ceph/src/mds/Server.h

   1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
   2 // vim: ts=8 sw=2 smarttab
   3 /*
   4  * Ceph - scalable distributed file system
   5  *
   6  * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
   7  *
   8  * This is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License version 2.1, as published by the Free Software
  11  * Foundation.  See file COPYING.
  12  *
  13  */
  14
  15 #ifndef CEPH_MDS_SERVER_H
  16 #define CEPH_MDS_SERVER_H
  17
  18 #include <string_view>
  19
  20 #include <common/DecayCounter.h>
  21
  22 #include "include/common_fwd.h"
  23
  24 #include "messages/MClientReconnect.h"
  25 #include "messages/MClientReply.h"
  26 #include "messages/MClientRequest.h"
  27 #include "messages/MClientSession.h"
  28 #include "messages/MClientSnap.h"
  29 #include "messages/MClientReclaim.h"
  30 #include "messages/MClientReclaimReply.h"
  31 #include "messages/MLock.h"
  32
  33 #include "MDSRank.h"
  34 #include "Mutation.h"
  35 #include "MDSContext.h"
  36
  37 class OSDMap;
  38 class LogEvent;
  39 class EMetaBlob;
  40 class EUpdate;
  41 class MDLog;
  42 struct SnapInfo;
  43
  44 enum {
  45   l_mdss_first = 1000,
  46   l_mdss_dispatch_client_request,
  47   l_mdss_dispatch_slave_request,
  48   l_mdss_handle_client_request,
  49   l_mdss_handle_client_session,
  50   l_mdss_handle_slave_request,
  51   l_mdss_req_create_latency,
  52   l_mdss_req_getattr_latency,
  53   l_mdss_req_getfilelock_latency,
  54   l_mdss_req_link_latency,
  55   l_mdss_req_lookup_latency,
  56   l_mdss_req_lookuphash_latency,
  57   l_mdss_req_lookupino_latency,
  58   l_mdss_req_lookupname_latency,
  59   l_mdss_req_lookupparent_latency,
  60   l_mdss_req_lookupsnap_latency,
  61   l_mdss_req_lssnap_latency,
  62   l_mdss_req_mkdir_latency,
  63   l_mdss_req_mknod_latency,
  64   l_mdss_req_mksnap_latency,
  65   l_mdss_req_open_latency,
  66   l_mdss_req_readdir_latency,
  67   l_mdss_req_rename_latency,
  68   l_mdss_req_renamesnap_latency,
  69   l_mdss_req_rmdir_latency,
  70   l_mdss_req_rmsnap_latency,
  71   l_mdss_req_rmxattr_latency,
  72   l_mdss_req_setattr_latency,
  73   l_mdss_req_setdirlayout_latency,
  74   l_mdss_req_setfilelock_latency,
  75   l_mdss_req_setlayout_latency,
  76   l_mdss_req_setxattr_latency,
  77   l_mdss_req_symlink_latency,
  78   l_mdss_req_unlink_latency,
  79   l_mdss_cap_revoke_eviction,
  80   l_mdss_cap_acquisition_throttle,
  81   l_mdss_last,
  82 };
  83
  84 class Server {
  85 public:
  86   using clock = ceph::coarse_mono_clock;
  87   using time = ceph::coarse_mono_time;
  88
  89   enum class RecallFlags : uint64_t {
  90     NONE = 0,
  91     STEADY = (1<<0),
  92     ENFORCE_MAX = (1<<1),
  93     TRIM = (1<<2),
  94     ENFORCE_LIVENESS = (1<<3),
  95   };
  96   explicit Server(MDSRank *m);
  97   ~Server() {
  98     g_ceph_context->get_perfcounters_collection()->remove(logger);
  99     delete logger;
 100     delete reconnect_done;
 101   }
 102
 103   void create_logger();
 104
 105   // message handler
 106   void dispatch(const cref_t<Message> &m);
 107
 108   void handle_osd_map();
 109
 110   // -- sessions and recovery --
 111   bool waiting_for_reconnect(client_t c) const;
 112   void dump_reconnect_status(Formatter *f) const;
 113
 114   time last_recalled() const {
 115     return last_recall_state;
 116   }
 117
 118   void handle_client_session(const cref_t<MClientSession> &m);
 119   void _session_logged(Session *session, uint64_t state_seq,
 120                        bool open, version_t pv, const interval_set<inodeno_t>& inos,version_t piv,
 121                        const interval_set<inodeno_t>& purge_inos, LogSegment *ls);
 122   version_t prepare_force_open_sessions(map<client_t,entity_inst_t> &cm,
 123                                         map<client_t,client_metadata_t>& cmm,
 124                                         map<client_t,pair<Session*,uint64_t> >& smap);
 125   void finish_force_open_sessions(const map<client_t,pair<Session*,uint64_t> >& smap,
