Merge pull request #7765 from athanatos/wip-lost-unfound-repop

Repop and lost-unfound overhaul

Reviewed-by: Sage Weil <sage@redhat.com>

src/common/dout.h

@@ -39,6 +39,14 @@ inline std::ostream& operator<<(std::ostream& out, _bad_endl_use_dendl_t) {
   return out;
 }
 
+class DoutPrefixProvider {
+public:
+  virtual string gen_prefix() const = 0;
+  virtual CephContext *get_cct() const = 0;
+  virtual unsigned get_subsys() const = 0;
+  virtual ~DoutPrefixProvider() {}
+};
+
 // generic macros
 #define dout_prefix *_dout
 
@@ -58,6 +66,8 @@ inline std::ostream& operator<<(std::ostream& out, _bad_endl_use_dendl_t) {
 #define ldout(cct, v)  dout_impl(cct, dout_subsys, v) dout_prefix
 #define lderr(cct) dout_impl(cct, ceph_subsys_, -1) dout_prefix
 
+#define ldpp_dout(dpp, v) if (dpp) dout_impl(dpp->get_cct(), dpp->get_subsys(), v) (*_dout << dpp->gen_prefix())
+
 #define lgeneric_subdout(cct, sub, v) dout_impl(cct, ceph_subsys_##sub, v) *_dout
 #define lgeneric_dout(cct, v) dout_impl(cct, ceph_subsys_, v) *_dout
 #define lgeneric_derr(cct) dout_impl(cct, ceph_subsys_, -1) *_dout

src/include/ceph_features.h

@@ -73,6 +73,7 @@
 #define CEPH_FEATURE_MON_STATEFUL_SUB (1ULL<<57) /* stateful mon subscription */
 #define CEPH_FEATURE_MON_ROUTE_OSDMAP (1ULL<<57) /* peon sends osdmaps */
 #define CEPH_FEATURE_OSDSUBOP_NO_SNAPCONTEXT (1ULL<<57) /* overlap, drop unused SnapContext in v12 */
+#define CEPH_FEATURE_SERVER_JEWEL (1ULL<<57)   /* overlap, features introduced in jewel */
 #define CEPH_FEATURE_CRUSH_TUNABLES5	(1ULL<<58) /* chooseleaf stable mode */
 // duplicated since it was introduced at the same time as CEPH_FEATURE_CRUSH_TUNABLES5
 #define CEPH_FEATURE_NEW_OSDOPREPLY_ENCODING   (1ULL<<58) /* New, v7 encoding */
@@ -172,6 +173,7 @@ static inline unsigned long long ceph_sanitize_features(unsigned long long f) {
 	 CEPH_FEATURE_MON_STATEFUL_SUB |	 \
 	 CEPH_FEATURE_MON_ROUTE_OSDMAP |	 \
 	 CEPH_FEATURE_CRUSH_TUNABLES5 |	    \
+	 CEPH_FEATURE_SERVER_JEWEL |  \
 	 0ULL)
 
 #define CEPH_FEATURES_SUPPORTED_DEFAULT  CEPH_FEATURES_ALL

src/include/rados.h

@@ -144,6 +144,7 @@ extern const char *ceph_osd_state_name(int s);
 #define CEPH_OSDMAP_NOTIERAGENT (1<<13) /* disable tiering agent */
 #define CEPH_OSDMAP_NOREBALANCE (1<<14) /* block osd backfill unless pg is degraded */
 #define CEPH_OSDMAP_SORTBITWISE (1<<15) /* use bitwise hobject_t sort */
+#define CEPH_OSDMAP_REQUIRE_JEWEL (1<<16) /* require jewel for booting osds */
 
