mds: move predirty_nested into locker, and update with file size

diff --git a/src/client/Client.cc b/src/client/Client.cc
index 09bb2d813cb1b7856dd35587dc23b3b208308128..7590f1efa4cd60102b229194aeaf9edb24cd9da2 100644 (file)
--- a/src/client/Client.cc
+++ b/src/client/Client.cc
@@ -2244,7 +2244,10 @@ int Client::fill_stat(Inode *in, struct stat *st)
   st->st_ctime = MAX(in->inode.ctime, in->inode.mtime);
   st->st_atime = in->inode.atime;
   st->st_mtime = in->inode.mtime;
-  st->st_size = in->inode.size;
+  if (in->inode.is_dir())
+    st->st_size = in->inode.nested.rbytes;
+  else
+    st->st_size = in->inode.size;
   st->st_blksize = MAX(ceph_file_layout_su(in->inode.layout), 4096);
   st->st_blocks = in->inode.size ? DIV_ROUND_UP(in->inode.size, st->st_blksize):0;
   return in->lease_mask;

diff --git a/src/mds/CInode.cc b/src/mds/CInode.cc
index 93a460c45e9f6b633c460c5723645e4551b61372..0d826621646fab68e14490932b50449a77fafce7 100644 (file)
--- a/src/mds/CInode.cc
+++ b/src/mds/CInode.cc
@@ -70,6 +70,9 @@ ostream& operator<<(ostream& out, CInode& in)
   if (in.is_freezing_inode()) out << " FREEZING=" << in.auth_pin_freeze_allowance;
   if (in.is_frozen_inode()) out << " FROZEN";
 
+  out << " s=" << in.inode.size;
+  out << " rb=" << in.inode.nested.rbytes << "/" << in.inode.accounted_nested.rbytes;
+
   // locks
   out << " " << in.authlock;
   out << " " << in.linklock;
@@ -409,7 +412,7 @@ void CInode::name_stray_dentry(string& dname)
 
 version_t CInode::pre_dirty()
 {    
-  assert(parent);
+  assert(parent || projected_parent);
   version_t pv;
   if (projected_parent)
     pv = projected_parent->pre_dirty(get_projected_version());

diff --git a/src/mds/CInode.h b/src/mds/CInode.h
index 8ac19a921c92bc0041d77e09b70648439cc3dfc1..5fe837c0a1acb750d11e36328c05d4f8a9ccbb15 100644 (file)
--- a/src/mds/CInode.h
+++ b/src/mds/CInode.h
@@ -287,6 +287,7 @@ public:
   inodeno_t ino() const { return inode.ino; }
   inode_t& get_inode() { return inode; }
   CDentry* get_parent_dn() { return parent; }
+  CDentry* get_projected_parent_dn() { return projected_parent ? projected_parent:parent; }
   CDir *get_parent_dir();
   CInode *get_parent_inode();
   

diff --git a/src/mds/Locker.cc b/src/mds/Locker.cc
index ae0d83edb0c4d6609733aa8f3a77ae7dda69ba54..459118aedac67498d430346998e822461d531b26 100644 (file)
--- a/src/mds/Locker.cc
+++ b/src/mds/Locker.cc
@@ -471,22 +471,32 @@ version_t Locker::issue_file_data_version(CInode *in)
 struct C_Locker_FileUpdate_finish : public Context {
   Locker *locker;
   CInode *in;
+  list<CInode*> nest_updates;
   LogSegment *ls;
   bool share;
-  C_Locker_FileUpdate_finish(Locker *l, CInode *i, LogSegment *s, bool e=false) : 
+  C_Locker_FileUpdate_finish(Locker *l, CInode *i, LogSegment *s, list<CInode*> &ls, bool e=false) : 
     locker(l), in(i), ls(s), share(e) {
+    nest_updates.swap(ls);
     in->get(CInode::PIN_PTRWAITER);
   }
   void finish(int r) {
-    locker->file_update_finish(in, ls, share);
+    locker->file_update_finish(in, ls, nest_updates, share);
   }
 };
 