 126                                   bool dec_import=true);
 127   void flush_client_sessions(set<client_t>& client_set, MDSGatherBuilder& gather);
 128   void finish_flush_session(Session *session, version_t seq);
 129   void terminate_sessions();
 130   void find_idle_sessions();
 131   void kill_session(Session *session, Context *on_safe, bool need_purge_inos = false);
 132   size_t apply_blacklist(const std::set<entity_addr_t> &blacklist);
 133   void journal_close_session(Session *session, int state, Context *on_safe, bool need_purge_inos = false);
 134
 135   size_t get_num_pending_reclaim() const { return client_reclaim_gather.size(); }
 136   Session *find_session_by_uuid(std::string_view uuid);
 137   void reclaim_session(Session *session, const cref_t<MClientReclaim> &m);
 138   void finish_reclaim_session(Session *session, const ref_t<MClientReclaimReply> &reply=nullptr);
 139   void handle_client_reclaim(const cref_t<MClientReclaim> &m);
 140
 141   void reconnect_clients(MDSContext *reconnect_done_);
 142   void handle_client_reconnect(const cref_t<MClientReconnect> &m);
 143   void infer_supported_features(Session *session, client_metadata_t& client_metadata);
 144   void update_required_client_features();
 145
 146   //void process_reconnect_cap(CInode *in, int from, ceph_mds_cap_reconnect& capinfo);
 147   void reconnect_gather_finish();
 148   void reconnect_tick();
 149   void recover_filelocks(CInode *in, bufferlist locks, int64_t client);
 150
 151   std::pair<bool, uint64_t> recall_client_state(MDSGatherBuilder* gather, RecallFlags=RecallFlags::NONE);
 152   void force_clients_readonly();
 153
 154   // -- requests --
 155   void handle_client_request(const cref_t<MClientRequest> &m);
 156
 157   void journal_and_reply(MDRequestRef& mdr, CInode *tracei, CDentry *tracedn,
 158                          LogEvent *le, MDSLogContextBase *fin);
 159   void submit_mdlog_entry(LogEvent *le, MDSLogContextBase *fin,
 160                           MDRequestRef& mdr, std::string_view event);
 161   void dispatch_client_request(MDRequestRef& mdr);
 162   void perf_gather_op_latency(const cref_t<MClientRequest> &req, utime_t lat);
 163   void early_reply(MDRequestRef& mdr, CInode *tracei, CDentry *tracedn);
 164   void respond_to_request(MDRequestRef& mdr, int r = 0);
 165   void set_trace_dist(const ref_t<MClientReply> &reply, CInode *in, CDentry *dn,
 166                       MDRequestRef& mdr);
 167
 168   void handle_slave_request(const cref_t<MMDSSlaveRequest> &m);
 169   void handle_slave_request_reply(const cref_t<MMDSSlaveRequest> &m);
 170   void dispatch_slave_request(MDRequestRef& mdr);
 171   void handle_slave_auth_pin(MDRequestRef& mdr);
 172   void handle_slave_auth_pin_ack(MDRequestRef& mdr, const cref_t<MMDSSlaveRequest> &ack);
 173
 174   // some helpers
 175   bool check_fragment_space(MDRequestRef& mdr, CDir *in);
 176   bool check_access(MDRequestRef& mdr, CInode *in, unsigned mask);
 177   bool _check_access(Session *session, CInode *in, unsigned mask, int caller_uid, int caller_gid, int setattr_uid, int setattr_gid);
 178   CDentry *prepare_stray_dentry(MDRequestRef& mdr, CInode *in);
 179   CInode* prepare_new_inode(MDRequestRef& mdr, CDir *dir, inodeno_t useino, unsigned mode,
 180                             file_layout_t *layout=NULL);
 181   void journal_allocated_inos(MDRequestRef& mdr, EMetaBlob *blob);
 182   void apply_allocated_inos(MDRequestRef& mdr, Session *session);
 183
 184   CInode* rdlock_path_pin_ref(MDRequestRef& mdr, bool want_auth,
 185                               bool no_want_auth=false);
 186   CDentry* rdlock_path_xlock_dentry(MDRequestRef& mdr, bool create,
 187                                     bool okexist=false, bool want_layout=false);
 188   std::pair<CDentry*, CDentry*>
 189             rdlock_two_paths_xlock_destdn(MDRequestRef& mdr, bool xlock_srcdn);
 190
 191   CDir* try_open_auth_dirfrag(CInode *diri, frag_t fg, MDRequestRef& mdr);
 192
 193   // requests on existing inodes.
