mds: fixed various predirty_nested problems, rename, straydir locking. temp fragstat...

diff --git a/src/TODO b/src/TODO
index 9e61c4836211af45ed626e097f005fd41df24975..9aae898bf8472e92cf22cf871accc0552d51b95f 100644 (file)
--- a/src/TODO
+++ b/src/TODO
@@ -74,9 +74,6 @@ mds mustfix
 - anchor_destroy needs to xlock linklock.. which means it needs a Mutation wrapper?
   - ... when it gets a caller.. someday..
 
-- fix rm -r vs mds exports
-  - use new dir size values in dirstat
-
 - replay of dir fragmentation  (dont want dir frozen, pins, etc.?)
 - rename slave in-memory rollback on failure
 

diff --git a/src/client/SyntheticClient.cc b/src/client/SyntheticClient.cc
index 0af362c43d47b0634a69cb936750214eb933be0c..d4ec5c57f918763abd35d9eda9118cd1973599a3 100644 (file)
--- a/src/client/SyntheticClient.cc
+++ b/src/client/SyntheticClient.cc
@@ -1489,8 +1489,9 @@ int SyntheticClient::full_walk(string& basedir)
       }
     }
 
-    if (actual.nsubdirs != expect.nsubdirs ||
-       actual.nfiles != expect.nfiles) {
+    if (dir != "" &&
+       (actual.nsubdirs != expect.nsubdirs ||
+        actual.nfiles != expect.nfiles)) {
       dout(0) << dir << ": expected " << expect << dendl;
       dout(0) << dir << ":      got " << actual << dendl;
     }
@@ -2857,6 +2858,7 @@ int SyntheticClient::thrash_links(const char *basedir, int dirs, int files, int
 
   if (time_to_stop()) return 0;
 
+  srand(0);
   if (1) {
     for (int k=0; k<n; k++) {
       

diff --git a/src/mds/CDir.cc b/src/mds/CDir.cc
index f8f72e8a830b8c16890e12e0bd677e13d1abc0fd..f2678eee69dc8728b2f080c9c8fb287a9ab2fab4 100644 (file)
--- a/src/mds/CDir.cc
+++ b/src/mds/CDir.cc
@@ -1583,6 +1583,49 @@ void CDir::adjust_nested_anchors(int by)
   inode->adjust_nested_anchors(by);
 }
 
+void CDir::verify_fragstat()
+{
+  assert(is_complete());
+  if (inode->is_stray())
+    return;
+
+  frag_info_t c;
+  memset(&c, 0, sizeof(c));
+
+  for (map_t::iterator it = items.begin();
+       it != items.end();
+       it++) {
+    CDentry *dn = it->second;
+    if (dn->is_null())
+      continue;
+    
+    dout(10) << " " << *dn << dendl;
+    if (dn->is_primary())
+      dout(10) << "     " << *dn->inode << dendl;
+
+    if (dn->is_primary()) {
+      if (dn->inode->is_dir())
+       c.nsubdirs++;
+      else
+       c.nfiles++;
+    }
+    if (dn->is_remote()) {
+      if (dn->get_remote_d_type() == (S_IFDIR >> 12))
+       c.nsubdirs++;
+      else
+       c.nfiles++;
+    }
+  }
+
+  if (c.nsubdirs != fnode.fragstat.nsubdirs ||
+      c.nfiles != fnode.fragstat.nfiles) {
+    dout(0) << "verify_fragstat failed " << fnode.fragstat << " on " << *this << dendl;
+    dout(0) << "               i count " << c << dendl;
+    assert(0);
+  } else {
+    dout(0) << "verify_fragstat ok " << fnode.fragstat << " on " << *this << dendl;
+  }
+}
 
