio, struct buffer_head *bh)
@@ -372,10 +454,11 @@ static inline int bio_add_buffer(struct bio *bio, struct buffer_head *bh)
  * assumes that all buffers on the page are started at the same time.
  *
  * The fix is two passes across the ioend list - one to start writeback on the
- * bufferheads, and then the second one submit them for I/O.
+ * buffer_heads, and then submit them for I/O on the second pass.
  */
 STATIC void
 xfs_submit_ioend(
+       struct writeback_control *wbc,
        xfs_ioend_t             *ioend)
 {
        xfs_ioend_t             *head = ioend;
@@ -404,20 +487,20 @@ xfs_submit_ioend(
  retry:
                                bio = xfs_alloc_ioend_bio(bh);
                        } else if (bh->b_blocknr != lastblock + 1) {
-                               xfs_submit_ioend_bio(ioend, bio);
+                               xfs_submit_ioend_bio(wbc, ioend, bio);
                                goto retry;
                        }
 
                        if (bio_add_buffer(bio, bh) != bh->b_size) {
-                               xfs_submit_ioend_bio(ioend, bio);
+                               xfs_submit_ioend_bio(wbc, ioend, bio);
                                goto retry;
                        }
 
                        lastblock = bh->b_blocknr;
                }
                if (bio)
-                       xfs_submit_ioend_bio(ioend, bio);
-               xfs_finish_ioend(ioend);
+                       xfs_submit_ioend_bio(wbc, ioend, bio);
+               xfs_finish_ioend(ioend, 0);
        } while ((ioend = next) != NULL);
 }
 
@@ -442,7 +525,7 @@ xfs_cancel_ioend(
                        unlock_buffer(bh);
                } while ((bh = next_bh) != NULL);
 
-               vn_iowake(ioend->io_vnode);
+               xfs_ioend_wake(XFS_I(ioend->io_inode));
                mempool_free(ioend, xfs_ioend_pool);
        } while ((ioend = next) != NULL);
 }
@@ -485,19 +568,23 @@ xfs_add_to_ioend(
 
 STATIC void
 xfs_map_buffer(
+       struct inode            *inode,
        struct buffer_head      *bh,
-       xfs_iomap_t             *mp,
-       xfs_off_t               offset,
-       uint                    block_bits)
+       struct xfs_bmbt_irec    *imap,
+       xfs_off_t               offset)
 {
        sector_t                bn;
+       struct xfs_mount        *m = XFS_I(inode)->i_mount;
+       xfs_off_t               iomap_offset = XFS_FSB_TO_B(m, imap->br_startoff);
+       xfs_daddr_t             iomap_bn = xfs_fsb_to_db(XFS_I(inode), imap->br_startblock);
 
-       ASSERT(mp->iomap_bn != IOMAP_DADDR_NULL);
+       ASSERT(imap->br_startblock != HOLESTARTBLOCK);
+       ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
 
-       bn = (mp->iomap_bn >> (block_bits - BBSHIFT)) +
-             ((offset - mp->iomap_offset) >> block_bits);
+       bn = (iomap_bn >> (inode->i_blkbits - BBSHIFT)) +
+             ((offset - iomap_offset) >> inode->i_blkbits);
 
-       ASSERT(bn || (mp->iomap_flags & IOMAP_REALTIME));
+       ASSERT(bn || XFS_IS_REALTIME_INODE(XFS_I(inode)));
 
        bh->b_blocknr = bn;
        set_buffer_mapped(bh);
@@ -505,17 +592,17 @@ xfs_map_buffer(
 
 STATIC void
 xfs_map_at_offset(
+       struct inode            *inode,
        struct buffer_head      *bh,
-       loff_t                  offset,
-       int                     block_bits,
-       xfs_iomap_t             *iomapp)
+       struct xfs_bmbt_irec    *imap,
+       xfs_off_t               offset)
 {
-       ASSERT(!(iomapp->iomap_flags & IOMAP_HOLE));
-       ASSERT(!(iomapp->iomap_flags & IOMAP_DELAY));
+       ASSERT(imap->br_startblock != HOLESTARTBLOCK);
+       ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
 
        lock_buffer(bh);
-       xfs_map_buffer(bh, iomapp, offset, block_bits);
-       bh->b_bdev = iomapp->iomap_target->bt_bdev;
+       xfs_map_buffer(inode, bh, imap, offset);
+       bh->b_bdev = xfs_find_bdev_for_inode(inode);
        set_buffer_mapped(bh);
        clear_buffer_delay(bh);
        clear_buffer_unwritten(bh);
@@ -592,7 +679,7 @@ xfs_probe_cluster(
 
                for (i = 0; i < pagevec_count(&pvec); i++) {
                        struct page *page = pvec.pages[i];
-                       size_t pg_offset, len = 0;
+                       size_t pg_offset, pg_len = 0;
 
                        if (tindex == tlast) {
                                pg_offset =
@@ -604,17 +691,17 @@ xfs_probe_cluster(
                        } else
                                pg_offset = PAGE_CACHE_SIZE;
 
