Merge PR ceph#65392 into main

* refs/pull/65392/head:

Reviewed-by: Venky Shankar <[email protected]>
Reviewed-by: Dhairya Parmar <[email protected]>

Author: vshankar

src/global/signal_handler.cc

@@ -23,6 +23,7 @@
 #include "common/Clock.h" // for ceph_clock_now()
 #include "common/BackTrace.h"
 #include "common/debug.h"
+#include "common/Formatter.h"
 #include "common/safe_io.h"
 #include "common/version.h"
 

src/mds/MDBalancer.cc

@@ -18,7 +18,6 @@
 #include "include/compat.h"
 #include "mdstypes.h"
 
-#include "mon/MonClient.h"
 #include "osdc/Objecter.h"
 #include "MDSRank.h"
 #include "MDSMap.h"

src/mds/MDCache.cc

@@ -13,6 +13,7 @@
  */
 
 #include "MDCache.h"
+#include "Mutation.h"
 #include "RetryMessage.h"
 #include "RetryRequest.h"
 

src/mds/MDCache.h

@@ -37,9 +37,10 @@
 #include "StrayManager.h"
 #include "OpenFileTable.h"
 #include "MDSContext.h"
-#include "Mutation.h"
 #include "LogSegmentRef.h"
 
+#include <boost/intrusive_ptr.hpp>
+
 class EMetaBlob;
 class MCacheExpire;
 class MClientRequest;
@@ -68,6 +69,11 @@ class Session;
 
 class ESubtreeMap;
 
+struct MutationImpl;
+struct MDRequestImpl;
+typedef boost::intrusive_ptr<MutationImpl> MutationRef;
+typedef boost::intrusive_ptr<MDRequestImpl> MDRequestRef;
+
 enum {
   l_mdc_first = 3000,
 

src/mds/MDSRank.cc

@@ -16,6 +16,7 @@
 #include "osdc/Journaler.h"
 
 #include <typeinfo>
+#include "common/DecayCounter.h"
 #include "common/debug.h"
 #include "common/errno.h"
 #include "common/fair_mutex.h"
@@ -34,6 +35,7 @@
 
 #include "mgr/MgrClient.h"
 
+#include "Beacon.h"
 #include "MDCache.h"
 #include "MDLog.h"
 #include "MDSDaemon.h"
@@ -898,6 +900,15 @@ MDSTableServer *MDSRank::get_table_server(int t)
   }
 }
 
+MDSMap::DaemonState MDSRank::get_want_state() const
+{
+  return beacon.get_want_state();
+}
+
+uint64_t MDSRank::get_global_id() const {
+  return monc->get_global_id();
+}
+
 void MDSRank::suicide()
 {
   if (suicide_hook) {
@@ -943,6 +954,11 @@ void MDSRank::damaged_unlocked()
   damaged();
 }
 
+double MDSRank::last_cleared_laggy() const
+{
+  return beacon.last_cleared_laggy();
+}
+
 void MDSRank::handle_write_error(int err)
 {
   if (err == -EBLOCKLISTED) {

src/mds/MDSRank.h

@@ -19,26 +19,24 @@
 #include <string_view>
 
 #include "common/admin_socket.h" // for asok_finisher
-#include "common/DecayCounter.h"
 #include "common/LogClient.h"
 #include "common/TrackedOp.h" // for class OpTracker
 
 #include "include/common_fwd.h"
 
-#include "Beacon.h"
 #include "DamageTable.h"
 #include "MDSMap.h"
 #include "SessionMap.h"
 #include "PurgeQueue.h"
 #include "MetricsHandler.h"
-#include "mon/MonClient.h"
 
 // Full .h import instead of forward declaration for PerfCounter, for the
 // benefit of those including this header and using MDSRank::logger
 #include "common/perf_counters.h"
 
 #include <boost/intrusive_ptr.hpp>
 
+class DecayCounter;
 class MDSContext;
 class MDSMetaRequest;
 class MMDSMap;
@@ -134,6 +132,7 @@ namespace ceph {
 template <class Mutex>
 class CommonSafeTimer;
 
+class Beacon;
 class Locker;
 class MDCache;
 class MDLog;
@@ -209,7 +208,7 @@ class MDSRank {
     Session *get_session(const cref_t<Message> &m);
 
     MDSMap::DaemonState get_state() const { return state; } 
-    MDSMap::DaemonState get_want_state() const { return beacon.get_want_state(); } 
+    MDSMap::DaemonState get_want_state() const;
 
     bool is_creating() const { return state == MDSMap::STATE_CREATING; }
     bool is_starting() const { return state == MDSMap::STATE_STARTING; }
@@ -259,9 +258,7 @@ class MDSRank {
       progress_thread.signal();
     }
 
