mds: remote link/unlink behaving well enough. various bugfixes.

diff --git a/src/client/SyntheticClient.cc b/src/client/SyntheticClient.cc
index 51cccb4f4d10f9d06cc8063e22d27308056acca2..1761b306ea0e4143e5462739a265d95facf396ca 100644 (file)
--- a/src/client/SyntheticClient.cc
+++ b/src/client/SyntheticClient.cc
@@ -2641,6 +2641,28 @@ void SyntheticClient::make_dir_mess(const char *basedir, int n)
 
 void SyntheticClient::foo()
 {
+  if (1) {
+    // make 2 parallel dirs, link/unlink between them.
+    char a[100], b[100];
+    client->mkdir("/a", 0755);
+    client->mkdir("/b", 0755);
+    for (int i=0; i<10; i++) {
+      sprintf(a, "/a/%d", i);
+      client->mknod(a, 0644);
+    }
+    while (1) {
+      for (int i=0; i<10; i++) {
+       sprintf(a, "/a/%d", i);
+       sprintf(b, "/b/%d", i);
+       client->link(a, b);
+      }
+      for (int i=0; i<10; i++) {
+       sprintf(b, "/b/%d", i);
+       client->unlink(b);
+      }
+    }
+    return;
+  }
   if (1) {
     // bug1.cpp
     const char *fn = "blah";

diff --git a/src/mds/CDir.cc b/src/mds/CDir.cc
index cb9842f6708d0c457d4ee9c176daec6fee56506e..631b4848c7456578624946395f35cd3fdb8c98d3 100644 (file)
--- a/src/mds/CDir.cc
+++ b/src/mds/CDir.cc
@@ -1367,6 +1367,7 @@ void CDir::finish_export(utime_t now)
 void CDir::decode_import(bufferlist::iterator& blp)
 {
   ::decode(fnode, blp);
+  projected_version = fnode.version;
   ::decode(committed_version, blp);
   ::decode(committed_version_equivalent, blp);
   committing_version = committed_version;

diff --git a/src/mds/CDir.h b/src/mds/CDir.h
index 3eac0af84331c6efc734ec39f62193705c03673b..386cf50ae4f0aee15137a8a7070b3189bc0a91a2 100644 (file)
--- a/src/mds/CDir.h
+++ b/src/mds/CDir.h
@@ -183,8 +183,10 @@ public:
   version_t pre_dirty(version_t min=0);
   void _mark_dirty(LogSegment *ls);
   void _set_dirty_flag() {
-    state_set(STATE_DIRTY);
-    get(PIN_DIRTY);
+    if (!state_test(STATE_DIRTY)) {
+      state_set(STATE_DIRTY);
+      get(PIN_DIRTY);
+    }
   }
   void mark_dirty(version_t pv, LogSegment *ls);
   void mark_clean();

diff --git a/src/mds/CInode.cc b/src/mds/CInode.cc
index 8122f0634ae18ce65a7164dcb57de261521fefe2..453b8059cb8c5369f87cb5466d109a9efe04673a 100644 (file)
--- a/src/mds/CInode.cc
+++ b/src/mds/CInode.cc
@@ -622,7 +622,8 @@ void CInode::decode_lock_state(int type, bufferlist& bl)
                     << *dir << dendl;
            fnode_t *pf = dir->get_projected_fnode();
            pf->accounted_fragstat = fragstat;
-           dir->_set_dirty_flag();         // bit of a hack
+           if (dir->is_auth())
+             dir->_set_dirty_flag();       // bit of a hack
          }
        }
       }

diff --git a/src/mds/Locker.cc b/src/mds/Locker.cc
index dbb035c93bebbf041f75f38976f7485cfced2a7d..ed2cebdf5c0cd1db8c4ef06e62dcde23c014c2b7 100644 (file)
--- a/src/mds/Locker.cc
+++ b/src/mds/Locker.cc
@@ -1215,9 +1215,12 @@ void Locker::revoke_client_leases(SimpleLock *lock)
 
 void Locker::predirty_nested(Mutation *mut, EMetaBlob *blob,
                             CInode *in, CDir *parent,
-                            bool primary_dn, bool do_parent, int linkunlink,
+                            int flags, int linkunlink,
                             EMetaBlob *rollback)
 {
+  bool primary_dn = flags & PREDIRTY_PRIMARY;
+  bool do_parent = flags & PREDIRTY_DIR;
+
   dout(10) << "predirty_nested"
           << (do_parent ? " do_parent_mtime":"")
           << " linkunlink=" <<  linkunlink
@@ -1238,7 +1241,7 @@ void Locker::predirty_nested(Mutation *mut, EMetaBlob *blob,
   list<CInode*> lsi;
   CInode *cur = in;
   while (parent) {
-    assert(cur->is_auth());
+    assert(cur->is_auth() || !primary_dn);
     assert(parent->is_auth());
     