-void Locker::file_update_finish(CInode *in, LogSegment *ls, bool share)
+void Locker::file_update_finish(CInode *in, LogSegment *ls, list<CInode*> &nest_updates, bool share)
 {
   dout(10) << "file_update_finish on " << *in << dendl;
   in->pop_and_dirty_projected_inode(ls);
   in->put(CInode::PIN_PTRWAITER);
+
+  for (list<CInode*>::iterator p = nest_updates.begin();
+       p != nest_updates.end();
+       p++) {
+    (*p)->pop_and_dirty_projected_inode(ls);
+    scatter_wrlock_finish(&(*p)->dirlock, 0);
+  }
+
   file_wrlock_finish(&in->filelock);
   if (share && in->is_auth() && in->filelock.is_stable())
     share_inode_max_size(in);
@@ -842,11 +852,13 @@ bool Locker::check_inode_max_size(CInode *in, bool forcewrlock)
   pi->max_size = new_max;
   EOpen *le = new EOpen(mds->mdlog);
   le->metablob.add_dir_context(in->get_parent_dir());
+  list<CInode*> nest_updates;
+  predirty_nested(&le->metablob, in, nest_updates);
   le->metablob.add_primary_dentry(in->parent, true, 0, pi);
   LogSegment *ls = mds->mdlog->get_current_segment();
   le->add_ino(in->ino());
   ls->open_files.push_back(&in->xlist_open_file);
-  mds->mdlog->submit_entry(le, new C_Locker_FileUpdate_finish(this, in, ls, true));
+  mds->mdlog->submit_entry(le, new C_Locker_FileUpdate_finish(this, in, ls, nest_updates, true));
   file_wrlock_start(&in->filelock, forcewrlock);  // wrlock for duration of journal
   return true;
 }
@@ -1036,9 +1048,11 @@ void Locker::handle_client_file_caps(MClientFileCaps *m)
       pi->time_warp_seq = m->get_time_warp_seq();
     }
     le->metablob.add_dir_context(in->get_parent_dir());
+    list<CInode*> nest_updates;
+    predirty_nested(&le->metablob, in, nest_updates);
     le->metablob.add_primary_dentry(in->parent, true, 0, pi);
     LogSegment *ls = mds->mdlog->get_current_segment();
-    mds->mdlog->submit_entry(le, new C_Locker_FileUpdate_finish(this, in, ls, change_max));
+    mds->mdlog->submit_entry(le, new C_Locker_FileUpdate_finish(this, in, ls, nest_updates, change_max));
     file_wrlock_start(&in->filelock);  // wrlock for duration of journal
   }
 
@@ -1209,6 +1223,92 @@ void Locker::revoke_client_leases(SimpleLock *lock)
 
 // nested ---------------------------------------------------------------
 
+void Locker::predirty_nested(EMetaBlob *blob, CInode *in, list<CInode*> &ls)
+{
+  assert(ls.empty());
+
+  CDir *parent = in->get_projected_parent_dn()->get_dir();
+  blob->add_dir_context(parent);
+
+  // initial diff from *in
+  inode_t *curi = in->get_projected_inode();
+  __u64 drbytes;
+  __u64 drfiles;
+  utime_t rctime;
+  if (in->is_dir()) {
+    drbytes = curi->nested.rbytes - curi->accounted_nested.rbytes;
+    drfiles = curi->nested.rfiles - curi->accounted_nested.rfiles;
+    rctime = MAX(curi->ctime, curi->nested.rctime);
+  } else {
+    drbytes = curi->size - curi->accounted_nested.rbytes;
+    drfiles = 1 - curi->accounted_nested.rfiles;
+    rctime = curi->ctime;
+  }
+  
+  dout(10) << "predirty_nested delta " << drbytes << " bytes / " << drfiles << " files from " << *in << dendl;
+  
+  // build list of inodes to wrlock, dirty, and update
+  CInode *cur = in;
+  while (parent) {
+    assert(cur->is_auth());
+    assert(parent->is_auth());
+    
+    // opportunistically adjust parent dirfrag
+    CInode *pin = parent->get_inode();
+
+    dout(10) << "predirty_nested delta " << drbytes << " bytes / " << drfiles << " files for " << *pin << dendl;
+    if (pin->is_base())
+      break;
+
+    if (!scatter_wrlock_try(&pin->dirlock)) {
+      dout(10) << "predirty_nested can't wrlock " << pin->dirlock << " on " << *pin << dendl;
+      break;
+    }
+
+    ls.push_back(pin);
+
+    // FIXME
+    if (!pin->is_auth()) {
+      assert(0);
+      break;
+    }
+
+    // project update
+    version_t ppv = pin->pre_dirty();
+    inode_t *pi = pin->project_inode();
+    pi->version = ppv;
+    pi->nested.rbytes += drbytes;
+    pi->nested.rfiles += drfiles;
+    pi->nested.rctime = rctime;
+
+    frag_t fg = parent->dirfrag().frag;
+    pin->dirfrag_nested[fg].rbytes += drbytes;
+    pin->dirfrag_nested[fg].rfiles += drfiles;
+    pin->dirfrag_nested[fg].rctime = rctime;
+    
+    curi->accounted_nested.rbytes += drbytes;
+    curi->accounted_nested.rfiles += drfiles;
+    curi->accounted_nested.rctime = rctime;
+
+    cur = pin;
+    curi = pi;
+    parent = cur->get_projected_parent_dn()->get_dir();
+
+    drbytes = curi->nested.rbytes - curi->accounted_nested.rbytes;
+    drfiles = curi->nested.rfiles - curi->accounted_nested.rfiles;
+    rctime = MAX(curi->ctime, curi->nested.rctime);
+  }
+
+  // now, stick it in the blob
+  for (list<CInode*>::iterator p = ls.begin();
+       p != ls.end();
+       p++) {
+    CInode *cur = *p;
+    inode_t *pi = cur->get_projected_inode();
+    blob->add_primary_dentry(cur->get_parent_dn(), true, 0, pi);
+  }
+}
+
 