-                       if (page->index == tindex && !TestSetPageLocked(page)) {
-                               len = xfs_probe_page(page, pg_offset, mapped);
+                       if (page->index == tindex && trylock_page(page)) {
+                               pg_len = xfs_probe_page(page, pg_offset, mapped);
                                unlock_page(page);
                        }
 
-                       if (!len) {
+                       if (!pg_len) {
                                done = 1;
                                break;
                        }
 
-                       total += len;
+                       total += pg_len;
                        tindex++;
                }
 
@@ -644,11 +731,11 @@ xfs_is_delayed_page(
                bh = head = page_buffers(page);
                do {
                        if (buffer_unwritten(bh))
-                               acceptable = (type == IOMAP_UNWRITTEN);
+                               acceptable = (type == IO_UNWRITTEN);
                        else if (buffer_delay(bh))
-                               acceptable = (type == IOMAP_DELAY);
+                               acceptable = (type == IO_DELAY);
                        else if (buffer_dirty(bh) && buffer_mapped(bh))
-                               acceptable = (type == 0);
+                               acceptable = (type == IO_NEW);
                        else
                                break;
                } while ((bh = bh->b_this_page) != head);
@@ -671,7 +758,7 @@ xfs_convert_page(
        struct inode            *inode,
        struct page             *page,
        loff_t                  tindex,
-       xfs_iomap_t             *mp,
+       struct xfs_bmbt_irec    *imap,
        xfs_ioend_t             **ioendp,
        struct writeback_control *wbc,
        int                     startio,
@@ -681,14 +768,13 @@ xfs_convert_page(
        xfs_off_t               end_offset;
        unsigned long           p_offset;
        unsigned int            type;
-       int                     bbits = inode->i_blkbits;
        int                     len, page_dirty;
        int                     count = 0, done = 0, uptodate = 1;
        xfs_off_t               offset = page_offset(page);
 
        if (page->index != tindex)
                goto fail;
-       if (TestSetPageLocked(page))
+       if (!trylock_page(page))
                goto fail;
        if (PageWriteback(page))
                goto fail_unlock_page;
@@ -699,7 +785,7 @@ xfs_convert_page(
 
        /*
         * page_dirty is initially a count of buffers on the page before
-        * EOF and is decrememted as we move each into a cleanable state.
+        * EOF and is decremented as we move each into a cleanable state.
         *
         * Derivation:
         *
@@ -733,19 +819,19 @@ xfs_convert_page(
 
                if (buffer_unwritten(bh) || buffer_delay(bh)) {
                        if (buffer_unwritten(bh))
-                               type = IOMAP_UNWRITTEN;
+                               type = IO_UNWRITTEN;
                        else
-                               type = IOMAP_DELAY;
+                               type = IO_DELAY;
 
-                       if (!xfs_iomap_valid(mp, offset)) {
+                       if (!xfs_imap_valid(inode, imap, offset)) {
                                done = 1;
                                continue;
                        }
 
-                       ASSERT(!(mp->iomap_flags & IOMAP_HOLE));
-                       ASSERT(!(mp->iomap_flags & IOMAP_DELAY));
+                       ASSERT(imap->br_startblock != HOLESTARTBLOCK);
+                       ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
 
-                       xfs_map_at_offset(bh, offset, bbits, mp);
+                       xfs_map_at_offset(inode, bh, imap, offset);
                        if (startio) {
                                xfs_add_to_ioend(inode, bh, offset,
                                                type, ioendp, done);
@@ -757,7 +843,7 @@ xfs_convert_page(
                        page_dirty--;
                        count++;
                } else {
-                       type = 0;
+                       type = IO_NEW;
                        if (buffer_mapped(bh) && all_bh && startio) {
                                lock_buffer(bh);
                                xfs_add_to_ioend(inode, bh, offset,
@@ -775,18 +861,11 @@ xfs_convert_page(
 
        if (startio) {
                if (count) {
-                       struct backing_dev_info *bdi;
-
-                       bdi = inode->i_mapping->backing_dev_info;
                        wbc->nr_to_write--;
-                       if (bdi_write_congested(bdi)) {
-                               wbc->encountered_congestion = 1;
-                               done = 1;
-                       } else if (wbc->nr_to_write <= 0) {
+                       if (wbc->nr_to_write <= 0)
                                done = 1;
-                       }
                }
-               xfs_start_page_writeback(page, wbc, !page_dirty, count);
+               xfs_start_page_writeback(page, !page_dirty, count);
        }
 
        return done;
@@ -804,7 +883,7 @@ STATIC void
 xfs_cluster_write(
        struct inode            *inode,
        pgoff_t                 tindex,
-       xfs_iomap_t             *iomapp,
+       struct xfs_bmbt_irec    *imap,
        xfs_ioend_t             **ioendp,
        struct writeback_control *wbc,
        int                     startio,
@@ -823,7 +902,7 @@ xfs_cluster_write(
 
                for (i = 0; i < pagevec_count(&pvec); i++) {
                        done = xfs_convert_page(inode, pvec.pages[i], tindex++,
-                                       iomapp, ioendp, wbc, startio, all_bh);
+                                       imap, ioendp, wbc, startio, all_bh);
                        if (done)
                                break;
                }
@@ -833,6 +912,125 @@ xfs_cluster_write(
        }
 }
 
