mds: .snap mostly working

diff --git a/src/client/Client.cc b/src/client/Client.cc
index dd6371c77ea9a48eb99808e5280e79365182e3c9..a1a644de28f18bfbf968df9dfd7d9b013ac691ff 100644 (file)
--- a/src/client/Client.cc
+++ b/src/client/Client.cc
@@ -85,6 +85,7 @@ ostream& operator<<(ostream &out, Inode &in)
       << " cap_refs=" << in.cap_refs
       << " open=" << in.open_by_mode
       << " ref=" << in.ref
+      << " parent=" << in.dn
       << ")";
   return out;
 }
@@ -540,13 +541,16 @@ Inode* Client::insert_trace(MClientReply *reply, utime_t from)
   }
 
   __u16 numi, numd;
+  __s16 snapdirpos;
   ::decode(numi, p);
   ::decode(numd, p);
-  dout(10) << "insert_trace got " << numi << " inodes, " << numd << " dentries" << dendl;
+  ::decode(snapdirpos, p);
+  dout(10) << "insert_trace got " << numi << " inodes, " << numd << " dentries, snapdir at " << snapdirpos << dendl;
+  int icount = 0;
 
   // decode
-  LeaseStat ilease[numi];
-  InodeStat ist[numi];
+  LeaseStat ilease[numi], snapdirlease;
+  InodeStat ist[numi], snapdirst;
   DirStat dst[numd];
   string dname[numd];
   LeaseStat dlease[numd];
@@ -561,7 +565,11 @@ Inode* Client::insert_trace(MClientReply *reply, utime_t from)
 
  inode:
   if (!ileft) goto done;
-  ileft--;
+  ileft--; icount++;
+  if (icount == snapdirpos) {
+    snapdirst.decode(p);
+    ::decode(snapdirlease, p);
+  }
   ist[ileft].decode(p);
   ::decode(ilease[ileft], p);
 
@@ -590,6 +598,7 @@ Inode* Client::insert_trace(MClientReply *reply, utime_t from)
     curi = inode_map[vino];
   }
   update_inode(curi, &ist[0], &ilease[0], from);
+  dout(10) << " (base) curi " << *curi << dendl;
 
   for (unsigned i=0; i<numd; i++) {
     Dir *dir = curi->open_dir();
@@ -603,6 +612,20 @@ Inode* Client::insert_trace(MClientReply *reply, utime_t from)
     }
     
     curi = insert_dentry_inode(dir, dname[i], &dlease[i], &ist[i+1], &ilease[i+1], from);
+    dout(10) << " curi " << *curi << dendl;
+
+    if ((int)i == numi-snapdirpos-1) {
+      Inode *snapdiri = open_snapdir(curi);
+      dout(10) << " snapdir " << *snapdiri << dendl;
+      char s[20];
+      sprintf(s, "%llu", (unsigned long long)snapdirst.vino.snapid);
+      string snapname = s;
+      Dir *snapdir = snapdiri->open_dir();
+      curi = insert_dentry_inode(snapdir, snapname, &snapdirlease, // FIXME
+                                &snapdirst, &snapdirlease, from);
+      dout(10) << " snapped diri " << *curi << dendl;
+    }
+
     update_dir_dist(curi, &dst[i]);  // dir stat info is attached to inode...
   }
   assert(p.end());
@@ -2767,7 +2790,7 @@ int Client::_readdir_get_frag(DirResult *dirp)
   dout(10) << "_readdir_get_frag " << dirp << " on " << dirp->path << " fg " << fg << dendl;
 
   int op = CEPH_MDS_OP_READDIR;
-  if (dirp->inode && dirp->inode->snapid == SNAPDIR)
+  if (dirp->inode && dirp->inode->snapid == CEPH_SNAPDIR)
     op = CEPH_MDS_OP_LSSNAP;
 
   MClientRequest *req = new MClientRequest(op);
@@ -3801,6 +3824,28 @@ int Client::_rmsnap(const filepath& path, const char *name, int uid, int gid)
 #define FUSE_SET_ATTR_ATIME    (1 << 4)
 #define FUSE_SET_ATTR_MTIME    (1 << 5)
 
+
+Inode *Client::open_snapdir(Inode *diri)
+{
+  Inode *in;
+  vinodeno_t vino(diri->ino(), CEPH_SNAPDIR);
+  if (!inode_map.count(vino)) {
+    in = new Inode(vino, &diri->inode.layout);
+    in->inode = diri->inode;
+    in->snapid = CEPH_SNAPDIR;
+    in->inode.mode = S_IFDIR | 0600;
+    in->dirfragtree.clear();
+    inode_map[vino] = in;
+    in->snapdir_parent = diri;
+    diri->get();
+    dout(10) << "open_snapdir created snapshot inode " << *in << dendl;
+  } else {
+    in = inode_map[vino];
+    dout(10) << "open_snapdir had snapshot inode " << *in << dendl;
+  }
+  return in;
+}
+
 int Client::ll_lookup(vinodeno_t parent, const char *name, struct stat *attr, int uid, int gid)
 {
   Mutex::Locker lock(client_lock);
@@ -3832,21 +3877,7 @@ int Client::ll_lookup(vinodeno_t parent, const char *name, struct stat *attr, in
   // .snapshot?
   if (dname == g_conf.client_snapdir &&
       diri->snapid == CEPH_NOSNAP) {
-    vinodeno_t vino(parent.ino, SNAPDIR);
-    if (!inode_map.count(vino)) {
-      in = new Inode(vino, &diri->inode.layout);
-      in->inode = diri->inode;
-      in->snapid = SNAPDIR;
-      in->inode.mode = S_IFDIR | 0600;
-      in->dirfragtree.clear();
-      inode_map[vino] = in;
-      in->snapdir_parent = diri;
-      diri->get();
-      dout(10) << " created snapshot inode " << *in << dendl;
-    } else {
-      in = inode_map[vino];
-      dout(10) << " had snapshot inode " << *in << dendl;
-    }
+    in = open_snapdir(diri);
   } else {
     // get the inode
     if (diri->dir &&
@@ -3862,7 +3893,16 @@ int Client::ll_lookup(vinodeno_t parent, const char *name, struct stat *attr, in
       filepath path;
       diri->make_path(path);
       path.push_dentry(name);
-      _do_lstat(path, 0, &in, uid, gid);
+      dout(10) << "ll_lookup on " << path << dendl;
+      //_do_lstat(path, 0, &in, uid, gid);
+      MClientRequest *req = new MClientRequest(CEPH_MDS_OP_LSTAT);
+      req->head.args.stat.mask = 0;
+      req->set_filepath(path);
+
+      MClientReply *reply = make_request(req, uid, gid, &in, 0);
+      r = reply->get_result();
+      dout(10) << "ll_lookup lstat res is " << r << dendl;
+      delete reply;
     }
   }
 
