From: alex <alex>
Date: Sat, 24 Sep 2005 00:33:55 +0000 (+0000)
Subject: - re-fill size/blocks attributes in open reply if ea lov has been changed
X-Git-Tag: v1_7_100~606
X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=commitdiff_plain;h=b5745c153b081b42e3806f6aaf35b96ce1eb195f

- re-fill size/blocks attributes in open reply if ea lov has been changed
- yank granted update locks in the same case in order to invalidate
  cached size/blocks attributes on clients
---

diff --git a/lustre/mds/mds_open.c b/lustre/mds/mds_open.c
index 52aef76..beda3a3 100644
--- a/lustre/mds/mds_open.c
+++ b/lustre/mds/mds_open.c
@@ -718,7 +718,8 @@ int accmode(int flags)
 /* Handles object creation, actual opening, and I/O epoch */
 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
                            struct mds_body *body, int flags, void **handle,
-                           struct mds_update_record *rec, struct ldlm_reply *rep)
+                           struct mds_update_record *rec, struct ldlm_reply *rep,
+                           int *cancel_update_lock)
 {
         struct obd_device *obd = req->rq_export->exp_obd;
         struct mds_obd *mds = mds_req2mds(req);
@@ -729,6 +730,8 @@ static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
         int rc = 0, reply_off;
         ENTRY;
 
+        *cancel_update_lock = 0;
+
         /* atomically create objects if necessary */
         down(&dchild->d_inode->i_sem);
         mode = dchild->d_inode->i_mode;
@@ -764,9 +767,16 @@ static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
                 if (S_ISREG(dchild->d_inode->i_mode) &&
                     (body->valid & OBD_MD_FLEASIZE)) {
                         rc = mds_revalidate_lov_ea(obd,dchild,req->rq_repmsg,2);
-                        if (!rc)
+                        if (!rc) {
                                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
                                                  dchild->d_inode, 0, 0);
+                                /* after ost add/delete, lov ea can change
+                                 * along with size/blocks, which we have
+                                 * to update as well */
+                                body->size = dchild->d_inode->i_size;
+                                body->blocks = dchild->d_inode->i_blocks;
+                                *cancel_update_lock = 1;
+                        }
                         if (rc) {
                                 up(&dchild->d_inode->i_sem);
                                 RETURN(rc);
@@ -846,7 +856,7 @@ static int mds_open_by_id(struct ptlrpc_request *req,
         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
         struct dentry *dchild;
         char idname[LL_ID_NAMELEN];
-        int idlen = 0, rc;
+        int idlen = 0, rc, fake;
         void *handle = NULL;
         ENTRY;
 
@@ -889,7 +899,7 @@ static int mds_open_by_id(struct ptlrpc_request *req,
         intent_set_disposition(rep, DISP_LOOKUP_POS);
 
  open:
-        rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
+        rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep,&fake);
         rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
                                 req, rc, rep ? rep->lock_policy_res1 : 0);
         /* XXX what do we do here if mds_finish_transno itself failed? */
@@ -970,6 +980,7 @@ int mds_open(struct mds_update_record *rec, int offset,
         int child_mode = LCK_PR;
         struct lustre_id sid;
         __u64 fid = 0;
+        int cancel_update_lock;
         ENTRY;
 
         DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
@@ -1398,7 +1409,7 @@ got_child:
 
         /* Step 5: mds_open it */
         rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
-                             rec, rep);
+                             rec, rep, &cancel_update_lock);
         if (rc)
                 GOTO(cleanup, rc);
 
@@ -1418,7 +1429,7 @@ got_child:
          * isize/iblocks from mds anymore.
          * FIXME: can cause a deadlock, use mds_get_parent_child_locked()
          * XXX: optimization is to do this for first writer only */
-        if (accmode(rec->ur_flags) & MAY_WRITE) {
+        if ((accmode(rec->ur_flags) & MAY_WRITE) || cancel_update_lock) {
                 struct ldlm_res_id child_res_id = { .name = {0}};
                 ldlm_policy_data_t sz_policy;
                 struct lustre_handle sz_lockh;