staging/lustre/clio: don't ignore layout on writeback

In some cases such as kernel writeback, we shouldn't ignore the
layout, otherwise, it could race with layout change undergoing.

Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-3160
Lustre-change: http://review.whamcloud.com/6154
Signed-off-by: Niu Yawei <yawei.niu@intel.com>
Signed-off-by: Jinshan Xiong <jinshan.xiong@intel.com>
Reviewed-by: Andreas Dilger <andreas.dilger@intel.com>
Reviewed-by: Fan Yong <fan.yong@intel.com>
Reviewed-by: Oleg Drokin <oleg.drokin@intel.com>
Signed-off-by: Peng Tao <tao.peng@emc.com>
Signed-off-by: Andreas Dilger <andreas.dilger@intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

drivers/staging/lustre/lustre/llite/file.c

@@ -2175,7 +2175,7 @@ int ll_flush(struct file *file, fl_owner_t id)
  * Return how many pages have been written.
  */
 int cl_sync_file_range(struct inode *inode, loff_t start, loff_t end,
-		       enum cl_fsync_mode mode)
+		       enum cl_fsync_mode mode, int ignore_layout)
 {
 	struct cl_env_nest nest;
 	struct lu_env *env;
@@ -2197,7 +2197,7 @@ int cl_sync_file_range(struct inode *inode, loff_t start, loff_t end,
 
 	io = ccc_env_thread_io(env);
 	io->ci_obj = cl_i2info(inode)->lli_clob;
-	io->ci_ignore_layout = 1;
+	io->ci_ignore_layout = ignore_layout;
 
 	/* initialize parameters for sync */
 	fio = &io->u.ci_fsync;
@@ -2270,7 +2270,7 @@ int ll_fsync(struct file *file, loff_t start, loff_t end, int datasync)
 		struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
 
 		err = cl_sync_file_range(inode, 0, OBD_OBJECT_EOF,
-				CL_FSYNC_ALL);
+				CL_FSYNC_ALL, 0);
 		if (rc == 0 && err < 0)
 			rc = err;
 		if (rc < 0)

drivers/staging/lustre/lustre/llite/llite_internal.h

@@ -465,6 +465,7 @@ struct ll_sb_info {
 	struct lu_fid	     ll_root_fid; /* root object fid */
 
 	int		       ll_flags;
+	int			  ll_umounting:1;
 	struct list_head		ll_conn_chain; /* per-conn chain of SBs */
 	struct lustre_client_ocd  ll_lco;
 
@@ -1419,7 +1420,7 @@ static inline int cl_merge_lvb(const struct lu_env *env, struct inode *inode)
 struct obd_capa *cl_capa_lookup(struct inode *inode, enum cl_req_type crt);
 
 int cl_sync_file_range(struct inode *inode, loff_t start, loff_t end,
-		       enum cl_fsync_mode mode);
+		       enum cl_fsync_mode mode, int ignore_layout);
 
 /** direct write pages */
 struct ll_dio_pages {

drivers/staging/lustre/lustre/llite/llite_lib.c

@@ -724,8 +724,10 @@ void ll_kill_super(struct super_block *sb)
 	/* we need restore s_dev from changed for clustred NFS before put_super
 	 * because new kernels have cached s_dev and change sb->s_dev in
 	 * put_super not affected real removing devices */
-	if (sbi)
+	if (sbi) {
 		sb->s_dev = sbi->ll_sdev_orig;
+		sbi->ll_umounting = 1;
+	}
 	EXIT;
 }
 
@@ -1856,7 +1858,8 @@ void ll_delete_inode(struct inode *inode)
 	if (S_ISREG(inode->i_mode) && lli->lli_clob != NULL)
 		/* discard all dirty pages before truncating them, required by
 		 * osc_extent implementation at LU-1030. */
-		cl_sync_file_range(inode, 0, OBD_OBJECT_EOF, CL_FSYNC_DISCARD);
+		cl_sync_file_range(inode, 0, OBD_OBJECT_EOF,
+				   CL_FSYNC_DISCARD, 1);
 
 	truncate_inode_pages(&inode->i_data, 0);
 
@@ -2026,7 +2029,6 @@ void ll_umount_begin(struct super_block *sb)
 		OBD_FREE_PTR(ioc_data);
 	}
 
-
 	/* Really, we'd like to wait until there are no requests outstanding,
 	 * and then continue.  For now, we just invalidate the requests,
 	 * schedule() and sleep one second if needed, and hope.

drivers/staging/lustre/lustre/llite/rw.c

@@ -1213,7 +1213,7 @@ int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
 		 * PageWriteback or clean the page. */
 		result = cl_sync_file_range(inode, offset,
 					    offset + PAGE_CACHE_SIZE - 1,
-					    CL_FSYNC_LOCAL);
+					    CL_FSYNC_LOCAL, 1);
 		if (result > 0) {
 			/* actually we may have written more than one page.
 			 * decreasing this page because the caller will count
@@ -1240,11 +1240,13 @@ int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
 {
 	struct inode *inode = mapping->host;
+	struct ll_sb_info *sbi = ll_i2sbi(inode);
 	loff_t start;
 	loff_t end;
 	enum cl_fsync_mode mode;
 	int range_whole = 0;
 	int result;
+	int ignore_layout = 0;
 	ENTRY;
 
 	if (wbc->range_cyclic) {
@@ -1263,7 +1265,12 @@ int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
 	if (wbc->sync_mode == WB_SYNC_ALL)
 		mode = CL_FSYNC_LOCAL;
 
-	result = cl_sync_file_range(inode, start, end, mode);
+	if (sbi->ll_umounting)
+		/* if the mountpoint is being umounted, all pages have to be
+		 * evicted to avoid hitting LBUG when truncate_inode_pages()
+		 * is called later on. */
+		ignore_layout = 1;
+	result = cl_sync_file_range(inode, start, end, mode, ignore_layout);
 	if (result > 0) {
 		wbc->nr_to_write -= result;
 		result = 0;

drivers/staging/lustre/lustre/llite/vvp_io.c

@@ -1160,8 +1160,19 @@ int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
 	/* Enqueue layout lock and get layout version. We need to do this
 	 * even for operations requiring to open file, such as read and write,
 	 * because it might not grant layout lock in IT_OPEN. */
-	if (result == 0 && !io->ci_ignore_layout)
+	if (result == 0 && !io->ci_ignore_layout) {
 		result = ll_layout_refresh(inode, &cio->cui_layout_gen);
+		if (result == -ENOENT)
+			/* If the inode on MDS has been removed, but the objects
+			 * on OSTs haven't been destroyed (async unlink), layout
+			 * fetch will return -ENOENT, we'd ingore this error
+			 * and continue with dirty flush. LU-3230. */
+			result = 0;
+		if (result < 0)
+			CERROR("%s: refresh file layout " DFID " error %d.\n",
+				ll_get_fsname(inode->i_sb, NULL, 0),
+				PFID(lu_object_fid(&obj->co_lu)), result);
+	}
 
 	RETURN(result);
 }