ceph/src/client/Inode.cc

   1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
   2 // vim: ts=8 sw=2 smarttab
   3
   4 #include "Client.h"
   5 #include "Inode.h"
   6 #include "Dentry.h"
   7 #include "Dir.h"
   8 #include "Fh.h"
   9 #include "MetaSession.h"
  10 #include "ClientSnapRealm.h"
  11
  12 #include "mds/flock.h"
  13
  14 Inode::~Inode()
  15 {
  16   cap_item.remove_myself();
  17   snaprealm_item.remove_myself();
  18
  19   if (snapdir_parent) {
  20     snapdir_parent->flags &= ~I_SNAPDIR_OPEN;
  21     snapdir_parent.reset();
  22   }
  23
  24   if (!oset.objects.empty()) {
  25     lsubdout(client->cct, client, 0) << __func__ << ": leftover objects on inode 0x"
  26       << std::hex << ino << std::dec << dendl;
  27     assert(oset.objects.empty());
  28   }
  29
  30   delete fcntl_locks;
  31   delete flock_locks;
  32 }
  33
  34 ostream& operator<<(ostream &out, const Inode &in)
  35 {
  36   out << in.vino() << "("
  37       << "faked_ino=" << in.faked_ino
  38       << " ref=" << in._ref
  39       << " ll_ref=" << in.ll_ref
  40       << " cap_refs=" << in.cap_refs
  41       << " open=" << in.open_by_mode
  42       << " mode=" << oct << in.mode << dec
  43       << " size=" << in.size << "/" << in.max_size
  44       << " mtime=" << in.mtime
  45       << " caps=" << ccap_string(in.caps_issued());
  46   if (!in.caps.empty()) {
  47     out << "(";
  48     for (auto p = in.caps.begin(); p != in.caps.end(); ++p) {
  49       if (p != in.caps.begin())
  50         out << ',';
  51       out << p->first << '=' << ccap_string(p->second->issued);
  52     }
  53     out << ")";
  54   }
  55   if (in.dirty_caps)
  56     out << " dirty_caps=" << ccap_string(in.dirty_caps);
  57   if (in.flushing_caps)
  58     out << " flushing_caps=" << ccap_string(in.flushing_caps);
  59
  60   if (in.flags & I_COMPLETE)
  61     out << " COMPLETE";
  62
  63   if (in.is_file())
  64     out << " " << in.oset;
  65
  66   if (!in.dn_set.empty())
  67     out << " parents=" << in.dn_set;
  68
  69   if (in.is_dir() && in.has_dir_layout())
  70     out << " has_dir_layout";
  71
  72   if (in.quota.is_enable())
  73     out << " " << in.quota;
  74
  75   out << ' ' << &in << ")";
  76   return out;
  77 }
  78
  79
  80 void Inode::make_long_path(filepath& p)
  81 {
  82   if (!dn_set.empty()) {
  83     assert((*dn_set.begin())->dir && (*dn_set.begin())->dir->parent_inode);
  84     (*dn_set.begin())->dir->parent_inode->make_long_path(p);
  85     p.push_dentry((*dn_set.begin())->name);
  86   } else if (snapdir_parent) {
  87     snapdir_parent->make_nosnap_relative_path(p);
  88     string empty;
  89     p.push_dentry(empty);
  90   } else
  91     p = filepath(ino);
  92 }
  93
  94 /*
  95  * make a filepath suitable for an mds request:
  96  *  - if we are non-snapped/live, the ino is sufficient, e.g. #1234
  97  *  - if we are snapped, make filepath relative to first non-snapped parent.