 /*****************************************************************************
  * FREEZING

diff --git a/src/mds/CDir.h b/src/mds/CDir.h
index 5e87f83af261d11349c09400c4ac4ee5e43b9836..9835409834cfb8e8d379bbcebe8cedd7d45b1a6b 100644 (file)
--- a/src/mds/CDir.h
+++ b/src/mds/CDir.h
@@ -446,7 +446,9 @@ public:
   int get_nested_auth_pins() { return nested_auth_pins; }
   void auth_pin();
   void auth_unpin();
+
   void adjust_nested_auth_pins(int inc);
+  void verify_fragstat();
 
   int get_nested_anchors() { return nested_anchors; }
   void adjust_nested_anchors(int by);

diff --git a/src/mds/Locker.cc b/src/mds/Locker.cc
index c09caa0a6a9090e5b16cc9cbd8cc292117f27f05..c5c89032fedc97c418adbd6eb019e4bbbe8a8717 100644 (file)
--- a/src/mds/Locker.cc
+++ b/src/mds/Locker.cc
@@ -1220,10 +1220,10 @@ void Locker::predirty_nested(Mutation *mut, EMetaBlob *blob,
                             EMetaBlob *rollback)
 {
   bool primary_dn = flags & PREDIRTY_PRIMARY;
-  bool do_parent = flags & PREDIRTY_DIR;
+  bool do_parent_mtime = flags & PREDIRTY_DIR;
 
   dout(10) << "predirty_nested"
-          << (do_parent ? " do_parent_mtime":"")
+          << (do_parent_mtime ? " do_parent_mtime":"")
           << " linkunlink=" <<  linkunlink
           << (primary_dn ? " primary_dn":" remote_dn")
           << " " << *in << dendl;
@@ -1233,8 +1233,8 @@ void Locker::predirty_nested(Mutation *mut, EMetaBlob *blob,
     parent = in->get_projected_parent_dn()->get_dir();
   }
 
-  if (flags == 0) {
-    dout(10) << " no flags, just adding dir context to blob(s)" << dendl;
+  if (flags == 0 && linkunlink == 0) {
+    dout(10) << " no flags/linkunlink, just adding dir context to blob(s)" << dendl;
     blob->add_dir_context(parent);
     if (rollback)
       rollback->add_dir_context(parent);
@@ -1256,8 +1256,9 @@ void Locker::predirty_nested(Mutation *mut, EMetaBlob *blob,
     // opportunistically adjust parent dirfrag
     CInode *pin = parent->get_inode();
 
-    if (do_parent) {
-      assert(mut->wrlocks.count(&pin->dirlock));
+    if (do_parent_mtime || linkunlink) {
+      assert(mut->wrlocks.count(&pin->dirlock) ||
+            mut->is_slave());   // we are slave.  master will have wrlocked the dir.
     }
 
     // inode -> dirfrag
@@ -1266,17 +1267,18 @@ void Locker::predirty_nested(Mutation *mut, EMetaBlob *blob,
     fnode_t *pf = parent->project_fnode();
     pf->version = parent->pre_dirty();
 
-    if (do_parent) {
-      dout(10) << "predirty_nested updating mtime/size on " << *parent << dendl;
+    if (do_parent_mtime) {
+      dout(10) << "predirty_nested updating mtime on " << *parent << dendl;
       pf->fragstat.mtime = mut->now;
       if (mut->now > pf->fragstat.rctime)
        pf->fragstat.rctime = mut->now;
-      if (linkunlink) {
-       if (in->is_dir())
-         pf->fragstat.nsubdirs += linkunlink;
-       else
-         pf->fragstat.nfiles += linkunlink;
-      }
+    }
+    if (linkunlink) {
+      dout(10) << "predirty_nested updating size on " << *parent << dendl;
+      if (in->is_dir())
+       pf->fragstat.nsubdirs += linkunlink;
+      else
+       pf->fragstat.nfiles += linkunlink;
     }
     if (primary_dn) {
       if (linkunlink == 0) {
@@ -1352,7 +1354,7 @@ void Locker::predirty_nested(Mutation *mut, EMetaBlob *blob,
     curi = pi;
     parent = cur->get_projected_parent_dn()->get_dir();
     linkunlink = 0;
-    do_parent = false;
+    do_parent_mtime = false;
     primary_dn = true;
   }
 