+STATIC void
+xfs_vm_invalidatepage(
+       struct page             *page,
+       unsigned long           offset)
+{
+       trace_xfs_invalidatepage(page->mapping->host, page, offset);
+       block_invalidatepage(page, offset);
+}
+
+/*
+ * If the page has delalloc buffers on it, we need to punch them out before we
+ * invalidate the page. If we don't, we leave a stale delalloc mapping on the
+ * inode that can trip a BUG() in xfs_get_blocks() later on if a direct IO read
+ * is done on that same region - the delalloc extent is returned when none is
+ * supposed to be there.
+ *
+ * We prevent this by truncating away the delalloc regions on the page before
+ * invalidating it. Because they are delalloc, we can do this without needing a
+ * transaction. Indeed - if we get ENOSPC errors, we have to be able to do this
+ * truncation without a transaction as there is no space left for block
+ * reservation (typically why we see a ENOSPC in writeback).
+ *
+ * This is not a performance critical path, so for now just do the punching a
+ * buffer head at a time.
+ */
+STATIC void
+xfs_aops_discard_page(
+       struct page             *page)
+{
+       struct inode            *inode = page->mapping->host;
+       struct xfs_inode        *ip = XFS_I(inode);
+       struct buffer_head      *bh, *head;
+       loff_t                  offset = page_offset(page);
+       ssize_t                 len = 1 << inode->i_blkbits;
+
+       if (!xfs_is_delayed_page(page, IO_DELAY))
+               goto out_invalidate;
+
+       if (XFS_FORCED_SHUTDOWN(ip->i_mount))
+               goto out_invalidate;
+
+       xfs_fs_cmn_err(CE_ALERT, ip->i_mount,
+               "page discard on page %p, inode 0x%llx, offset %llu.",
+                       page, ip->i_ino, offset);
+
+       xfs_ilock(ip, XFS_ILOCK_EXCL);
+       bh = head = page_buffers(page);
+       do {
+               int             done;
+               xfs_fileoff_t   offset_fsb;
+               xfs_bmbt_irec_t imap;
+               int             nimaps = 1;
+               int             error;
+               xfs_fsblock_t   firstblock;
+               xfs_bmap_free_t flist;
+
+               if (!buffer_delay(bh))
+                       goto next_buffer;
+
+               offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
+
+               /*
+                * Map the range first and check that it is a delalloc extent
+                * before trying to unmap the range. Otherwise we will be
+                * trying to remove a real extent (which requires a
+                * transaction) or a hole, which is probably a bad idea...
+                */
+               error = xfs_bmapi(NULL, ip, offset_fsb, 1,
+                               XFS_BMAPI_ENTIRE,  NULL, 0, &imap,
+                               &nimaps, NULL, NULL);
+
+               if (error) {
+                       /* something screwed, just bail */
+                       if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
+                               xfs_fs_cmn_err(CE_ALERT, ip->i_mount,
+                               "page discard failed delalloc mapping lookup.");
+                       }
+                       break;
+               }
+               if (!nimaps) {
+                       /* nothing there */
+                       goto next_buffer;
+               }
+               if (imap.br_startblock != DELAYSTARTBLOCK) {
+                       /* been converted, ignore */
+                       goto next_buffer;
+               }
+               WARN_ON(imap.br_blockcount == 0);
+
+               /*
+                * Note: while we initialise the firstblock/flist pair, they
+                * should never be used because blocks should never be
+                * allocated or freed for a delalloc extent and hence we need
+                * don't cancel or finish them after the xfs_bunmapi() call.
+                */
+               xfs_bmap_init(&flist, &firstblock);
+               error = xfs_bunmapi(NULL, ip, offset_fsb, 1, 0, 1, &firstblock,
+                                       &flist, NULL, &done);
+
+               ASSERT(!flist.xbf_count && !flist.xbf_first);
+               if (error) {
+                       /* something screwed, just bail */
+                       if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
+                               xfs_fs_cmn_err(CE_ALERT, ip->i_mount,
+                       "page discard unable to remove delalloc mapping.");
+                       }
+                       break;
+               }
+next_buffer:
+               offset += len;
+
+       } while ((bh = bh->b_this_page) != head);
+
+       xfs_iunlock(ip, XFS_ILOCK_EXCL);
+out_invalidate:
+       xfs_vm_invalidatepage(page, 0);
+       return;
+}
+
 /*
  * Calling this without startio set means we are being asked to make a dirty
  * page ready for freeing it's buffers.  When called with startio set then
@@ -842,7 +1040,7 @@ xfs_cluster_write(
  * page if possible.
  * The bh->b_state's cannot know if any of the blocks or which block for
  * that matter are dirty due to mmap writes, and therefore bh uptodate is
- * only vaild if the page itself isn't completely uptodate.  Some layers
+ * only valid if the page itself isn't completely uptodate.  Some layers
  * may clear the page dirty flag prior to calling write page, under the
  * assumption the entire page will be written out; by not writing out the
  * whole page the page can be reused before all valid dirty data is
@@ -861,21 +1059,23 @@ xfs_page_state_convert(
        int             unmapped) /* also implies page uptodate */
 {
        struct buffer_head      *bh, *head;
-       xfs_iomap_t             iomap;
+       struct xfs_bmbt_irec    imap;
        xfs_ioend_t             *ioend = NULL, *iohead = NULL;
        loff_t                  offset;
        unsigned long           p_offset = 0;
        unsigned int            type;
        __uint64_t              end_offset;
-       pgoff_t                 end_index, last_index, tlast;
+       pgoff_t                 end_index, last_index;
        ssize_t                 size, len;
-       int                     flags, err, iomap_valid = 0, uptodate = 1;
-       int                     page_dirty, count = 0, trylock_flag = 0;
+       int                     flags, err, imap_valid = 0, uptodate = 1;
+       int                     page_dirty, count = 0;
+       int                     trylock = 0;
        int                     all_bh = unmapped;
 
