ceph/src/mds/Capability.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 #ifndef CEPH_CAPABILITY_H
  17 #define CEPH_CAPABILITY_H
  18
  19 #include "include/counter.h"
  20 #include "include/buffer_fwd.h"
  21 #include "include/xlist.h"
  22
  23 #include "common/config.h"
  24
  25 #include "mdstypes.h"
  26
  27 /*
  28
  29   Capability protocol notes.
  30
  31 - two types of cap events from mds -> client:
  32   - cap "issue" in a MClientReply, or an MClientCaps IMPORT op.
  33   - cap "update" (revocation or grant) .. an MClientCaps message.
  34 - if client has cap, the mds should have it too.
  35
  36 - if client has no dirty data, it can release it without waiting for an mds ack.
  37   - client may thus get a cap _update_ and not have the cap.  ignore it.
  38
  39 - mds should track seq of last issue.  any release
  40   attempt will only succeed if the client has seen the latest.
  41
  42 - a UPDATE updates the clients issued caps, wanted, etc.  it may also flush dirty metadata.
  43   - 'caps' are which caps the client retains.
  44     - if 0, client wishes to release the cap
  45   - 'wanted' is which caps the client wants.
  46   - 'dirty' is which metadata is to be written.
  47     - client gets a FLUSH_ACK with matching dirty flags indicating which caps were written.
  48
  49 - a FLUSH_ACK acks a FLUSH.
  50   - 'dirty' is the _original_ FLUSH's dirty (i.e., which metadata was written back)
  51   - 'seq' is the _original_ FLUSH's seq.
  52   - 'caps' is the _original_ FLUSH's caps (not actually important)
  53   - client can conclude that (dirty & ~caps) bits were successfully cleaned.
  54
  55 - a FLUSHSNAP flushes snapshot metadata.
  56   - 'dirty' indicates which caps, were dirty, if any.
  57   - mds writes metadata.  if dirty!=0, replies with FLUSHSNAP_ACK.
  58
  59  */
  60
  61 class CInode;
  62
  63 namespace ceph {
  64   class Formatter;
  65 }
  66
  67 class Capability : public Counter<Capability> {
  68 public:
  69   struct Export {
  70     int64_t cap_id;
  71     int32_t wanted;
  72     int32_t issued;
  73     int32_t pending;
  74     snapid_t client_follows;
  75     ceph_seq_t seq;
  76     ceph_seq_t mseq;
  77     utime_t last_issue_stamp;
  78     Export() : cap_id(0), wanted(0), issued(0), pending(0), seq(0), mseq(0) {}
  79     Export(int64_t id, int w, int i, int p, snapid_t cf, ceph_seq_t s, ceph_seq_t m, utime_t lis) :
  80       cap_id(id), wanted(w), issued(i), pending(p), client_follows(cf),
  81       seq(s), mseq(m), last_issue_stamp(lis) {}
  82     void encode(bufferlist &bl) const;
  83     void decode(bufferlist::iterator &p);
  84     void dump(Formatter *f) const;
  85     static void generate_test_instances(list<Export*>& ls);
  86   };
  87   struct Import {
  88     int64_t cap_id;
  89     ceph_seq_t issue_seq;
  90     ceph_seq_t mseq;
  91     Import() : cap_id(0), issue_seq(0), mseq(0) {}
  92     Import(int64_t i, ceph_seq_t s, ceph_seq_t m) : cap_id(i), issue_seq(s), mseq(m) {}
  93     void encode(bufferlist &bl) const;
  94     void decode(bufferlist::iterator &p);
  95     void dump(Formatter *f) const;
  96   };
  97   struct revoke_info {
  98     __u32 before;
  99     ceph_seq_t seq, last_issue;
 100     revoke_info() : before(0), seq(0), last_issue(0) {}
 101     revoke_info(__u32 b, ceph_seq_t s, ceph_seq_t li) : before(b), seq(s), last_issue(li) {}
 102     void encode(bufferlist& bl) const;
 103     void decode(bufferlist::iterator& bl);
 104     void dump(Formatter *f) const;
 105     static void generate_test_instances(list<revoke_info*>& ls);
 106   };
 107
 108
 109   const static unsigned STATE_STALE             = (1<<0);
 110   const static unsigned STATE_NEW               = (1<<1);
 111   const static unsigned STATE_IMPORTING         = (1<<2);
 112
 113
 114   Capability(CInode *i = NULL, uint64_t id = 0, client_t c = 0) :
 115     client_follows(0), client_xattr_version(0),
 116     client_inline_version(0),
 117     last_rbytes(0), last_rsize(0),
 118     item_session_caps(this), item_snaprealm_caps(this),
 119     item_revoking_caps(this), item_client_revoking_caps(this),
 120     inode(i), client(c),
 121     cap_id(id),
 122     _wanted(0), num_revoke_warnings(0),
 123     _pending(0), _issued(0),
 124     last_sent(0),
 125     last_issue(0),
 126     mseq(0),
 127     suppress(0), state(0) {
 128   }
 129   Capability(const Capability& other);  // no copying
 130
 131   const Capability& operator=(const Capability& other);  // no copying
 132
 133   int pending() { return _pending; }
 134   int issued() { return _issued; }
 135   bool is_null() { return !_pending && _revokes.empty(); }
 136
 137   ceph_seq_t issue(unsigned c) {
 138     if (_pending & ~c) {
 139       // revoking (and maybe adding) bits.  note caps prior to this revocation
 140       _revokes.push_back(revoke_info(_pending, last_sent, last_issue));
 141       _pending = c;
 142       _issued |= c;
 143     } else if (~_pending & c) {
 144       // adding bits only.  remove obsolete revocations?