 /*
  * The error code to return when an OSD can't handle a write

src/messages/MOSDPGUpdateLogMissing.h

@@ -0,0 +1,82 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+/*
+ * Ceph - scalable distributed file system
+ *
+ * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
+ *
+ * This is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License version 2.1, as published by the Free Software
+ * Foundation.  See file COPYING.
+ *
+ */
+
+
+#ifndef CEPH_MOSDPGUPDATELOGMISSING_H
+#define CEPH_MOSDPGUPDATELOGMISSING_H
+
+#include "msg/Message.h"
+
+class MOSDPGUpdateLogMissing : public Message {
+
+  static const int HEAD_VERSION = 1;
+  static const int COMPAT_VERSION = 1;
+
+
+public:
+  epoch_t map_epoch;
+  spg_t pgid;
+  shard_id_t from;
+  ceph_tid_t rep_tid;
+  list<pg_log_entry_t> entries;
+
+  epoch_t get_epoch() const { return map_epoch; }
+  spg_t get_pgid() const { return pgid; }
+  epoch_t get_query_epoch() const { return map_epoch; }
+  ceph_tid_t get_tid() const { return rep_tid; }
+
+  MOSDPGUpdateLogMissing() :
+    Message(MSG_OSD_PG_UPDATE_LOG_MISSING, HEAD_VERSION, COMPAT_VERSION) { }
+  MOSDPGUpdateLogMissing(
+    const list<pg_log_entry_t> &entries,
+    spg_t pgid,
+    shard_id_t from,
+    epoch_t epoch,
+    ceph_tid_t rep_tid)
+    : Message(MSG_OSD_PG_UPDATE_LOG_MISSING, HEAD_VERSION, COMPAT_VERSION),
+      map_epoch(epoch),
+      pgid(pgid),
+      from(from),
+      rep_tid(rep_tid),
+      entries(entries) {}
+
+private:
+  ~MOSDPGUpdateLogMissing() {}
+
+public:
+  const char *get_type_name() const { return "PGUpdateLogMissing"; }
+  void print(ostream& out) const {
+    out << "pg_update_log_missing(" << pgid << " epoch " << map_epoch
+	<< " rep_tid " << rep_tid
+	<< " entries " << entries << ")";
+  }
+
+  void encode_payload(uint64_t features) {
+    ::encode(map_epoch, payload);
+    ::encode(pgid, payload);
+    ::encode(from, payload);
+    ::encode(rep_tid, payload);
+    ::encode(entries, payload);
+  }
+  void decode_payload() {
+    bufferlist::iterator p = payload.begin();
+    ::decode(map_epoch, p);
+    ::decode(pgid, p);
+    ::decode(from, p);
+    ::decode(rep_tid, p);
+    ::decode(entries, p);
+  }
+};
+
+#endif

src/messages/MOSDPGUpdateLogMissingReply.h

@@ -0,0 +1,87 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+/*
+ * Ceph - scalable distributed file system
+ *
+ * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
+ *
+ * This is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License version 2.1, as published by the Free Software
+ * Foundation.  See file COPYING.
+ *
+ */
+
+
+#ifndef CEPH_MOSDPGUPDATELOGMISSINGREPLY_H
+#define CEPH_MOSDPGUPDATELOGMISSINGREPLY_H
+
+#include "msg/Message.h"
+
+class MOSDPGUpdateLogMissingReply : public Message {
+
+  static const int HEAD_VERSION = 1;
+  static const int COMPAT_VERSION = 1;
+
+
+public:
+  epoch_t map_epoch;
+  spg_t pgid;
+  shard_id_t from;
+  ceph_tid_t rep_tid;
+
+  epoch_t get_epoch() const { return map_epoch; }
+  spg_t get_pgid() const { return pgid; }
+  epoch_t get_query_epoch() const { return map_epoch; }
+  ceph_tid_t get_tid() const { return rep_tid; }
+  pg_shard_t get_from() const {
+    return pg_shard_t(get_source().num(), from);
+  }
+
+  MOSDPGUpdateLogMissingReply() :
+    Message(
+      MSG_OSD_PG_UPDATE_LOG_MISSING_REPLY,
+      HEAD_VERSION,
+      COMPAT_VERSION)
+      {}
+  MOSDPGUpdateLogMissingReply(
+    spg_t pgid,
+    shard_id_t from,
+    epoch_t epoch,
+    ceph_tid_t rep_tid)
+    : Message(
+        MSG_OSD_PG_UPDATE_LOG_MISSING_REPLY,
+        HEAD_VERSION,
+        COMPAT_VERSION),
+      map_epoch(epoch),
+      pgid(pgid),
+      from(from),
+      rep_tid(rep_tid)
+    {}
+
+private:
+  ~MOSDPGUpdateLogMissingReply() {}
+
+public:
+  const char *get_type_name() const { return "PGUpdateLogMissingReply"; }
+  void print(ostream& out) const {
+    out << "pg_update_log_missing_reply(" << pgid << " epoch " << map_epoch
+	<< " rep_tid " << rep_tid << ")";
+  }
+
+  void encode_payload(uint64_t features) {
+    ::encode(map_epoch, payload);
+    ::encode(pgid, payload);
+    ::encode(from, payload);
+    ::encode(rep_tid, payload);
+  }
+  void decode_payload() {
+    bufferlist::iterator p = payload.begin();
+    ::decode(map_epoch, p);
+    ::decode(pgid, p);
+    ::decode(from, p);
+    ::decode(rep_tid, p);
+  }
+};
+
+#endif