 
 // locks ----------------------------------------------------------------
@@ -1791,11 +1891,8 @@ void Locker::scatter_rdlock_finish(ScatterLock *lock, MDRequest *mdr)
 }
 
 
-bool Locker::scatter_wrlock_start(ScatterLock *lock, MDRequest *mdr)
+bool Locker::scatter_wrlock_try(ScatterLock *lock)
 {
-  dout(7) << "scatter_wrlock_start  on " << *lock
-         << " on " << *lock->get_parent() << dendl;  
-  
   // pre-twiddle?
   if (lock->get_parent()->is_auth() &&
       !lock->get_parent()->is_replicated() &&
@@ -1809,6 +1906,18 @@ bool Locker::scatter_wrlock_start(ScatterLock *lock, MDRequest *mdr)
   // can wrlock?
   if (lock->can_wrlock()) {
     lock->get_wrlock();
+    return true;
+  }
+
+  return false;
+}
+
+bool Locker::scatter_wrlock_start(ScatterLock *lock, MDRequest *mdr)
+{
+  dout(7) << "scatter_wrlock_start  on " << *lock
+         << " on " << *lock->get_parent() << dendl;  
+  
+  if (scatter_wrlock_try(lock)) {
     mdr->wrlocks.insert(lock);
     mdr->locks.insert(lock);
     return true;

diff --git a/src/mds/Locker.h b/src/mds/Locker.h
index 4195f4a25a40b5204cab3475d6a42ba6cff9bf23..0c09a348d5da8c425275e309a7e2e7e739c7c5d3 100644 (file)
--- a/src/mds/Locker.h
+++ b/src/mds/Locker.h
@@ -135,9 +135,8 @@ protected:
   void scatter_tempsync(ScatterLock *lock);
   bool scatter_rdlock_start(ScatterLock *lock, MDRequest *mdr);
   void scatter_rdlock_finish(ScatterLock *lock, MDRequest *mdr);
-public:
-  bool scatter_wrlock_start(ScatterLock *lock, MDRequest *mdr);  // public for Server's predirty_nested
-protected:
+  bool scatter_wrlock_try(ScatterLock *lock);
+  bool scatter_wrlock_start(ScatterLock *lock, MDRequest *mdr);
   void scatter_wrlock_finish(ScatterLock *lock, MDRequest *mdr);
 
   void scatter_writebehind(ScatterLock *lock);
@@ -153,6 +152,9 @@ protected:
   };
   void scatter_writebehind_finish(ScatterLock *lock, LogSegment *ls);
 
+public:
+  void predirty_nested(class EMetaBlob *blob, CInode *in, list<CInode*> &ls);
+
   // local
 protected:
   bool local_wrlock_start(LocalLock *lock, MDRequest *mdr);
@@ -197,7 +199,7 @@ protected:
   void request_inode_file_caps(CInode *in);
   void handle_inode_file_caps(class MInodeFileCaps *m);
 
