]> git.karo-electronics.de Git - linux-beck.git/commitdiff
ceph: properly handle aborted mds requests
authorSage Weil <sage@newdream.net>
Mon, 25 Jan 2010 19:33:08 +0000 (11:33 -0800)
committerSage Weil <sage@newdream.net>
Mon, 25 Jan 2010 19:49:51 +0000 (11:49 -0800)
Previously, if the MDS request was interrupted, we would unregister the
request and ignore any reply.  This could cause the caps or other cache
state to become out of sync.  (For instance, aborting dbench and doing
rm -r on clients would complain about a non-empty directory because the
client didn't realize it's aborted file create request completed.)

Even we don't unregister, we still can't process the reply normally because
we are no longer holding the caller's locks (like the dir i_mutex).

So, mark aborted operations with r_aborted, and in the reply handler, be
sure to process all the caps.  Do not process the namespace changes,
though, since we no longer will hold the dir i_mutex.  The dentry lease
state can also be ignored as it's more forgiving.

Signed-off-by: Sage Weil <sage@newdream.net>
fs/ceph/inode.c
fs/ceph/mds_client.c
fs/ceph/mds_client.h

index 518beb628f0920afff3b063251ce26a12af46245..71e107fb4dbcdc9dacf5c2e248db890cb692e28f 100644 (file)
@@ -915,6 +915,16 @@ int ceph_fill_trace(struct super_block *sb, struct ceph_mds_request *req,
        }
 
        if (rinfo->head->is_dentry) {
+               struct inode *dir = req->r_locked_dir;
+
+               err = fill_inode(dir, &rinfo->diri, rinfo->dirfrag,
+                                session, req->r_request_started, -1,
+                                &req->r_caps_reservation);
+               if (err < 0)
+                       return err;
+       }
+
+       if (rinfo->head->is_dentry && !req->r_aborted) {
                /*
                 * lookup link rename   : null -> possibly existing inode
                 * mknod symlink mkdir  : null -> new inode
@@ -932,12 +942,6 @@ int ceph_fill_trace(struct super_block *sb, struct ceph_mds_request *req,
                BUG_ON(ceph_snap(dir) !=
                       le64_to_cpu(rinfo->diri.in->snapid));
 
-               err = fill_inode(dir, &rinfo->diri, rinfo->dirfrag,
-                                session, req->r_request_started, -1,
-                                &req->r_caps_reservation);
-               if (err < 0)
-                       return err;
-
                /* do we have a lease on the whole dir? */
                have_dir_cap =
                        (le32_to_cpu(rinfo->diri.in->cap.caps) &
index 6e08f488a30f0d8535dcb43779855531e82bef38..623c67cd484ba35f8dc43c323a8a3c381ebd5a2d 100644 (file)
@@ -1624,11 +1624,29 @@ int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
                err = PTR_ERR(req->r_reply);
                req->r_reply = NULL;
 
-               /* clean up */
-               __unregister_request(mdsc, req);
-               if (!list_empty(&req->r_unsafe_item))
-                       list_del_init(&req->r_unsafe_item);
-               complete(&req->r_safe_completion);
+               if (err == -ERESTARTSYS) {
+                       /* aborted */
+                       req->r_aborted = true;
+
+                       if (req->r_locked_dir &&
+                           (req->r_op & CEPH_MDS_OP_WRITE)) {
+                               struct ceph_inode_info *ci =
+                                       ceph_inode(req->r_locked_dir);
+
+                               dout("aborted, clearing I_COMPLETE on %p\n", 
+                                    req->r_locked_dir);
+                               spin_lock(&req->r_locked_dir->i_lock);
+                               ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
+                               ci->i_release_count++;
+                               spin_unlock(&req->r_locked_dir->i_lock);
+                       }
+               } else {
+                       /* clean up this request */
+                       __unregister_request(mdsc, req);
+                       if (!list_empty(&req->r_unsafe_item))
+                               list_del_init(&req->r_unsafe_item);
+                       complete(&req->r_safe_completion);
+               }
        } else if (req->r_err) {
                err = req->r_err;
        } else {
index b1c2025227c5d45c9b74e2238fe1dbe84a7dabd5..ee71495e27c4c6d7a6d0aa24e9aa41131e860258 100644 (file)
@@ -188,6 +188,7 @@ struct ceph_mds_request {
        struct ceph_msg  *r_reply;
        struct ceph_mds_reply_info_parsed r_reply_info;
        int r_err;
+       bool r_aborted;
 
        unsigned long r_timeout;  /* optional.  jiffies */
        unsigned long r_started;  /* start time to measure timeout against */