diff --git a/fs/ceph/file.c b/fs/ceph/file.c
index 3c68e6a..8e924b7 100644
--- a/fs/ceph/file.c
+++ b/fs/ceph/file.c
@@ -397,8 +397,9 @@
 }
 
 enum {
-	CHECK_EOF = 1,
-	READ_INLINE = 2,
+	HAVE_RETRIED = 1,
+	CHECK_EOF =    2,
+	READ_INLINE =  3,
 };
 
 /*
@@ -411,17 +412,14 @@
 static int striped_read(struct inode *inode,
 			u64 off, u64 len,
 			struct page **pages, int num_pages,
-			int *checkeof, bool o_direct,
-			unsigned long buf_align)
+			int *checkeof)
 {
 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 	struct ceph_inode_info *ci = ceph_inode(inode);
 	u64 pos, this_len, left;
-	int io_align, page_align;
-	int pages_left;
-	int read;
+	int page_align, pages_left;
+	int read, ret;
 	struct page **page_pos;
-	int ret;
 	bool hit_stripe, was_short;
 
 	/*
@@ -432,13 +430,9 @@
 	page_pos = pages;
 	pages_left = num_pages;
 	read = 0;
-	io_align = off & ~PAGE_MASK;
 
 more:
-	if (o_direct)
-		page_align = (pos - io_align + buf_align) & ~PAGE_MASK;
-	else
-		page_align = pos & ~PAGE_MASK;
+	page_align = pos & ~PAGE_MASK;
 	this_len = left;
 	ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
 				  &ci->i_layout, pos, &this_len,
@@ -457,8 +451,7 @@
 		if (was_short && (pos + ret < inode->i_size)) {
 			int zlen = min(this_len - ret,
 				       inode->i_size - pos - ret);
-			int zoff = (o_direct ? buf_align : io_align) +
-				    read + ret;
+			int zoff = (off & ~PAGE_MASK) + read + ret;
 			dout(" zero gap %llu to %llu\n",
 				pos + ret, pos + ret + zlen);
 			ceph_zero_page_vector_range(zoff, zlen, pages);
@@ -521,54 +514,28 @@
 	if (ret < 0)
 		return ret;
 
-	if (iocb->ki_flags & IOCB_DIRECT) {
-		while (iov_iter_count(i)) {
-			size_t start;
-			ssize_t n;
+	num_pages = calc_pages_for(off, len);
+	pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
+	if (IS_ERR(pages))
+		return PTR_ERR(pages);
+	ret = striped_read(inode, off, len, pages,
+				num_pages, checkeof);
+	if (ret > 0) {
+		int l, k = 0;
+		size_t left = ret;
 
-			n = dio_get_pagev_size(i);
-			pages = dio_get_pages_alloc(i, n, &start, &num_pages);
-			if (IS_ERR(pages))
-				return PTR_ERR(pages);
-
-			ret = striped_read(inode, off, n,
-					   pages, num_pages, checkeof,
-					   1, start);
-
-			ceph_put_page_vector(pages, num_pages, true);
-
-			if (ret <= 0)
-				break;
-			off += ret;
-			iov_iter_advance(i, ret);
-			if (ret < n)
+		while (left) {
+			size_t page_off = off & ~PAGE_MASK;
+			size_t copy = min_t(size_t, left,
+					    PAGE_SIZE - page_off);
+			l = copy_page_to_iter(pages[k++], page_off, copy, i);
+			off += l;
+			left -= l;
+			if (l < copy)
 				break;
 		}
-	} else {
-		num_pages = calc_pages_for(off, len);
-		pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
-		if (IS_ERR(pages))
-			return PTR_ERR(pages);
-		ret = striped_read(inode, off, len, pages,
-					num_pages, checkeof, 0, 0);
-		if (ret > 0) {
-			int l, k = 0;
-			size_t left = ret;
-
-			while (left) {
-				size_t page_off = off & ~PAGE_MASK;
-				size_t copy = min_t(size_t,
-						    PAGE_SIZE - page_off, left);
-				l = copy_page_to_iter(pages[k++], page_off,
-						      copy, i);
-				off += l;
-				left -= l;
-				if (l < copy)
-					break;
-			}
-		}
-		ceph_release_page_vector(pages, num_pages);
 	}
+	ceph_release_page_vector(pages, num_pages);
 
 	if (off > iocb->ki_pos) {
 		ret = off - iocb->ki_pos;
@@ -579,6 +546,113 @@
 	return ret;
 }
 
