ceph/src/mds/CDentry.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
  16
  17 #ifndef CEPH_CDENTRY_H
  18 #define CEPH_CDENTRY_H
  19
  20 #include <string>
  21 #include <set>
  22
  23 #include "include/counter.h"
  24 #include "include/types.h"
  25 #include "include/buffer_fwd.h"
  26 #include "include/lru.h"
  27 #include "include/elist.h"
  28 #include "include/filepath.h"
  29
  30 #include "MDSCacheObject.h"
  31 #include "SimpleLock.h"
  32 #include "LocalLock.h"
  33 #include "ScrubHeader.h"
  34
  35 class CInode;
  36 class CDir;
  37 class Locker;
  38 class Message;
  39 class CDentry;
  40 class LogSegment;
  41
  42 class Session;
  43
  44
  45
  46 // define an ordering
  47 bool operator<(const CDentry& l, const CDentry& r);
  48
  49 // dentry
  50 class CDentry : public MDSCacheObject, public LRUObject, public Counter<CDentry> {
  51 public:
  52   MEMPOOL_CLASS_HELPERS();
  53   friend class CDir;
  54
  55   struct linkage_t {
  56     CInode *inode;
  57     inodeno_t remote_ino;
  58     unsigned char remote_d_type;
  59
  60     linkage_t() : inode(0), remote_ino(0), remote_d_type(0) {}
  61
  62     // dentry type is primary || remote || null
  63     // inode ptr is required for primary, optional for remote, undefined for null
  64     bool is_primary() const { return remote_ino == 0 && inode != 0; }
  65     bool is_remote() const { return remote_ino > 0; }
  66     bool is_null() const { return remote_ino == 0 && inode == 0; }
  67
  68     CInode *get_inode() { return inode; }
  69     const CInode *get_inode() const { return inode; }
  70     inodeno_t get_remote_ino() const { return remote_ino; }
  71     unsigned char get_remote_d_type() const { return remote_d_type; }
  72     std::string get_remote_d_type_string() const;
  73
  74     void set_remote(inodeno_t ino, unsigned char d_type) {
  75       remote_ino = ino;
  76       remote_d_type = d_type;
  77       inode = 0;
  78     }
  79     void link_remote(CInode *in);
  80   };
  81
  82
  83   // -- state --
  84   static const int STATE_NEW =          (1<<0);
  85   static const int STATE_FRAGMENTING =  (1<<1);
  86   static const int STATE_PURGING =      (1<<2);
  87   static const int STATE_BADREMOTEINO = (1<<3);
  88   static const int STATE_EVALUATINGSTRAY = (1<<4);
  89   static const int STATE_PURGINGPINNED =  (1<<5);
  90   static const int STATE_BOTTOMLRU =    (1<<6);
  91   // stray dentry needs notification of releasing reference
  92   static const int STATE_STRAY =        STATE_NOTIFYREF;
  93   static const int MASK_STATE_IMPORT_KEPT = STATE_BOTTOMLRU;
  94
  95   // -- pins --
  96   static const int PIN_INODEPIN =     1;  // linked inode is pinned
  97   static const int PIN_FRAGMENTING = -2;  // containing dir is refragmenting
  98   static const int PIN_PURGING =      3;
  99   static const int PIN_SCRUBPARENT =  4;
 100
 101   static const unsigned EXPORT_NONCE = 1;
 102
 103
 104   CDentry(const std::string& n, __u32 h,
 105           snapid_t f, snapid_t l) :
 106     name(n), hash(h),
 107     first(f), last(l),
 108     item_dirty(this),
 109     lock(this, &lock_type),
 110     versionlock(this, &versionlock_type),
 111     dir(0),
 112     version(0), projected_version(0) {
 113   }
 114   CDentry(const std::string& n, __u32 h, inodeno_t ino, unsigned char dt,
 115           snapid_t f, snapid_t l) :
 116     name(n), hash(h),
 117     first(f), last(l),
 118     item_dirty(this),
 119     lock(this, &lock_type),
 120     versionlock(this, &versionlock_type),
 121     dir(0),
 122     version(0), projected_version(0) {
 123     linkage.remote_ino = ino;
 124     linkage.remote_d_type = dt;
 125   }
 126
 127   const char *pin_name(int p) const override {
 128     switch (p) {
 129     case PIN_INODEPIN: return "inodepin";
 130     case PIN_FRAGMENTING: return "fragmenting";
 131     case PIN_PURGING: return "purging";
 132     case PIN_SCRUBPARENT: return "scrubparent";
 133     default: return generic_pin_name(p);
 134     }
 135   }
 136
 137   // -- wait --
 138   //static const int WAIT_LOCK_OFFSET = 8;
 139
 140   void add_waiter(uint64_t tag, MDSInternalContextBase *c) override;
