LU-3160 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.

Test-Parameters: envdefinitions=DURATION=7200  clientdistro=el6 serverdistro=el6 clientcount=4  osscount=2 mdscount=2 austeroptions=-R failover=true  useiscsi=true testlist=recovery-random-scale
Signed-off-by: Niu Yawei <yawei.niu@intel.com>
Signed-off-by: Jinshan Xiong <jinshan.xiong@intel.com>
Change-Id: Ib9d0aa581de90711c92db4c631c52f1950ad5b67
Reviewed-on: http://review.whamcloud.com/6154
Tested-by: Hudson
Reviewed-by: Andreas Dilger <andreas.dilger@intel.com>
Tested-by: Maloo <whamcloud.maloo@gmail.com>
Reviewed-by: Fan Yong <fan.yong@intel.com>
Reviewed-by: Oleg Drokin <oleg.drokin@intel.com>

diff --git a/lustre/llite/file.c b/lustre/llite/file.c
index e61b696..c3dade8 100644 (file)
--- a/lustre/llite/file.c
+++ b/lustre/llite/file.c
@@ -2374,7 +2374,7 @@ 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,
  * 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;
 {
        struct cl_env_nest nest;
        struct lu_env *env;


@@ -2396,7 +2396,7 @@ 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 = 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;
 
        /* initialize parameters for sync */
        fio = &io->u.ci_fsync;


@@ -2484,7 +2484,7 @@ int ll_fsync(struct file *file, struct dentry *dentry, int datasync)
                struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
 
                err = cl_sync_file_range(inode, 0, OBD_OBJECT_EOF,
                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)
                if (rc == 0 && err < 0)
                        rc = err;
                if (rc < 0)

diff --git a/lustre/llite/llite_internal.h b/lustre/llite/llite_internal.h
index cc7fc04..0d51446 100644 (file)
--- a/lustre/llite/llite_internal.h
+++ b/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;
         struct lu_fid             ll_root_fid; /* root object fid */
 
         int                       ll_flags;

+       int                       ll_umounting:1;

         cfs_list_t                ll_conn_chain; /* per-conn chain of SBs */
         struct lustre_client_ocd  ll_lco;
 
         cfs_list_t                ll_conn_chain; /* per-conn chain of SBs */
         struct lustre_client_ocd  ll_lco;
 


@@ -1471,7 +1472,7 @@ 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,
 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 {
 
 /** direct write pages */
 struct ll_dio_pages {

diff --git a/lustre/llite/llite_lib.c b/lustre/llite/llite_lib.c
index 744ed25..e8e31bf 100644 (file)
--- a/lustre/llite/llite_lib.c
+++ b/lustre/llite/llite_lib.c
@@ -736,9 +736,11 @@ 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 */
         /* 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)
-                sb->s_dev = sbi->ll_sdev_orig;
-        EXIT;
+       if (sbi) {
+               sb->s_dev = sbi->ll_sdev_orig;
+               sbi->ll_umounting = 1;
+       }
+       EXIT;

 }
 
 char *ll_read_opt(const char *opt, char *data)
 }
 
 char *ll_read_opt(const char *opt, char *data)


@@ -1919,7 +1921,8 @@ 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. */
        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);
 
 
         truncate_inode_pages(&inode->i_data, 0);
 


@@ -2103,7 +2106,6 @@ void ll_umount_begin(struct super_block *sb)
                OBD_FREE_PTR(ioc_data);
        }
 
                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.
         /* 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.

diff --git a/lustre/llite/rw.c b/lustre/llite/rw.c
index 0cf7ab2..201f60d 100644 (file)
--- a/lustre/llite/rw.c
+++ b/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 + CFS_PAGE_SIZE - 1,
                 * PageWriteback or clean the page. */
                result = cl_sync_file_range(inode, offset,
                                            offset + CFS_PAGE_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
                if (result > 0) {
                        /* actually we may have written more than one page.
                         * decreasing this page because the caller will count


@@ -1240,11 +1240,13 @@ out:
 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
 {
        struct inode *inode = mapping->host;
 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;
        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) {
        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;
 
        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;
        if (result > 0) {
                wbc->nr_to_write -= result;
                result = 0;

diff --git a/lustre/llite/vvp_io.c b/lustre/llite/vvp_io.c
index db53842..27f58fc 100644 (file)
--- a/lustre/llite/vvp_io.c
+++ b/lustre/llite/vvp_io.c
@@ -1214,8 +1214,19 @@ 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. */
        /* 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);
                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);
 }
 
        RETURN(result);
 }