mds: wait the linkmerge/migrate to finish after unlink

If one inode has more than one link and after one of its dentries
being unlinked it will be moved to stray directory. Before the
linkmerge/migrate finises if a link request comes it will fail
with -EXDEV.

While in non-multiple link case it's also possible that the clients
could pass one invalidate ino, which is still under unlinking.

Just wait the linkmerge/migrate or purge to finish.

https://tracker.ceph.com/issues/56695
Signed-off-by: Xiubo Li <xiubli@redhat.com>
(cherry picked from commit 48f9a8990b402daa0f0f9109bc4b0087c01b33c6)

diff --git a/src/mds/CInode.h b/src/mds/CInode.h
index a2b034ec647c9061fcbe0d12f60dcdeda05198a2..3f18ab0fe864b199c0dc1f5f973bd77b3feec8d8 100644 (file)
--- a/src/mds/CInode.h
+++ b/src/mds/CInode.h
@@ -398,7 +398,8 @@ class CInode : public MDSCacheObject, public InodeStoreBase, public Counter<CIno
   static const uint64_t WAIT_FROZEN      = (1<<1);
   static const uint64_t WAIT_TRUNC       = (1<<2);
   static const uint64_t WAIT_FLOCK       = (1<<3);
-  
+  static const uint64_t WAIT_UNLINK      = (1<<4);
+
   static const uint64_t WAIT_ANY_MASK  = (uint64_t)(-1);
 
   // misc

diff --git a/src/mds/Server.cc b/src/mds/Server.cc
index fbac61e0870514881fca4b3d76ecf13eae290cb1..eaaec5d7bfd691b0ee32270567b78f9834e5581b 100644 (file)
--- a/src/mds/Server.cc
+++ b/src/mds/Server.cc
@@ -7209,6 +7209,13 @@ void Server::handle_client_link(MDRequestRef& mdr)
   if (target_pin != dir->inode &&
       target_realm->get_subvolume_ino() !=
       dir->inode->find_snaprealm()->get_subvolume_ino()) {
+    if (target_pin->is_stray()) {
+      mds->locker->drop_locks(mdr.get());
+      targeti->add_waiter(CInode::WAIT_UNLINK,
+                          new C_MDS_RetryRequest(mdcache, mdr));
+      mdlog->flush();
+      return;
+    }
     dout(7) << "target is in different subvolume, failing..." << dendl;
     respond_to_request(mdr, -CEPHFS_EXDEV);
     return;
@@ -9024,6 +9031,12 @@ void Server::handle_client_rename(MDRequestRef& mdr)
   C_MDS_rename_finish *fin = new C_MDS_rename_finish(this, mdr, srcdn, destdn, straydn);
 
   journal_and_reply(mdr, srci, destdn, le, fin);
+
+  // trigger to flush mdlog in case reintegrating or migrating the stray dn,
+  // because the link requests maybe waiting.
+  if (srcdn->get_dir()->inode->is_stray()) {
+    mdlog->flush();
+  }
   mds->balancer->maybe_fragment(destdn->get_dir(), false);
 }
 
@@ -9624,6 +9637,14 @@ void Server::_rename_apply(MDRequestRef& mdr, CDentry *srcdn, CDentry *destdn, C
 
   srcdn->get_dir()->unlink_inode(srcdn);
 
+  // After the stray dn being unlinked from the corresponding inode in case of
+  // reintegrate_stray/migrate_stray, just wake up the waitiers.
+  MDSContext::vec finished;
+  in->take_waiting(CInode::WAIT_UNLINK, finished);
+  if (!finished.empty()) {
+    mds->queue_waiters(finished);
+  }
+
   // dest
   if (srcdn_was_remote) {
     if (!linkmerge) {

diff --git a/src/mds/StrayManager.cc b/src/mds/StrayManager.cc
index ab8f0ddb49e260c599dbe508d1aebb61afe145ec..4aae3edead8f1f0c50a15133bb2eaff982230c80 100644 (file)
--- a/src/mds/StrayManager.cc
+++ b/src/mds/StrayManager.cc
@@ -281,6 +281,17 @@ void StrayManager::_purge_stray_logged(CDentry *dn, version_t pdv, MutationRef&
     dir->remove_dentry(dn);
   }
 
+  // Once we are here normally the waiter list are mostly empty
+  // but in corner case that the clients pass a invalidate ino,
+  // which maybe under unlinking, the link caller will add the
+  // request to the waiter list. We need try to wake them up
+  // anyway.
+  MDSContext::vec finished;
+  in->take_waiting(CInode::WAIT_UNLINK, finished);
+  if (!finished.empty()) {
+    mds->queue_waiters(finished);
+  }
+
   // drop inode
   inodeno_t ino = in->ino();
   if (in->is_dirty())