+struct ceph_aio_request {
+	struct kiocb *iocb;
+	size_t total_len;
+	int write;
+	int error;
+	struct list_head osd_reqs;
+	unsigned num_reqs;
+	atomic_t pending_reqs;
+	struct ceph_cap_flush *prealloc_cf;
+};
+
+static void ceph_aio_complete(struct inode *inode,
+			      struct ceph_aio_request *aio_req)
+{
+	struct ceph_inode_info *ci = ceph_inode(inode);
+	int ret;
+
+	if (!atomic_dec_and_test(&aio_req->pending_reqs))
+		return;
+
+	ret = aio_req->error;
+	if (!ret)
+		ret = aio_req->total_len;
+
+	dout("ceph_aio_complete %p rc %d\n", inode, ret);
+
+	if (ret >= 0 && aio_req->write) {
+		int dirty;
+
+		loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
+		if (endoff > i_size_read(inode)) {
+			if (ceph_inode_set_size(inode, endoff))
+				ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
+		}
+
+		spin_lock(&ci->i_ceph_lock);
+		ci->i_inline_version = CEPH_INLINE_NONE;
+		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
+					       &aio_req->prealloc_cf);
+		spin_unlock(&ci->i_ceph_lock);
+		if (dirty)
+			__mark_inode_dirty(inode, dirty);
+
+	}
+
+	ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
+						CEPH_CAP_FILE_RD));
+
+	aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
+
+	ceph_free_cap_flush(aio_req->prealloc_cf);
+	kfree(aio_req);
+}
+
+static void ceph_aio_complete_req(struct ceph_osd_request *req,
+				  struct ceph_msg *msg)
+{
+	int rc = req->r_result;
+	struct inode *inode = req->r_inode;
+	struct ceph_aio_request *aio_req = req->r_priv;
+	struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
+	int num_pages = calc_pages_for((u64)osd_data->alignment,
+				       osd_data->length);
+
+	dout("ceph_aio_complete_req %p rc %d bytes %llu\n",
+	     inode, rc, osd_data->length);
+
+	if (rc == -EOLDSNAPC) {
+		BUG_ON(1);
+	}
+
+	if (!aio_req->write) {
+		if (rc == -ENOENT)
+			rc = 0;
+		if (rc >= 0 && osd_data->length > rc) {
+			int zoff = osd_data->alignment + rc;
+			int zlen = osd_data->length - rc;
+			/*
+			 * If read is satisfied by single OSD request,
+			 * it can pass EOF. Otherwise read is within
+			 * i_size.
+			 */
+			if (aio_req->num_reqs == 1) {
+				loff_t i_size = i_size_read(inode);
+				loff_t endoff = aio_req->iocb->ki_pos + rc;
+				if (endoff < i_size)
+					zlen = min_t(size_t, zlen,
+						     i_size - endoff);
+				aio_req->total_len = rc + zlen;
+			}
+
+			if (zlen > 0)
+				ceph_zero_page_vector_range(zoff, zlen,
+							    osd_data->pages);
+		}
+	}
+
+	ceph_put_page_vector(osd_data->pages, num_pages, false);
+	ceph_osdc_put_request(req);
+
+	if (rc < 0)
+		cmpxchg(&aio_req->error, 0, rc);
+
+	ceph_aio_complete(inode, aio_req);
+	return;
+}
+
 /*
  * Write commit request unsafe callback, called to tell us when a
  * request is unsafe (that is, in flight--has been handed to the
@@ -612,16 +686,10 @@
 }
 
 
-/*
- * Synchronous write, straight from __user pointer or user pages.
- *
- * If write spans object boundary, just do multiple writes.  (For a
- * correct atomic write, we should e.g. take write locks on all
- * objects, rollback on failure, etc.)
- */
 static ssize_t
-ceph_sync_direct_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
-		       struct ceph_snap_context *snapc)
+ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
+		       struct ceph_snap_context *snapc,
+		       struct ceph_cap_flush **pcf)
 {
 	struct file *file = iocb->ki_filp;
 	struct inode *inode = file_inode(file);
@@ -630,44 +698,52 @@
 	struct ceph_vino vino;
 	struct ceph_osd_request *req;
 	struct page **pages;
-	int num_pages;
-	int written = 0;
+	struct ceph_aio_request *aio_req = NULL;
+	int num_pages = 0;
 	int flags;
-	int check_caps = 0;
 	int ret;
 	struct timespec mtime = CURRENT_TIME;
-	size_t count = iov_iter_count(from);
+	size_t count = iov_iter_count(iter);
+	loff_t pos = iocb->ki_pos;
+	bool write = iov_iter_rw(iter) == WRITE;
 
