mds: wait reintegrate to finish when unlinking

When unlinking a remote dentry while primary dentry has been just
unlinked and was trying to reintegrate, which will finally be a
RENAME client request to current MDS, the stray dentry.

If the RENAME request is sleeping and waiting for some resources
just after authpin the CInode, so the unlinking request will mark
the remote dentry as UNLINKING state and then add itself to the
wait list just waiting for the CInode to be authpined.

Then after the RENAME request is retried later it will add to a
wait list again and waiting for the UNLINKING state to be cleared.
And finally the RENAME and UNLINK requests will be deadlock.

This fix will block the UNLINK request if the reintegrate_stray
is going on for current dentry.

Fixes: https://tracker.ceph.com/issues/58340
Signed-off-by: Xiubo Li <xiubli@redhat.com>
(cherry picked from commit c3b36723defae32b0a3ef1ac37b58ec57d356b65)

diff --git a/src/mds/CDentry.cc b/src/mds/CDentry.cc
index f76ba6579535fbb9ab1b0ac7d40941a570fe1608..28d684c33596c42719834ce5f3ef36727728f828 100644 (file)
--- a/src/mds/CDentry.cc
+++ b/src/mds/CDentry.cc
@@ -107,6 +107,8 @@ ostream& operator<<(ostream& out, const CDentry& dn)
   out << " state=" << dn.get_state();
   if (dn.is_new()) out << "|new";
   if (dn.state_test(CDentry::STATE_BOTTOMLRU)) out << "|bottomlru";
+  if (dn.state_test(CDentry::STATE_UNLINKING)) out << "|unlinking";
+  if (dn.state_test(CDentry::STATE_REINTEGRATING)) out << "|reintegrating";
 
   if (dn.get_num_ref()) {
     out << " |";

diff --git a/src/mds/CDentry.h b/src/mds/CDentry.h
index 83f32cd2534832046778612dcb9cb8226d5ccbd4..66fc64a3969e62421f36f2015406e63258400786 100644 (file)
--- a/src/mds/CDentry.h
+++ b/src/mds/CDentry.h
@@ -88,6 +88,7 @@ public:
   static const int STATE_PURGINGPINNED =  (1<<5);
   static const int STATE_BOTTOMLRU =    (1<<6);
   static const int STATE_UNLINKING =    (1<<7);
+  static const int STATE_REINTEGRATING = (1<<8);
   // stray dentry needs notification of releasing reference
   static const int STATE_STRAY =       STATE_NOTIFYREF;
   static const int MASK_STATE_IMPORT_KEPT = STATE_BOTTOMLRU;
@@ -101,8 +102,9 @@ public:
 
   static const unsigned EXPORT_NONCE = 1;
 
-  const static uint64_t WAIT_UNLINK_STATE  = (1<<0);
-  const static uint64_t WAIT_UNLINK_FINISH = (1<<1);
+  const static uint64_t WAIT_UNLINK_STATE       = (1<<0);
+  const static uint64_t WAIT_UNLINK_FINISH      = (1<<1);
+  const static uint64_t WAIT_REINTEGRATE_FINISH = (1<<2);
   uint32_t replica_unlinking_ref = 0;
 
   CDentry(std::string_view n, __u32 h,

diff --git a/src/mds/MDSRank.cc b/src/mds/MDSRank.cc
index d1294c8ce97103412f92d96552b91e14271525bc..82c73652fbf1b3f8ee40c6ce0d330fbd2a8b3fc3 100644 (file)
--- a/src/mds/MDSRank.cc
+++ b/src/mds/MDSRank.cc
@@ -1177,6 +1177,7 @@ bool MDSRank::is_valid_message(const cref_t<Message> &m) {
       type == CEPH_MSG_CLIENT_RECONNECT ||
       type == CEPH_MSG_CLIENT_RECLAIM ||
       type == CEPH_MSG_CLIENT_REQUEST ||
+      type == CEPH_MSG_CLIENT_REPLY ||
       type == MSG_MDS_PEER_REQUEST ||
       type == MSG_MDS_HEARTBEAT ||
       type == MSG_MDS_TABLE_REQUEST ||
@@ -1230,6 +1231,7 @@ void MDSRank::handle_message(const cref_t<Message> &m)
       ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_CLIENT);
       // fall-thru
     case CEPH_MSG_CLIENT_REQUEST:
+    case CEPH_MSG_CLIENT_REPLY:
       server->dispatch(m);
       break;
     case MSG_MDS_PEER_REQUEST:

diff --git a/src/mds/MDSRank.h b/src/mds/MDSRank.h
index fb2061eb548ab06168715920bc83fa121de35305..16e0a6433af2981440b013f837d87d08b7a626b4 100644 (file)
--- a/src/mds/MDSRank.h
+++ b/src/mds/MDSRank.h
@@ -151,6 +151,29 @@ class Finisher;
 class ScrubStack;
 class C_ExecAndReply;
 