  98  */
  99 void Inode::make_nosnap_relative_path(filepath& p)
 100 {
 101   if (snapid == CEPH_NOSNAP) {
 102     p = filepath(ino);
 103   } else if (snapdir_parent) {
 104     snapdir_parent->make_nosnap_relative_path(p);
 105     string empty;
 106     p.push_dentry(empty);
 107   } else if (!dn_set.empty()) {
 108     assert((*dn_set.begin())->dir && (*dn_set.begin())->dir->parent_inode);
 109     (*dn_set.begin())->dir->parent_inode->make_nosnap_relative_path(p);
 110     p.push_dentry((*dn_set.begin())->name);
 111   } else {
 112     p = filepath(ino);
 113   }
 114 }
 115
 116 void Inode::get_open_ref(int mode)
 117 {
 118   open_by_mode[mode]++;
 119 }
 120
 121 bool Inode::put_open_ref(int mode)
 122 {
 123   //cout << "open_by_mode[" << mode << "] " << open_by_mode[mode] << " -> " << (open_by_mode[mode]-1) << std::endl;
 124   if (--open_by_mode[mode] == 0)
 125     return true;
 126   return false;
 127 }
 128
 129 void Inode::get_cap_ref(int cap)
 130 {
 131   int n = 0;
 132   while (cap) {
 133     if (cap & 1) {
 134       int c = 1 << n;
 135       cap_refs[c]++;
 136       //cout << "inode " << *this << " get " << cap_string(c) << " " << (cap_refs[c]-1) << " -> " << cap_refs[c] << std::endl;
 137     }
 138     cap >>= 1;
 139     n++;
 140   }
 141 }
 142
 143 int Inode::put_cap_ref(int cap)
 144 {
 145   // if cap is always a single bit (which it seems to be)
 146   // all this logic is equivalent to:
 147   // if (--cap_refs[c]) return false; else return true;
 148   int last = 0;
 149   int n = 0;
 150   while (cap) {
 151     if (cap & 1) {
 152       int c = 1 << n;
 153       if (cap_refs[c] <= 0) {
 154         lderr(client->cct) << "put_cap_ref " << ccap_string(c) << " went negative on " << *this << dendl;
 155         assert(cap_refs[c] > 0);
 156       }
 157       if (--cap_refs[c] == 0)
 158         last |= c;
 159       //cout << "inode " << *this << " put " << cap_string(c) << " " << (cap_refs[c]+1) << " -> " << cap_refs[c] << std::endl;
 160     }
 161     cap >>= 1;
 162     n++;
 163   }
 164   return last;
 165 }
 166
 167 bool Inode::is_any_caps()
 168 {
 169   return !caps.empty() || snap_caps;
 170 }
 171
 172 bool Inode::cap_is_valid(Cap* cap) const
 173 {
 174   /*cout << "cap_gen     " << cap->session-> cap_gen << std::endl
 175     << "session gen " << cap->gen << std::endl
 176     << "cap expire  " << cap->session->cap_ttl << std::endl
 177     << "cur time    " << ceph_clock_now(cct) << std::endl;*/
 178   if ((cap->session->cap_gen <= cap->gen)
 179       && (ceph_clock_now() < cap->session->cap_ttl)) {
 180     return true;
 181   }
 182   return false;
 183 }
 184
 185 int Inode::caps_issued(int *implemented) const
 186 {
 187   int c = snap_caps;
 188   int i = 0;
 189   for (map<mds_rank_t,Cap*>::const_iterator it = caps.begin();
 190        it != caps.end();
 191        ++it)
 192     if (cap_is_valid(it->second)) {
 193       c |= it->second->issued;
 194       i |= it->second->implemented;
 195     }
 196   if (implemented)
 197     *implemented = i;
 198   return c;
 199 }
 200
 201 void Inode::touch_cap(Cap *cap)
 202 {
 203   // move to back of LRU
 204   cap->session->caps.push_back(&cap->cap_item);
 205 }
 206
 207 void Inode::try_touch_cap(mds_rank_t mds)
 208 {
 209   if (caps.count(mds))
 210     touch_cap(caps[mds]);
 211 }
 212
 213 bool Inode::caps_issued_mask(unsigned mask)
 214 {
 215   int c = snap_caps;
 216   if ((c & mask) == mask)
 217     return true;
 218   // prefer auth cap
 219   if (auth_cap &&
 220       cap_is_valid(auth_cap) &&
 221       (auth_cap->issued & mask) == mask) {
 222     touch_cap(auth_cap);
 223     return true;