@@ -2028,8 +2030,9 @@ void Locker::scatter_wrlock_finish(ScatterLock *lock, Mutation *mut)
     mut->wrlocks.erase(lock);
     mut->locks.erase(lock);
   }
-  
-  scatter_eval_gather(lock);
+
+  if (!lock->is_wrlocked())
+    scatter_eval_gather(lock);
 }
 
 
@@ -2395,17 +2398,17 @@ void Locker::scatter_sync(ScatterLock *lock)
     break; // do it.
 
   case LOCK_SCATTER:
-    // lock first.  this is the slow way, incidentally.
-    if (lock->get_parent()->is_replicated()) {
-      send_lock_message(lock, LOCK_AC_LOCK);
-      lock->init_gather();
-    } else {
-      if (!lock->is_wrlocked()) {
-       break; // do it now, we're fine
-      }
-    }
+    if (!lock->get_parent()->is_replicated() &&
+       !lock->is_wrlocked())
+      break; // do it now
+
     lock->set_state(LOCK_GLOCKC);
     lock->get_parent()->auth_pin();
+
+    if (lock->get_parent()->is_replicated()) {
+      lock->init_gather();
+      send_lock_message(lock, LOCK_AC_LOCK);
+    }    
     return;
 
   default:
@@ -2561,12 +2564,14 @@ void Locker::scatter_tempsync(ScatterLock *lock)
       break; // do it.
     }
     
+    lock->set_state(LOCK_GTEMPSYNCC);
+    lock->get_parent()->auth_pin();
+
     if (lock->get_parent()->is_replicated()) {
-      send_lock_message(lock, LOCK_AC_LOCK);
       lock->init_gather();
+      send_lock_message(lock, LOCK_AC_LOCK);
     }
-    lock->set_state(LOCK_GTEMPSYNCC);
-    lock->get_parent()->auth_pin();
+
     return;
 
   case LOCK_TEMPSYNC:

diff --git a/src/mds/MDCache.cc b/src/mds/MDCache.cc
index 87159039fe32686f96a2039d90b0942b6b0cf408..c5276b6c03ab8c195dad180a62142d652edd659a 100644 (file)
--- a/src/mds/MDCache.cc
+++ b/src/mds/MDCache.cc
@@ -4017,6 +4017,10 @@ int MDCache::path_traverse(MDRequest *mdr, Message *req,     // who
     }
     assert(curdir);
 
+    // HACK
+    if (curdir->is_complete())
+      curdir->verify_fragstat();
+
     // frozen?
     /*
     if (curdir->is_frozen()) {

diff --git a/src/mds/MDCache.h b/src/mds/MDCache.h
index b442cef1b52e497a1b3af66286d9a3491764291a..a218618e7b2fe2ab46e70bd3bad365a6badc2751 100644 (file)
--- a/src/mds/MDCache.h
+++ b/src/mds/MDCache.h
@@ -79,6 +79,9 @@ struct Mutation {
   LogSegment *ls;  // the log segment i'm committing to
   utime_t now;
 
+  // flag mutation as slave
+  int slave_to_mds;                // this is a slave request if >= 0.
+
   // -- my pins and locks --
   // cache pins (so things don't expire)
   set< MDSCacheObject* > pins;
@@ -107,11 +110,16 @@ struct Mutation {
 
   Mutation() : ls(0),
               done_locking(false), committing(false), aborted(false) {}
-  Mutation(metareqid_t ri) : reqid(ri),
-                            ls(0),
-                            done_locking(false), committing(false), aborted(false) {}
+  Mutation(metareqid_t ri, int slave_to=-1) : 
+    reqid(ri),
+    ls(0),
+    slave_to_mds(slave_to), 
+    done_locking(false), committing(false), aborted(false) {}
   virtual ~Mutation() {}
 
+  bool is_master() { return slave_to_mds < 0; }
+  bool is_slave() { return slave_to_mds >= 0; }
+
   // pin items in cache
   void pin(MDSCacheObject *o) {
     if (pins.count(o) == 0) {
@@ -214,7 +222,6 @@ struct MDRequest : public Mutation {
 
   // -- i am a slave request
   MMDSSlaveRequest *slave_request; // slave request (if one is pending; implies slave == true)
-  int slave_to_mds;                // this is a slave request if >= 0.
 