@@ -3870,7 +3910,7 @@ int Client::ll_lookup(vinodeno_t parent, const char *name, struct stat *attr, in
     fill_stat(in, attr);
     _ll_get(in);
   } else {
-    r = -ENOENT;
+    if (!r) r = -ENOENT;
     attr->st_ino = 0;
   }
 
@@ -4385,7 +4425,7 @@ int Client::ll_opendir(vinodeno_t vino, void **dirpp, int uid, int gid)
   assert(diri);
 
   filepath path;
-  if (vino.snapid == SNAPDIR) {
+  if (vino.snapid == CEPH_SNAPDIR) {
     Inode *livediri = inode_map[vinodeno_t(vino.ino, CEPH_NOSNAP)];
     assert(livediri);
     livediri->make_path(path);
@@ -4394,7 +4434,7 @@ int Client::ll_opendir(vinodeno_t vino, void **dirpp, int uid, int gid)
   dout(10) << " ino path is " << path << dendl;
 
   int r = 0;
-  if (vino.snapid == SNAPDIR) {
+  if (vino.snapid == CEPH_SNAPDIR) {
     *dirpp = new DirResult(path, diri);
   } else {
     r = _opendir(path.c_str(), (DirResult**)dirpp);

diff --git a/src/client/Client.h b/src/client/Client.h
index 7d8cb9f8e2f9cb0e03abba12e6296d7404c077f8..c2e3b66f3c7675f102d1f7b22ea1329cee7eca4c 100644 (file)
--- a/src/client/Client.h
+++ b/src/client/Client.h
@@ -50,9 +50,6 @@ using std::fstream;
 using namespace __gnu_cxx;
 
 
-#define SNAPDIR (CEPH_NOSNAP-1)
-
-
 
 class MStatfsReply;
 class MClientSession;
@@ -628,6 +625,9 @@ protected:
     }
   }
 
+  Inode *open_snapdir(Inode *diri);
+
+
   // file handles, etc.
   filepath cwd;
   interval_set<int> free_fd_set;  // unused fds

diff --git a/src/include/nstring.h b/src/include/nstring.h
index ee7d448d4323ba58f0fb054dd86a1767b415bc16..01812dc1b6249ba7f2ca9680038e05b869fcde42 100644 (file)
--- a/src/include/nstring.h
+++ b/src/include/nstring.h
@@ -9,6 +9,10 @@ typedef tstring nstring;
 typedef cstring nstring;
 #endif
 
+static inline bool operator==(const nstring &l, const char *s) {
+  return strcmp(l.c_str(), s) == 0;
+}
+
 static inline bool operator==(const nstring &l, const nstring &r) {
   return l.length() == r.length() && memcmp(l.data(), r.data(), l.length()) == 0;
 }

diff --git a/src/include/types.h b/src/include/types.h
index bdf65acf198195cad1ecf32c68ce11479c11dfc1..03a7c3d3a701b37afbe8f31e13e102e55994f112 100644 (file)
--- a/src/include/types.h
+++ b/src/include/types.h
@@ -154,6 +154,8 @@ inline void decode(snapid_t &i, bufferlist::iterator &p) { decode(i.val, p); }
 inline ostream& operator<<(ostream& out, snapid_t s) {
   if (s == CEPH_NOSNAP)
     return out << "head";
+  else if (s == CEPH_SNAPDIR)
+    return out << "snapdir";
   else
     return out << s.val;
 }

diff --git a/src/mds/CDir.cc b/src/mds/CDir.cc
index 06d071ddfe21fdb7b26b4041fef31cad68fa2833..96e398b23e1f3ad07415f12a64fb5a59d9deeb13 100644 (file)
--- a/src/mds/CDir.cc
+++ b/src/mds/CDir.cc
@@ -179,7 +179,20 @@ CDir::CDir(CInode *in, frag_t fg, MDCache *mdcache, bool auth) :
 
 
 
-
+CDentry *CDir::lookup(const char *name, snapid_t snap)
+{ 
+  dout(20) << "lookup (" << snap << ", '" << name << "')" << dendl;
+  map_t::iterator iter = items.lower_bound(dentry_key_t(snap, name));
+  if (iter == items.end())
+    return 0;
+  if (iter->second->name == name &&
+      iter->second->first <= snap) {
+    dout(20) << "  hit -> " << iter->first << dendl;
+    return iter->second;
+  }
+  dout(20) << "  miss -> " << iter->first << dendl;
+  return 0;
+}
 