 224   }
 225   // try any cap
 226   for (map<mds_rank_t,Cap*>::iterator it = caps.begin();
 227        it != caps.end();
 228        ++it) {
 229     if (cap_is_valid(it->second)) {
 230       if ((it->second->issued & mask) == mask) {
 231         touch_cap(it->second);
 232         return true;
 233       }
 234       c |= it->second->issued;
 235     }
 236   }
 237   if ((c & mask) == mask) {
 238     // bah.. touch them all
 239     for (map<mds_rank_t,Cap*>::iterator it = caps.begin();
 240          it != caps.end();
 241          ++it)
 242       touch_cap(it->second);
 243     return true;
 244   }
 245   return false;
 246 }
 247
 248 int Inode::caps_used()
 249 {
 250   int w = 0;
 251   for (map<int,int>::iterator p = cap_refs.begin();
 252        p != cap_refs.end();
 253        ++p)
 254     if (p->second)
 255       w |= p->first;
 256   return w;
 257 }
 258
 259 int Inode::caps_file_wanted()
 260 {
 261   int want = 0;
 262   for (map<int,int>::iterator p = open_by_mode.begin();
 263        p != open_by_mode.end();
 264        ++p)
 265     if (p->second)
 266       want |= ceph_caps_for_mode(p->first);
 267   return want;
 268 }
 269
 270 int Inode::caps_wanted()
 271 {
 272   int want = caps_file_wanted() | caps_used();
 273   if (want & CEPH_CAP_FILE_BUFFER)
 274     want |= CEPH_CAP_FILE_EXCL;
 275   return want;
 276 }
 277
 278 int Inode::caps_mds_wanted()
 279 {
 280   int want = 0;
 281   for (auto it = caps.begin(); it != caps.end(); ++it)
 282     want |= it->second->wanted;
 283   return want;
 284 }
 285
 286 int Inode::caps_dirty()
 287 {
 288   return dirty_caps | flushing_caps;
 289 }
 290
 291 const UserPerm* Inode::get_best_perms()
 292 {
 293   const UserPerm *perms = NULL;
 294   for (const auto ci : caps) {
 295     const UserPerm& iperm = ci.second->latest_perms;
 296     if (!perms) { // we don't have any, take what's present
 297       perms = &iperm;
 298     } else if (iperm.uid() == uid) {
 299       if (iperm.gid() == gid) { // we have the best possible, return
 300         return &iperm;
 301       }
 302       if (perms->uid() != uid) { // take uid > gid every time
 303         perms = &iperm;
 304       }
 305     } else if (perms->uid() != uid && iperm.gid() == gid) {
 306       perms = &iperm; // a matching gid is better than nothing
 307     }
 308   }
 309   return perms;
 310 }
 311
 312 bool Inode::have_valid_size()
 313 {
 314   // RD+RDCACHE or WR+WRBUFFER => valid size
 315   if (caps_issued() & (CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL))
 316     return true;
 317   return false;
 318 }
 319
 320 // open Dir for an inode.  if it's not open, allocated it (and pin dentry in memory).
 321 Dir *Inode::open_dir()
 322 {
 323   if (!dir) {
 324     dir = new Dir(this);
 325     lsubdout(client->cct, client, 15) << "open_dir " << dir << " on " << this << dendl;
 326     assert(dn_set.size() < 2); // dirs can't be hard-linked
 327     if (!dn_set.empty())
 328       (*dn_set.begin())->get();      // pin dentry
 329     get();                  // pin inode
 330   }
 331   return dir;
 332 }
 333
 334 bool Inode::check_mode(const UserPerm& perms, unsigned want)
 335 {
 336   if (uid == perms.uid()) {
 337     // if uid is owner, owner entry determines access
 338     want = want << 6;
 339   } else if (perms.gid_in_groups(gid)) {
 340     // if a gid or sgid matches the owning group, group entry determines access
 341     want = want << 3;
 342   }
 343
 344   return (mode & want) == want;
 345 }
 346
 347 void Inode::get() {
 348   _ref++;
 349   lsubdout(client->cct, client, 15) << "inode.get on " << this << " " <<  ino << '.' << snapid