-	if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
+	if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
 		return -EROFS;
 
-	dout("sync_direct_write on file %p %lld~%u\n", file, pos,
-	     (unsigned)count);
+	dout("sync_direct_read_write (%s) on file %p %lld~%u\n",
+	     (write ? "write" : "read"), file, pos, (unsigned)count);
 
 	ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
 	if (ret < 0)
 		return ret;
 
-	ret = invalidate_inode_pages2_range(inode->i_mapping,
-					    pos >> PAGE_CACHE_SHIFT,
-					    (pos + count) >> PAGE_CACHE_SHIFT);
-	if (ret < 0)
-		dout("invalidate_inode_pages2_range returned %d\n", ret);
+	if (write) {
+		ret = invalidate_inode_pages2_range(inode->i_mapping,
+					pos >> PAGE_CACHE_SHIFT,
+					(pos + count) >> PAGE_CACHE_SHIFT);
+		if (ret < 0)
+			dout("invalidate_inode_pages2_range returned %d\n", ret);
 
-	flags = CEPH_OSD_FLAG_ORDERSNAP |
-		CEPH_OSD_FLAG_ONDISK |
-		CEPH_OSD_FLAG_WRITE;
+		flags = CEPH_OSD_FLAG_ORDERSNAP |
+			CEPH_OSD_FLAG_ONDISK |
+			CEPH_OSD_FLAG_WRITE;
+	} else {
+		flags = CEPH_OSD_FLAG_READ;
+	}
 
-	while (iov_iter_count(from) > 0) {
-		u64 len = dio_get_pagev_size(from);
-		size_t start;
-		ssize_t n;
+	while (iov_iter_count(iter) > 0) {
+		u64 size = dio_get_pagev_size(iter);
+		size_t start = 0;
+		ssize_t len;
 
 		vino = ceph_vino(inode);
 		req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
-					    vino, pos, &len, 0,
-					    2,/*include a 'startsync' command*/
-					    CEPH_OSD_OP_WRITE, flags, snapc,
+					    vino, pos, &size, 0,
+					    /*include a 'startsync' command*/
+					    write ? 2 : 1,
+					    write ? CEPH_OSD_OP_WRITE :
+						    CEPH_OSD_OP_READ,
+					    flags, snapc,
 					    ci->i_truncate_seq,
 					    ci->i_truncate_size,
 					    false);
@@ -676,10 +752,8 @@
 			break;
 		}
 
-		osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
-
-		n = len;
-		pages = dio_get_pages_alloc(from, len, &start, &num_pages);
+		len = size;
+		pages = dio_get_pages_alloc(iter, len, &start, &num_pages);
 		if (IS_ERR(pages)) {
 			ceph_osdc_put_request(req);
 			ret = PTR_ERR(pages);
@@ -687,47 +761,126 @@
 		}
 
 		/*
-		 * throw out any page cache pages in this range. this
-		 * may block.
+		 * To simplify error handling, allow AIO when IO within i_size
+		 * or IO can be satisfied by single OSD request.
 		 */
-		truncate_inode_pages_range(inode->i_mapping, pos,
-				   (pos+n) | (PAGE_CACHE_SIZE-1));
-		osd_req_op_extent_osd_data_pages(req, 0, pages, n, start,
-						false, false);
+		if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
+		    (len == count || pos + count <= i_size_read(inode))) {
+			aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
+			if (aio_req) {
+				aio_req->iocb = iocb;
+				aio_req->write = write;
+				INIT_LIST_HEAD(&aio_req->osd_reqs);
+				if (write) {
+					swap(aio_req->prealloc_cf, *pcf);
+				}
+			}
+			/* ignore error */
+		}
 
-		/* BUG_ON(vino.snap != CEPH_NOSNAP); */
+		if (write) {
+			/*
+			 * throw out any page cache pages in this range. this
+			 * may block.
+			 */
+			truncate_inode_pages_range(inode->i_mapping, pos,
+					(pos+len) | (PAGE_CACHE_SIZE - 1));
+
+			osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
+		}
+
+
+		osd_req_op_extent_osd_data_pages(req, 0, pages, len, start,
+						 false, false);
+
 		ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime);
 