-  void file_update_finish(CInode *in, LogSegment *ls, bool share);
+  void file_update_finish(CInode *in, LogSegment *ls, list<CInode*> &nest_updates, bool share);
 public:
   bool check_inode_max_size(CInode *in, bool forcewrlock=false);
 private:

diff --git a/src/mds/Server.cc b/src/mds/Server.cc
index febbe399c451e6e7c1964ad68ecdcd29d09376e8..33fde1bec8dcb93dec03bf2641b26542a5637c38 100644 (file)
--- a/src/mds/Server.cc
+++ b/src/mds/Server.cc
@@ -1539,86 +1539,23 @@ void Server::dirty_dn_diri(MDRequest *mdr, CDentry *dn, version_t dirpv)
   }
 }
 
-void Server::predirty_nested(MDRequest *mdr, EMetaBlob *blob, CInode *in, CDir *parent)
+void Server::predirty_nested(MDRequest *mdr, EMetaBlob *blob, CInode *in)
 {
-  if (!parent)
-    parent = in->get_parent_dir();
-
-  // initial diff from *in
-  inode_t *curi = in->get_projected_inode();
-  __u64 drbytes;
-  __u64 drfiles;
-  utime_t rctime;
-  if (in->is_dir()) {
-    drbytes = curi->nested.rbytes - curi->accounted_nested.rbytes;
-    drfiles = curi->nested.rfiles - curi->accounted_nested.rfiles;
-    rctime = MAX(curi->ctime, curi->nested.rctime);
-  } else {
-    drbytes = curi->size - curi->accounted_nested.rbytes;
-    drfiles = 1 - curi->accounted_nested.rfiles;
-    rctime = curi->ctime;
-  }
-  
-  blob->add_dir_context(in->get_parent_dir());
-
-  // build list of inodes to wrlock, dirty, and update
   list<CInode*> ls;
-  CInode *cur = in;
-  while (parent) {
-    assert(cur->is_auth());
-    assert(parent->is_auth());
-    
-    // opportunistically adjust parent dirfrag
-    CInode *pin = parent->get_inode();
-    if (!pin->dirlock.can_wrlock()) {
-      dout(10) << " can't wrlock " << pin->dirlock << " on " << *pin << dendl;
-      break;
-    }
-    bool r = mds->locker->scatter_wrlock_start(&pin->dirlock, mdr);
-    assert(r);
-
-    if (!pin->is_auth()) {
-      break;
-    }
-
-    // project update
-    version_t ppv = pin->pre_dirty();
-    inode_t *pi = pin->project_inode();
-    pi->version = ppv;
-    pi->nested.rbytes += drbytes;
-    pi->nested.rfiles += drfiles;
-    pi->nested.rctime = rctime;
-    mdr->add_projected_inode(pin);
-    ls.push_back(pin);
-
-    frag_t fg = parent->dirfrag().frag;
-    pin->dirfrag_nested[fg].rbytes += drbytes;
-    pin->dirfrag_nested[fg].rfiles += drfiles;
-    pin->dirfrag_nested[fg].rctime = rctime;
-    
-    curi->accounted_nested.rbytes += drbytes;
-    curi->accounted_nested.rfiles += drfiles;
-    curi->accounted_nested.rctime = rctime;
-
-    cur = pin;
-    curi = pi;
-    parent = cur->get_parent_dir();
-  }
+  mds->locker->predirty_nested(blob, in, ls);
 
-  // now, stick it in the blob
   for (list<CInode*>::iterator p = ls.begin();
        p != ls.end();
        p++) {
-    CInode *cur = *p;
-    inode_t *pi = cur->get_projected_inode();
-    blob->add_primary_dentry(cur->get_parent_dn(), true, 0, pi);
+    SimpleLock *lock = &(*p)->dirlock;
+    mdr->wrlocks.insert(lock);
+    mdr->locks.insert(lock);
+    mdr->add_projected_inode(*p);
   }
 }
 
 
 
-
-
 // ===============================================================================
 // STAT
 
@@ -2092,7 +2029,8 @@ public:
 
     // dir inode's mtime
     mds->server->dirty_dn_diri(mdr, dn, dirpv);
-    
+    mdr->pop_and_dirty_projected_inodes();
+
     // hit pop
     mds->balancer->hit_inode(mdr->now, newi, META_POP_IWR);
     //mds->balancer->hit_dir(mdr->now, dn->get_dir(), META_POP_DWR);
@@ -2116,6 +2054,7 @@ void Server::handle_client_mknod(MDRequest *mdr)
   CInode *newi = prepare_new_inode(mdr, dn->dir);
   assert(newi);
 