-       /* wait for other IO threads? */
-       if (startio && (wbc->sync_mode == WB_SYNC_NONE && wbc->nonblocking))
-               trylock_flag |= BMAPI_TRYLOCK;
+       if (startio) {
+               if (wbc->sync_mode == WB_SYNC_NONE && wbc->nonblocking)
+                       trylock |= BMAPI_TRYLOCK;
+       }
 
        /* Is this page beyond the end of the file? */
        offset = i_size_read(inode);
@@ -892,7 +1092,7 @@ xfs_page_state_convert(
 
        /*
         * page_dirty is initially a count of buffers on the page before
-        * EOF and is decrememted as we move each into a cleanable state.
+        * EOF and is decremented as we move each into a cleanable state.
         *
         * Derivation:
         *
@@ -913,8 +1113,8 @@ xfs_page_state_convert(
 
        bh = head = page_buffers(page);
        offset = page_offset(page);
-       flags = -1;
-       type = 0;
+       flags = BMAPI_READ;
+       type = IO_NEW;
 
        /* TODO: cleanup count and page_dirty */
 
@@ -928,12 +1128,12 @@ xfs_page_state_convert(
                         * the iomap is actually still valid, but the ioend
                         * isn't.  shouldn't happen too often.
                         */
-                       iomap_valid = 0;
+                       imap_valid = 0;
                        continue;
                }
 
-               if (iomap_valid)
-                       iomap_valid = xfs_iomap_valid(&iomap, offset);
+               if (imap_valid)
+                       imap_valid = xfs_imap_valid(inode, &imap, offset);
 
                /*
                 * First case, map an unwritten extent and prepare for
@@ -944,31 +1144,40 @@ xfs_page_state_convert(
                 *
                 * Third case, an unmapped buffer was found, and we are
                 * in a path where we need to write the whole page out.
-                */
+                */
                if (buffer_unwritten(bh) || buffer_delay(bh) ||
                    ((buffer_uptodate(bh) || PageUptodate(page)) &&
                     !buffer_mapped(bh) && (unmapped || startio))) {
-                       /*
+                       int new_ioend = 0;
+
+                       /*
                         * Make sure we don't use a read-only iomap
                         */
-                       if (flags == BMAPI_READ)
-                               iomap_valid = 0;
+                       if (flags == BMAPI_READ)
+                               imap_valid = 0;
 
                        if (buffer_unwritten(bh)) {
-                               type = IOMAP_UNWRITTEN;
-                               flags = BMAPI_WRITE|BMAPI_IGNSTATE;
+                               type = IO_UNWRITTEN;
+                               flags = BMAPI_WRITE | BMAPI_IGNSTATE;
                        } else if (buffer_delay(bh)) {
-                               type = IOMAP_DELAY;
-                               flags = BMAPI_ALLOCATE;
-                               if (!startio)
-                                       flags |= trylock_flag;
+                               type = IO_DELAY;
+                               flags = BMAPI_ALLOCATE | trylock;
                        } else {
-                               type = IOMAP_NEW;
-                               flags = BMAPI_WRITE|BMAPI_MMAP;
+                               type = IO_NEW;
+                               flags = BMAPI_WRITE | BMAPI_MMAP;
                        }
 
-                       if (!iomap_valid) {
-                               if (type == IOMAP_NEW) {
+                       if (!imap_valid) {
+                               /*
+                                * if we didn't have a valid mapping then we
+                                * need to ensure that we put the new mapping
+                                * in a new ioend structure. This needs to be
+                                * done to ensure that the ioends correctly
+                                * reflect the block mappings at io completion
+                                * for unwritten extent conversion.
+                                */
+                               new_ioend = 1;
+                               if (type == IO_NEW) {
                                        size = xfs_probe_cluster(inode,
                                                        page, bh, head, 0);
                                } else {
@@ -976,18 +1185,18 @@ xfs_page_state_convert(
                                }
 