 
 

diff --git a/src/mds/CDir.h b/src/mds/CDir.h
index c96caf159cd7dc286b7666e5d4758e0f202e1532..17073b8b90ce821063b9632f4d00fd5bdca6282e 100644 (file)
--- a/src/mds/CDir.h
+++ b/src/mds/CDir.h
@@ -197,7 +197,7 @@ public:
 public:
   //typedef hash_map<string, CDentry*> map_t;   // there is a bug somewhere, valgrind me.
   //typedef map<const char *, CDentry*, ltstr> map_t;
-  typedef map<dentry_key_t, CDentry*, ltdentrykey> map_t;
+  typedef map<dentry_key_t, CDentry*> map_t;
 protected:
 
   // contents
@@ -292,13 +292,7 @@ protected:
   CDentry* lookup(const nstring& ns, snapid_t snap=CEPH_NOSNAP) {
     return lookup(ns.c_str(), snap);
   }
-  CDentry* lookup(const char *n, snapid_t snap=CEPH_NOSNAP) {
-    map_t::iterator iter = items.find(dentry_key_t(snap, n));
-    if (iter == items.end()) 
-      return 0;
-    else
-      return iter->second;
-  }
+  CDentry* lookup(const char *n, snapid_t snap=CEPH_NOSNAP);
 
   CDentry* add_null_dentry(const nstring& dname, 
                           snapid_t first=1, snapid_t last=CEPH_NOSNAP);

diff --git a/src/mds/CInode.cc b/src/mds/CInode.cc
index bf6bd65708b6c28cdcda1661b243dfc113ef3b9d..46ea94888cc2365501f8213ec503445c125028a1 100644 (file)
--- a/src/mds/CInode.cc
+++ b/src/mds/CInode.cc
@@ -1051,7 +1051,7 @@ void CInode::encode_inodestat(bufferlist& bl, snapid_t snapid)
   // pick a version!
   inode_t *i = &inode;
   bufferlist xbl;
-  if (!old_inodes.empty()) {
+  if (snapid && !old_inodes.empty()) {
     map<snapid_t,old_inode_t>::iterator p = old_inodes.lower_bound(snapid);
     if (p != old_inodes.end()) {
       assert(p->second.first <= snapid && snapid <= p->first);
@@ -1066,7 +1066,7 @@ void CInode::encode_inodestat(bufferlist& bl, snapid_t snapid)
   struct ceph_mds_reply_inode e;
   memset(&e, 0, sizeof(e));
   e.ino = i->ino;
-  e.snapid = snapid;
+  e.snapid = snapid ? (__u64)snapid:CEPH_NOSNAP;  // 0 -> NOSNAP
   e.version = i->version;
   e.layout = i->layout;
   i->ctime.encode_timeval(&e.ctime);

diff --git a/src/mds/MDCache.cc b/src/mds/MDCache.cc
index dddd053cdb4fcd0aec33fe284abbd190832d0e57..ec7669fa7facebc15569dc8b775a52870ab0808e 100644 (file)
--- a/src/mds/MDCache.cc
+++ b/src/mds/MDCache.cc
@@ -1010,8 +1010,8 @@ void MDCache::journal_dirty_inode(EMetaBlob *metablob, CInode *in, snapid_t foll
               << *in << dendl;
     }
   } else {
-    // is dn within current snap?
-    if (follows < dn->first) {
+    // are we within the current snap?
+    if (follows < in->first) {
       metablob->add_primary_dentry(dn, true, in, in->get_projected_inode());
     } else {
       snapid_t oldfirst = dn->first;
@@ -1023,7 +1023,12 @@ void MDCache::journal_dirty_inode(EMetaBlob *metablob, CInode *in, snapid_t foll
       dout(10) << " olddn " << *olddn << dendl;
 
       metablob->add_primary_dentry(olddn, true);
-      metablob->add_primary_dentry(dn, true, in, in->get_projected_inode());
+      inode_t *pi;
+      if (in->is_projected())
+       pi = in->get_previous_projected_inode();  // mkdir/mknod/symlink don't bother to project new inodes
+      else
+       pi = &in->inode;        
+      metablob->add_primary_dentry(dn, true, in, pi);
     }   
   }
 }    
@@ -4196,7 +4201,7 @@ Context *MDCache::_get_waiter(MDRequest *mdr, Message *req)
 int MDCache::path_traverse(MDRequest *mdr, Message *req,     // who
                           filepath& origpath,               // what
                            vector<CDentry*>& trace,          // result
-                          snapid_t *psnapid,
+                          snapid_t *psnapid, CInode **psnapdiri,
                            bool follow_trailing_symlink,     // how
                            int onfail)
 {
@@ -4209,6 +4214,8 @@ int MDCache::path_traverse(MDRequest *mdr, Message *req,     // who
   set< pair<CInode*, string> > symlinks_resolved; 
 
   snapid_t snapid = CEPH_NOSNAP;
+  if (psnapdiri)
+    *psnapdiri = 0;
 