 141
 142   bool is_lt(const MDSCacheObject *r) const override {
 143     return *this < *static_cast<const CDentry*>(r);
 144   }
 145
 146   dentry_key_t key() {
 147     return dentry_key_t(last, name.c_str(), hash);
 148   }
 149
 150   const CDir *get_dir() const { return dir; }
 151   CDir *get_dir() { return dir; }
 152   const std::string& get_name() const { return name; }
 153
 154   __u32 get_hash() const { return hash; }
 155
 156   // linkage
 157   const linkage_t *get_linkage() const { return &linkage; }
 158   linkage_t *get_linkage() { return &linkage; }
 159
 160   linkage_t *_project_linkage() {
 161     projected.push_back(linkage_t());
 162     return &projected.back();
 163   }
 164   void push_projected_linkage();
 165   void push_projected_linkage(inodeno_t ino, char d_type) {
 166     linkage_t *p = _project_linkage();
 167     p->remote_ino = ino;
 168     p->remote_d_type = d_type;
 169   }
 170   void push_projected_linkage(CInode *inode);
 171   linkage_t *pop_projected_linkage();
 172
 173   bool is_projected() const { return !projected.empty(); }
 174
 175   linkage_t *get_projected_linkage() {
 176     if (!projected.empty())
 177       return &projected.back();
 178     return &linkage;
 179   }
 180
 181   const linkage_t *get_projected_linkage() const {
 182     if (!projected.empty())
 183       return &projected.back();
 184     return &linkage;
 185   }
 186
 187   CInode *get_projected_inode() {
 188     return get_projected_linkage()->inode;
 189   }
 190
 191   bool use_projected(client_t client, const MutationRef& mut) const {
 192     return lock.can_read_projected(client) ||
 193       lock.get_xlock_by() == mut;
 194   }
 195   linkage_t *get_linkage(client_t client, const MutationRef& mut) {
 196     return use_projected(client, mut) ? get_projected_linkage() : get_linkage();
 197   }
 198
 199   // ref counts: pin ourselves in the LRU when we're pinned.
 200   void first_get() override {
 201     lru_pin();
 202   }
 203   void last_put() override {
 204     lru_unpin();
 205   }
 206   void _put() override;
 207
 208   // auth pins
 209   bool can_auth_pin() const override;
 210   void auth_pin(void *by) override;
 211   void auth_unpin(void *by) override;
 212   void adjust_nested_auth_pins(int adjustment, int diradj, void *by);
 213   bool is_frozen() const override;
 214   bool is_freezing() const override;
 215   int get_num_dir_auth_pins() const;
 216
 217   // remote links
 218   void link_remote(linkage_t *dnl, CInode *in);
 219   void unlink_remote(linkage_t *dnl);
 220
 221   // copy cons
 222   CDentry(const CDentry& m);
 223   const CDentry& operator= (const CDentry& right);
 224
 225   // misc
 226   void make_path_string(std::string& s, bool projected=false) const;
 227   void make_path(filepath& fp, bool projected=false) const;
 228
 229   // -- version --
 230   version_t get_version() const { return version; }
 231   void set_version(version_t v) { projected_version = version = v; }
 232   version_t get_projected_version() const { return projected_version; }
 233   void set_projected_version(version_t v) { projected_version = v; }
 234
 235   mds_authority_t authority() const override;
 236
 237   version_t pre_dirty(version_t min=0);
 238   void _mark_dirty(LogSegment *ls);
 239   void mark_dirty(version_t projected_dirv, LogSegment *ls);
 240   void mark_clean();
 241
 242   void mark_new();
 243   bool is_new() const { return state_test(STATE_NEW); }
 244   void clear_new() { state_clear(STATE_NEW); }
 245
 246   // -- replication
 247   void encode_replica(mds_rank_t mds, bufferlist& bl) {
 248     if (!is_replicated())
 249       lock.replicate_relax();
 250
 251     __u32 nonce = add_replica(mds);
 252     ::encode(nonce, bl);
 253     ::encode(first, bl);
 254     ::encode(linkage.remote_ino, bl);
 255     ::encode(linkage.remote_d_type, bl);
 256     __s32 ls = lock.get_replica_state();
 257     ::encode(ls, bl);
 258   }
 259   void decode_replica(bufferlist::iterator& p, bool is_new);
 260
 261   // -- exporting
 262   // note: this assumes the dentry already exists.