                                err = xfs_map_blocks(inode, offset, size,
-                                               &iomap, flags);
+                                               &imap, flags);
                                if (err)
                                        goto error;
-                               iomap_valid = xfs_iomap_valid(&iomap, offset);
+                               imap_valid = xfs_imap_valid(inode, &imap,
+                                                           offset);
                        }
-                       if (iomap_valid) {
-                               xfs_map_at_offset(bh, offset,
-                                               inode->i_blkbits, &iomap);
+                       if (imap_valid) {
+                               xfs_map_at_offset(inode, bh, &imap, offset);
                                if (startio) {
                                        xfs_add_to_ioend(inode, bh, offset,
                                                        type, &ioend,
-                                                       !iomap_valid);
+                                                       new_ioend);
                                } else {
                                        set_buffer_dirty(bh);
                                        unlock_buffer(bh);
@@ -1002,32 +1211,41 @@ xfs_page_state_convert(
                         * That means it must already have extents allocated
                         * underneath it. Map the extent by reading it.
                         */
-                       if (!iomap_valid || type != 0) {
+                       if (!imap_valid || flags != BMAPI_READ) {
                                flags = BMAPI_READ;
                                size = xfs_probe_cluster(inode, page, bh,
                                                                head, 1);
                                err = xfs_map_blocks(inode, offset, size,
-                                               &iomap, flags);
+                                               &imap, flags);
                                if (err)
                                        goto error;
-                               iomap_valid = xfs_iomap_valid(&iomap, offset);
+                               imap_valid = xfs_imap_valid(inode, &imap,
+                                                           offset);
                        }
 
-                       type = 0;
-                       if (!test_and_set_bit(BH_Lock, &bh->b_state)) {
+                       /*
+                        * We set the type to IO_NEW in case we are doing a
+                        * small write at EOF that is extending the file but
+                        * without needing an allocation. We need to update the
+                        * file size on I/O completion in this case so it is
+                        * the same case as having just allocated a new extent
+                        * that we are writing into for the first time.
+                        */
+                       type = IO_NEW;
+                       if (trylock_buffer(bh)) {
                                ASSERT(buffer_mapped(bh));
-                               if (iomap_valid)
+                               if (imap_valid)
                                        all_bh = 1;
                                xfs_add_to_ioend(inode, bh, offset, type,
-                                               &ioend, !iomap_valid);
+                                               &ioend, !imap_valid);
                                page_dirty--;
                                count++;
                        } else {
-                               iomap_valid = 0;
+                               imap_valid = 0;
                        }
                } else if ((buffer_uptodate(bh) || PageUptodate(page)) &&
                           (unmapped || startio)) {
-                       iomap_valid = 0;
+                       imap_valid = 0;
                }
 
                if (!iohead)
@@ -1039,18 +1257,29 @@ xfs_page_state_convert(
                SetPageUptodate(page);
 
        if (startio)
-               xfs_start_page_writeback(page, wbc, 1, count);
-
-       if (ioend && iomap_valid) {
-               offset = (iomap.iomap_offset + iomap.iomap_bsize - 1) >>
-                                       PAGE_CACHE_SHIFT;
-               tlast = min_t(pgoff_t, offset, last_index);
-               xfs_cluster_write(inode, page->index + 1, &iomap, &ioend,
-                                       wbc, startio, all_bh, tlast);
+               xfs_start_page_writeback(page, 1, count);
+
+       if (ioend && imap_valid) {
+               xfs_off_t               end_index;
+
+               end_index = imap.br_startoff + imap.br_blockcount;
+
+               /* to bytes */
+               end_index <<= inode->i_blkbits;
+
+               /* to pages */
+               end_index = (end_index - 1) >> PAGE_CACHE_SHIFT;
+
+               /* check against file size */
+               if (end_index > last_index)
+                       end_index = last_index;
+
+               xfs_cluster_write(inode, page->index + 1, &imap, &ioend,
+                                       wbc, startio, all_bh, end_index);
        }
 
        if (iohead)
-               xfs_submit_ioend(iohead);
+               xfs_submit_ioend(wbc, iohead);
 
        return page_dirty;
 
@@ -1065,7 +1294,7 @@ error:
         */
        if (err != -EAGAIN) {
                if (!unmapped)
-                       block_invalidatepage(page, 0);
+                       xfs_aops_discard_page(page);
                ClearPageUptodate(page);
        }
        return err;
@@ -1101,7 +1330,22 @@ xfs_vm_writepage(
        int                     delalloc, unmapped, unwritten;
        struct inode            *inode = page->mapping->host;
 
-       xfs_page_trace(XFS_WRITEPAGE_ENTER, inode, page, 0);
+       trace_xfs_writepage(inode, page, 0);
+
+       /*
+        * Refuse to write the page out if we are called from reclaim context.
+        *
+        * This is primarily to avoid stack overflows when called from deep
+        * used stacks in random callers for direct reclaim, but disabling
+        * reclaim for kswap is a nice side-effect as kswapd causes rather
+        * suboptimal I/O patters, too.
+        *
+        * This should really be done by the core VM, but until that happens
+        * filesystems like XFS, btrfs and ext4 have to take care of this
+        * by themselves.
+        */
+       if (current->flags & PF_MEMALLOC)
+               goto out_fail;
 