 350                                     << " now " << _ref << dendl;
 351 }
 352
 353 //private method to put a reference; see Client::put_inode()
 354 int Inode::_put(int n) {
 355   _ref -= n;
 356   lsubdout(client->cct, client, 15) << "inode.put on " << this << " " << ino << '.' << snapid
 357                                     << " now " << _ref << dendl;
 358   assert(_ref >= 0);
 359   return _ref;
 360 }
 361
 362
 363 void Inode::dump(Formatter *f) const
 364 {
 365   f->dump_stream("ino") << ino;
 366   f->dump_stream("snapid") << snapid;
 367   if (rdev)
 368     f->dump_unsigned("rdev", rdev);
 369   f->dump_stream("ctime") << ctime;
 370   f->dump_stream("btime") << btime;
 371   f->dump_stream("mode") << '0' << std::oct << mode << std::dec;
 372   f->dump_unsigned("uid", uid);
 373   f->dump_unsigned("gid", gid);
 374   f->dump_unsigned("nlink", nlink);
 375
 376   f->dump_int("size", size);
 377   f->dump_int("max_size", max_size);
 378   f->dump_int("truncate_seq", truncate_seq);
 379   f->dump_int("truncate_size", truncate_size);
 380   f->dump_stream("mtime") << mtime;
 381   f->dump_stream("atime") << atime;
 382   f->dump_int("time_warp_seq", time_warp_seq);
 383   f->dump_int("change_attr", change_attr);
 384
 385   f->dump_object("layout", layout);
 386   if (is_dir()) {
 387     f->open_object_section("dir_layout");
 388     ::dump(dir_layout, f);
 389     f->close_section();
 390
 391     f->dump_bool("complete", flags & I_COMPLETE);
 392     f->dump_bool("ordered", flags & I_DIR_ORDERED);
 393
 394     /* FIXME when wip-mds-encoding is merged ***
 395     f->open_object_section("dir_stat");
 396     dirstat.dump(f);
 397     f->close_section();
 398
 399     f->open_object_section("rstat");
 400     rstat.dump(f);
 401     f->close_section();
 402     */
 403   }
 404
 405   f->dump_unsigned("version", version);
 406   f->dump_unsigned("xattr_version", xattr_version);
 407   f->dump_unsigned("flags", flags);
 408
 409   if (is_dir()) {
 410     if (!dir_contacts.empty()) {
 411       f->open_object_section("dir_contants");
 412       for (set<int>::iterator p = dir_contacts.begin(); p != dir_contacts.end(); ++p)
 413         f->dump_int("mds", *p);
 414       f->close_section();
 415     }
 416     f->dump_int("dir_hashed", (int)dir_hashed);
 417     f->dump_int("dir_replicated", (int)dir_replicated);
 418   }
 419
 420   f->open_array_section("caps");
 421   for (map<mds_rank_t,Cap*>::const_iterator p = caps.begin(); p != caps.end(); ++p) {
 422     f->open_object_section("cap");
 423     f->dump_int("mds", p->first);
 424     if (p->second == auth_cap)
 425       f->dump_int("auth", 1);
 426     p->second->dump(f);
 427     f->close_section();
 428   }
 429   f->close_section();
 430   if (auth_cap)
 431     f->dump_int("auth_cap", auth_cap->session->mds_num);
 432
 433   f->dump_stream("dirty_caps") << ccap_string(dirty_caps);
 434   if (flushing_caps) {
 435     f->dump_stream("flushings_caps") << ccap_string(flushing_caps);
 436     f->open_object_section("flushing_cap_tid");
 437     for (map<ceph_tid_t, int>::const_iterator p = flushing_cap_tids.begin();
 438          p != flushing_cap_tids.end();
 439          ++p) {
 440       string n(ccap_string(p->second));
 441       f->dump_unsigned(n.c_str(), p->first);
 442     }
 443     f->close_section();
 444   }
 445   f->dump_int("shared_gen", shared_gen);
 446   f->dump_int("cache_gen", cache_gen);
 447   if (snap_caps) {
 448     f->dump_int("snap_caps", snap_caps);
 449     f->dump_int("snap_cap_refs", snap_cap_refs);
 450   }
 451
 452   f->dump_stream("hold_caps_until") << hold_caps_until;
 453
 454   if (snaprealm) {