   // root
   CInode *cur = get_inode(origpath.get_ino());
@@ -4235,7 +4242,8 @@ int MDCache::path_traverse(MDRequest *mdr, Message *req,     // who
 
   unsigned depth = 0;
   while (depth < path.depth()) {
-    dout(12) << "traverse: path seg depth " << depth << " = " << path[depth] << dendl;
+    dout(12) << "traverse: path seg depth " << depth << " '" << path[depth]
+            << "' snapid " << snapid << dendl;
     
     if (!cur->is_dir()) {
       dout(7) << "traverse: " << *cur << " not a dir " << dendl;
@@ -4247,6 +4255,8 @@ int MDCache::path_traverse(MDRequest *mdr, Message *req,     // who
       dout(10) << "traverse: snapdir" << dendl;
       snapid = CEPH_SNAPDIR;
       depth++;
+      assert(psnapdiri);
+      *psnapdiri = cur;
       continue;
     }
     if (snapid == CEPH_SNAPDIR) {
@@ -4502,7 +4512,7 @@ int MDCache::path_traverse(MDRequest *mdr, Message *req,     // who
   
   // success.
   if (psnapid)
-    *psnapid = CEPH_NOSNAP;
+    *psnapid = snapid;
   if (mds->logger) mds->logger->inc("thit");
   return 0;
 }
@@ -6252,8 +6262,8 @@ void MDCache::handle_dir_update(MDirUpdate *m)
       dout(5) << "trying discover on dir_update for " << path << dendl;
 
       int r = path_traverse(0, m,
-                           path, trace, NULL, true,
-                            MDS_TRAVERSE_DISCOVER);
+                           path, trace, NULL, NULL,
+                            true, MDS_TRAVERSE_DISCOVER);
       if (r > 0)
         return;
       assert(r == 0);

diff --git a/src/mds/MDCache.h b/src/mds/MDCache.h
index b4485dd5cb343f5bb92487d2c278daf69701bbc4..22d4d30177e88d7cc74134e98f2bafc5bdf3d543 100644 (file)
--- a/src/mds/MDCache.h
+++ b/src/mds/MDCache.h
@@ -225,6 +225,8 @@ struct MDRequest : public Mutation {
 
   vector<CDentry*> trace;  // original path traversal.
   CInode *ref;             // reference inode.  if there is only one, and its path is pinned.
+  CInode *ref_snapdiri;
+  snapid_t ref_snapid;
 
   // -- i am a slave request
   MMDSSlaveRequest *slave_request; // slave request (if one is pending; implies slave == true)
@@ -276,19 +278,19 @@ struct MDRequest : public Mutation {
 
   // ---------------------------------------------------
   MDRequest() : 
-    session(0), client_request(0), ref(0), 
+    session(0), client_request(0), ref(0), ref_snapdiri(0), ref_snapid(CEPH_NOSNAP),
     slave_request(0),
     internal_op(-1),
     _more(0) {}
   MDRequest(metareqid_t ri, MClientRequest *req) : 
     Mutation(ri),
-    session(0), client_request(req), ref(0), 
+    session(0), client_request(req), ref(0), ref_snapdiri(0),
     slave_request(0),
     internal_op(-1),
     _more(0) {}
   MDRequest(metareqid_t ri, int by) : 
     Mutation(ri, by),
-    session(0), client_request(0), ref(0),
+    session(0), client_request(0), ref(0), ref_snapdiri(0),
     slave_request(0),
     internal_op(-1),
     _more(0) {}
@@ -742,7 +744,7 @@ public:
 
   Context *_get_waiter(MDRequest *mdr, Message *req);
   int path_traverse(MDRequest *mdr, Message *req, filepath& path, 
-                   vector<CDentry*>& trace, snapid_t *psnap,
+                   vector<CDentry*>& trace, snapid_t *psnap, CInode **psnapdiri,
                    bool follow_trailing_sym,
                     int onfail);
   bool path_is_mine(filepath& path);

diff --git a/src/mds/Migrator.cc b/src/mds/Migrator.cc
index aae86f38a6ed909ad27863db8653064237b09a58..6cc4e1dc0015fb950252aa5d56eddc0f327b42b5 100644 (file)
--- a/src/mds/Migrator.cc
+++ b/src/mds/Migrator.cc
@@ -986,7 +986,7 @@ int Migrator::encode_export_dir(bufferlist& exportbl,
     dout(7) << "encode_export_dir exporting " << *dn << dendl;
     
     // dn name
-    ::encode(it->first, exportbl);
+    ::encode(dn->name, exportbl);
     
     // state
     dn->encode_export(exportbl);
@@ -1402,7 +1402,7 @@ void Migrator::handle_export_discover(MExportDirDiscover *m)
     // must discover it!
     filepath fpath(m->get_path());
     vector<CDentry*> trace;
-    int r = cache->path_traverse(0, m, fpath, trace, NULL, true, MDS_TRAVERSE_DISCOVER);
+    int r = cache->path_traverse(0, m, fpath, trace, NULL, NULL, true, MDS_TRAVERSE_DISCOVER);
     if (r > 0) return; // wait
     if (r < 0) {
       dout(7) << "handle_export_discover_2 failed to discover or not dir " << m->get_path() << ", NAK" << dendl;

diff --git a/src/mds/Server.cc b/src/mds/Server.cc
index 573471672ad31c6397557da940502e4c1a1a1653..6f65c155b580f5c8f60ab0f7deeedd07b3582828 100644 (file)
--- a/src/mds/Server.cc
+++ b/src/mds/Server.cc
@@ -555,8 +555,12 @@ void Server::reply_request(MDRequest *mdr, MClientReply *reply, CInode *tracei,
 
