mds: avoid journaling overhead for ceph.dir.subvolume for no-op case

In preparation for acquiring the xlock on the directory inode, the MDS
must journal a few events before continuing on with the setvxattr. This
can cause significant delays in the volumes ceph-mgr module which needs
to regularly enable this vxattr from multiple code paths. We could cache
in that module whether the vxattr is set but it's also pretty easy to
adjust the MDS to acquire a rdlock on the directory to check if the
subvolume flag is already set. That is much lighter weight and the lock
is generally readily available.

Fixes: https://tracker.ceph.com/issues/51276
Signed-off-by: Patrick Donnelly <pdonnell@redhat.com>

diff --git a/src/mds/Mutation.h b/src/mds/Mutation.h
index 9ce55270e599374b71f6d30dee0b781e1c7e7287..09ebe7052a90f79655a57a1358a67f31d40e2bbf 100644 (file)
--- a/src/mds/Mutation.h
+++ b/src/mds/Mutation.h
@@ -308,6 +308,7 @@ struct MDRequestImpl : public MutationImpl {
     bool is_ambiguous_auth = false;
     bool is_remote_frozen_authpin = false;
     bool is_inode_exporter = false;
+    bool rdonly_checks = false;
 
     std::map<client_t, std::pair<Session*, uint64_t> > imported_session_map;
     std::map<CInode*, std::map<client_t,Capability::Export> > cap_imports;

diff --git a/src/mds/Server.cc b/src/mds/Server.cc
index 63ebc303a044eb377af1fce195f7dc4f3913ef33..856d06130f8604457ff079317b9ca08261084183 100644 (file)
--- a/src/mds/Server.cc
+++ b/src/mds/Server.cc
@@ -5689,9 +5689,36 @@ void Server::handle_set_vxattr(MDRequestRef& mdr, CInode *cur)
       return;
     }
 
-    if (!xlock_policylock(mdr, cur, false, true))
-      return;
+    /* Verify it's not already a subvolume with lighter weight
+     * rdlock.
+     */
+    if (!mdr->more()->rdonly_checks) {
+      if (!(mdr->locking_state & MutationImpl::ALL_LOCKED)) {
+        MutationImpl::LockOpVec lov;
+        lov.add_rdlock(&cur->snaplock);
+        if (!mds->locker->acquire_locks(mdr, lov))
+          return;
+        mdr->locking_state |= MutationImpl::ALL_LOCKED;
+      }
+      SnapRealm *realm = cur->find_snaprealm();
+      const auto srnode = cur->get_projected_srnode();
+      if (val == (srnode && srnode->is_subvolume())) {
+        dout(20) << "already marked subvolume" << dendl;
+        respond_to_request(mdr, 0);
+        return;
+      }
+      mdr->more()->rdonly_checks = true;
+    }
+
+    if ((mdr->locking_state & MutationImpl::ALL_LOCKED) && !mdr->is_xlocked(&cur->snaplock)) {
+      /* drop the rdlock and acquire xlocks */
+      dout(20) << "dropping rdlocks" << dendl;
+      mds->locker->drop_locks(mdr.get());
+      if (!xlock_policylock(mdr, cur, false, true))
+        return;
+    }
 
+    /* repeat rdonly checks in case changed between rdlock -> xlock */
     SnapRealm *realm = cur->find_snaprealm();
     if (val) {
       inodeno_t subvol_ino = realm->get_subvolume_ino();