mds: put a cap in each inode stat in trace; simplify snaprealm trace encoding

diff --git a/src/include/ceph_fs.h b/src/include/ceph_fs.h
index da46971ef9efbafd6f5c1a33868825e4bbf08607..af8d73b33b4f705300617b601730965969f25175 100644 (file)
--- a/src/include/ceph_fs.h
+++ b/src/include/ceph_fs.h
@@ -803,9 +803,6 @@ struct ceph_mds_reply_head {
        ceph_tid_t tid;
        __le32 op;
        __le32 result;
-       __le32 file_caps;
-       __le32 file_caps_seq;
-       __le32 file_caps_mseq;
        __le32 mdsmap_epoch;
        __u8 safe;
 } __attribute__ ((packed));
@@ -821,17 +818,23 @@ struct ceph_frag_tree_head {
        struct ceph_frag_tree_split splits[];
 } __attribute__ ((packed));
 
+struct ceph_mds_reply_cap {
+       __le32 caps;
+       __le32 seq, mseq;
+} __attribute__ ((packed));
+
 struct ceph_mds_reply_inode {
        __le64 ino;
        __le64 snapid;
+       __le32 rdev;
        __le64 version;
+       struct ceph_mds_reply_cap cap;
        struct ceph_file_layout layout;
        struct ceph_timespec ctime, mtime, atime;
        __le64 time_warp_seq;
-       __le32 rdev;
+       __le64 size, max_size, truncate_seq;
        __le32 mode, uid, gid;
        __le32 nlink;
-       __le64 size, max_size, truncate_seq;
        __le64 files, subdirs, rbytes, rfiles, rsubdirs;  /* dir stats */
        struct ceph_timespec rctime;
        struct ceph_frag_tree_head fragtree;

diff --git a/src/mds/CInode.cc b/src/mds/CInode.cc
index a1e42fb3d2721424a80a379a263dd257354ce531..d9716077d495f9eb23a6ac1ef46a3130b6d2bd65 100644 (file)
--- a/src/mds/CInode.cc
+++ b/src/mds/CInode.cc
@@ -1277,7 +1277,7 @@ void CInode::decode_snap_blob(bufferlist& snapbl)
 }
 
 
-bool CInode::encode_inodestat(bufferlist& bl, snapid_t snapid, bool projected)
+bool CInode::encode_inodestat(bufferlist& bl, Capability *cap, snapid_t snapid, bool projected)
 {
   bool valid = true;
 
@@ -1349,6 +1349,24 @@ bool CInode::encode_inodestat(bufferlist& bl, snapid_t snapid, bool projected)
     ::encode(xattrs, xbl);
   ::encode(xbl, bl);
 
+
+  // include capability?
+  if (snapid != CEPH_NOSNAP) {
+    e.cap.caps = valid ? get_caps_allowed(false) : CEPH_STAT_CAP_INODE;
+    e.cap.seq = 0;
+    e.cap.mseq = 0;
+  } else {
+    if (cap && valid) {
+      e.cap.caps = cap->pending();
+      e.cap.seq = cap->get_last_seq();
+      e.cap.mseq = cap->get_mseq();
+    } else {
+      e.cap.caps = 0;
+      e.cap.seq = 0;
+      e.cap.mseq = 0;
+    }
+  }
+
   return valid;
 }
 

diff --git a/src/mds/CInode.h b/src/mds/CInode.h
index 92c7ce18093037d5d776df5a9eb732e46f65f844..a014c5e152fdf8bb838e7d569a75ac294e8df760 100644 (file)
--- a/src/mds/CInode.h
+++ b/src/mds/CInode.h
@@ -419,7 +419,7 @@ private:
   
 
   // for giving to clients
-  bool encode_inodestat(bufferlist& bl, snapid_t snapid=CEPH_NOSNAP, bool projected=false);
+  bool encode_inodestat(bufferlist& bl, Capability *cap, snapid_t snapid=CEPH_NOSNAP, bool projected=false);
 
 
   // -- locks --

diff --git a/src/mds/Capability.h b/src/mds/Capability.h
index b9a56ae61da6f7918bece510f58393893fc1cbeb..0cdb1d7a60e522b95bf84294d6011517942e01da 100644 (file)
--- a/src/mds/Capability.h
+++ b/src/mds/Capability.h
@@ -24,11 +24,6 @@ using namespace std;
 
 #include "config.h"
 