 
   // infer tracei/tracedn from mdr?
+  snapid_t snapid = CEPH_NOSNAP;
+  CInode *snapdiri = 0;
   if (!tracei && !tracedn && mdr->ref) {
     tracei = mdr->ref;
+    snapdiri = mdr->ref_snapdiri;
+    snapid = mdr->ref_snapid;
     dout(20) << "inferring tracei to be " << *tracei << dendl;
     if (!mdr->trace.empty()) {
       tracedn = mdr->trace.back();
@@ -578,7 +582,7 @@ void Server::reply_request(MDRequest *mdr, MClientReply *reply, CInode *tracei,
   } else {
     // send reply, with trace, and possible leases
     if (tracei || tracedn)
-      set_trace_dist(session, reply, tracei, tracedn);
+      set_trace_dist(session, reply, tracei, tracedn, snapid, snapdiri);
     messenger->send_message(reply, client_inst);
   }
   
@@ -595,13 +599,15 @@ void Server::reply_request(MDRequest *mdr, MClientReply *reply, CInode *tracei,
  *
  * trace is in reverse order (i.e. root inode comes last)
  */
-void Server::set_trace_dist(Session *session, MClientReply *reply, CInode *in, CDentry *dn)
+void Server::set_trace_dist(Session *session, MClientReply *reply, CInode *in, CDentry *dn,
+                           snapid_t snapid, CInode *snapdiri)
 {
   // inode, dentry, dir, ..., inode
   bufferlist bl;
   int whoami = mds->get_nodeid();
   int client = session->get_client();
   __u16 numi = 0, numdn = 0;
+  __s16 snapdirpos = -1;
 
   // choose lease duration
   utime_t now = g_clock.now();
@@ -615,10 +621,19 @@ void Server::set_trace_dist(Session *session, MClientReply *reply, CInode *in, C
   }
 
  inode:
-  in->encode_inodestat(bl);
-  lmask = mds->locker->issue_client_lease(in, client, bl, now, session);
   numi++;
-  dout(20) << " trace added " << lmask << " " << *in << dendl;
+  in->encode_inodestat(bl, snapid);
+  lmask = mds->locker->issue_client_lease(in, client, bl, now, session);
+  dout(20) << " trace added " << lmask << " snapid " << snapid << " " << *in << dendl;
+
+  if (snapid != CEPH_NOSNAP && in == snapdiri) {
+    snapid = CEPH_NOSNAP;
+    snapdirpos = numi;
+    dout(10) << " snapdiri at pos " << snapdirpos << dendl;
+    in->encode_inodestat(bl, snapid);
+    lmask = mds->locker->issue_client_lease(in, client, bl, now, session);
+    dout(20) << " trace added " << lmask << " snapid " << snapid << " " << *in << dendl;
+  }
 
   if (!dn)
     dn = in->get_parent_dn();
@@ -629,7 +644,7 @@ void Server::set_trace_dist(Session *session, MClientReply *reply, CInode *in, C
   ::encode(dn->get_name(), bl);
   lmask = mds->locker->issue_client_lease(dn, client, bl, now, session);
   numdn++;
-  dout(20) << " trace added " << lmask << " " << *dn << dendl;
+  dout(20) << " trace added " << lmask << " snapid " << snapid << " " << *dn << dendl;
   
   // dir
 #ifdef MDS_VERIFY_FRAGSTAT
@@ -649,6 +664,7 @@ done:
   bufferlist fbl;
   ::encode(numi, fbl);
   ::encode(numdn, fbl);
+  ::encode(snapdirpos, fbl);
   fbl.claim_append(bl);
   reply->set_trace(fbl);
 }
@@ -1249,8 +1265,8 @@ CDir *Server::traverse_to_auth_dir(MDRequest *mdr, vector<CDentry*> &trace, file
   // traverse to parent dir
   snapid_t snapid;
   int r = mdcache->path_traverse(mdr, mdr->client_request,
-                                refpath, trace, &snapid, false,
-                                MDS_TRAVERSE_FORWARD);
+                                refpath, trace, &snapid, &mdr->ref_snapdiri,
+                                false, MDS_TRAVERSE_FORWARD);
   if (r > 0) return 0; // delayed
   if (r < 0) {
     reply_request(mdr, r);
@@ -1277,7 +1293,7 @@ CDir *Server::traverse_to_auth_dir(MDRequest *mdr, vector<CDentry*> &trace, file
 
 
 
-CInode* Server::rdlock_path_pin_ref(MDRequest *mdr, snapid_t *psnapid,
+CInode* Server::rdlock_path_pin_ref(MDRequest *mdr,
                                    bool want_auth, bool rdlock_dft)
 {
   dout(10) << "rdlock_path_pin_ref " << *mdr << dendl;
@@ -1299,8 +1315,8 @@ CInode* Server::rdlock_path_pin_ref(MDRequest *mdr, snapid_t *psnapid,
   vector<CDentry*> trace;
   int r = mdcache->path_traverse(mdr, req,
                                 refpath, 
-                                trace, psnapid, req->follow_trailing_symlink(),
-                                MDS_TRAVERSE_FORWARD);
+                                trace, &mdr->ref_snapid, &mdr->ref_snapdiri,
+                                req->follow_trailing_symlink(), MDS_TRAVERSE_FORWARD);
   if (r > 0) return false; // delayed
   if (r < 0) {  // error
     reply_request(mdr, r);
@@ -1844,8 +1860,7 @@ void Server::handle_client_readdir(MDRequest *mdr)
 {
   MClientRequest *req = mdr->client_request;
   int client = req->get_orig_source().num();
-  snapid_t snapid;
-  CInode *diri = rdlock_path_pin_ref(mdr, &snapid, false, true);  // rdlock dirfragtreelock!
+  CInode *diri = rdlock_path_pin_ref(mdr, false, true);  // rdlock dirfragtreelock!
   if (!diri) return;
 
