6 files changed, 174 insertions, 162 deletions

diff --git a/fs/ceph/caps.c b/fs/ceph/caps.c
index f3496db4bb3e..94c026bba2c2 100644
--- a/fs/ceph/caps.c
+++ b/fs/ceph/caps.c
@@ -657,6 +657,9 @@ void ceph_add_cap(struct inode *inode,
                 session->s_nr_caps++;
                 spin_unlock(&session->s_cap_lock);
         } else {
+                if (cap->cap_gen < session->s_cap_gen)
+                        cap->issued = cap->implemented = CEPH_CAP_PIN;
+
                 /*
                  * auth mds of the inode changed. we received the cap export
                  * message, but still haven't received the cap import message.
@@ -1855,14 +1858,17 @@ retry_locked:
                         retain |= CEPH_CAP_ANY;       /* be greedy */
                 } else if (S_ISDIR(inode->i_mode) &&
                            (issued & CEPH_CAP_FILE_SHARED) &&
-                            __ceph_dir_is_complete(ci)) {
+                           __ceph_dir_is_complete(ci)) {
                         /*
                          * If a directory is complete, we want to keep
                          * the exclusive cap. So that MDS does not end up
                          * revoking the shared cap on every create/unlink
                          * operation.
                          */
-                        want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
+                        if (IS_RDONLY(inode))
+                                want = CEPH_CAP_ANY_SHARED;
+                        else
+                                want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
                         retain |= want;
                 } else {
 
@@ -1970,8 +1976,7 @@ retry_locked:
                         goto ack;
 
                 /* things we might delay */
-                if ((cap->issued & ~retain) == 0 &&
-                    cap->mds_wanted == want)
+                if ((cap->issued & ~retain) == 0)
                         continue;     /* nope, all good */
 
                 if (no_delay)
@@ -3048,7 +3053,8 @@ static void handle_cap_grant(struct inode *inode,
         int used, wanted, dirty;
         u64 size = le64_to_cpu(grant->size);
         u64 max_size = le64_to_cpu(grant->max_size);
-        int check_caps = 0;
+        unsigned char check_caps = 0;
+        bool was_stale = cap->cap_gen < session->s_cap_gen;
         bool wake = false;
         bool writeback = false;
         bool queue_trunc = false;
@@ -3063,21 +3069,6 @@ static void handle_cap_grant(struct inode *inode,
 
 
         /*
-         * auth mds of the inode changed. we received the cap export message,
-         * but still haven't received the cap import message. handle_cap_export
-         * updated the new auth MDS' cap.
-         *
-         * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
-         * that was sent before the cap import message. So don't remove caps.
-         */
-        if (ceph_seq_cmp(seq, cap->seq) <= 0) {
-                WARN_ON(cap != ci->i_auth_cap);
-                WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
-                seq = cap->seq;
-                newcaps |= cap->issued;
-        }
-
-        /*
          * If CACHE is being revoked, and we have no dirty buffers,
          * try to invalidate (once).  (If there are dirty buffers, we
          * will invalidate _after_ writeback.)
@@ -3096,6 +3087,24 @@ static void handle_cap_grant(struct inode *inode,
                 }
         }
 
+        if (was_stale)
+                cap->issued = cap->implemented = CEPH_CAP_PIN;
+
+        /*
+         * auth mds of the inode changed. we received the cap export message,
+         * but still haven't received the cap import message. handle_cap_export
+         * updated the new auth MDS' cap.
+         *
+         * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
+         * that was sent before the cap import message. So don't remove caps.
+         */
+        if (ceph_seq_cmp(seq, cap->seq) <= 0) {
+                WARN_ON(cap != ci->i_auth_cap);
+                WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
+                seq = cap->seq;
+                newcaps |= cap->issued;
+        }
+
         /* side effects now are allowed */
         cap->cap_gen = session->s_cap_gen;
         cap->seq = seq;
@@ -3200,13 +3209,20 @@ static void handle_cap_grant(struct inode *inode,
              ceph_cap_string(wanted),
              ceph_cap_string(used),
              ceph_cap_string(dirty));
-        if (wanted != le32_to_cpu(grant->wanted)) {
-                dout("mds wanted %s -> %s\n",
-                     ceph_cap_string(le32_to_cpu(grant->wanted)),
-                     ceph_cap_string(wanted));
-                /* imported cap may not have correct mds_wanted */
-                if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT)
-                        check_caps = 1;
+
+        if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
+            (wanted & ~(cap->mds_wanted | newcaps))) {
+                /*
+                 * If mds is importing cap, prior cap messages that update
+                 * 'wanted' may get dropped by mds (migrate seq mismatch).
+                 *
+                 * We don't send cap message to update 'wanted' if what we
+                 * want are already issued. If mds revokes caps, cap message
+                 * that releases caps also tells mds what we want. But if
+                 * caps got revoked by mds forcedly (session stale). We may
+                 * haven't told mds what we want.
+                 */
+                check_caps = 1;
         }
 
         /* revocation, grant, or no-op? */
@@ -3539,9 +3555,9 @@ retry:
                 goto out_unlock;
 
         if (target < 0) {
-                __ceph_remove_cap(cap, false);
-                if (!ci->i_auth_cap)
+                if (cap->mds_wanted | cap->issued)
                         ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
+                __ceph_remove_cap(cap, false);
                 goto out_unlock;
         }
 
@@ -3569,7 +3585,6 @@ retry:
                         tcap->cap_id = t_cap_id;
                         tcap->seq = t_seq - 1;
                         tcap->issue_seq = t_seq - 1;
-                        tcap->mseq = t_mseq;
                         tcap->issued |= issued;
                         tcap->implemented |= issued;
                         if (cap == ci->i_auth_cap)
diff --git a/fs/ceph/inode.c b/fs/ceph/inode.c
index 79dd5e6ed755..9d1f34d46627 100644
--- a/fs/ceph/inode.c
+++ b/fs/ceph/inode.c
@@ -1098,8 +1098,9 @@ out_unlock:
  * splice a dentry to an inode.
  * caller must hold directory i_mutex for this to be safe.
  */
-static struct dentry *splice_dentry(struct dentry *dn, struct inode *in)
+static int splice_dentry(struct dentry **pdn, struct inode *in)
 {
+        struct dentry *dn = *pdn;
         struct dentry *realdn;
 
         BUG_ON(d_inode(dn));
@@ -1132,28 +1133,23 @@ static struct dentry *splice_dentry(struct dentry *dn, struct inode *in)
         if (IS_ERR(realdn)) {
                 pr_err("splice_dentry error %ld %p inode %p ino %llx.%llx\n",
                        PTR_ERR(realdn), dn, in, ceph_vinop(in));
-                dn = realdn;
-                /*
-                 * Caller should release 'dn' in the case of error.
-                 * If 'req->r_dentry' is passed to this function,
-                 * caller should leave 'req->r_dentry' untouched.
-                 */
-                goto out;
-        } else if (realdn) {
+                return PTR_ERR(realdn);
+        }
+
+        if (realdn) {
                 dout("dn %p (%d) spliced with %p (%d) "
                      "inode %p ino %llx.%llx\n",
                      dn, d_count(dn),
                      realdn, d_count(realdn),
                      d_inode(realdn), ceph_vinop(d_inode(realdn)));