     // opportunistically adjust parent dirfrag
@@ -1461,6 +1464,9 @@ void Locker::handle_simple_lock(SimpleLock *lock, MLock *m)
 {
   int from = m->get_asker();
   
+  dout(10) << "handle_simple_lock " << *m
+          << " on " << *lock << " " << *lock->get_parent() << dendl;
+
   if (mds->is_rejoin()) {
     if (lock->get_parent()->is_rejoining()) {
       dout(7) << "handle_simple_lock still rejoining " << *lock->get_parent()
@@ -1497,9 +1503,11 @@ void Locker::handle_simple_lock(SimpleLock *lock, MLock *m)
     //||           lock->get_state() == LOCK_GLOCKR);
     
     // wait for readers to finish?
-    if (lock->is_rdlocked()) {
-      dout(7) << "handle_simple_lock has reader, waiting before ack on " << *lock
+    if (lock->is_rdlocked() ||
+       lock->get_num_client_lease()) {
+      dout(7) << "handle_simple_lock has reader|leases, waiting before ack on " << *lock
              << " on " << *lock->get_parent() << dendl;
+      revoke_client_leases(lock);
       lock->set_state(LOCK_GLOCKR);
     } else {
       // update lock and reply

diff --git a/src/mds/Locker.h b/src/mds/Locker.h
index 42772b979e41ee8f19f05f703780bd7a3a08c2b5..c6a9f2bb0d3d780d67d5ff9f738eda4c3c5d8cad 100644 (file)
--- a/src/mds/Locker.h
+++ b/src/mds/Locker.h
@@ -54,6 +54,9 @@ class ScatterLock;
 class LocalLock;
 class MDCache;
 
+static const int PREDIRTY_PRIMARY = 1;    // primary dn, adjust nested accounting
+static const int PREDIRTY_DIR = 2;        // update parent dir mtime/size
+
 class Locker {
 private:
   MDS *mds;
@@ -159,7 +162,7 @@ protected:
 
 public:
   void predirty_nested(Mutation *mut, EMetaBlob *blob, CInode *in, CDir *dir,
-                      bool do_nested, bool do_parent, int linkunlink=0,
+                      int flags, int linkunlink=0,
                       EMetaBlob *rollback=0);
 
   // local

diff --git a/src/mds/MDCache.cc b/src/mds/MDCache.cc
index 1cd08e9d6aeed75959d970b162b3ae761d56371e..380453c647ac7e98c02087569b262d73e60dff43 100644 (file)
--- a/src/mds/MDCache.cc
+++ b/src/mds/MDCache.cc
@@ -4500,8 +4500,14 @@ void MDCache::open_remote_ino_2(inodeno_t ino,
     return;
   }
 
-  if (!dir && in->is_auth())
+  if (!dir && in->is_auth()) {
+    if (dir->is_frozen_dir()) {
+      dout(7) << "traverse: " << *dir << " is frozen_dir, waiting" << dendl;
+      dir->add_waiter(CDir::WAIT_UNFREEZE, _get_waiter(mdr, 0));
+      return;
+    }
     dir = in->get_or_open_dirfrag(this, frag);
+  }
   
   assert(dir);
   if (dir->is_auth()) {
@@ -4551,7 +4557,6 @@ MDRequest *MDCache::request_start(MClientRequest *req)
     if (mdr->is_slave()) {
       dout(10) << "request_start already had " << *mdr << ", cleaning up" << dendl;
       request_cleanup(mdr);
-      delete mdr;
     } else {
       dout(10) << "request_start already processing " << *mdr << ", dropping new msg" << dendl;
       delete req;
@@ -5215,6 +5220,12 @@ void MDCache::discover_ino(CDir *base,
          << (want_xlocked ? " want_xlocked":"")
          << dendl;
   