 194   void handle_client_getattr(MDRequestRef& mdr, bool is_lookup);
 195   void handle_client_lookup_ino(MDRequestRef& mdr,
 196                                 bool want_parent, bool want_dentry);
 197   void _lookup_snap_ino(MDRequestRef& mdr);
 198   void _lookup_ino_2(MDRequestRef& mdr, int r);
 199   void handle_client_readdir(MDRequestRef& mdr);
 200   void handle_client_file_setlock(MDRequestRef& mdr);
 201   void handle_client_file_readlock(MDRequestRef& mdr);
 202
 203   bool xlock_policylock(MDRequestRef& mdr, CInode *in,
 204                         bool want_layout=false, bool xlock_snaplock=false);
 205   CInode* try_get_auth_inode(MDRequestRef& mdr, inodeno_t ino);
 206   void handle_client_setattr(MDRequestRef& mdr);
 207   void handle_client_setlayout(MDRequestRef& mdr);
 208   void handle_client_setdirlayout(MDRequestRef& mdr);
 209
 210   int parse_quota_vxattr(string name, string value, quota_info_t *quota);
 211   void create_quota_realm(CInode *in);
 212   int parse_layout_vxattr(string name, string value, const OSDMap& osdmap,
 213                           file_layout_t *layout, bool validate=true);
 214   int check_layout_vxattr(MDRequestRef& mdr,
 215                           string name,
 216                           string value,
 217                           file_layout_t *layout);
 218   void handle_set_vxattr(MDRequestRef& mdr, CInode *cur);
 219   void handle_remove_vxattr(MDRequestRef& mdr, CInode *cur);
 220   void handle_client_setxattr(MDRequestRef& mdr);
 221   void handle_client_removexattr(MDRequestRef& mdr);
 222
 223   void handle_client_fsync(MDRequestRef& mdr);
 224
 225   // open
 226   void handle_client_open(MDRequestRef& mdr);
 227   void handle_client_openc(MDRequestRef& mdr);  // O_CREAT variant.
 228   void do_open_truncate(MDRequestRef& mdr, int cmode);  // O_TRUNC variant.