        /*
         * We need a transaction if:
@@ -1127,7 +1371,7 @@ xfs_vm_writepage(
         * then mark the page dirty again and leave the page
         * as is.
         */
-       if (PFLAGS_TEST_FSTRANS() && need_trans)
+       if (current_test_flags(PF_FSTRANS) && need_trans)
                goto out_fail;
 
        /*
@@ -1158,6 +1402,15 @@ out_unlock:
        return error;
 }
 
+STATIC int
+xfs_vm_writepages(
+       struct address_space    *mapping,
+       struct writeback_control *wbc)
+{
+       xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED);
+       return generic_writepages(mapping, wbc);
+}
+
 /*
  * Called to move a page into cleanable state - and from there
  * to be released. Possibly the page is already clean. We always
@@ -1189,7 +1442,7 @@ xfs_vm_releasepage(
                .nr_to_write = 1,
        };
 
-       xfs_page_trace(XFS_RELEASEPAGE_ENTER, inode, page, gfp_mask);
+       trace_xfs_releasepage(inode, page, 0);
 
        if (!page_has_buffers(page))
                return 0;
@@ -1204,7 +1457,7 @@ xfs_vm_releasepage(
        /* If we are already inside a transaction or the thread cannot
         * do I/O, we cannot release this page.
         */
-       if (PFLAGS_TEST_FSTRANS())
+       if (current_test_flags(PF_FSTRANS))
                return 0;
 
        /*
@@ -1223,67 +1476,72 @@ free_buffers:
 }
 
 STATIC int
-__xfs_get_block(
+__xfs_get_blocks(
        struct inode            *inode,
        sector_t                iblock,
-       unsigned long           blocks,
        struct buffer_head      *bh_result,
        int                     create,
        int                     direct,
        bmapi_flags_t           flags)
 {
-       vnode_t                 *vp = vn_from_inode(inode);
-       xfs_iomap_t             iomap;
+       struct xfs_bmbt_irec    imap;
        xfs_off_t               offset;
        ssize_t                 size;
-       int                     retpbbm = 1;
+       int                     nimap = 1;
+       int                     new = 0;
        int                     error;
 
        offset = (xfs_off_t)iblock << inode->i_blkbits;
-       if (blocks)
-               size = (ssize_t) min_t(xfs_off_t, LONG_MAX,
-                                       (xfs_off_t)blocks << inode->i_blkbits);
-       else
-               size = 1 << inode->i_blkbits;
+       ASSERT(bh_result->b_size >= (1 << inode->i_blkbits));
+       size = bh_result->b_size;
+
+       if (!create && direct && offset >= i_size_read(inode))
+               return 0;
 
-       VOP_BMAP(vp, offset, size,
-               create ? flags : BMAPI_READ, &iomap, &retpbbm, error);
+       error = xfs_iomap(XFS_I(inode), offset, size,
+                            create ? flags : BMAPI_READ, &imap, &nimap, &new);
        if (error)
                return -error;
-
-       if (retpbbm == 0)
+       if (nimap == 0)
                return 0;
 
-       if (iomap.iomap_bn != IOMAP_DADDR_NULL) {
+       if (imap.br_startblock != HOLESTARTBLOCK &&
+           imap.br_startblock != DELAYSTARTBLOCK) {
                /*
                 * For unwritten extents do not report a disk address on
                 * the read case (treat as if we're reading into a hole).
                 */
-               if (create || !(iomap.iomap_flags & IOMAP_UNWRITTEN)) {
-                       xfs_map_buffer(bh_result, &iomap, offset,
-                                      inode->i_blkbits);
-               }
-               if (create && (iomap.iomap_flags & IOMAP_UNWRITTEN)) {
+               if (create || !ISUNWRITTEN(&imap))
+                       xfs_map_buffer(inode, bh_result, &imap, offset);
+               if (create && ISUNWRITTEN(&imap)) {
                        if (direct)
                                bh_result->b_private = inode;
                        set_buffer_unwritten(bh_result);
-                       set_buffer_delay(bh_result);
                }
        }
 
-       /* If this is a realtime file, data might be on a new device */
-       bh_result->b_bdev = iomap.iomap_target->bt_bdev;
+       /*
+        * If this is a realtime file, data may be on a different device.
+        * to that pointed to from the buffer_head b_bdev currently.
+        */
+       bh_result->b_bdev = xfs_find_bdev_for_inode(inode);
 