+  if (base->is_ambiguous_auth()) {
+    dout(10) << " waiting for single auth on " << *base << dendl;
+    base->add_waiter(CDir::WAIT_SINGLEAUTH, onfinish);
+    return;
+  } 
+
   if (!base->is_waiting_for_ino(want_ino)) {
     MDiscover *dis = new MDiscover(mds->get_nodeid(),
                                   base->dirfrag(),

diff --git a/src/mds/Server.cc b/src/mds/Server.cc
index b6c07101274011a21046068e32e4593464960c88..854756ecec299670ae6376a5878ee1009e008df7 100644 (file)
--- a/src/mds/Server.cc
+++ b/src/mds/Server.cc
@@ -2048,7 +2048,7 @@ void Server::handle_client_mknod(MDRequest *mdr)
   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, true, true, 1);
+  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);
   
@@ -2093,7 +2093,7 @@ void Server::handle_client_mkdir(MDRequest *mdr)
   EUpdate *le = new EUpdate(mdlog, "mkdir");
   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, true, true, 1);
+  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);
   le->metablob.add_dir(newdir, true, true); // dirty AND complete
   
@@ -2145,7 +2145,7 @@ void Server::handle_client_symlink(MDRequest *mdr)
   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, true, true, 1);
+  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);
 
   // log + wait
@@ -2291,8 +2291,8 @@ void Server::_link_local(MDRequest *mdr, CDentry *dn, CInode *targeti)
   // log + wait
   EUpdate *le = new EUpdate(mdlog, "link_local");
   le->metablob.add_client_req(mdr->reqid);
-  mds->locker->predirty_nested(mdr, &le->metablob, targeti, dn->dir, false, true, 1); // new dn
-  mds->locker->predirty_nested(mdr, &le->metablob, targeti, 0, true, true);           // targeti
+  mds->locker->predirty_nested(mdr, &le->metablob, targeti, dn->dir, PREDIRTY_DIR, 1); // new dn
+  mds->locker->predirty_nested(mdr, &le->metablob, targeti, 0, PREDIRTY_PRIMARY);           // targeti
   le->metablob.add_remote_dentry(dn, true, targeti->ino(), 
                                 MODE_TO_DT(targeti->inode.mode));  // new remote
   le->metablob.add_primary_dentry(targeti->parent, true, targeti, pi);  // update old primary
@@ -2370,7 +2370,7 @@ void Server::_link_remote(MDRequest *mdr, CDentry *dn, CInode *targeti)
   mdr->ls = mdlog->get_current_segment();
   EUpdate *le = new EUpdate(mdlog, "link_remote");
   le->metablob.add_client_req(mdr->reqid);
-  mds->locker->predirty_nested(mdr, &le->metablob, targeti, dn->dir, false, true, 1);
+  mds->locker->predirty_nested(mdr, &le->metablob, targeti, dn->dir, PREDIRTY_DIR, 1);
   le->metablob.add_remote_dentry(dn, true, targeti->ino(), 
                                 MODE_TO_DT(targeti->inode.mode));  // new remote
 
@@ -2433,6 +2433,8 @@ void Server::handle_slave_link_prep(MDRequest *mdr)
 
   mdr->now = mdr->slave_request->now;
 
+  mdr->auth_pin(targeti);
+
   // anchor?
   if (mdr->slave_request->get_op() == MMDSSlaveRequest::OP_LINKPREP) {
     if (targeti->is_anchored() && !targeti->is_unanchoring()) {
@@ -2469,7 +2471,7 @@ void Server::handle_slave_link_prep(MDRequest *mdr)
   dout(10) << " projected inode " << pi << " v " << pi->version << dendl;
 
   // commit case
-  mds->locker->predirty_nested(mdr, &le->commit, dn->inode, 0, true, false, 0, &le->rollback);
+  mds->locker->predirty_nested(mdr, &le->commit, dn->inode, 0, PREDIRTY_PRIMARY, 0, &le->rollback);
   le->commit.add_primary_dentry(dn, true, targeti, pi);  // update old primary
   le->rollback.add_primary_dentry(dn, true, targeti, oldi);
 
@@ -2502,6 +2504,7 @@ void Server::_logged_slave_link(MDRequest *mdr, CInode *targeti, utime_t old_cti
 
   // update the target
   targeti->pop_and_dirty_projected_inode(mdr->ls);
+  mdr->apply();
 
   // hit pop
   mds->balancer->hit_inode(mdr->now, targeti, META_POP_IWR);
@@ -2738,18 +2741,17 @@ void Server::_unlink_local(MDRequest *mdr, CDentry *dn, CDentry *straydn)
   pi->nlink--;
   pi->ctime = mdr->now;
 