   // it's a directory, right?
@@ -1893,6 +1908,9 @@ void Server::handle_client_readdir(MDRequest *mdr)
 
   mdr->now = g_clock.real_now();
 
+  snapid_t snapid = mdr->ref_snapid;
+  dout(10) << "snapid " << snapid << dendl;
+
   // build dir contents
   bufferlist dirbl, dnbl;
   dir->encode_dirstat(dirbl, mds->get_nodeid());
@@ -1930,13 +1948,13 @@ void Server::handle_client_readdir(MDRequest *mdr)
     }
     assert(in);
 
-    dout(12) << "including inode " << *in << dendl;
-    
     // dentry
+    dout(12) << "including    dn " << *dn << dendl;
     ::encode(dn->name, dnbl);
     mds->locker->issue_client_lease(dn, client, dnbl, mdr->now, mdr->session);
 
     // inode
+    dout(12) << "including inode " << *in << dendl;
     in->encode_inodestat(dnbl);
     mds->locker->issue_client_lease(in, client, dnbl, mdr->now, mdr->session);
     numfiles++;
@@ -1972,9 +1990,10 @@ class C_MDS_mknod_finish : public Context {
   MDRequest *mdr;
   CDentry *dn;
   CInode *newi;
+  snapid_t follows;
 public:
-  C_MDS_mknod_finish(MDS *m, MDRequest *r, CDentry *d, CInode *ni) :
-    mds(m), mdr(r), dn(d), newi(ni) {}
+  C_MDS_mknod_finish(MDS *m, MDRequest *r, CDentry *d, CInode *ni, snapid_t f) :
+    mds(m), mdr(r), dn(d), newi(ni), follows(f) {}
   void finish(int r) {
     assert(r == 0);
 
@@ -2012,7 +2031,9 @@ void Server::handle_client_mknod(MDRequest *mdr)
   CDentry *dn = rdlock_path_xlock_dentry(mdr, false, false);
   if (!dn) return;
 
+  snapid_t follows = dn->dir->inode->find_snaprealm()->get_latest_snap();
   mdr->now = g_clock.real_now();
+
   CInode *newi = prepare_new_inode(mdr, dn->dir);
   assert(newi);
 
@@ -2025,7 +2046,8 @@ void Server::handle_client_mknod(MDRequest *mdr)
   newi->inode.dirstat.rfiles = 1;
 
   newi->projected_parent = dn;
-  
+  dn->first = newi->first = follows+1;
+    
   dout(10) << "mknod mode " << newi->inode.mode << " rdev " << newi->inode.rdev << dendl;
 
   // prepare finisher
@@ -2033,12 +2055,12 @@ void Server::handle_client_mknod(MDRequest *mdr)
   EUpdate *le = new EUpdate(mdlog, "mknod");
   le->metablob.add_client_req(req->get_reqid());
   le->metablob.add_allocated_ino(newi->ino(), mds->idalloc->get_version());
-
+  
   mds->locker->predirty_nested(mdr, &le->metablob, newi, dn->dir, PREDIRTY_PRIMARY|PREDIRTY_DIR, 1);
-  mdcache->journal_dirty_inode(&le->metablob, newi, dn->dir->inode->find_snaprealm()->get_latest_snap());
+  mdcache->journal_dirty_inode(&le->metablob, newi, follows);
   
   // log + wait
-  mdlog->submit_entry(le, new C_MDS_mknod_finish(mds, mdr, dn, newi));
+  mdlog->submit_entry(le, new C_MDS_mknod_finish(mds, mdr, dn, newi, follows));
 }
 
 
@@ -2053,7 +2075,9 @@ void Server::handle_client_mkdir(MDRequest *mdr)
   if (!dn) return;
 
   // new inode
+  snapid_t follows = dn->dir->inode->find_snaprealm()->get_latest_snap();
   mdr->now = g_clock.real_now();
+
   CInode *newi = prepare_new_inode(mdr, dn->dir);  
   assert(newi);
 
@@ -2066,6 +2090,8 @@ void Server::handle_client_mkdir(MDRequest *mdr)
   newi->inode.version = dn->pre_dirty() - 1;
   newi->inode.dirstat.rsubdirs = 1;
 
+  dn->first = newi->first = follows+1;
+
   // ...and that new dir is empty.
   CDir *newdir = newi->get_or_open_dirfrag(mds->mdcache, frag_t());
   newdir->mark_complete();
@@ -2080,11 +2106,11 @@ void Server::handle_client_mkdir(MDRequest *mdr)
   le->metablob.add_allocated_ino(newi->ino(), mds->idalloc->get_version());
   mds->locker->predirty_nested(mdr, &le->metablob, newi, dn->dir, PREDIRTY_PRIMARY|PREDIRTY_DIR, 1);
   //le->metablob.add_primary_dentry(dn, true, newi, &newi->inode);
-  mdcache->journal_dirty_inode(&le->metablob, newi, dn->dir->inode->find_snaprealm()->get_latest_snap());
+  mdcache->journal_dirty_inode(&le->metablob, newi, follows);
   le->metablob.add_dir(newdir, true, true); // dirty AND complete
   
   // log + wait
-  mdlog->submit_entry(le, new C_MDS_mknod_finish(mds, mdr, dn, newi));
+  mdlog->submit_entry(le, new C_MDS_mknod_finish(mds, mdr, dn, newi, follows));
 }
 