src/messages/Makefile.am

@@ -91,6 +91,8 @@ noinst_HEADERS += \
 	messages/MOSDPGQuery.h \
 	messages/MOSDPGRemove.h \
 	messages/MOSDPGScan.h \
+	messages/MOSDPGUpdateLogMissing.h \
+	messages/MOSDPGUpdateLogMissingReply.h \
 	messages/MOSDECSubOpWrite.h \
 	messages/MOSDECSubOpWriteReply.h \
 	messages/MOSDECSubOpRead.h \

src/mon/OSDMonitor.cc

@@ -125,6 +125,9 @@ void OSDMonitor::create_initial()
   // new clusters should sort bitwise by default.
   newmap.set_flag(CEPH_OSDMAP_SORTBITWISE);
 
+  // new cluster should require jewel by default
+  newmap.set_flag(CEPH_OSDMAP_REQUIRE_JEWEL);
+
   // encode into pending incremental
   newmap.encode(pending_inc.fullmap, mon->quorum_features | CEPH_FEATURE_RESERVED);
   pending_inc.full_crc = newmap.get_crc();
@@ -1871,6 +1874,16 @@ bool OSDMonitor::preprocess_boot(MonOpRequestRef op)
     goto ignore;
   }
 
+  if (osdmap.test_flag(CEPH_OSDMAP_REQUIRE_JEWEL) &&
+      !(m->get_connection()->get_features() & CEPH_FEATURE_SERVER_JEWEL)) {
+    mon->clog->info() << "disallowing boot of OSD "
+		      << m->get_orig_source_inst()
+		      << " because the osdmap requires"
+		      << " CEPH_FEATURE_SERVER_JEWEL"
+		      << " but the osd lacks CEPH_FEATURE_SERVER_JEWEL\n";
+    goto ignore;
+  }
+
   if (osdmap.test_flag(CEPH_OSDMAP_SORTBITWISE) &&
       !(m->osd_features & CEPH_FEATURE_OSD_BITWISE_HOBJ_SORT)) {
     mon->clog->info() << "disallowing boot of OSD "
@@ -6152,6 +6165,13 @@ bool OSDMonitor::prepare_command_impl(MonOpRequestRef op,
 	ss << "not all up OSDs have OSD_BITWISE_HOBJ_SORT feature";
 	err = -EPERM;
       }
+    } else if (key == "require_jewel_osds") {
+      if (osdmap.get_up_osd_features() & CEPH_FEATURE_SERVER_JEWEL) {
+	return prepare_set_flag(op, CEPH_OSDMAP_REQUIRE_JEWEL);
+      } else {
+	ss << "not all up OSDs have CEPH_FEATURE_SERVER_JEWEL feature";
+	err = -EPERM;
+      }
     } else {
       ss << "unrecognized flag '" << key << "'";
       err = -EINVAL;

src/msg/Message.cc

@@ -170,6 +170,9 @@ using namespace std;
 #include "messages/MOSDECSubOpRead.h"
 #include "messages/MOSDECSubOpReadReply.h"
 
