- replay i/o epochs along with other attributes to improve isize consistency

  without involving OST

diff --git a/lustre/include/linux/lustre_idl.h b/lustre/include/linux/lustre_idl.h
index 67a1784..3b0e1ea 100644 (file)
--- a/lustre/include/linux/lustre_idl.h
+++ b/lustre/include/linux/lustre_idl.h
@@ -818,6 +818,7 @@ struct mds_rec_create {
         __u64            cr_rdev;
         __u64            cr_enkey;
         __u64            cr_mac;
+        __u64            cr_ioepoch;
 };
 
 extern void lustre_swab_mds_rec_create (struct mds_rec_create *cr);

diff --git a/lustre/include/linux/lustre_mds.h b/lustre/include/linux/lustre_mds.h
index 07b957c..ee98a4e 100644 (file)
--- a/lustre/include/linux/lustre_mds.h
+++ b/lustre/include/linux/lustre_mds.h
@@ -71,6 +71,7 @@ struct mds_update_record {
         __u32               ur_mode;
         __u64               ur_time;
         __u32               ur_flags;
+        __u64               ur_ioepoch;
 };
 
 #define ur_fsuid    ur_uc.luc_fsuid

diff --git a/lustre/mdc/mdc_lib.c b/lustre/mdc/mdc_lib.c
index 36dc34c..8130e88 100644 (file)
--- a/lustre/mdc/mdc_lib.c
+++ b/lustre/mdc/mdc_lib.c
@@ -70,6 +70,7 @@ void mdc_open_pack(struct lustre_msg *msg, int offset,
         rec->cr_time = op_data->mod_time;
         rec->cr_mode = mode;
         rec->cr_rdev = rdev;
+        rec->cr_ioepoch = 0;
 
         if (op_data->name) {
                 tmp = lustre_msg_buf(msg, offset + 1,

diff --git a/lustre/mdc/mdc_request.c b/lustre/mdc/mdc_request.c
index 24a9920..4078860 100644 (file)
--- a/lustre/mdc/mdc_request.c
+++ b/lustre/mdc/mdc_request.c
@@ -672,6 +672,7 @@ int mdc_set_open_replay_data(struct obd_export *exp,
         mod->mod_open_req = ptlrpc_request_addref(open_req);
 
         memcpy(&rec->cr_replayid, &body->id1, sizeof rec->cr_replayid);
+        memcpy(&rec->cr_ioepoch, &body->io_epoch, sizeof rec->cr_ioepoch);
         open_req->rq_replay_cb = mdc_replay_open;
         open_req->rq_commit_cb = mdc_commit_open;
         open_req->rq_cb_data = mod;

diff --git a/lustre/mds/mds_lib.c b/lustre/mds/mds_lib.c
index b97f854..bc0fb6f 100644 (file)
--- a/lustre/mds/mds_lib.c
+++ b/lustre/mds/mds_lib.c
@@ -606,6 +606,7 @@ static int mds_open_unpack(struct ptlrpc_request *req, int offset,
         r->ur_rdev = rec->cr_rdev;
         r->ur_time = rec->cr_time;
         r->ur_flags = rec->cr_flags;
+        r->ur_ioepoch = rec->cr_ioepoch;
 
         LASSERT_REQSWAB(req, offset + 1);
         r->ur_name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);

diff --git a/lustre/mds/mds_open.c b/lustre/mds/mds_open.c
index 482ef7c..f96835e 100644 (file)
--- a/lustre/mds/mds_open.c
+++ b/lustre/mds/mds_open.c
@@ -168,13 +168,16 @@ static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
                 rc = -ENOMEM;
                 goto out;
         }
-        if (epoch > mds->mds_io_epoch)
+        if (epoch > mds->mds_io_epoch) {
                 mds->mds_io_epoch = epoch;
-        else
+                CDEBUG(D_INODE, "repair MDS epoch "LPU64" for ino %lu/%u\n",
+                       mds->mds_io_epoch, inode->i_ino, inode->i_generation);
+        } else {
                 mds->mds_io_epoch++;
+                CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
+                       mds->mds_io_epoch, inode->i_ino, inode->i_generation);
+        }
         MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
-        CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
-               mds->mds_io_epoch, inode->i_ino, inode->i_generation);
  out:
         if (rc == 0)
                 atomic_inc(&inode->i_writecount);
@@ -234,7 +237,8 @@ int mds_query_write_access(struct inode *inode)
 /* This replaces the VFS dentry_open, it manages mfd and writecount */
 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
                                              struct vfsmount *mnt, int flags,
-                                             struct ptlrpc_request *req)
+                                             struct ptlrpc_request *req,
+                                             struct mds_update_record *rec)
 {
         struct mds_export_data *med = &req->rq_export->exp_mds_data;
         struct mds_obd *mds = mds_req2mds(req);
@@ -253,7 +257,7 @@ static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
 
         if (flags & FMODE_WRITE) {
                 /* FIXME: in recovery, need to pass old epoch here */
-                rc = mds_get_write_access(mds, dentry->d_inode, 0);
+                rc = mds_get_write_access(mds, dentry->d_inode, rec->ur_ioepoch);
                 if (rc)
                         GOTO(cleanup_mfd, rc);
                 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
@@ -667,7 +671,7 @@ static void reconstruct_open(struct mds_update_record *rec, int offset,
                 mntget(mds->mds_vfsmnt);
                 CERROR("Re-opened file \n");
                 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
-                                      rec->ur_flags & ~MDS_OPEN_TRUNC, req);
+                                      rec->ur_flags & ~MDS_OPEN_TRUNC, req, rec);
                 if (!mfd) {
                         CERROR("mds: out of memory\n");
                         GOTO(out_dput, req->rq_status = -ENOMEM);
@@ -807,7 +811,7 @@ static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
         up(&dchild->d_inode->i_sem);
 
         intent_set_disposition(rep, DISP_OPEN_OPEN);
-        mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
+        mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req, rec);
         if (IS_ERR(mfd))
                 RETURN(PTR_ERR(mfd));
 

diff --git a/lustre/ptlrpc/pack_generic.c b/lustre/ptlrpc/pack_generic.c
index 5fc436a..f07d95b 100644 (file)
--- a/lustre/ptlrpc/pack_generic.c
+++ b/lustre/ptlrpc/pack_generic.c
@@ -789,6 +789,7 @@ void lustre_swab_mds_rec_create(struct mds_rec_create *cr)
         lustre_swab_lustre_id(&cr->cr_replayid);
         __swab64s(&cr->cr_time);
         __swab64s(&cr->cr_rdev);
+        __swab64s(&cr->cr_ioepoch);
 }
 
 void lustre_swab_mds_rec_link(struct mds_rec_link *lk)