+struct MDSMetaRequest {
+private:
+  int _op;
+  CDentry *_dentry;
+  ceph_tid_t _tid;
+public:
+  explicit MDSMetaRequest(int op, CDentry *dn, ceph_tid_t tid) :
+    _op(op), _dentry(dn), _tid(tid) {
+    if (_dentry) {
+      _dentry->get(CDentry::PIN_PURGING);
+    }
+  }
+  ~MDSMetaRequest() {
+    if (_dentry) {
+      _dentry->put(CDentry::PIN_PURGING);
+    }
+  }
+
+  CDentry *get_dentry() { return _dentry; }
+  int get_op() { return _op; }
+  ceph_tid_t get_tid() { return _tid; }
+};
+
 /**
  * The public part of this class's interface is what's exposed to all
  * the various subsystems (server, mdcache, etc), such as pointers
@@ -412,6 +435,8 @@ class MDSRank {
     PerfCounters *logger = nullptr, *mlogger = nullptr;
     OpTracker op_tracker;
 
+    std::map<ceph_tid_t, MDSMetaRequest> internal_client_requests;
+
     // The last different state I held before current
     MDSMap::DaemonState last_state = MDSMap::STATE_BOOT;
     // The state assigned to me by the MDSMap

diff --git a/src/mds/Server.cc b/src/mds/Server.cc
index fbac61e0870514881fca4b3d76ecf13eae290cb1..d12accb6cf5630d8dc14bc2c7336f19c12f795cb 100644 (file)
--- a/src/mds/Server.cc
+++ b/src/mds/Server.cc
@@ -31,6 +31,7 @@
 #include "Mutation.h"
 #include "MetricsHandler.h"
 #include "cephfs_features.h"
+#include "MDSContext.h"
 
 #include "msg/Messenger.h"
 
@@ -357,6 +358,9 @@ void Server::dispatch(const cref_t<Message> &m)
   case CEPH_MSG_CLIENT_REQUEST:
     handle_client_request(ref_cast<MClientRequest>(m));
     return;
+  case CEPH_MSG_CLIENT_REPLY:
+    handle_client_reply(ref_cast<MClientReply>(m));
+    return;
   case CEPH_MSG_CLIENT_RECLAIM:
     handle_client_reclaim(ref_cast<MClientReclaim>(m));
     return;
@@ -2287,6 +2291,10 @@ void Server::reply_client_request(MDRequestRef& mdr, const ref_t<MClientReply> &
     mds->send_message_client(reply, session);
   }
 
+  if (client_inst.name.is_mds() && reply->get_op() == CEPH_MDS_OP_RENAME) {
+    mds->send_message(reply, mdr->client_request->get_connection());
+  }
+
   if (req->is_queued_for_replay() &&
       (mdr->has_completed || reply->get_result() < 0)) {
     if (reply->get_result() < 0) {
@@ -2518,6 +2526,38 @@ void Server::handle_client_request(const cref_t<MClientRequest> &req)
   return;
 }
 
+void Server::handle_client_reply(const cref_t<MClientReply> &reply)
+{
+  dout(4) << "handle_client_reply " << *reply << dendl;
+
+  ceph_assert(reply->is_safe());
+  ceph_tid_t tid = reply->get_tid();
+
+  if (mds->internal_client_requests.count(tid) == 0) {
+    dout(1) << " no pending request on tid " << tid << dendl;
+    return;
+  }
+
+  auto &req = mds->internal_client_requests.at(tid);
+  CDentry *dn = req.get_dentry();
+
+  switch (reply->get_op()) {
+  case CEPH_MDS_OP_RENAME:
+    if (dn) {
+      dn->state_clear(CDentry::STATE_REINTEGRATING);
+
+      MDSContext::vec finished;
+      dn->take_waiting(CDentry::WAIT_REINTEGRATE_FINISH, finished);
+      mds->queue_waiters(finished);
+    }
+    break;
+  default:
+    dout(5) << " unknown client op " << reply->get_op() << dendl;
+  }
+
+  mds->internal_client_requests.erase(tid);
+}
+
 void Server::handle_osd_map()
 {
   /* Note that we check the OSDMAP_FULL flag directly rather than
@@ -6773,6 +6813,45 @@ void Server::wait_for_pending_unlink(CDentry *dn, MDRequestRef& mdr)
   dn->add_waiter(CDentry::WAIT_UNLINK_FINISH, new C_WaitUnlinkToFinish(mdcache, dn, fin));
 }
 