-		ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
+		if (aio_req) {
+			aio_req->total_len += len;
+			aio_req->num_reqs++;
+			atomic_inc(&aio_req->pending_reqs);
+
+			req->r_callback = ceph_aio_complete_req;
+			req->r_inode = inode;
+			req->r_priv = aio_req;
+			list_add_tail(&req->r_unsafe_item, &aio_req->osd_reqs);
+
+			pos += len;
+			iov_iter_advance(iter, len);
+			continue;
+		}
+
+		ret = ceph_osdc_start_request(req->r_osdc, req, false);
 		if (!ret)
 			ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
 
+		size = i_size_read(inode);
+		if (!write) {
+			if (ret == -ENOENT)
+				ret = 0;
+			if (ret >= 0 && ret < len && pos + ret < size) {
+				int zlen = min_t(size_t, len - ret,
+						 size - pos - ret);
+				ceph_zero_page_vector_range(start + ret, zlen,
+							    pages);
+				ret += zlen;
+			}
+			if (ret >= 0)
+				len = ret;
+		}
+
 		ceph_put_page_vector(pages, num_pages, false);
 
 		ceph_osdc_put_request(req);
-		if (ret)
+		if (ret < 0)
 			break;
-		pos += n;
-		written += n;
-		iov_iter_advance(from, n);
 
-		if (pos > i_size_read(inode)) {
-			check_caps = ceph_inode_set_size(inode, pos);
-			if (check_caps)
+		pos += len;
+		iov_iter_advance(iter, len);
+
+		if (!write && pos >= size)
+			break;
+
+		if (write && pos > size) {
+			if (ceph_inode_set_size(inode, pos))
 				ceph_check_caps(ceph_inode(inode),
 						CHECK_CAPS_AUTHONLY,
 						NULL);
 		}
 	}
 
-	if (ret != -EOLDSNAPC && written > 0) {
+	if (aio_req) {
+		if (aio_req->num_reqs == 0) {
+			kfree(aio_req);
+			return ret;
+		}
+
+		ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
+					      CEPH_CAP_FILE_RD);
+
+		while (!list_empty(&aio_req->osd_reqs)) {
+			req = list_first_entry(&aio_req->osd_reqs,
+					       struct ceph_osd_request,
+					       r_unsafe_item);
+			list_del_init(&req->r_unsafe_item);
+			if (ret >= 0)
+				ret = ceph_osdc_start_request(req->r_osdc,
+							      req, false);
+			if (ret < 0) {
+				req->r_result = ret;
+				ceph_aio_complete_req(req, NULL);
+			}
+		}
+		return -EIOCBQUEUED;
+	}
+
+	if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
+		ret = pos - iocb->ki_pos;
 		iocb->ki_pos = pos;
-		ret = written;
 	}
 	return ret;
 }
 
-
 /*
  * Synchronous write, straight from __user pointer or user pages.
  *
@@ -897,8 +1050,14 @@
 		     ceph_cap_string(got));
 
 		if (ci->i_inline_version == CEPH_INLINE_NONE) {
-			/* hmm, this isn't really async... */
-			ret = ceph_sync_read(iocb, to, &retry_op);
+			if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
+				ret = ceph_direct_read_write(iocb, to,
+							     NULL, NULL);
+				if (ret >= 0 && ret < len)
+					retry_op = CHECK_EOF;
+			} else {
+				ret = ceph_sync_read(iocb, to, &retry_op);
+			}
 		} else {
 			retry_op = READ_INLINE;
 		}
@@ -916,7 +1075,7 @@
 		pinned_page = NULL;
 	}
 	ceph_put_cap_refs(ci, got);
-	if (retry_op && ret >= 0) {
+	if (retry_op > HAVE_RETRIED && ret >= 0) {
 		int statret;
 		struct page *page = NULL;
 		loff_t i_size;
@@ -973,7 +1132,7 @@
 
 			read += ret;
 			len -= ret;
-			retry_op = 0;
+			retry_op = HAVE_RETRIED;
 			goto again;
 		}
 	}
@@ -1088,8 +1247,8 @@
 		/* we might need to revert back to that point */
 		data = *from;
 		if (iocb->ki_flags & IOCB_DIRECT)
-			written = ceph_sync_direct_write(iocb, &data, pos,
-							 snapc);
+			written = ceph_direct_read_write(iocb, &data, snapc,
+							 &prealloc_cf);
 		else
 			written = ceph_sync_write(iocb, &data, pos, snapc);
 		if (written == -EOLDSNAPC) {