mds/rename: Handle referent remote linkmerge case

For a file having hardlinks, when the primary file is
removed first before the secondary, stray reintegration
makes any secondary link available in the cache to be
primary. This is referred as linkmerge.

When the linkmerge happens, the referent inode of the
secondary needs to be removed otherwise it causes the
referent inode object leak in the data pool. This PR
handles the linkmerge scenario and removed the referent
inode of the secondary during the linkmerge.

Also, this patch does the following.
  - removes the corresponding referent inode from the
    referent_inodes list of primary/real inode
  - adds is_referent_remote() check in required places
    for linkmerge in the rename path.

Fixes: https://tracker.ceph.com/issues/54205
Signed-off-by: Kotresh HR <khiremat@redhat.com>

diff --git a/src/mds/Server.cc b/src/mds/Server.cc
index 0901f7cff2d6591cb3622e9ae711d74ccc89bb9a..541158853e71f8f5010c91bf3a2c5d183a945cf4 100644 (file)
--- a/src/mds/Server.cc
+++ b/src/mds/Server.cc
@@ -9286,7 +9286,7 @@ void Server::handle_client_rename(const MDRequestRef& mdr)
   if (mdr->reqid.name.is_mds() &&
       !(MDS_INO_IS_STRAY(srcpath.get_ino()) &&
        MDS_INO_IS_STRAY(destpath.get_ino())) &&
-      !(destdnl->is_remote() &&
+      !((destdnl->is_remote() || destdnl->is_referent_remote()) &&
        destdnl->get_remote_ino() == srci->ino())) {
     respond_to_request(mdr, -EINVAL);  // actually, this won't reply, but whatev.
     return;
@@ -9365,6 +9365,12 @@ void Server::handle_client_rename(const MDRequestRef& mdr)
     if (!straydn)
       return;
     dout(10) << " straydn is " << *straydn << dendl;
+  } else if (destdnl->is_referent_remote()) { //both linkmerge and non linkmerge case
+    straydn = prepare_stray_dentry(mdr, destdnl->get_referent_inode());
+    if (!straydn)
+      return;
+    dout(10) << __func__ << (linkmerge ? " linkmerge:yes ": " linkmerge:no ")
+             << " referent straydn is " << *straydn << dendl;
   } else if (mdr->straydn) {
     mdr->unpin(mdr->straydn);
     mdr->straydn = NULL;
@@ -9414,7 +9420,7 @@ void Server::handle_client_rename(const MDRequestRef& mdr)
   }
 
   if (linkmerge)
-    ceph_assert(srcdir->inode->is_stray() && srcdnl->is_primary() && destdnl->is_remote());
+    ceph_assert(srcdir->inode->is_stray() && srcdnl->is_primary() && (destdnl->is_remote() || destdnl->is_referent_remote()));
 
   if ((!mdr->has_more() || mdr->more()->witnessed.empty())) {
     if (!check_access(mdr, srcdir->get_inode(), MAY_WRITE))
@@ -9562,10 +9568,10 @@ void Server::handle_client_rename(const MDRequestRef& mdr)
     srcdn->list_replicas(witnesses);
   else
     witnesses.insert(srcdn->authority().first);
-  if (srcdnl->is_remote() && !srci->is_auth())
+  if ((srcdnl->is_remote() || srcdnl->is_referent_remote()) && !srci->is_auth())
     witnesses.insert(srci->authority().first);
   destdn->list_replicas(witnesses);
-  if (destdnl->is_remote() && !oldin->is_auth())
+  if ((destdnl->is_remote() || destdnl->is_referent_remote()) && !oldin->is_auth())
     witnesses.insert(oldin->authority().first);
   dout(10) << " witnesses " << witnesses << ", have " << mdr->more()->witnessed << dendl;
 
@@ -9699,7 +9705,7 @@ void Server::_rename_finish(const MDRequestRef& mdr, CDentry *srcdn, CDentry *de
   
   // bump popularity
   mds->balancer->hit_dir(srcdn->get_dir(), META_POP_IWR);
-  if (destdnl->is_remote() && in->is_auth())
+  if ((destdnl->is_remote() || destdnl->is_referent_remote()) && in->is_auth())
     mds->balancer->hit_inode(in, META_POP_IWR);
 