-  mds->locker->predirty_nested(mdr, &le->metablob, dn->inode, dn->dir,
-                              dn->is_primary(), true, -1);
-  
   if (dn->is_primary()) {
     // primary link.  add stray dentry.
     assert(straydn);
-    mds->locker->predirty_nested(mdr, &le->metablob, dn->inode, straydn->dir, true, true, 1);
+    mds->locker->predirty_nested(mdr, &le->metablob, dn->inode, dn->dir, PREDIRTY_PRIMARY|PREDIRTY_DIR, -1);
+    mds->locker->predirty_nested(mdr, &le->metablob, dn->inode, straydn->dir, PREDIRTY_PRIMARY|PREDIRTY_DIR, 1);
     //le->metablob.add_dir_context(straydn->dir);
     le->metablob.add_primary_dentry(straydn, true, dn->inode, pi);
   } else {
     // remote link.  update remote inode.
-    mds->locker->predirty_nested(mdr, &le->metablob, dn->inode, 0, true, true);
+    mds->locker->predirty_nested(mdr, &le->metablob, dn->inode, dn->dir, PREDIRTY_DIR, -1);
+    mds->locker->predirty_nested(mdr, &le->metablob, dn->inode, 0, PREDIRTY_PRIMARY);
     le->metablob.add_primary_dentry(dn->inode->parent, true, dn->inode);
   }
 
@@ -2859,7 +2861,7 @@ void Server::_unlink_remote(MDRequest *mdr, CDentry *dn)
   le->metablob.add_client_req(mdr->reqid);
 
   // the unlinked dentry
-  mds->locker->predirty_nested(mdr, &le->metablob, dn->inode, dn->dir, false, true, -1);
+  mds->locker->predirty_nested(mdr, &le->metablob, dn->inode, dn->dir, PREDIRTY_DIR, -1);
   dn->pre_dirty();
   le->metablob.add_null_dentry(dn, true);
 
@@ -3456,19 +3458,18 @@ void Server::_rename_prepare(MDRequest *mdr,
   if (mdr->is_master()) {
     // sub off target
     if (!linkmerge && destdn->is_primary())
-      mds->locker->predirty_nested(mdr, metablob, destdn->inode, destdn->dir,
-                                  true, true, -1);
+      mds->locker->predirty_nested(mdr, metablob, destdn->inode, destdn->dir, PREDIRTY_PRIMARY|PREDIRTY_DIR, -1);
     if (destdn->dir == srcdn->dir) {
       // same dir.  don't update nested info or adjust counts.
       mds->locker->predirty_nested(mdr, metablob, srcdn->inode, srcdn->dir,
                                   false, true);
     } else {
       // different dir.  update nested accounting.
+      int flags = srcdn->is_primary() ? PREDIRTY_PRIMARY:0;
+      flags |= PREDIRTY_DIR;
       if (srcdn->is_auth())
-       mds->locker->predirty_nested(mdr, metablob, srcdn->inode, srcdn->dir,
-                                    srcdn->is_primary(), true, -1);
-      mds->locker->predirty_nested(mdr, metablob, srcdn->inode, destdn->dir,
-                                  srcdn->is_primary(), true, 1);
+       mds->locker->predirty_nested(mdr, metablob, srcdn->inode, srcdn->dir, flags, -1);
+      mds->locker->predirty_nested(mdr, metablob, srcdn->inode, destdn->dir, flags, 1);
     }
   }
 
@@ -4345,7 +4346,7 @@ void Server::handle_client_openc(MDRequest *mdr)
   EUpdate *le = new EUpdate(mdlog, "openc");
   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, true, true, 1);
+  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);
   
   // log + wait

diff --git a/src/vstartnew.sh b/src/vstartnew.sh
index 793f57eedecfc483ab30ddb62678eea42a209958..208625094d2adac7a514e4a31081a7263870ca25 100755 (executable)
--- a/src/vstartnew.sh
+++ b/src/vstartnew.sh
@@ -48,7 +48,9 @@ do
 done
 
 # mds
-$CEPH_BIN/cmds $ARGS --debug_ms 1 --debug_mds 20 --mds_thrash_fragments 0 #--debug_ms 20
+$CEPH_BIN/cmds $ARGS --debug_ms 1 --debug_mds 20 --mds_thrash_fragments 0 --mds_thrash_exports 1 #--debug_ms 20
+$CEPH_BIN/cmds $ARGS --debug_ms 1 --debug_mds 20 --mds_thrash_fragments 0 --mds_thrash_exports 1 #--debug_ms 20
+./cmonctl mds set_max_mds 2
 
 echo "started.  stop.sh to stop.  see out/* (e.g. 'tail -f out/????') for debug output."