 
   // break rarely-used fields into a separately allocated structure 
@@ -252,25 +259,22 @@ struct MDRequest : public Mutation {
   // ---------------------------------------------------
   MDRequest() : 
     session(0), client_request(0), ref(0), 
-    slave_request(0), slave_to_mds(-1), 
+    slave_request(0),
     _more(0) {}
   MDRequest(metareqid_t ri, MClientRequest *req) : 
     Mutation(ri),
     session(0), client_request(req), ref(0), 
-    slave_request(0), slave_to_mds(-1), 
+    slave_request(0),
     _more(0) {}
   MDRequest(metareqid_t ri, int by) : 
-    Mutation(ri),
+    Mutation(ri, by),
     session(0), client_request(0), ref(0),
-    slave_request(0), slave_to_mds(by), 
+    slave_request(0),
     _more(0) {}
   ~MDRequest() {
     delete _more;
   }
   
-  bool is_master() { return slave_to_mds < 0; }
-  bool is_slave() { return slave_to_mds >= 0; }
-
   More* more() { 
     if (!_more) _more = new More();
     return _more;

diff --git a/src/mds/Migrator.cc b/src/mds/Migrator.cc
index e9965486c42e0980eae7f0f0e04286a94713e5c0..53eaaddf48b8b86d6e1f8f97912b189c1c3757b1 100644 (file)
--- a/src/mds/Migrator.cc
+++ b/src/mds/Migrator.cc
@@ -958,6 +958,9 @@ int Migrator::encode_export_dir(bufferlist& exportbl,
   
   assert(dir->get_projected_version() == dir->get_version());
 
+  if (dir->is_complete())
+    dir->verify_fragstat();
+
   // dir 
   dirfrag_t df = dir->dirfrag();
   ::encode(df, exportbl);
@@ -2166,6 +2169,9 @@ int Migrator::decode_import_dir(bufferlist::iterator& blp,
       le->metablob.add_dentry(dn, dn->is_dirty());
   }
   
+  if (dir->is_complete())
+    dir->verify_fragstat();
+  
   dout(7) << "decode_import_dir done " << *dir << dendl;
   return num_imported;
 }

diff --git a/src/mds/Server.cc b/src/mds/Server.cc
index 33e9daccb88c52e7b8ee044b7e9b809b05aa3f0c..11758cc6015d30a59a5cbf5ba2ff0466a2221703 100644 (file)
--- a/src/mds/Server.cc
+++ b/src/mds/Server.cc
@@ -629,6 +629,8 @@ void Server::set_trace_dist(Session *session, MClientReply *reply, CInode *in, C
   dout(20) << " trace added " << lmask << " " << *dn << dendl;
   
   // dir
+  if (dn->get_dir()->is_complete())
+    dn->get_dir()->verify_fragstat();
   DirStat::encode(bl, dn->get_dir(), whoami);
   dout(20) << " trace added " << *dn->get_dir() << dendl;
 
@@ -1846,6 +1848,8 @@ void Server::handle_client_readdir(MDRequest *mdr)
     return;
   }
 
+  dir->verify_fragstat();
+
   mdr->now = g_clock.real_now();
 
   // build dir contents
@@ -3033,21 +3037,32 @@ void Server::handle_client_rename(MDRequest *mdr)
 
   dout(10) << " destdn " << *destdn << dendl;
 