   // did we import srci?  if so, explicitly ack that import that, before we unlock and reply.
@@ -9854,7 +9860,7 @@ void Server::_rename_prepare(const MDRequestRef& mdr,
   // primary+remote link merge?
   bool linkmerge = (srci == oldin);
   if (linkmerge)
-    ceph_assert(srcdnl->is_primary() && destdnl->is_remote());
+    ceph_assert(srcdnl->is_primary() && (destdnl->is_remote() || destdnl->is_referent_remote()));
   bool silent = srcdn->get_dir()->inode->is_stray();
 
   bool force_journal_dest = false;
@@ -9892,6 +9898,7 @@ void Server::_rename_prepare(const MDRequestRef& mdr,
   // prepare
   CInode::mempool_inode *spi = 0;    // renamed inode
   CInode::mempool_inode *tpi = 0;  // target/overwritten inode
+  CInode::mempool_inode *trpi = 0;  // target/overwritten referent inode
   
   // target inode
   if (!linkmerge) {
@@ -9913,6 +9920,21 @@ void Server::_rename_prepare(const MDRequestRef& mdr,
         tpi = pi.inode.get();
       }
     }
+  } else if (linkmerge && destdnl->is_referent_remote()) {
+      CInode *oldrefi = destdnl->get_referent_inode();
+      dout(10) << __func__ << " linkmerge and destdnl is referent. oldrefi= " << *oldrefi << dendl;
+      ceph_assert(straydn && oldrefi);  // moving referent inode to straydn.
+
+      // link--, and move.
+      if (destdn->is_auth()) {
+       auto pi= oldrefi->project_inode(mdr); //project_snaprealm
+       pi.inode->version = straydn->pre_dirty(pi.inode->version);
+       //backtrace updation is not required
+        trpi = pi.inode.get();
+        trpi->nlink = 0;
+        oldrefi->state_set(CInode::STATE_ORPHAN);
+      }
+      straydn->push_projected_linkage(oldrefi);
   }
 
   // dest
@@ -9964,6 +9986,16 @@ void Server::_rename_prepare(const MDRequestRef& mdr,
       pi.inode->version = mdr->more()->pvmap[destdn] = destdn->pre_dirty(oldpv);
       pi.inode->update_backtrace();
       spi = pi.inode.get();
+
+      // Remove referent inode from primary link on linkmerge
+      if (linkmerge && destdnl->is_referent_remote()) {
+        CInode *oldrefi = destdnl->get_referent_inode();
+        spi->remove_referent_ino(oldrefi->ino());
+        dout(20) << __func__ << " linkmerge referent_inodes " << std::hex
+                 << spi->get_referent_inodes() << " referent ino removed "
+                << oldrefi->ino() << dendl;
+      }
+
     }
     destdn->push_projected_linkage(srci);
   }
@@ -9997,6 +10029,18 @@ void Server::_rename_prepare(const MDRequestRef& mdr,
       if (tpi->nlink == 0)
        oldin->state_set(CInode::STATE_ORPHAN);
     }
+    if (trpi) {
+      trpi->ctime = mdr->get_op_stamp();
+      if (mdr->get_op_stamp() > trpi->rstat.rctime)
+       trpi->rstat.rctime = mdr->get_op_stamp();
+      trpi->change_attr++;
+      {
+        std::string t;
+        destdn->make_path_string(t, true);
+        dout(20) << __func__ << " referent stray_prior_path = " << t << dendl;
+        trpi->stray_prior_path = std::move(t);
+      }
+    }
   }
 
   // prepare nesting, mtime updates
@@ -10010,9 +10054,15 @@ void Server::_rename_prepare(const MDRequestRef& mdr,
       metablob->add_dir_context(straydn->get_dir());
       metablob->add_dir(straydn->get_dir(), true);
     }
+  } else if (destdnl->is_referent_remote()) { //both linkmerge and non linkmerge case
+    ceph_assert(straydn);
+    if (straydn->is_auth()) {
+      metablob->add_dir_context(straydn->get_dir());
+      metablob->add_dir(straydn->get_dir(), true);
+    }
   }
 