 263   // i.e., the name is already extracted... so we just need the other state.
 264   void encode_export(bufferlist& bl) {
 265     ::encode(first, bl);
 266     ::encode(state, bl);
 267     ::encode(version, bl);
 268     ::encode(projected_version, bl);
 269     ::encode(lock, bl);
 270     ::encode(get_replicas(), bl);
 271     get(PIN_TEMPEXPORTING);
 272   }
 273   void finish_export() {
 274     // twiddle
 275     clear_replica_map();
 276     replica_nonce = EXPORT_NONCE;
 277     state_clear(CDentry::STATE_AUTH);
 278     if (is_dirty())
 279       mark_clean();
 280     put(PIN_TEMPEXPORTING);
 281   }
 282   void abort_export() {
 283     put(PIN_TEMPEXPORTING);
 284   }
 285   void decode_import(bufferlist::iterator& blp, LogSegment *ls) {
 286     ::decode(first, blp);
 287     __u32 nstate;
 288     ::decode(nstate, blp);
 289     ::decode(version, blp);
 290     ::decode(projected_version, blp);
 291     ::decode(lock, blp);
 292     ::decode(get_replicas(), blp);
 293
 294     // twiddle
 295     state &= MASK_STATE_IMPORT_KEPT;
 296     state_set(CDentry::STATE_AUTH);
 297     if (nstate & STATE_DIRTY)
 298       _mark_dirty(ls);
 299     if (is_replicated())
 300       get(PIN_REPLICATED);
 301     replica_nonce = 0;
 302   }
 303
 304   // -- locking --
 305   SimpleLock* get_lock(int type) override {
 306     assert(type == CEPH_LOCK_DN);
 307     return &lock;
 308   }
 309   void set_object_info(MDSCacheObjectInfo &info) override;
 310   void encode_lock_state(int type, bufferlist& bl) override;
 311   void decode_lock_state(int type, bufferlist& bl) override;
 312
 313   // ---------------------------------------------
 314   // replicas (on clients)
 315
 316   bool is_any_leases() const {
 317     return !client_lease_map.empty();
 318   }
 319   const ClientLease *get_client_lease(client_t c) const {
 320     if (client_lease_map.count(c))
 321       return client_lease_map.find(c)->second;
 322     return 0;
 323   }
 324   ClientLease *get_client_lease(client_t c) {
 325     if (client_lease_map.count(c))
 326       return client_lease_map.find(c)->second;
 327     return 0;
 328   }
 329   bool have_client_lease(client_t c) const {
 330     const ClientLease *l = get_client_lease(c);
 331     if (l)
 332       return true;
 333     else
 334       return false;
 335   }
 336
 337   ClientLease *add_client_lease(client_t c, Session *session);
 338   void remove_client_lease(ClientLease *r, Locker *locker);  // returns remaining mask (if any), and kicks locker eval_gathers
 339   void remove_client_leases(Locker *locker);
 340
 341   ostream& print_db_line_prefix(ostream& out) override;
 342   void print(ostream& out) override;
 343   void dump(Formatter *f) const;
 344
 345
 346   std::string name;
 347   __u32 hash;
 348   snapid_t first, last;
 349
 350   elist<CDentry*>::item item_dirty;
 351   elist<CDentry*>::item item_stray;
 352
 353   // lock
 354   static LockType lock_type;
 355   static LockType versionlock_type;
 356
 357   SimpleLock lock;
 358   LocalLock versionlock;
 359
 360   map<client_t,ClientLease*> client_lease_map;
 361
 362
 363 protected:
 364   friend class Migrator;
 365   friend class Locker;
 366   friend class MDCache;
 367   friend class StrayManager;
 368   friend class CInode;
 369   friend class C_MDC_XlockRequest;
 370
 371   CDir *dir;     // containing dirfrag
 372   linkage_t linkage;
 373   list<linkage_t> projected;
 374
 375   version_t version;  // dir version when last touched.
 376   version_t projected_version;  // what it will be when i unlock/commit.
 377 };
 378
 379 ostream& operator<<(ostream& out, const CDentry& dn);
 380
 381
 382 #endif