 
@@ -2098,6 +2124,7 @@ void Server::handle_client_symlink(MDRequest *mdr)
   if (!dn) return;
 
   mdr->now = g_clock.real_now();
+  snapid_t follows = dn->dir->inode->find_snaprealm()->get_latest_snap();
 
   CInode *newi = prepare_new_inode(mdr, dn->dir);
   assert(newi);
@@ -2112,16 +2139,18 @@ void Server::handle_client_symlink(MDRequest *mdr)
   newi->inode.version = dn->pre_dirty() - 1;
   newi->inode.dirstat.rfiles = 1;
 
+  dn->first = newi->first = follows+1;
+
   // prepare finisher
   mdr->ls = mdlog->get_current_segment();
   EUpdate *le = new EUpdate(mdlog, "symlink");
   le->metablob.add_client_req(req->get_reqid());
   le->metablob.add_allocated_ino(newi->ino(), mds->idalloc->get_version());
   mds->locker->predirty_nested(mdr, &le->metablob, newi, dn->dir, PREDIRTY_PRIMARY|PREDIRTY_DIR, 1);
-  mdcache->journal_dirty_inode(&le->metablob, newi, dn->dir->inode->find_snaprealm()->get_latest_snap());
+  mdcache->journal_dirty_inode(&le->metablob, newi, follows);
 
   // log + wait
-  mdlog->submit_entry(le, new C_MDS_mknod_finish(mds, mdr, dn, newi));
+  mdlog->submit_entry(le, new C_MDS_mknod_finish(mds, mdr, dn, newi, follows));
 }
 
 
@@ -2151,8 +2180,8 @@ void Server::handle_client_link(MDRequest *mdr)
   dout(7) << "handle_client_link discovering target " << targetpath << dendl;
   vector<CDentry*> targettrace;
   int r = mdcache->path_traverse(mdr, req,
-                                targetpath, targettrace, NULL, false,
-                                MDS_TRAVERSE_DISCOVER);
+                                targetpath, targettrace, NULL, NULL,
+                                false, MDS_TRAVERSE_DISCOVER);
   if (r > 0) return; // wait
   if (targettrace.empty()) r = -EINVAL; 
   if (r < 0) {
@@ -2690,8 +2719,8 @@ void Server::handle_client_unlink(MDRequest *mdr)
   // traverse to path
   vector<CDentry*> trace;
   int r = mdcache->path_traverse(mdr, req, 
-                                req->get_filepath(), trace, NULL, false,
-                                MDS_TRAVERSE_FORWARD);
+                                req->get_filepath(), trace, NULL, NULL,
+                                false, MDS_TRAVERSE_FORWARD);
   if (r > 0) return;
   if (trace.empty()) r = -EINVAL;   // can't unlink root
   if (r < 0) {
@@ -3039,8 +3068,8 @@ void Server::handle_client_rename(MDRequest *mdr)
   // traverse to src
   vector<CDentry*> srctrace;
   int r = mdcache->path_traverse(mdr, req,
-                                srcpath, srctrace, NULL, false,
-                                MDS_TRAVERSE_DISCOVER);
+                                srcpath, srctrace, NULL, NULL,
+                                false, MDS_TRAVERSE_DISCOVER);
   if (r > 0) return;
   if (r < 0) {
     reply_request(mdr, r);
@@ -3743,8 +3772,8 @@ void Server::handle_slave_rename_prep(MDRequest *mdr)
   dout(10) << " dest " << destpath << dendl;
   vector<CDentry*> trace;
   int r = mdcache->path_traverse(mdr, mdr->slave_request, 
-                                destpath, trace, NULL, false,
-                                MDS_TRAVERSE_DISCOVERXLOCK);
+                                destpath, trace, NULL, NULL,
+                                false, MDS_TRAVERSE_DISCOVERXLOCK);
   if (r > 0) return;
   assert(r == 0);  // we shouldn't get an error here!
       
@@ -3756,8 +3785,8 @@ void Server::handle_slave_rename_prep(MDRequest *mdr)
   filepath srcpath(mdr->slave_request->srcdnpath);
   dout(10) << " src " << srcpath << dendl;
   r = mdcache->path_traverse(mdr, mdr->slave_request,
-                            srcpath, trace, NULL, false,  
-                            MDS_TRAVERSE_DISCOVERXLOCK);
+                            srcpath, trace, NULL, NULL,
+                            false, MDS_TRAVERSE_DISCOVERXLOCK);
   if (r > 0) return;
   assert(r == 0);
       
@@ -4567,9 +4596,10 @@ class C_MDS_openc_finish : public Context {
   MDRequest *mdr;
   CDentry *dn;
   CInode *newi;
+  snapid_t follows;
 public:
-  C_MDS_openc_finish(MDS *m, MDRequest *r, CDentry *d, CInode *ni) :
-    mds(m), mdr(r), dn(d), newi(ni) {}
+  C_MDS_openc_finish(MDS *m, MDRequest *r, CDentry *d, CInode *ni, snapid_t f) :
+    mds(m), mdr(r), dn(d), newi(ni), follows(f) {}
   void finish(int r) {
     assert(r == 0);
 
@@ -4624,6 +4654,8 @@ void Server::handle_client_openc(MDRequest *mdr)
     
   // create inode.
   mdr->now = g_clock.real_now();
+  snapid_t follows = dn->dir->inode->find_snaprealm()->get_latest_snap();
+
   CInode *in = prepare_new_inode(mdr, dn->dir);
   assert(in);
   