+struct C_WaitReintegrateToFinish : public MDSContext {
+protected:
+  MDCache *mdcache;
+  CDentry *dn;
+  MDSContext *fin;
+
+  MDSRank *get_mds() override
+  {
+    ceph_assert(mdcache != NULL);
+    return mdcache->mds;
+  }
+
+public:
+  C_WaitReintegrateToFinish(MDCache *m, CDentry *d, MDSContext *f) :
+    mdcache(m), dn(d), fin(f) {}
+  void finish(int r) override {
+    fin->complete(r);
+    dn->put(CDentry::PIN_PURGING);
+  }
+};
+
+bool Server::is_reintegrate_pending(CDentry *dn)
+{
+  CDentry::linkage_t *dnl = dn->get_projected_linkage();
+  if (!dnl->is_null() && dn->state_test(CDentry::STATE_REINTEGRATING)) {
+      return true;
+  }
+  return false;
+}
+
+void Server::wait_for_pending_reintegrate(CDentry *dn, MDRequestRef& mdr)
+{
+  dout(20) << __func__ << " dn " << *dn << dendl;
+  mds->locker->drop_locks(mdr.get());
+  auto fin = new C_MDS_RetryRequest(mdcache, mdr);
+  dn->get(CDentry::PIN_PURGING);
+  dn->add_waiter(CDentry::WAIT_REINTEGRATE_FINISH, new C_WaitReintegrateToFinish(mdcache, dn, fin));
+}
+
 // MKNOD
 
 class C_MDS_mknod_finish : public ServerLogContext {
@@ -7828,6 +7907,11 @@ void Server::handle_client_unlink(MDRequestRef& mdr)
   if (!dn)
     return;
 
+  if (is_reintegrate_pending(dn)) {
+    wait_for_pending_reintegrate(dn, mdr);
+    return;
+  }
+
   // notify replica MDSes the dentry is under unlink
   if (!dn->state_test(CDentry::STATE_UNLINKING)) {
     dn->state_set(CDentry::STATE_UNLINKING);

diff --git a/src/mds/Server.h b/src/mds/Server.h
index a2d8f69670871d043fb824fa10d406124aa6bc68..eafa5108474de87e5d52b88d7a907f454728278e 100644 (file)
--- a/src/mds/Server.h
+++ b/src/mds/Server.h
@@ -158,6 +158,7 @@ public:
 
   // -- requests --
   void handle_client_request(const cref_t<MClientRequest> &m);
+  void handle_client_reply(const cref_t<MClientReply> &m);
 
   void journal_and_reply(MDRequestRef& mdr, CInode *tracei, CDentry *tracedn,
                         LogEvent *le, MDSLogContextBase *fin);
@@ -237,6 +238,9 @@ public:
   bool is_unlink_pending(CDentry *dn);
   void wait_for_pending_unlink(CDentry *dn, MDRequestRef& mdr);
 
+  bool is_reintegrate_pending(CDentry *dn);
+  void wait_for_pending_reintegrate(CDentry *dn, MDRequestRef& mdr);
+
   // open
   void handle_client_open(MDRequestRef& mdr);
   void handle_client_openc(MDRequestRef& mdr);  // O_CREAT variant.

diff --git a/src/mds/StrayManager.cc b/src/mds/StrayManager.cc
index 61e4f23fd7a3b42fe163e08cecfb401ca7db734b..412192b3a3acf991208bc37c3c95b4f9872fc2df 100644 (file)
--- a/src/mds/StrayManager.cc
+++ b/src/mds/StrayManager.cc
@@ -670,19 +670,27 @@ void StrayManager::reintegrate_stray(CDentry *straydn, CDentry *rdn)
   dout(10) << __func__ << " " << *straydn << " to " << *rdn << dendl;
 
   logger->inc(l_mdc_strays_reintegrated);
-  
+
   // rename it to remote linkage .
   filepath src(straydn->get_name(), straydn->get_dir()->ino());
   filepath dst(rdn->get_name(), rdn->get_dir()->ino());
 
+  ceph_tid_t tid = mds->issue_tid();
+
   auto req = make_message<MClientRequest>(CEPH_MDS_OP_RENAME);
   req->set_filepath(dst);
   req->set_filepath2(src);
-  req->set_tid(mds->issue_tid());
+  req->set_tid(tid);
+
+  rdn->state_set(CDentry::STATE_REINTEGRATING);
+  mds->internal_client_requests.emplace(std::piecewise_construct,
+                                        std::make_tuple(tid),
+                                        std::make_tuple(CEPH_MDS_OP_RENAME,
+                                                        rdn, tid));
 
   mds->send_message_mds(req, rdn->authority().first);
 }
- 
+
 void StrayManager::migrate_stray(CDentry *dn, mds_rank_t to)
 {
   dout(10) << __func__ << " " << *dn << " to mds." << to << dendl;
@@ -696,10 +704,17 @@ void StrayManager::migrate_stray(CDentry *dn, mds_rank_t to)
   filepath src(dn->get_name(), dirino);
   filepath dst(dn->get_name(), MDS_INO_STRAY(to, MDS_INO_STRAY_INDEX(dirino)));
 
+  ceph_tid_t tid = mds->issue_tid();
+
   auto req = make_message<MClientRequest>(CEPH_MDS_OP_RENAME);
   req->set_filepath(dst);
   req->set_filepath2(src);
-  req->set_tid(mds->issue_tid());
+  req->set_tid(tid);
+
+  mds->internal_client_requests.emplace(std::piecewise_construct,
+                                        std::make_tuple(tid),
+                                        std::make_tuple(CEPH_MDS_OP_RENAME,
+                                                        nullptr, tid));
 
   mds->send_message_mds(req, to);
 }