-  if (!linkmerge && destdnl->is_remote() && oldin->is_auth()) {
+  if (!linkmerge && (destdnl->is_remote() || destdnl->is_referent_remote()) && oldin->is_auth()) {
     CDir *oldin_dir = oldin->get_projected_parent_dir();
     if (oldin_dir != srcdn->get_dir() && oldin_dir != destdn->get_dir())
       mdcache->predirty_journal_parents(mdr, metablob, oldin, oldin_dir, PREDIRTY_PRIMARY);
@@ -10026,10 +10076,16 @@ void Server::_rename_prepare(const MDRequestRef& mdr,
       ceph_assert(straydn);
       mdcache->predirty_journal_parents(mdr, metablob, oldin, straydn->get_dir(),
                                        PREDIRTY_PRIMARY|PREDIRTY_DIR, 1);
+    } else if (destdnl->is_referent_remote()) { // both linkmerge and non linkmerge case
+      CInode *oldrefi = destdnl->get_referent_inode();
+      ceph_assert(straydn);
+      ceph_assert(oldrefi);
+      mdcache->predirty_journal_parents(mdr, metablob, oldrefi, straydn->get_dir(),
+                                       PREDIRTY_PRIMARY|PREDIRTY_DIR, 1);
     }
   }
 
-  if (srcdnl->is_remote() && srci->is_auth()) {
+  if ((srcdnl->is_remote() || srcdnl->is_referent_remote()) && srci->is_auth()) {
     CDir *srci_dir = srci->get_projected_parent_dir();
     if (srci_dir != srcdn->get_dir() && srci_dir != destdn->get_dir())
       mdcache->predirty_journal_parents(mdr, metablob, srci, srci_dir, PREDIRTY_PRIMARY);
@@ -10063,7 +10119,7 @@ void Server::_rename_prepare(const MDRequestRef& mdr,
        metablob->add_dir_context(straydn->get_dir());
        metablob->add_primary_dentry(straydn, oldin, true);
       }
-    } else if (destdnl->is_remote()) {
+    } else if (destdnl->is_remote() || destdnl->is_referent_remote()) {
       if (oldin->is_auth()) {
        sr_t *new_srnode = NULL;
        if (mdr->peer_request) {
@@ -10087,10 +10143,25 @@ void Server::_rename_prepare(const MDRequestRef& mdr,
        metablob->add_primary_dentry(oldin_pdn, oldin, true);
       }
     }
+  } else if (linkmerge && destdnl->is_referent_remote()) {
+    CInode *oldrefi = destdnl->get_referent_inode();
+    ceph_assert(straydn);
+    ceph_assert(oldrefi);
+    if (destdn->is_auth()) {
+      // No snapshot on referent inode - Ignoring snaprealm projection
+      // Updating first is required here ? Mostly not required, keeping
+      // it for now.
+      straydn->first = mdcache->get_global_snaprealm()->get_newest_seq() + 1;
+      metablob->add_primary_dentry(straydn, oldrefi, true, true);
+    } else if (force_journal_stray) {
+      dout(10) << __func__ << " forced journaling referent straydn " << *straydn << dendl;
+      metablob->add_dir_context(straydn->get_dir());
+      metablob->add_primary_dentry(straydn, oldrefi, true);
+    }
   }
 
   // dest
-  if (srcdnl->is_remote()) {
+  if (srcdnl->is_remote() || srcdnl->is_referent_remote()) {
     ceph_assert(!linkmerge);
     if (destdn->is_auth() && !destdnl->is_null())
       mdcache->journal_cow_dentry(mdr.get(), metablob, destdn, CEPH_NOSNAP, 0, destdnl);
@@ -10179,8 +10250,14 @@ void Server::_rename_prepare(const MDRequestRef& mdr,
   }
   // make stray inode first track the straydn
   if (straydn && straydn->is_auth()) {
-    ceph_assert(oldin->first <= straydn->first);
-    oldin->first = straydn->first;
+    if (destdnl->is_referent_remote()) { //both linkmerge and non-linkmerge case
+      CInode *oldrefi = destdnl->get_referent_inode();
+      ceph_assert(oldrefi->first <= straydn->first);
+      oldrefi->first = straydn->first;
+    } else {
+      ceph_assert(oldin->first <= straydn->first);
+      oldin->first = straydn->first;
+    }
   }
 