 455     f->open_object_section("snaprealm");
 456     snaprealm->dump(f);
 457     f->close_section();
 458   }
 459   if (!cap_snaps.empty()) {
 460     for (const auto &p : cap_snaps) {
 461       f->open_object_section("cap_snap");
 462       f->dump_stream("follows") << p.first;
 463       p.second.dump(f);
 464       f->close_section();
 465     }
 466   }
 467
 468   // open
 469   if (!open_by_mode.empty()) {
 470     f->open_array_section("open_by_mode");
 471     for (map<int,int>::const_iterator p = open_by_mode.begin(); p != open_by_mode.end(); ++p) {
 472       f->open_object_section("ref");
 473       f->dump_unsigned("mode", p->first);
 474       f->dump_unsigned("refs", p->second);
 475       f->close_section();
 476     }
 477     f->close_section();
 478   }
 479   if (!cap_refs.empty()) {
 480     f->open_array_section("cap_refs");
 481     for (map<int,int>::const_iterator p = cap_refs.begin(); p != cap_refs.end(); ++p) {
 482       f->open_object_section("cap_ref");
 483       f->dump_stream("cap") << ccap_string(p->first);
 484       f->dump_int("refs", p->second);
 485       f->close_section();
 486     }
 487     f->close_section();
 488   }
 489
 490   f->dump_unsigned("reported_size", reported_size);
 491   if (wanted_max_size != max_size)
 492     f->dump_unsigned("wanted_max_size", wanted_max_size);
 493   if (requested_max_size != max_size)
 494     f->dump_unsigned("requested_max_size", requested_max_size);
 495
 496   f->dump_int("ref", _ref);
 497   f->dump_int("ll_ref", ll_ref);
 498
 499   if (!dn_set.empty()) {
 500     f->open_array_section("parents");
 501     for (set<Dentry*>::const_iterator p = dn_set.begin(); p != dn_set.end(); ++p) {
 502       f->open_object_section("dentry");
 503       f->dump_stream("dir_ino") << (*p)->dir->parent_inode->ino;
 504       f->dump_string("name", (*p)->name);
 505       f->close_section();
 506     }
 507     f->close_section();
 508   }
 509 }
 510
 511 void Cap::dump(Formatter *f) const
 512 {
 513   f->dump_int("mds", session->mds_num);
 514   f->dump_stream("ino") << inode->ino;
 515   f->dump_unsigned("cap_id", cap_id);
 516   f->dump_stream("issued") << ccap_string(issued);
 517   if (implemented != issued)
 518     f->dump_stream("implemented") << ccap_string(implemented);
 519   f->dump_stream("wanted") << ccap_string(wanted);
 520   f->dump_unsigned("seq", seq);
 521   f->dump_unsigned("issue_seq", issue_seq);
 522   f->dump_unsigned("mseq", mseq);
 523   f->dump_unsigned("gen", gen);
 524 }
 525
 526 void CapSnap::dump(Formatter *f) const
 527 {
 528   f->dump_stream("ino") << in->ino;
 529   f->dump_stream("issued") << ccap_string(issued);
 530   f->dump_stream("dirty") << ccap_string(dirty);
 531   f->dump_unsigned("size", size);
 532   f->dump_stream("ctime") << ctime;
 533   f->dump_stream("mtime") << mtime;
 534   f->dump_stream("atime") << atime;
 535   f->dump_int("time_warp_seq", time_warp_seq);
 536   f->dump_stream("mode") << '0' << std::oct << mode << std::dec;
 537   f->dump_unsigned("uid", uid);
 538   f->dump_unsigned("gid", gid);
 539   if (!xattrs.empty()) {
 540     f->open_object_section("xattr_lens");
 541     for (map<string,bufferptr>::const_iterator p = xattrs.begin(); p != xattrs.end(); ++p)
 542       f->dump_int(p->first.c_str(), p->second.length());
 543     f->close_section();
 544   }
 545   f->dump_unsigned("xattr_version", xattr_version);
 546   f->dump_int("writing", (int)writing);
 547   f->dump_int("dirty_data", (int)dirty_data);
 548   f->dump_unsigned("flush_tid", flush_tid);
 549 }
 550
 551 void Inode::set_async_err(int r)
 552 {
 553   for (const auto &fh : fhs) {
 554     fh->async_err = r;
 555   }
 556 }
 557