+#include "messages/MOSDPGUpdateLogMissing.h"
+#include "messages/MOSDPGUpdateLogMissingReply.h"
+
 #define DEBUGLVL  10    // debug level of output
 
 #define dout_subsys ceph_subsys_ms
@@ -437,6 +440,12 @@ Message *decode_message(CephContext *cct, int crcflags,
   case MSG_OSD_REPOPREPLY:
     m = new MOSDRepOpReply();
     break;
+  case MSG_OSD_PG_UPDATE_LOG_MISSING:
+    m = new MOSDPGUpdateLogMissing();
+    break;
+  case MSG_OSD_PG_UPDATE_LOG_MISSING_REPLY:
+    m = new MOSDPGUpdateLogMissingReply();
+    break;
 
   case CEPH_MSG_OSD_MAP:
     m = new MOSDMap;

src/msg/Message.h

@@ -116,6 +116,8 @@
 
 #define MSG_OSD_REPOP         112
 #define MSG_OSD_REPOPREPLY    113
+#define MSG_OSD_PG_UPDATE_LOG_MISSING  114
+#define MSG_OSD_PG_UPDATE_LOG_MISSING_REPLY  115
 
 
 // *** MDS ***

src/os/ObjectStore.h

@@ -408,6 +408,8 @@ class ObjectStore {
 
       OP_SETALLOCHINT = 39,  // cid, oid, object_size, write_size
       OP_COLL_HINT = 40, // cid, type, bl
+
+      OP_TRY_RENAME = 41,   // oldcid, oldoid, newoid
     };
 
     // Transaction hint type
@@ -1415,6 +1417,23 @@ class ObjectStore {
       }
       data.ops++;
     }
+    void try_rename(coll_t cid, const ghobject_t& oldoid,
+                    const ghobject_t& oid) {
+      if (use_tbl) {
+        __u32 op = OP_TRY_RENAME;
+        ::encode(op, tbl);
+        ::encode(cid, tbl);
+        ::encode(oldoid, tbl);
+        ::encode(oid, tbl);
+      } else {
+        Op* _op = _get_next_op();
+        _op->op = OP_TRY_RENAME;
+        _op->cid = _get_coll_id(cid);
+        _op->oid = _get_object_id(oldoid);
+        _op->dest_oid = _get_object_id(oid);
+      }
+      data.ops++;
+    }
 
     // NOTE: Collection attr operations are all DEPRECATED.  new
     // backends need not implement these at all.

src/os/Transaction.cc

@@ -323,6 +323,20 @@ void ObjectStore::Transaction::_build_actions_from_tbl()
       }
       break;
 
+    case Transaction::OP_TRY_RENAME:
+      {
+	coll_t cid;
+	ghobject_t oldoid;
+	ghobject_t newoid;
+
+	::decode(cid, p);
+	::decode(oldoid, p);
+	::decode(newoid, p);
+
+	try_rename(cid, oldoid, newoid);
+      }
+      break;
+
     case Transaction::OP_COLL_SETATTR:
       {
 	coll_t cid;

src/os/bluestore/BlueStore.cc

@@ -4492,6 +4492,17 @@ void BlueStore::_txc_add_transaction(TransContext *txc, Transaction *t)
       }
       break;
 
+    case Transaction::OP_TRY_RENAME:
+      {
+	const ghobject_t& noid = i.get_oid(op->dest_oid);
+	OnodeRef no = c->get_onode(noid, true);
+	r = _rename(txc, c, o, no, noid);
+	if (r == -ENOENT)
+	  r = 0;
+	o.reset();
+      }
+      break;
+
     case Transaction::OP_OMAP_CLEAR:
       {
 	r = _omap_clear(txc, c, o);