-
-// heuristics
-//#define CEPH_CAP_DELAYFLUSH  32
-
-
 class CInode;
 
 class Capability {

diff --git a/src/mds/MDCache.h b/src/mds/MDCache.h
index 0c5b8aa2e50ac104a410e2f3a2d6ab0244d9e5e6..056f61972f76431fe5dc76731a3cd0e0a0177552 100644 (file)
--- a/src/mds/MDCache.h
+++ b/src/mds/MDCache.h
@@ -260,7 +260,6 @@ struct MDRequest : public Mutation {
 
   Capability *cap;
   int snap_caps;
-  bufferlist snapbl;
   bool did_early_reply;
 
   // -- i am a slave request

diff --git a/src/mds/Server.cc b/src/mds/Server.cc
index faa05214bade4b41d5cc43e818c9caffb75516a9..ae14f5e89fb6778afd0bf8e3738fbcdea7e5725c 100644 (file)
--- a/src/mds/Server.cc
+++ b/src/mds/Server.cc
@@ -516,23 +516,6 @@ void Server::reply_request(MDRequest *mdr, int r, CInode *tracei, CDentry *trace
   reply_request(mdr, new MClientReply(mdr->client_request, r), tracei, tracedn);
 }
 
-void Server::include_cap_in_reply(MDRequest *mdr, MClientReply *reply)
-{
-  // include cap / snapbl?
-  //  (only once!)
-  if (mdr->snapbl.length())
-    reply->snapbl.claim(mdr->snapbl);
-  if (mdr->cap) {
-    reply->set_file_caps(mdr->cap->pending());
-    reply->set_file_caps_seq(mdr->cap->get_last_seq());
-    reply->set_file_caps_mseq(mdr->cap->get_mseq());
-    mdr->cap = 0;
-  }
-  if (mdr->snap_caps)
-    reply->set_file_caps(mdr->snap_caps);
-}
-
-
 void Server::early_reply(MDRequest *mdr, CInode *tracei, CDentry *tracedn)
 {
   if (!g_conf.mds_early_reply)
@@ -560,9 +543,6 @@ void Server::early_reply(MDRequest *mdr, CInode *tracei, CDentry *tracedn)
   if (tracei || tracedn)
     set_trace_dist(mdr->session, reply, tracei, tracedn, snapid, snapdiri, true);
 
-  // include cap info?
-  include_cap_in_reply(mdr, reply);
-
   mdr->did_early_reply = true;
 
   messenger->send_message(reply, client_inst);
@@ -607,9 +587,6 @@ void Server::reply_request(MDRequest *mdr, MClientReply *reply, CInode *tracei,
 
   reply->set_mdsmap_epoch(mds->mdsmap->get_epoch());
 
-  // include cap info?
-  include_cap_in_reply(mdr, reply);
-
   // infer tracei/tracedn from mdr?
   snapid_t snapid = CEPH_NOSNAP;
   CInode *snapdiri = 0;
@@ -700,6 +677,11 @@ void Server::set_trace_dist(Session *session, MClientReply *reply, CInode *in, C
   utime_t now = g_clock.now();
   int lmask = 0;
 
+  // snapbl
+  SnapRealm *realm = in->find_snaprealm();
+  reply->snapbl = realm->get_snap_trace();
+  dout(10) << "set_trace_dist snaprealm " << *realm << dendl;
+
   // start with dentry or inode?
   if (!in) {
     assert(dn);
@@ -709,18 +691,13 @@ void Server::set_trace_dist(Session *session, MClientReply *reply, CInode *in, C
 
  inode:
   numi++;
-  if (in->encode_inodestat(bl, snapid, projected))
-    lmask = mds->locker->issue_client_lease(in, client, bl, now, session);
-  else {
-    lmask = CEPH_STAT_CAP_INODE; // immutable bits
-    encode_null_lease(bl);
-  }
-  dout(20) << " trace added " << lmask << " snapid " << snapid << " " << *in << dendl;
+  in->encode_inodestat(bl, in->get_client_cap(client), snapid, projected);
+  dout(20) << "set_trace_dist added snapid " << snapid << " " << *in << dendl;
 
   if (snapid != CEPH_NOSNAP && in == snapdiri) {
     // do the snap name dentry
     const string& snapname = in->find_snaprealm()->get_snapname(snapid, in->ino());
-    dout(10) << " snapname " << snapname << dendl;
+    dout(10) << "set_trace_dist snapname " << snapname << dendl;
     ::encode(snapname, bl);
     encode_infinite_lease(bl);
     numdn++;
@@ -728,13 +705,12 @@ void Server::set_trace_dist(Session *session, MClientReply *reply, CInode *in, C
 