   if (oldin && oldin->is_dir()) {
@@ -10206,7 +10283,7 @@ void Server::_rename_apply(const MDRequestRef& mdr, CDentry *srcdn, CDentry *des
   // primary+remote link merge?
   bool linkmerge = (srcdnl->get_inode() == oldin);
   if (linkmerge)
-    ceph_assert(srcdnl->is_primary() && destdnl->is_remote());
+    ceph_assert(srcdnl->is_primary() && (destdnl->is_remote() || destdnl->is_referent_remote()));
 
   bool new_in_snaprealm = false;
   bool new_oldin_snaprealm = false;
@@ -10312,8 +10389,28 @@ void Server::_rename_apply(const MDRequestRef& mdr, CDentry *srcdn, CDentry *des
     }
   } else { // primary
     if (linkmerge) {
-      dout(10) << "merging primary onto remote link" << dendl;
-      destdn->get_dir()->unlink_inode(destdn, false);
+      if (destdnl->is_referent_remote()) {
+        dout(10) << __func__ << " merging primary onto referent link" << dendl;
+        CInode *oldrefi = destdnl->get_referent_inode();
+        ceph_assert(straydn);
+        ceph_assert(oldrefi);
+        dout(10) << __func__ << " linkmerge referent straydn is " << *straydn << dendl;
+        // no need to handle any snaprealm related stuff for referent inode
+        destdn->get_dir()->unlink_inode(destdn, false);
+
+        straydn->pop_projected_linkage();
+        if (mdr->is_peer() && !mdr->more()->peer_update_journaled)
+          ceph_assert(!straydn->is_projected()); // no other projected
+
+        // nlink-- referent inode
+        if (destdn->is_auth())
+          oldrefi->pop_and_dirty_projected_inode(mdr->ls, mdr);
+
+        mdcache->touch_dentry_bottom(straydn);  // drop dn as quickly as possible.
+      } else if (destdnl->is_remote()) {
+        dout(10) << __func__ << " merging primary onto remote link" << dendl;
+        destdn->get_dir()->unlink_inode(destdn, false);
+      }
     }
     destdnl = destdn->pop_projected_linkage();
     if (mdr->is_peer() && !mdr->more()->peer_update_journaled)
@@ -10481,10 +10578,12 @@ void Server::handle_peer_rename_prep(const MDRequestRef& mdr)
   // stray?
   bool linkmerge = srcdnl->get_inode() == destdnl->get_inode();
   if (linkmerge)
-    ceph_assert(srcdnl->is_primary() && destdnl->is_remote());
+    ceph_assert(srcdnl->is_primary() && (destdnl->is_remote() || destdnl->is_referent_remote()));
   CDentry *straydn = mdr->straydn;
   if (destdnl->is_primary() && !linkmerge)
     ceph_assert(straydn);
+  if (destdnl->is_referent_remote()) //both linkmerge and non-linkmerge case
+    ceph_assert(straydn);
 
   mdr->set_op_stamp(mdr->peer_request->op_stamp);
   mdr->more()->srcdn_auth_mds = srcdn->authority().first;
@@ -10597,7 +10696,7 @@ void Server::handle_peer_rename_prep(const MDRequestRef& mdr)
   if (srcdnl->is_primary())
     rollback.orig_src.ino = srcdnl->get_inode()->ino();
   else {
-    ceph_assert(srcdnl->is_remote());
+    ceph_assert(srcdnl->is_remote() || srcdnl->is_referent_remote());
     rollback.orig_src.remote_ino = srcdnl->get_remote_ino();
     rollback.orig_src.remote_d_type = srcdnl->get_remote_d_type();
   }
@@ -10608,7 +10707,7 @@ void Server::handle_peer_rename_prep(const MDRequestRef& mdr)
   rollback.orig_dest.dname = destdn->get_name();
   if (destdnl->is_primary())
     rollback.orig_dest.ino = destdnl->get_inode()->ino();
-  else if (destdnl->is_remote()) {
+  else if (destdnl->is_remote() || destdnl->is_referent_remote()) {
     rollback.orig_dest.remote_ino = destdnl->get_remote_ino();
     rollback.orig_dest.remote_d_type = destdnl->get_remote_d_type();
   }