 145       _pending |= c;
 146       _issued |= c;
 147       // drop old _revokes with no bits we don't have
 148       while (!_revokes.empty() &&
 149              (_revokes.back().before & ~_pending) == 0)
 150         _revokes.pop_back();
 151     } else {
 152       // no change.
 153       assert(_pending == c);
 154     }
 155     //last_issue =
 156     ++last_sent;
 157     return last_sent;
 158   }
 159   ceph_seq_t issue_norevoke(unsigned c) {
 160     _pending |= c;
 161     _issued |= c;
 162     //check_rdcaps_list();
 163     ++last_sent;
 164     return last_sent;
 165   }
 166   void _calc_issued() {
 167     _issued = _pending;
 168     for (list<revoke_info>::iterator p = _revokes.begin(); p != _revokes.end(); ++p)
 169       _issued |= p->before;
 170   }
 171   void confirm_receipt(ceph_seq_t seq, unsigned caps) {
 172     if (seq == last_sent) {
 173       _revokes.clear();
 174       _issued = caps;
 175       // don't add bits
 176       _pending &= caps;
 177     } else {
 178       // can i forget any revocations?
 179       while (!_revokes.empty() && _revokes.front().seq < seq)
 180         _revokes.pop_front();
 181       if (!_revokes.empty()) {
 182         if (_revokes.front().seq == seq)
 183           _revokes.begin()->before = caps;
 184         _calc_issued();
 185       } else {
 186         // seq < last_sent
 187         _issued = caps | _pending;
 188       }
 189     }
 190
 191     if (_issued == _pending) {
 192       item_revoking_caps.remove_myself();
 193       item_client_revoking_caps.remove_myself();
 194     }
 195     //check_rdcaps_list();
 196   }
 197   // we may get a release racing with revocations, which means our revokes will be ignored
 198   // by the client.  clean them out of our _revokes history so we don't wait on them.
 199   void clean_revoke_from(ceph_seq_t li) {
 200     bool changed = false;
 201     while (!_revokes.empty() && _revokes.front().last_issue <= li) {
 202       _revokes.pop_front();
 203       changed = true;
 204     }
 205     if (changed) {
 206       _calc_issued();
 207       if (_issued == _pending) {
 208         item_revoking_caps.remove_myself();
 209         item_client_revoking_caps.remove_myself();
 210       }
 211     }
 212   }
 213   ceph_seq_t get_mseq() { return mseq; }
 214   void inc_mseq() { mseq++; }
 215
 216   ceph_seq_t get_last_sent() { return last_sent; }
 217   utime_t get_last_issue_stamp() { return last_issue_stamp; }
 218   utime_t get_last_revoke_stamp() { return last_revoke_stamp; }
 219
 220   void set_last_issue() { last_issue = last_sent; }
 221   void set_last_issue_stamp(utime_t t) { last_issue_stamp = t; }
 222   void set_last_revoke_stamp(utime_t t) { last_revoke_stamp = t; }
 223   void reset_num_revoke_warnings() { num_revoke_warnings = 0; }
 224   void inc_num_revoke_warnings() { ++num_revoke_warnings; }
 225   unsigned get_num_revoke_warnings() { return num_revoke_warnings; }
 226
 227   void set_cap_id(uint64_t i) { cap_id = i; }
 228   uint64_t get_cap_id() { return cap_id; }
 229
 230   //ceph_seq_t get_last_issue() { return last_issue; }
 231
 232   bool is_suppress() { return suppress > 0; }
 233   void inc_suppress() { suppress++; }
 234   void dec_suppress() { suppress--; }
 235
 236   bool is_stale() { return state & STATE_STALE; }
 237   void mark_stale() { state |= STATE_STALE; }
 238   void clear_stale() { state &= ~STATE_STALE; }
 239   bool is_new() { return state & STATE_NEW; }
 240   void mark_new() { state |= STATE_NEW; }