     // back to the live tree
     snapid = CEPH_NOSNAP;
-    in->encode_inodestat(bl, snapid);
-    lmask = mds->locker->issue_client_lease(in, client, bl, now, session);
+    in->encode_inodestat(bl, in->get_client_cap(client), snapid, false);
     numi++;
-    dout(20) << " trace added " << lmask << " snapid " << snapid << " " << *in << dendl;
+    dout(20) << "set_trace_dist added snapid " << snapid << " " << *in << dendl;
 
     snapdirpos = numi;
-    dout(10) << " snapdiri at pos " << snapdirpos << dendl;
+    dout(10) << "set_trace_dist snapdiri at pos " << snapdirpos << dendl;
   }
 
   if (!dn) {
@@ -755,7 +731,7 @@ void Server::set_trace_dist(Session *session, MClientReply *reply, CInode *in, C
   else
     encode_null_lease(bl);
   numdn++;
-  dout(20) << " trace added " << lmask << " snapid " << snapid << " " << *dn << dendl;
+  dout(20) << "set_trace_dist added " << lmask << " snapid " << snapid << " " << *dn << dendl;
   
   // dir
 #ifdef MDS_VERIFY_FRAGSTAT
@@ -764,7 +740,7 @@ void Server::set_trace_dist(Session *session, MClientReply *reply, CInode *in, C
 #endif
 
   dn->get_dir()->encode_dirstat(bl, whoami);
-  dout(20) << " trace added " << *dn->get_dir() << dendl;
+  dout(20) << "set_trace_dist added " << *dn->get_dir() << dendl;
 
   in = dn->get_dir()->get_inode();
   dn = 0;
@@ -2254,9 +2230,8 @@ void Server::handle_client_readdir(MDRequest *mdr)
 
     // inode
     dout(12) << "including inode " << *in << dendl;
-    bool valid = in->encode_inodestat(dnbl, snapid);
+    bool valid = in->encode_inodestat(dnbl, in->get_client_cap(client), snapid);
     assert(valid);
-    mds->locker->issue_client_lease(in, client, dnbl, mdr->now, mdr->session);
     numfiles++;
 
     // touch dn
@@ -4914,10 +4889,6 @@ void Server::_do_open(MDRequest *mdr, CInode *cur)
     mds->balancer->hit_inode(mdr->now, cur, META_POP_IRD, 
                             mdr->client_request->get_orig_source().num());
   
-  SnapRealm *realm = cur->find_snaprealm();
-  realm->build_snap_trace(mdr->snapbl);
-  dout(10) << " snaprealm is " << *realm << " on " << *realm->inode << dendl;
-
   reply_request(mdr, 0);
 }
 
@@ -5108,7 +5079,6 @@ void Server::handle_client_openc(MDRequest *mdr)
 
   // stick cap, snapbl info in mdr
   mdr->cap = cap;
-  realm->build_snap_trace(mdr->snapbl);
 
   // make sure this inode gets into the journal
   le->metablob.add_opened_ino(in->ino());
@@ -5197,7 +5167,7 @@ void Server::handle_client_lssnap(MDRequest *mdr)
     else
       ::encode(p->second->get_long_name(), dnbl);
     encode_infinite_lease(dnbl);
-    diri->encode_inodestat(dnbl, p->first);
+    diri->encode_inodestat(dnbl, NULL, p->first);
     mds->locker->issue_client_lease(diri, client, dnbl, now, mdr->session);
     num++;
   }
@@ -5387,7 +5357,7 @@ void Server::_mksnap_finish(MDRequest *mdr, CInode *diri, SnapInfo &info)
   mdr->ref_snapid = snapid;
   mdr->ref_snapdiri = diri;
   MClientReply *reply = new MClientReply(mdr->client_request, 0);
-  diri->snaprealm->build_snap_trace(reply->snapbl);
+  reply->snapbl = diri->snaprealm->get_snap_trace();
   reply_request(mdr, reply);
 }
 

diff --git a/src/mds/Server.h b/src/mds/Server.h
index 6a83d52d5c37cf9e0f8420e40f8ad5de8fc937ff..4f5619dbd930d2f4e54d2ad69cf4d4d7bc534f81 100644 (file)
--- a/src/mds/Server.h
+++ b/src/mds/Server.h
@@ -81,7 +81,6 @@ public:
   void set_trace_dist(Session *session, MClientReply *reply, CInode *in, CDentry *dn,
                      snapid_t snapid, CInode *snapdiri,
                      bool projected = false);
-  void include_cap_in_reply(MDRequest *mdr, MClientReply *reply);
 