-       /* If we previously allocated a block out beyond eof and
-        * we are now coming back to use it then we will need to
-        * flag it as new even if it has a disk address.
+       /*
+        * If we previously allocated a block out beyond eof and we are now
+        * coming back to use it then we will need to flag it as new even if it
+        * has a disk address.
+        *
+        * With sub-block writes into unwritten extents we also need to mark
+        * the buffer as new so that the unwritten parts of the buffer gets
+        * correctly zeroed.
         */
        if (create &&
            ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
-            (offset >= i_size_read(inode)) || (iomap.iomap_flags & IOMAP_NEW)))
+            (offset >= i_size_read(inode)) ||
+            (new || ISUNWRITTEN(&imap))))
                set_buffer_new(bh_result);
 
-       if (iomap.iomap_flags & IOMAP_DELAY) {
+       if (imap.br_startblock == DELAYSTARTBLOCK) {
                BUG_ON(direct);
                if (create) {
                        set_buffer_uptodate(bh_result);
@@ -1292,26 +1550,36 @@ __xfs_get_block(
                }
        }
 
-       if (blocks) {
-               ASSERT(iomap.iomap_bsize - iomap.iomap_delta > 0);
-               offset = min_t(xfs_off_t,
-                               iomap.iomap_bsize - iomap.iomap_delta,
-                               (xfs_off_t)blocks << inode->i_blkbits);
-               bh_result->b_size = (u32) min_t(xfs_off_t, UINT_MAX, offset);
+       /*
+        * If this is O_DIRECT or the mpage code calling tell them how large
+        * the mapping is, so that we can avoid repeated get_blocks calls.
+        */
+       if (direct || size > (1 << inode->i_blkbits)) {
+               xfs_off_t               mapping_size;
+
+               mapping_size = imap.br_startoff + imap.br_blockcount - iblock;
+               mapping_size <<= inode->i_blkbits;
+
+               ASSERT(mapping_size > 0);
+               if (mapping_size > size)
+                       mapping_size = size;
+               if (mapping_size > LONG_MAX)
+                       mapping_size = LONG_MAX;
+
+               bh_result->b_size = mapping_size;
        }
 
        return 0;
 }
 
 int
-xfs_get_block(
+xfs_get_blocks(
        struct inode            *inode,
        sector_t                iblock,
        struct buffer_head      *bh_result,
        int                     create)
 {
-       return __xfs_get_block(inode, iblock,
-                               bh_result->b_size >> inode->i_blkbits,
+       return __xfs_get_blocks(inode, iblock,
                                bh_result, create, 0, BMAPI_WRITE);
 }
 
@@ -1319,12 +1587,11 @@ STATIC int
 xfs_get_blocks_direct(
        struct inode            *inode,
        sector_t                iblock,
-       unsigned long           max_blocks,
        struct buffer_head      *bh_result,
        int                     create)
 {
-       return __xfs_get_block(inode, iblock, max_blocks, bh_result,
-                                       create, 1, BMAPI_WRITE|BMAPI_DIRECT);
+       return __xfs_get_blocks(inode, iblock,
+                               bh_result, create, 1, BMAPI_WRITE|BMAPI_DIRECT);
 }
 
 STATIC void
@@ -1339,26 +1606,41 @@ xfs_end_io_direct(
        /*
         * Non-NULL private data means we need to issue a transaction to
         * convert a range from unwritten to written extents.  This needs
-        * to happen from process contect but aio+dio I/O completion
+        * to happen from process context but aio+dio I/O completion
         * happens from irq context so we need to defer it to a workqueue.
-        * This is not nessecary for synchronous direct I/O, but we do
+        * This is not necessary for synchronous direct I/O, but we do
         * it anyway to keep the code uniform and simpler.
         *
+        * Well, if only it were that simple. Because synchronous direct I/O
+        * requires extent conversion to occur *before* we return to userspace,
+        * we have to wait for extent conversion to complete. Look at the
+        * iocb that has been passed to us to determine if this is AIO or
+        * not. If it is synchronous, tell xfs_finish_ioend() to kick the
+        * workqueue and wait for it to complete.
+        *
         * The core direct I/O code might be changed to always call the
         * completion handler in the future, in which case all this can
         * go away.
         */
-       if (private && size > 0) {
-               ioend->io_offset = offset;
-               ioend->io_size = size;
-               xfs_finish_ioend(ioend);
+       ioend->io_offset = offset;
+       ioend->io_size = size;
+       if (ioend->io_type == IO_READ) {
+               xfs_finish_ioend(ioend, 0);
+       } else if (private && size > 0) {
+               xfs_finish_ioend(ioend, is_sync_kiocb(iocb));
        } else {
-               ASSERT(size >= 0);
-               xfs_destroy_ioend(ioend);
+               /*
+                * A direct I/O write ioend starts it's life in unwritten
+                * state in case they map an unwritten extent.  This write
+                * didn't map an unwritten extent so switch it's completion
+                * handler.
+                */
+               ioend->io_type = IO_NEW;
+               xfs_finish_ioend(ioend, 0);
        }
 
