Merge branch 'xfs-dio-extend-fix' into for-next

Conflicts:
	fs/xfs/xfs_file.c

3 files changed, 239 insertions, 82 deletions

diff --git a/fs/xfs/xfs_aops.c b/fs/xfs/xfs_aops.c
index 3a9b7a1b8704..598b259fda04 100644
--- a/fs/xfs/xfs_aops.c
+++ b/fs/xfs/xfs_aops.c
@@ -1233,6 +1233,117 @@ xfs_vm_releasepage(
         return try_to_free_buffers(page);
 }
 
+/*
+ * When we map a DIO buffer, we may need to attach an ioend that describes the
+ * type of write IO we are doing. This passes to the completion function the
+ * operations it needs to perform. If the mapping is for an overwrite wholly
+ * within the EOF then we don't need an ioend and so we don't allocate one.
+ * This avoids the unnecessary overhead of allocating and freeing ioends for
+ * workloads that don't require transactions on IO completion.
+ *
+ * If we get multiple mappings in a single IO, we might be mapping different
+ * types. But because the direct IO can only have a single private pointer, we
+ * need to ensure that:
+ *
+ * a) i) the ioend spans the entire region of unwritten mappings; or
+ *    ii) the ioend spans all the mappings that cross or are beyond EOF; and
+ * b) if it contains unwritten extents, it is *permanently* marked as such
+ *
+ * We could do this by chaining ioends like buffered IO does, but we only
+ * actually get one IO completion callback from the direct IO, and that spans
+ * the entire IO regardless of how many mappings and IOs are needed to complete
+ * the DIO. There is only going to be one reference to the ioend and its life
+ * cycle is constrained by the DIO completion code. hence we don't need
+ * reference counting here.
+ */
+static void
+xfs_map_direct(
+        struct inode            *inode,
+        struct buffer_head      *bh_result,
+        struct xfs_bmbt_irec    *imap,
+        xfs_off_t               offset)
+{
+        struct xfs_ioend        *ioend;
+        xfs_off_t               size = bh_result->b_size;
+        int                     type;
+
+        if (ISUNWRITTEN(imap))
+                type = XFS_IO_UNWRITTEN;
+        else
+                type = XFS_IO_OVERWRITE;
+
+        trace_xfs_gbmap_direct(XFS_I(inode), offset, size, type, imap);
+
+        if (bh_result->b_private) {
+                ioend = bh_result->b_private;
+                ASSERT(ioend->io_size > 0);
+                ASSERT(offset >= ioend->io_offset);
+                if (offset + size > ioend->io_offset + ioend->io_size)
+                        ioend->io_size = offset - ioend->io_offset + size;
+
+                if (type == XFS_IO_UNWRITTEN && type != ioend->io_type)
+                        ioend->io_type = XFS_IO_UNWRITTEN;
+
+                trace_xfs_gbmap_direct_update(XFS_I(inode), ioend->io_offset,
+                                              ioend->io_size, ioend->io_type,
+                                              imap);
+        } else if (type == XFS_IO_UNWRITTEN ||
+                   offset + size > i_size_read(inode)) {
+                ioend = xfs_alloc_ioend(inode, type);
+                ioend->io_offset = offset;
+                ioend->io_size = size;
+
+                bh_result->b_private = ioend;
+                set_buffer_defer_completion(bh_result);
+
+                trace_xfs_gbmap_direct_new(XFS_I(inode), offset, size, type,
+                                           imap);
+        } else {
+                trace_xfs_gbmap_direct_none(XFS_I(inode), offset, size, type,
+                                            imap);
+        }
+}
+
+/*
+ * 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 the mapping spans EOF, then we have to break the mapping up as the mapping
+ * for blocks beyond EOF must be marked new so that sub block regions can be
+ * correctly zeroed. We can't do this for mappings within EOF unless the mapping
+ * was just allocated or is unwritten, otherwise the callers would overwrite
+ * existing data with zeros. Hence we have to split the mapping into a range up
+ * to and including EOF, and a second mapping for beyond EOF.
+ */
+static void
+xfs_map_trim_size(
+        struct inode            *inode,
+        sector_t                iblock,
+        struct buffer_head      *bh_result,
+        struct xfs_bmbt_irec    *imap,
+        xfs_off_t               offset,
+        ssize_t                 size)
+{
+        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 (offset < i_size_read(inode) &&
+            offset + mapping_size >= i_size_read(inode)) {
+                /* limit mapping to block that spans EOF */
+                mapping_size = roundup_64(i_size_read(inode) - offset,
+                                          1 << inode->i_blkbits);
+        }
+        if (mapping_size > LONG_MAX)
+                mapping_size = LONG_MAX;
+
+        bh_result->b_size = mapping_size;
+}
+
 STATIC int
 __xfs_get_blocks(
         struct inode            *inode,
@@ -1321,31 +1432,37 @@ __xfs_get_blocks(
 
                         xfs_iunlock(ip, lockmode);
                 }
-
-                trace_xfs_get_blocks_alloc(ip, offset, size, 0, &imap);
+                trace_xfs_get_blocks_alloc(ip, offset, size,
+                                ISUNWRITTEN(&imap) ? XFS_IO_UNWRITTEN
+                                                   : XFS_IO_DELALLOC, &imap);
         } else if (nimaps) {
-                trace_xfs_get_blocks_found(ip, offset, size, 0, &imap);
+                trace_xfs_get_blocks_found(ip, offset, size,
+                                ISUNWRITTEN(&imap) ? XFS_IO_UNWRITTEN
+                                                   : XFS_IO_OVERWRITE, &imap);
                 xfs_iunlock(ip, lockmode);
         } else {
                 trace_xfs_get_blocks_notfound(ip, offset, size);
                 goto out_unlock;
         }
 
+        /* trim mapping down to size requested */
+        if (direct || size > (1 << inode->i_blkbits))
+                xfs_map_trim_size(inode, iblock, bh_result,
+                                  &imap, offset, size);
+
+        /*
+         * For unwritten extents do not report a disk address in the buffered
+         * read case (treat as if we're reading into a hole).
+         */
         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 || !ISUNWRITTEN(&imap))
-                        xfs_map_buffer(inode, bh_result, &imap, offset);
-                if (create && ISUNWRITTEN(&imap)) {
-                        if (direct) {
-                                bh_result->b_private = inode;
-                                set_buffer_defer_completion(bh_result);
-                        }
+            imap.br_startblock != DELAYSTARTBLOCK &&
+            (create || !ISUNWRITTEN(&imap))) {
+                xfs_map_buffer(inode, bh_result, &imap, offset);
+                if (ISUNWRITTEN(&imap))
                         set_buffer_unwritten(bh_result);
-                }
+                /* direct IO needs special help */
+                if (create && direct)
+                        xfs_map_direct(inode, bh_result, &imap, offset);
         }
 
         /*
@@ -1378,39 +1495,6 @@ __xfs_get_blocks(
                 }
         }
 
-        /*
-         * 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 the mapping spans EOF, then we have to break the mapping up as the
-         * mapping for blocks beyond EOF must be marked new so that sub block
-         * regions can be correctly zeroed. We can't do this for mappings within
-         * EOF unless the mapping was just allocated or is unwritten, otherwise