 229
 230   // namespace changes
 231   void handle_client_mknod(MDRequestRef& mdr);
 232   void handle_client_mkdir(MDRequestRef& mdr);
 233   void handle_client_symlink(MDRequestRef& mdr);
 234
 235   // link
 236   void handle_client_link(MDRequestRef& mdr);
 237   void _link_local(MDRequestRef& mdr, CDentry *dn, CInode *targeti, SnapRealm *target_realm);
 238   void _link_local_finish(MDRequestRef& mdr, CDentry *dn, CInode *targeti,
 239                           version_t, version_t, bool);
 240
 241   void _link_remote(MDRequestRef& mdr, bool inc, CDentry *dn, CInode *targeti);
 242   void _link_remote_finish(MDRequestRef& mdr, bool inc, CDentry *dn, CInode *targeti,
 243                            version_t);
 244
 245   void handle_slave_link_prep(MDRequestRef& mdr);
 246   void _logged_slave_link(MDRequestRef& mdr, CInode *targeti, bool adjust_realm);
 247   void _commit_slave_link(MDRequestRef& mdr, int r, CInode *targeti);
 248   void _committed_slave(MDRequestRef& mdr);  // use for rename, too
 249   void handle_slave_link_prep_ack(MDRequestRef& mdr, const cref_t<MMDSSlaveRequest> &m);
 250   void do_link_rollback(bufferlist &rbl, mds_rank_t master, MDRequestRef& mdr);
 251   void _link_rollback_finish(MutationRef& mut, MDRequestRef& mdr,
 252                              map<client_t,ref_t<MClientSnap>>& split);
 253
 254   // unlink
 255   void handle_client_unlink(MDRequestRef& mdr);
 256   bool _dir_is_nonempty_unlocked(MDRequestRef& mdr, CInode *rmdiri);
 257   bool _dir_is_nonempty(MDRequestRef& mdr, CInode *rmdiri);
 258   void _unlink_local(MDRequestRef& mdr, CDentry *dn, CDentry *straydn);
 259   void _unlink_local_finish(MDRequestRef& mdr,
 260                             CDentry *dn, CDentry *straydn,
 261                             version_t);
 262   bool _rmdir_prepare_witness(MDRequestRef& mdr, mds_rank_t who, vector<CDentry*>& trace, CDentry *straydn);
 263   void handle_slave_rmdir_prep(MDRequestRef& mdr);
 264   void _logged_slave_rmdir(MDRequestRef& mdr, CDentry *srcdn, CDentry *straydn);
 265   void _commit_slave_rmdir(MDRequestRef& mdr, int r, CDentry *straydn);
 266   void handle_slave_rmdir_prep_ack(MDRequestRef& mdr, const cref_t<MMDSSlaveRequest> &ack);
 267   void do_rmdir_rollback(bufferlist &rbl, mds_rank_t master, MDRequestRef& mdr);
 268   void _rmdir_rollback_finish(MDRequestRef& mdr, metareqid_t reqid, CDentry *dn, CDentry *straydn);
 269
 270   // rename
 271   void handle_client_rename(MDRequestRef& mdr);
 272   void _rename_finish(MDRequestRef& mdr,
 273                       CDentry *srcdn, CDentry *destdn, CDentry *straydn);
 274
 275   void handle_client_lssnap(MDRequestRef& mdr);
 276   void handle_client_mksnap(MDRequestRef& mdr);
 277   void _mksnap_finish(MDRequestRef& mdr, CInode *diri, SnapInfo &info);
 278   void handle_client_rmsnap(MDRequestRef& mdr);
 279   void _rmsnap_finish(MDRequestRef& mdr, CInode *diri, snapid_t snapid);
 280   void handle_client_renamesnap(MDRequestRef& mdr);
 281   void _renamesnap_finish(MDRequestRef& mdr, CInode *diri, snapid_t snapid);
 282
 283   // helpers
 284   bool _rename_prepare_witness(MDRequestRef& mdr, mds_rank_t who, set<mds_rank_t> &witnesse,
 285                                vector<CDentry*>& srctrace, vector<CDentry*>& dsttrace, CDentry *straydn);
 286   version_t _rename_prepare_import(MDRequestRef& mdr, CDentry *srcdn, bufferlist *client_map_bl);
 287   bool _need_force_journal(CInode *diri, bool empty);
 288   void _rename_prepare(MDRequestRef& mdr,
 289                        EMetaBlob *metablob, bufferlist *client_map_bl,
 290                        CDentry *srcdn, CDentry *destdn, CDentry *straydn);
 291   /* set not_journaling=true if you're going to discard the results --
 292    * this bypasses the asserts to make sure we're journaling the right
 293    * things on the right nodes */