        /*
-        * blockdev_direct_IO can return an error even afer the I/O
+        * blockdev_direct_IO can return an error even after the I/O
         * completion handler was called.  Thus we need to protect
         * against double-freeing.
         */
@@ -1375,37 +1657,37 @@ xfs_vm_direct_IO(
 {
        struct file     *file = iocb->ki_filp;
        struct inode    *inode = file->f_mapping->host;
-       vnode_t         *vp = vn_from_inode(inode);
-       xfs_iomap_t     iomap;
-       int             maps = 1;
-       int             error;
+       struct block_device *bdev;
        ssize_t         ret;
 
-       VOP_BMAP(vp, offset, 0, BMAPI_DEVICE, &iomap, &maps, error);
-       if (error)
-               return -error;
+       bdev = xfs_find_bdev_for_inode(inode);
 
-       iocb->private = xfs_alloc_ioend(inode, IOMAP_UNWRITTEN);
+       iocb->private = xfs_alloc_ioend(inode, rw == WRITE ?
+                                       IO_UNWRITTEN : IO_READ);
 
-       ret = blockdev_direct_IO_own_locking(rw, iocb, inode,
-               iomap.iomap_target->bt_bdev,
-               iov, offset, nr_segs,
-               xfs_get_blocks_direct,
-               xfs_end_io_direct);
+       ret = blockdev_direct_IO_no_locking(rw, iocb, inode, bdev, iov,
+                                           offset, nr_segs,
+                                           xfs_get_blocks_direct,
+                                           xfs_end_io_direct);
 
-       if (unlikely(ret <= 0 && iocb->private))
+       if (unlikely(ret != -EIOCBQUEUED && iocb->private))
                xfs_destroy_ioend(iocb->private);
        return ret;
 }
 
 STATIC int
-xfs_vm_prepare_write(
+xfs_vm_write_begin(
        struct file             *file,
-       struct page             *page,
-       unsigned int            from,
-       unsigned int            to)
+       struct address_space    *mapping,
+       loff_t                  pos,
+       unsigned                len,
+       unsigned                flags,
+       struct page             **pagep,
+       void                    **fsdata)
 {
-       return block_prepare_write(page, from, to, xfs_get_block);
+       *pagep = NULL;
+       return block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
+                                                               xfs_get_blocks);
 }
 
 STATIC sector_t
@@ -1414,15 +1696,13 @@ xfs_vm_bmap(
        sector_t                block)
 {
        struct inode            *inode = (struct inode *)mapping->host;
-       vnode_t                 *vp = vn_from_inode(inode);
-       int                     error;
+       struct xfs_inode        *ip = XFS_I(inode);
 
-       vn_trace_entry(vp, __FUNCTION__, (inst_t *)__return_address);
-
-       VOP_RWLOCK(vp, VRWLOCK_READ);
-       VOP_FLUSH_PAGES(vp, (xfs_off_t)0, -1, 0, FI_REMAPF, error);
-       VOP_RWUNLOCK(vp, VRWLOCK_READ);
-       return generic_block_bmap(mapping, block, xfs_get_block);
+       xfs_itrace_entry(XFS_I(inode));
+       xfs_ilock(ip, XFS_IOLOCK_SHARED);
+       xfs_flush_pages(ip, (xfs_off_t)0, -1, 0, FI_REMAPF);
+       xfs_iunlock(ip, XFS_IOLOCK_SHARED);
+       return generic_block_bmap(mapping, block, xfs_get_blocks);
 }
 
 STATIC int
@@ -1430,7 +1710,7 @@ xfs_vm_readpage(
        struct file             *unused,
        struct page             *page)
 {
-       return mpage_readpage(page, xfs_get_block);
+       return mpage_readpage(page, xfs_get_blocks);
 }
 
 STATIC int
@@ -1440,29 +1720,22 @@ xfs_vm_readpages(
        struct list_head        *pages,
        unsigned                nr_pages)
 {
-       return mpage_readpages(mapping, pages, nr_pages, xfs_get_block);
-}
-
-STATIC void
-xfs_vm_invalidatepage(
-       struct page             *page,
-       unsigned long           offset)
-{
-       xfs_page_trace(XFS_INVALIDPAGE_ENTER,
-                       page->mapping->host, page, offset);
-       block_invalidatepage(page, offset);
+       return mpage_readpages(mapping, pages, nr_pages, xfs_get_blocks);
 }
 
-struct address_space_operations xfs_address_space_operations = {
+const struct address_space_operations xfs_address_space_operations = {
        .readpage               = xfs_vm_readpage,
        .readpages              = xfs_vm_readpages,
        .writepage              = xfs_vm_writepage,
+       .writepages             = xfs_vm_writepages,
        .sync_page              = block_sync_page,
        .releasepage            = xfs_vm_releasepage,
        .invalidatepage         = xfs_vm_invalidatepage,
-       .prepare_write          = xfs_vm_prepare_write,
-       .commit_write           = generic_commit_write,
+       .write_begin            = xfs_vm_write_begin,
+       .write_end              = generic_write_end,
        .bmap                   = xfs_vm_bmap,
        .direct_IO              = xfs_vm_direct_IO,
        .migratepage            = buffer_migrate_page,
+       .is_partially_uptodate  = block_is_partially_uptodate,
+       .error_remove_page      = generic_error_remove_page,
 };