From: Trond Myklebust <Trond.Myklebust@netapp.com>
Date: Sun, 20 May 2007 10:18:27 -0400
NFS: Avoid a deadlock situation on write

When processes are allowed to attempt to lock a non-contiguous range of nfs
write requests, it is possible for generic_writepages to 'wrap round' the
address space, and call writepage() on a request that is already locked by
the same process.

We avoid the deadlock by checking if the page index is contiguous with the
list of nfs write requests that is already held in our
nfs_pageio_descriptor prior to attempting to lock a new request.

Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
---

 fs/nfs/pagelist.c        |   20 ++++++++++++++++++++
 fs/nfs/write.c           |    6 ++++--
 include/linux/nfs_page.h |    1 +
 3 files changed, 25 insertions(+), 2 deletions(-)

diff --git a/fs/nfs/pagelist.c b/fs/nfs/pagelist.c
index cbdd1c6..c5bb51a 100644
--- a/fs/nfs/pagelist.c
+++ b/fs/nfs/pagelist.c
@@ -355,6 +355,26 @@ void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
 	nfs_pageio_doio(desc);
 }
 
+/**
+ * nfs_pageio_cond_complete - Conditional I/O completion
+ * @desc: pointer to io descriptor
+ * @index: page index
+ *
+ * It is important to ensure that processes don't try to take locks
+ * on non-contiguous ranges of pages as that might deadlock. This
+ * function should be called before attempting to wait on a locked
+ * nfs_page. It will complete the I/O if the page index 'index'
+ * is not contiguous with the existing list of pages in 'desc'.
+ */
+void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
+{
+	if (!list_empty(&desc->pg_list)) {
+		struct nfs_page *prev = nfs_list_entry(desc->pg_list.prev);
+		if (index != prev->wb_index + 1)
+			nfs_pageio_doio(desc);
+	}
+}
+
 #define NFS_SCAN_MAXENTRIES 16
 /**
  * nfs_scan_list - Scan a list for matching requests
diff --git a/fs/nfs/write.c b/fs/nfs/write.c
index b084c03..af344a1 100644
--- a/fs/nfs/write.c
+++ b/fs/nfs/write.c
@@ -273,8 +273,6 @@ static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 		 *	 request as dirty (in which case we don't care).
 		 */
 		spin_unlock(req_lock);
-		/* Prevent deadlock! */
-		nfs_pageio_complete(pgio);
 		ret = nfs_wait_on_request(req);
 		nfs_release_request(req);
 		if (ret != 0)
@@ -321,6 +319,8 @@ static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc
 		pgio = &mypgio;
 	}
 
+	nfs_pageio_cond_complete(pgio, page->index);
+
 	err = nfs_page_async_flush(pgio, page);
 	if (err <= 0)
 		goto out;
@@ -329,6 +329,8 @@ static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc
 	if (!offset)
 		goto out;
 
+	nfs_pageio_cond_complete(pgio, page->index);
+
 	ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
 	if (ctx == NULL) {
 		err = -EBADF;
diff --git a/include/linux/nfs_page.h b/include/linux/nfs_page.h
index 41afab6..bd193af 100644
--- a/include/linux/nfs_page.h
+++ b/include/linux/nfs_page.h
@@ -81,6 +81,7 @@ extern	void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
 extern	int nfs_pageio_add_request(struct nfs_pageio_descriptor *,
 				   struct nfs_page *);
 extern	void nfs_pageio_complete(struct nfs_pageio_descriptor *desc);
+extern	void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *, pgoff_t);
 extern  int nfs_wait_on_request(struct nfs_page *);
 extern	void nfs_unlock_request(struct nfs_page *req);
 extern  int nfs_set_page_writeback_locked(struct nfs_page *req);