 294   void _rename_apply(MDRequestRef& mdr, CDentry *srcdn, CDentry *destdn, CDentry *straydn);
 295
 296   // slaving
 297   void handle_slave_rename_prep(MDRequestRef& mdr);
 298   void handle_slave_rename_prep_ack(MDRequestRef& mdr, const cref_t<MMDSSlaveRequest> &m);
 299   void handle_slave_rename_notify_ack(MDRequestRef& mdr, const cref_t<MMDSSlaveRequest> &m);
 300   void _slave_rename_sessions_flushed(MDRequestRef& mdr);
 301   void _logged_slave_rename(MDRequestRef& mdr, CDentry *srcdn, CDentry *destdn, CDentry *straydn);
 302   void _commit_slave_rename(MDRequestRef& mdr, int r, CDentry *srcdn, CDentry *destdn, CDentry *straydn);
 303   void do_rename_rollback(bufferlist &rbl, mds_rank_t master, MDRequestRef& mdr, bool finish_mdr=false);
 304   void _rename_rollback_finish(MutationRef& mut, MDRequestRef& mdr, CDentry *srcdn, version_t srcdnpv,
 305                                CDentry *destdn, CDentry *staydn, map<client_t,ref_t<MClientSnap>> splits[2],
 306                                bool finish_mdr);
 307
 308   void evict_cap_revoke_non_responders();
 309   void handle_conf_change(const std::set<std::string>& changed);
 310
 311   bool terminating_sessions = false;
 312
 313   set<client_t> client_reclaim_gather;
 314
 315 private:
 316   friend class MDSContinuation;
 317   friend class ServerContext;
 318   friend class ServerLogContext;
 319   friend class Batch_Getattr_Lookup;
 320
 321   void reply_client_request(MDRequestRef& mdr, const ref_t<MClientReply> &reply);
 322   void flush_session(Session *session, MDSGatherBuilder& gather);
 323
 324   MDSRank *mds;
 325   MDCache *mdcache;
 326   MDLog *mdlog;
 327   PerfCounters *logger = nullptr;
 328
 329   // OSDMap full status, used to generate ENOSPC on some operations
 330   bool is_full = false;
 331
 332   // State for while in reconnect
 333   MDSContext *reconnect_done = nullptr;
 334   int failed_reconnects = 0;
 335   bool reconnect_evicting = false;  // true if I am waiting for evictions to complete
 336                             // before proceeding to reconnect_gather_finish
 337   time reconnect_start = clock::zero();
 338   time reconnect_last_seen = clock::zero();
 339   set<client_t> client_reconnect_gather;  // clients i need a reconnect msg from.
 340
 341   feature_bitset_t supported_features;
 342   feature_bitset_t required_client_features;
 343
 344   bool forward_all_requests_to_auth = false;
 345   bool replay_unsafe_with_closed_session = false;
 346   double cap_revoke_eviction_timeout = 0;
 347   uint64_t max_snaps_per_dir = 100;
 348   unsigned delegate_inos_pct = 0;
 349
 350   DecayCounter recall_throttle;
 351   time last_recall_state;
 352
 353   // Cache cap acquisition throttle configs
 354   uint64_t max_caps_per_client;
 355   uint64_t cap_acquisition_throttle;
 356   double max_caps_throttle_ratio;
 357   double caps_throttle_retry_request_timeout;
 358 };
 359
 360 static inline constexpr auto operator|(Server::RecallFlags a, Server::RecallFlags b) {
 361   using T = std::underlying_type<Server::RecallFlags>::type;
 362   return static_cast<Server::RecallFlags>(static_cast<T>(a) | static_cast<T>(b));
 363 }
 364 static inline constexpr auto operator&(Server::RecallFlags a, Server::RecallFlags b) {
 365   using T = std::underlying_type<Server::RecallFlags>::type;
 366   return static_cast<Server::RecallFlags>(static_cast<T>(a) & static_cast<T>(b));
 367 }
 368 static inline std::ostream& operator<<(std::ostream& os, const Server::RecallFlags& f) {
 369   using T = std::underlying_type<Server::RecallFlags>::type;
 370   return os << "0x" << std::hex << static_cast<T>(f) << std::dec;
 371 }
 372 static inline constexpr bool operator!(const Server::RecallFlags& f) {
 373   using T = std::underlying_type<Server::RecallFlags>::type;
 374   return static_cast<T>(f) == static_cast<T>(0);
 375 }
 376 #endif