+  bool linkmerge = (srcdn->inode == destdn->inode &&
+                   (srcdn->is_primary() || destdn->is_primary()));
+  if (linkmerge)
+    dout(10) << " this is a link merge" << dendl;
+
+  // -- create stray dentry? --
+  CDentry *straydn = 0;
+  if (destdn->is_primary() && !linkmerge) {
+    straydn = mdcache->get_or_create_stray_dentry(destdn->inode);
+    mdr->pin(straydn);
+    dout(10) << "straydn is " << *straydn << dendl;
+  }
+
   // -- locks --
   set<SimpleLock*> rdlocks, wrlocks, xlocks;
 
+  // straydn?
+  if (straydn)
+    wrlocks.insert(&straydn->dir->inode->dirlock);
+
   // rdlock sourcedir path, xlock src dentry
   for (int i=0; i<(int)srctrace.size()-1; i++) 
     rdlocks.insert(&srctrace[i]->lock);
   xlocks.insert(&srcdn->lock);
   wrlocks.insert(&srcdn->dir->inode->dirlock);
 
-  /*
-   * no, this causes problems if the dftlock is scattered...
-   *  and what was i thinking anyway? 
-   * rdlocks.insert(&srcdn->dir->inode->dirfragtreelock);  // rd lock on srci dirfragtree.
-   */
-
   // rdlock destdir path, xlock dest dentry
   for (int i=0; i<(int)desttrace.size(); i++)
     rdlocks.insert(&desttrace[i]->lock);
@@ -3111,17 +3126,6 @@ void Server::handle_client_rename(MDRequest *mdr)
   if (mdr->now == utime_t())
     mdr->now = g_clock.real_now();
 
-  bool linkmerge = (srcdn->inode == destdn->inode &&
-                   (srcdn->is_primary() || destdn->is_primary()));
-
-  // -- create stray dentry? --
-  CDentry *straydn = 0;
-  if (destdn->is_primary() && !linkmerge) {
-    straydn = mdcache->get_or_create_stray_dentry(destdn->inode);
-    mdr->pin(straydn);
-    dout(10) << "straydn is " << *straydn << dendl;
-  }
-
   // -- prepare witnesses --
   /*
    * NOTE: we use _all_ replicas as witnesses.
@@ -3404,22 +3408,21 @@ void Server::_rename_prepare(MDRequest *mdr,
   }
 
   // prepare nesting, mtime updates
-  if (mdr->is_master()) {
-    int predirty_dir = silent ? 0:PREDIRTY_DIR;
-
-    // sub off target
-    if (!destdn->is_null())
-      mds->locker->predirty_nested(mdr, metablob, destdn->inode, destdn->dir,
-                                  (destdn->is_primary() ? PREDIRTY_PRIMARY:0)|predirty_dir, -1,
-                                  rollback);
-    
-    // move srcdn
-    int predirty_primary = (srcdn->is_primary() && srcdn->dir != destdn->dir) ? PREDIRTY_PRIMARY:0;
-    int flags = predirty_dir | predirty_primary;
-    if (srcdn->is_auth())
-      mds->locker->predirty_nested(mdr, metablob, srcdn->inode, srcdn->dir, flags, -1, rollback);
+  int predirty_dir = silent ? 0:PREDIRTY_DIR;
+  
+  // sub off target
+  if (destdn->is_auth() && !destdn->is_null())
+    mds->locker->predirty_nested(mdr, metablob, destdn->inode, destdn->dir,
+                                (destdn->is_primary() ? PREDIRTY_PRIMARY:0)|predirty_dir, -1,
+                                rollback);
+  
+  // move srcdn
+  int predirty_primary = (srcdn->is_primary() && srcdn->dir != destdn->dir) ? PREDIRTY_PRIMARY:0;
+  int flags = predirty_dir | predirty_primary;
+  if (srcdn->is_auth())
+    mds->locker->predirty_nested(mdr, metablob, srcdn->inode, srcdn->dir, flags, -1, rollback);
+  if (destdn->is_auth())
     mds->locker->predirty_nested(mdr, metablob, srcdn->inode, destdn->dir, flags, 1, rollback);
-  }
 
   // add it all to the metablob
   // target inode