 241   void clear_new() { state &= ~STATE_NEW; }
 242   bool is_importing() { return state & STATE_IMPORTING; }
 243   void mark_importing() { state |= STATE_IMPORTING; }
 244   void clear_importing() { state &= ~STATE_IMPORTING; }
 245
 246   CInode *get_inode() { return inode; }
 247   client_t get_client() const { return client; }
 248
 249   // caps this client wants to hold
 250   int wanted() { return _wanted; }
 251   void set_wanted(int w) {
 252     _wanted = w;
 253     //check_rdcaps_list();
 254   }
 255
 256   void inc_last_seq() { last_sent++; }
 257   ceph_seq_t get_last_seq() { return last_sent; }
 258   ceph_seq_t get_last_issue() { return last_issue; }
 259
 260   void reset_seq() {
 261     last_sent = 0;
 262     last_issue = 0;
 263   }
 264
 265   // -- exports --
 266   Export make_export() {
 267     return Export(cap_id, _wanted, issued(), pending(), client_follows, last_sent, mseq+1, last_issue_stamp);
 268   }
 269   void merge(Export& other, bool auth_cap) {
 270     if (!is_stale()) {
 271       // issued + pending
 272       int newpending = other.pending | pending();
 273       if (other.issued & ~newpending)
 274         issue(other.issued | newpending);
 275       else
 276         issue(newpending);
 277       last_issue_stamp = other.last_issue_stamp;
 278     } else {
 279       issue(CEPH_CAP_PIN);
 280     }
 281
 282     client_follows = other.client_follows;
 283
 284     // wanted
 285     _wanted = _wanted | other.wanted;
 286     if (auth_cap)
 287       mseq = other.mseq;
 288   }
 289   void merge(int otherwanted, int otherissued) {
 290     if (!is_stale()) {
 291       // issued + pending
 292       int newpending = pending();
 293       if (otherissued & ~newpending)
 294         issue(otherissued | newpending);
 295       else
 296         issue(newpending);
 297     } else {
 298       issue(CEPH_CAP_PIN);
 299     }
 300
 301     // wanted
 302     _wanted = _wanted | otherwanted;
 303   }
 304
 305   void revoke() {
 306     if (pending() & ~CEPH_CAP_PIN)
 307       issue(CEPH_CAP_PIN);
 308     confirm_receipt(last_sent, CEPH_CAP_PIN);
 309   }
 310
 311   // serializers
 312   void encode(bufferlist &bl) const;
 313   void decode(bufferlist::iterator &bl);
 314   void dump(Formatter *f) const;
 315   static void generate_test_instances(list<Capability*>& ls);
 316
 317   snapid_t client_follows;
 318   version_t client_xattr_version;
 319   version_t client_inline_version;
 320   int64_t last_rbytes;
 321   int64_t last_rsize;
 322
 323   xlist<Capability*>::item item_session_caps;
 324   xlist<Capability*>::item item_snaprealm_caps;
 325   xlist<Capability*>::item item_revoking_caps;
 326   xlist<Capability*>::item item_client_revoking_caps;
 327
 328 private:
 329   CInode *inode;
 330   client_t client;
 331
 332   uint64_t cap_id;
 333
 334   __u32 _wanted;     // what the client wants (ideally)
 335
 336   utime_t last_issue_stamp;
 337   utime_t last_revoke_stamp;
 338   unsigned num_revoke_warnings;
 339
 340   // track in-flight caps --------------
 341   //  - add new caps to _pending
 342   //  - track revocations in _revokes list
 343   __u32 _pending, _issued;
 344   list<revoke_info> _revokes;
 345
 346   ceph_seq_t last_sent;
 347   ceph_seq_t last_issue;
 348   ceph_seq_t mseq;
 349
 350   int suppress;
 351   unsigned state;
 352 };
 353
 354 WRITE_CLASS_ENCODER(Capability::Export)
 355 WRITE_CLASS_ENCODER(Capability::Import)
 356 WRITE_CLASS_ENCODER(Capability::revoke_info)
 357 WRITE_CLASS_ENCODER(Capability)
 358
 359
 360
 361 #endif