+  newi->projected_parent = dn;
   newi->inode.rdev = req->head.args.mknod.rdev;
   newi->inode.mode = req->head.args.mknod.mode;
   if ((newi->inode.mode & S_IFMT) == 0)
@@ -2131,6 +2070,7 @@ void Server::handle_client_mknod(MDRequest *mdr)
   le->metablob.add_allocated_ino(newi->ino(), mds->idalloc->get_version());
   version_t dirpv = predirty_dn_diri(mdr, dn, &le->metablob);  // dir mtime too
   le->metablob.add_dir_context(dn->dir);
+  predirty_nested(mdr, &le->metablob, newi);
   le->metablob.add_primary_dentry(dn, true, newi, &newi->inode);
   
   // log + wait
@@ -2154,6 +2094,7 @@ void Server::handle_client_mkdir(MDRequest *mdr)
   assert(newi);
 
   // it's a directory.
+  newi->projected_parent = dn;
   newi->inode.mode = req->head.args.mkdir.mode;
   newi->inode.mode &= ~S_IFMT;
   newi->inode.mode |= S_IFDIR;
@@ -2174,6 +2115,7 @@ void Server::handle_client_mkdir(MDRequest *mdr)
   le->metablob.add_allocated_ino(newi->ino(), mds->idalloc->get_version());
   version_t dirpv = predirty_dn_diri(mdr, dn, &le->metablob);  // dir mtime too
   le->metablob.add_dir_context(dn->dir);
+  predirty_nested(mdr, &le->metablob, newi);
   le->metablob.add_primary_dentry(dn, true, newi, &newi->inode);
   le->metablob.add_dir(newdir, true, true); // dirty AND complete
   
@@ -2211,6 +2153,7 @@ void Server::handle_client_symlink(MDRequest *mdr)
   assert(newi);
 
   // it's a symlink
+  newi->projected_parent = dn;
   newi->inode.mode &= ~S_IFMT;
   newi->inode.mode |= S_IFLNK;
   newi->inode.mode |= 0777;     // ?
@@ -2225,7 +2168,7 @@ void Server::handle_client_symlink(MDRequest *mdr)
   le->metablob.add_allocated_ino(newi->ino(), mds->idalloc->get_version());
   version_t dirpv = predirty_dn_diri(mdr, dn, &le->metablob);  // dir mtime too
   le->metablob.add_dir_context(dn->dir);
-  predirty_nested(mdr, &le->metablob, newi, dn->dir);
+  predirty_nested(mdr, &le->metablob, newi);
   le->metablob.add_primary_dentry(dn, true, newi, &newi->inode);
 
   // log + wait

diff --git a/src/mds/Server.h b/src/mds/Server.h
index c28a7726f2ee4b74b39faebfcfa4a9c2bc090e46..840492aae2710cea1ed950a1a6a7501db944f45c 100644 (file)
--- a/src/mds/Server.h
+++ b/src/mds/Server.h
@@ -102,7 +102,7 @@ public:
 
   version_t predirty_dn_diri(MDRequest *mdr, CDentry *dn, class EMetaBlob *blob);
   void dirty_dn_diri(MDRequest *mdr, CDentry *dn, version_t dirpv);
-  void predirty_nested(MDRequest *mdr, EMetaBlob *blob, CInode *in, CDir *parent);
+  void predirty_nested(MDRequest *mdr, EMetaBlob *blob, CInode *in);
 
 
   // requests on existing inodes.

diff --git a/src/mds/events/EMetaBlob.h b/src/mds/events/EMetaBlob.h
index 85974474e761f2ef3552b5d36a035d98580a8677..efa266fb1b6c08e3a59abc667004b41db941fc04 100644 (file)
--- a/src/mds/events/EMetaBlob.h
+++ b/src/mds/events/EMetaBlob.h
@@ -465,7 +465,7 @@ private:
     // journaled?
 
     // add parent dn
-    CDentry *parent = diri->get_parent_dn();
+    CDentry *parent = diri->get_projected_parent_dn();
     add_dir_context(parent->get_dir(), mode);
     add_dentry(parent, false);
   }