@@ -4634,6 +4666,9 @@ void Server::handle_client_openc(MDRequest *mdr)
   in->inode.version = dn->pre_dirty() - 1;
   in->inode.max_size = in->get_layout_size_increment();
   in->inode.dirstat.rfiles = 1;
+
+  in->projected_parent = dn;
+  dn->first = in->first = follows+1;
   
   // prepare finisher
   mdr->ls = mdlog->get_current_segment();
@@ -4641,10 +4676,10 @@ void Server::handle_client_openc(MDRequest *mdr)
   le->metablob.add_client_req(req->get_reqid());
   le->metablob.add_allocated_ino(in->ino(), mds->idalloc->get_version());
   mds->locker->predirty_nested(mdr, &le->metablob, in, dn->dir, PREDIRTY_PRIMARY|PREDIRTY_DIR, 1);
-  le->metablob.add_primary_dentry(dn, true, in, &in->inode);
+  mdcache->journal_dirty_inode(&le->metablob, in, follows);
   
   // log + wait
-  C_MDS_openc_finish *fin = new C_MDS_openc_finish(mds, mdr, dn, in);
+  C_MDS_openc_finish *fin = new C_MDS_openc_finish(mds, mdr, dn, in, follows);
   mdlog->submit_entry(le, fin);
 }
 
@@ -4680,8 +4715,8 @@ void Server::handle_client_lssnap(MDRequest *mdr)
   // traverse to path
   vector<CDentry*> trace;
   int r = mdcache->path_traverse(mdr, req, 
-                                req->get_filepath(), trace, NULL, false,
-                                MDS_TRAVERSE_FORWARD);
+                                req->get_filepath(), trace, NULL, NULL,
+                                false, MDS_TRAVERSE_FORWARD);
   if (r > 0) return;
   if (trace.empty()) r = -EINVAL;   // can't snap root
   if (r < 0) {
@@ -4766,8 +4801,8 @@ void Server::handle_client_mksnap(MDRequest *mdr)
   // traverse to path
   vector<CDentry*> trace;
   int r = mdcache->path_traverse(mdr, req, 
-                                req->get_filepath(), trace, NULL, false,
-                                MDS_TRAVERSE_FORWARD);
+                                req->get_filepath(), trace, NULL, NULL,
+                                false, MDS_TRAVERSE_FORWARD);
   if (r > 0) return;
   if (trace.empty()) r = -EINVAL;   // can't snap root
   if (r < 0) {

diff --git a/src/mds/Server.h b/src/mds/Server.h
index cd3e93b4be999334d31f0f04ac950f267be64a24..c62027999630b3cb84d65370e42339affcdae297 100644 (file)
--- a/src/mds/Server.h
+++ b/src/mds/Server.h
@@ -84,7 +84,8 @@ public:
   void dispatch_client_request(MDRequest *mdr);
   void reply_request(MDRequest *mdr, int r = 0, CInode *tracei = 0, CDentry *tracedn = 0);
   void reply_request(MDRequest *mdr, MClientReply *reply, CInode *tracei = 0, CDentry *tracedn = 0);
-  void set_trace_dist(Session *session, MClientReply *reply, CInode *in, CDentry *dn);
+  void set_trace_dist(Session *session, MClientReply *reply, CInode *in, CDentry *dn,
+                     snapid_t snapid, CInode *snapdiri);
 
 
   void handle_slave_request(MMDSSlaveRequest *m);
@@ -98,7 +99,7 @@ public:
   CDentry *prepare_null_dentry(MDRequest *mdr, CDir *dir, const string& dname, bool okexist=false);
   CInode* prepare_new_inode(MDRequest *mdr, CDir *dir);
 
-  CInode* rdlock_path_pin_ref(MDRequest *mdr, snapid_t *psnapid, bool want_auth, bool rdlock_dft=false);
+  CInode* rdlock_path_pin_ref(MDRequest *mdr, bool want_auth, bool rdlock_dft=false);
   CDentry* rdlock_path_xlock_dentry(MDRequest *mdr, bool okexist, bool mustexist);
 
   CDir* try_open_auth_dirfrag(CInode *diri, frag_t fg, MDRequest *mdr);

diff --git a/src/mds/mdstypes.h b/src/mds/mdstypes.h
index 8504f3b86b24e9f9cb295e71b029f10ff6d0530c..43da6c2fc828f07511790eab2f44e17515d5e743 100644 (file)
--- a/src/mds/mdstypes.h
+++ b/src/mds/mdstypes.h
@@ -321,18 +321,27 @@ WRITE_CLASS_ENCODER(fnode_t)
 // =======
 // dentries
 
-typedef pair<snapid_t, const char *> dentry_key_t;
+struct dentry_key_t {
+  snapid_t snapid;
+  const char *name;
+  dentry_key_t() : snapid(0), name(0) {}
+  dentry_key_t(snapid_t s, const char *n) : snapid(s), name(n) {}
+};
 
-struct ltdentrykey
+inline ostream& operator<<(ostream& out, const dentry_key_t &k)
 {
-  bool operator()(const dentry_key_t& k1,
-                 const dentry_key_t& k2) const
-  {
-    return 
-      k1.first < k2.first ||
-      (k1.first == k2.first && strcmp(k1.second, k2.second) < 0);
-  }
-};
+  return out << "(" << k.name << "," << k.snapid << ")";
+}
+
+inline bool operator<(const dentry_key_t& k1, const dentry_key_t& k2)
+{
+  /*
+   * order by name, then snap
+   */
+  int c = strcmp(k1.name, k2.name);
+  return 
+    c < 0 || (c == 0 && k1.snapid < k2.snapid);
+}