-    uint64_t get_global_id() const {
-      return monc->get_global_id();
-    }
+    uint64_t get_global_id() const;
 
     // Daemon lifetime functions: these guys break the abstraction
     // and call up into the parent MDSDaemon instance.  It's kind
@@ -308,9 +305,7 @@ class MDSRank {
      */
     void damaged_unlocked();
 
-    double last_cleared_laggy() const {
-      return beacon.last_cleared_laggy();
-    }
+    double last_cleared_laggy() const;
 
     double get_dispatch_queue_max_age(utime_t now) const;
 

src/mds/MDSRankQuiesce.cc

@@ -12,6 +12,7 @@
 
 #include "MDSRank.h"
 #include "MDCache.h"
+#include "mon/MonClient.h"
 
 #include "QuiesceDbManager.h"
 #include "QuiesceAgent.h"

src/mds/MDSTableClient.cc

@@ -16,6 +16,7 @@
 #include "MDSMap.h"
 
 #include "MDSContext.h"
+#include "mds_table_types.h"
 #include "RetryMessage.h"
 #include "msg/Messenger.h"
 

src/mds/MDSTableClient.h

@@ -17,8 +17,7 @@
 
 #include "include/buffer.h"
 #include "include/types.h"
-#include "mds_table_types.h"
-#include "mdstypes.h" // for mds_rank_t
+#include "include/cephfs/types.h" // for mds_rank_t
 #include "common/ref.h" // for cref_t
 #include "LogSegmentRef.h"
 

src/mds/MDSTableServer.cc

@@ -21,13 +21,65 @@
 #include "events/ETableServer.h"
 #include "common/debug.h"
 
+#include "messages/MMDSTableRequest.h"
+
 #define dout_context g_ceph_context
 #define dout_subsys ceph_subsys_mds
 #undef dout_prefix
 #define dout_prefix *_dout << "mds." << rank << ".tableserver(" << get_mdstable_name(table) << ") "
 
 using namespace std;
 
+struct MDSTableServer::notify_info_t {
+  notify_info_t() {}
+  std::set<mds_rank_t> notify_ack_gather;
+  mds_rank_t mds;
+  ref_t<MMDSTableRequest> reply = NULL;
+  MDSContext *onfinish = nullptr;
+};
+
+MDSTableServer::MDSTableServer(MDSRank *m, int tab) :
+    MDSTable(m, get_mdstable_name(tab), false), table(tab) {}
+
+MDSTableServer::~MDSTableServer() = default;
+
+MDSTableServer::MDSTableServer(const MDSTableServer &) = default;
+MDSTableServer &MDSTableServer::operator=(const MDSTableServer &) = default;
+
+void MDSTableServer::_note_prepare(mds_rank_t mds, uint64_t reqid, bool replay) {
+  version++;
+  if (replay)
+    projected_version = version;
+  pending_for_mds[version].mds = mds;
+  pending_for_mds[version].reqid = reqid;
+  pending_for_mds[version].tid = version;
+}
+
+void MDSTableServer::_note_commit(uint64_t tid, bool replay) {
+  version++;
+  if (replay)
+    projected_version = version;
+  pending_for_mds.erase(tid);
+}
+
+void MDSTableServer::_note_rollback(uint64_t tid, bool replay) {
+  version++;
+  if (replay)
+    projected_version = version;
+  pending_for_mds.erase(tid);
+}
+
+void MDSTableServer::_note_server_update(bufferlist& bl, bool replay) {
+  version++;
+  if (replay)
+    projected_version = version;
+}
+
+void MDSTableServer::reset_state() {
+  pending_for_mds.clear();
+  ++version;
+}
+
 void MDSTableServer::handle_request(const cref_t<MMDSTableRequest> &req)
 {
   ceph_assert(req->op >= 0);
@@ -297,6 +349,16 @@ void MDSTableServer::_do_server_recovery()
   recovered = true;
 }
 
+void MDSTableServer::encode_state(bufferlist& bl) const {
+  encode_server_state(bl);
+  encode(pending_for_mds, bl);
+}
+
+void MDSTableServer::decode_state(bufferlist::const_iterator& bl) {
+  decode_server_state(bl);
+  decode(pending_for_mds, bl);
+}
+
 void MDSTableServer::finish_recovery(set<mds_rank_t>& active)
 {
   dout(7) << __func__ << dendl;