   void encode_empty_dirstat(bufferlist& bl);
   void encode_infinite_lease(bufferlist& bl);

diff --git a/src/mds/snap.cc b/src/mds/snap.cc
index 13eddd35f45e7a05a7cce91413407a58b11adf6a..7a041751fccdf0f3cd530d0855efff8ff0bc3e75 100644 (file)
--- a/src/mds/snap.cc
+++ b/src/mds/snap.cc
@@ -192,6 +192,7 @@ void SnapRealm::check_cache()
   cached_seq = seq;
   build_snap_set(cached_snaps, cached_seq, cached_last_created, cached_last_destroyed,
                 0, CEPH_NOSNAP);
+  build_snap_trace(cached_snap_trace);
   
   dout(10) << "check_cache rebuilt " << cached_snaps
           << " seq " << seq
@@ -404,7 +405,11 @@ void SnapRealm::split_at(SnapRealm *child)
 
 }
 
-
+const bufferlist& SnapRealm::get_snap_trace()
+{
+  check_cache();
+  return cached_snap_trace;
+}
 
 void SnapRealm::build_snap_trace(bufferlist& snapbl)
 {

diff --git a/src/mds/snap.h b/src/mds/snap.h
index 2e701b412598ba75ca64404bc3fdad3f45649b96..743cc73fb6bcb118827c9f67d3201f82e2df660d 100644 (file)
--- a/src/mds/snap.h
+++ b/src/mds/snap.h
@@ -124,6 +124,8 @@ struct SnapRealm {
   set<snapid_t> cached_snaps;
   SnapContext cached_snap_context;
 
+  bufferlist cached_snap_trace;
+
   xlist<CInode*> inodes_with_caps;             // for efficient realm splits
   map<int, xlist<Capability*> > client_caps;   // to identify clients who need snap notifications
 
@@ -165,6 +167,7 @@ struct SnapRealm {
                      snapid_t first, snapid_t last);
   void get_snap_info(map<snapid_t,SnapInfo*>& infomap, snapid_t first=0, snapid_t last=CEPH_NOSNAP);
 
+  const bufferlist& get_snap_trace();
   void build_snap_trace(bufferlist& snapbl);
 
   const string& get_snapname(snapid_t snapid, inodeno_t atino);

diff --git a/src/messages/MClientCaps.h b/src/messages/MClientCaps.h
index 0cba925ae9e16f7bb1ce091e0bfe64eba6463ba1..fc832eeeec15d9f53b0ff0a295d6bf9045cbb487 100644 (file)
--- a/src/messages/MClientCaps.h
+++ b/src/messages/MClientCaps.h
@@ -80,6 +80,8 @@ class MClientCaps : public Message {
     head.gid = inode.gid;
     head.mode = inode.mode;
 
+    head.nlink = inode.nlink;
+
     head.xattr_len = 0; // FIXME
 
     head.layout = inode.layout;

diff --git a/src/messages/MClientReply.h b/src/messages/MClientReply.h
index 678b18396f29844d371c1a6f2ba5875f3d80964c..84bff4335d0367cdaaad1a25d4c4c660504aa50a 100644 (file)
--- a/src/messages/MClientReply.h
+++ b/src/messages/MClientReply.h
@@ -92,6 +92,8 @@ struct DirStat {
 struct InodeStat {
   vinodeno_t vino;
   version_t version;
+  ceph_mds_reply_cap cap;
+
   ceph_file_layout layout;
   unsigned mode, uid, gid, nlink, rdev;
   loff_t size, max_size;
@@ -179,16 +181,7 @@ public:
 
   int get_result() { return (__s32)(__u32)st.result; }
 
-  unsigned get_file_caps() { return st.file_caps; }
-  unsigned get_file_caps_seq() { return st.file_caps_seq; }
-  unsigned get_file_caps_mseq() { return st.file_caps_mseq; }
-  //uint64_t get_file_data_version() { return st.file_data_version; }
-  
   void set_result(int r) { st.result = r; }
-  void set_file_caps(unsigned char c) { st.file_caps = c; }
-  void set_file_caps_seq(capseq_t s) { st.file_caps_seq = s; }
-  void set_file_caps_mseq(capseq_t s) { st.file_caps_mseq = s; }
-  //void set_file_data_version(uint64_t v) { st.file_data_version = v; }
 
   void set_unsafe() { st.safe = 0; }