* be explicit about journal append ack/commit safety

git-svn-id: https://ceph.svn.sf.net/svnroot/ceph@1401 29311d96-e01e-0410-9327-a35deaab8ce9

diff --git a/branches/sage/cephmds2/TODO b/branches/sage/cephmds2/TODO
index cd972c0c85335c51cffe2c46ac787dd72c8bfe44..9134e424eb3dae2233f1b1fad96f12d92307ce90 100644 (file)
--- a/branches/sage/cephmds2/TODO
+++ b/branches/sage/cephmds2/TODO
@@ -113,7 +113,6 @@ sage mds
 - statfs?
 
 
-- fix lock caps gather ack versus ambiguous auth
 
 
 foreign rename
@@ -213,9 +212,13 @@ rados snapshots
 
 
 objecter
+- transaction prepare/commit
 - read+floor_lockout
 
 osd/rados
+- transaction prepare/commit
+  - rollback
+  - rollback logging (to fix slow prepare vs rollback race)
 - read+floor_lockout for clean STOGITH-like/fencing semantics after failover.
 - separate out replication code into a PG class, to pave way for RAID
 

diff --git a/branches/sage/cephmds2/config.cc b/branches/sage/cephmds2/config.cc
index c91c756cbd6f9324aeac66df560632d5f4996893..01314f5bfbfd2c685fffbe1c10cf167e65609c7d 100644 (file)
--- a/branches/sage/cephmds2/config.cc
+++ b/branches/sage/cephmds2/config.cc
@@ -150,6 +150,7 @@ md_config_t g_conf = {
 
   // --- journaler ---
   journaler_allow_split_entries: true,
+  journaler_safe: false,  // wait for COMMIT on journal writes
 
   // --- mds ---
   mds_cache_size: MDS_CACHE_SIZE,
@@ -165,7 +166,6 @@ md_config_t g_conf = {
   mds_log_max_trimming: 10000,
   mds_log_read_inc: 1<<20,
   mds_log_pad_entry: 128,//256,//64,
-  mds_log_before_reply: true,
   mds_log_flush_on_shutdown: true,
   mds_log_import_map_interval: 1024*1024,  // frequency (in bytes) of EImportMap in log
   mds_log_eopen_size: 100,   // # open inodes per log entry
@@ -232,7 +232,7 @@ md_config_t g_conf = {
   ebofs_cloneable: false,
   ebofs_verify: false,
   ebofs_commit_ms:      2000,       // 0 = no forced commit timeout (for debugging/tracing)
-  ebofs_idle_commit_ms: 100,        // 0 = no idle detection.  use this -or- bdev_idle_kick_after_ms
+  ebofs_idle_commit_ms: 20,         // 0 = no idle detection.  use this -or- bdev_idle_kick_after_ms
   ebofs_oc_size:        10000,      // onode cache
   ebofs_cc_size:        10000,      // cnode cache
   ebofs_bc_size:        (80 *256), // 4k blocks, *256 for MB
@@ -564,6 +564,9 @@ void parse_config_options(std::vector<char*>& args)
     else if (strcmp(args[i], "--objecter_buffer_uncommitted") == 0) 
       g_conf.objecter_buffer_uncommitted = atoi(args[++i]);
 
+    else if (strcmp(args[i], "--journaler_safe") == 0) 
+      g_conf.journaler_safe = atoi(args[++i]);
+
     else if (strcmp(args[i], "--mds_cache_size") == 0) 
       g_conf.mds_cache_size = atoi(args[++i]);
 
@@ -574,8 +577,6 @@ void parse_config_options(std::vector<char*>& args)
 
     else if (strcmp(args[i], "--mds_log") == 0) 
       g_conf.mds_log = atoi(args[++i]);
-    else if (strcmp(args[i], "--mds_log_before_reply") == 0) 
-      g_conf.mds_log_before_reply = atoi(args[++i]);
     else if (strcmp(args[i], "--mds_log_max_len") == 0) 
       g_conf.mds_log_max_len = atoi(args[++i]);
     else if (strcmp(args[i], "--mds_log_read_inc") == 0) 

diff --git a/branches/sage/cephmds2/config.h b/branches/sage/cephmds2/config.h
index a9fde4f7a040610beb063e5c1b6e5072a08cd894..b8c6d20d7d6a50ac99cbe32b4d24fd702250f1f4 100644 (file)
--- a/branches/sage/cephmds2/config.h
+++ b/branches/sage/cephmds2/config.h
@@ -151,7 +151,8 @@ struct md_config_t {
 
   // journaler
   bool  journaler_allow_split_entries;
-
+  bool  journaler_safe;
+  
   // mds
   int   mds_cache_size;
   float mds_cache_mid;
@@ -166,7 +167,6 @@ struct md_config_t {
   int mds_log_max_trimming;
   int mds_log_read_inc;
   int mds_log_pad_entry;
-  bool  mds_log_before_reply;
   bool  mds_log_flush_on_shutdown;
   off_t mds_log_import_map_interval;
   int mds_log_eopen_size;

diff --git a/branches/sage/cephmds2/mds/MDBalancer.cc b/branches/sage/cephmds2/mds/MDBalancer.cc
index 5e6bac91671edbe0386b42db1614a62c85e000c1..75e8872a4dc0a4c01e508bcd2a3a4f5766f00e5b 100644 (file)
--- a/branches/sage/cephmds2/mds/MDBalancer.cc
+++ b/branches/sage/cephmds2/mds/MDBalancer.cc
@@ -715,7 +715,7 @@ void MDBalancer::hit_inode(CInode *in, int type)
     anydom = in->popularity[MDS_POP_ANYDOM].pop[type].hit();
   }
 
-  dout(-20) << "hit_inode " << type << " pop " << me << " me, "
+  dout(20) << "hit_inode " << type << " pop " << me << " me, "
            << nested << " nested, "
            << curdom << " curdom, " 
            << anydom << " anydom" 
@@ -737,7 +737,7 @@ void MDBalancer::hit_dir(CDir *dir, int type)
   if (g_conf.num_mds > 2 &&             // FIXME >2 thing
       !dir->inode->is_root() &&        // not root (for now at least)
       dir->is_auth()) {
-    dout(-20) << "hit_dir " << type << " pop " << v << " me "
+    dout(20) << "hit_dir " << type << " pop " << v << " me "
              << *dir << endl;
 
     // hash this dir?  (later?)
@@ -766,7 +766,7 @@ void MDBalancer::hit_recursive(CDir *dir, int type)
   // replicate?
   float dir_pop = dir->popularity[MDS_POP_CURDOM].pop[type].get();    // hmm??
 
-  dout(-20) << "hit_recursive " << type << " pop " << dir_pop << " curdom " << *dir << endl;
+  dout(20) << "hit_recursive " << type << " pop " << dir_pop << " curdom " << *dir << endl;
 
   if (dir->is_auth()) {
     if (!dir->is_rep() &&
@@ -776,7 +776,7 @@ void MDBalancer::hit_recursive(CDir *dir, int type)
       rd_adj = rdp / mds->get_mds_map()->get_num_mds() - rdp; 
       rd_adj /= 2.0;  // temper somewhat
 
-      dout(1) << "replicating dir " << *dir << " pop " << dir_pop << " .. rdp " << rdp << " adj " << rd_adj << endl;
+      dout(2) << "replicating dir " << *dir << " pop " << dir_pop << " .. rdp " << rdp << " adj " << rd_adj << endl;
           
       dir->dir_rep = CDir::REP_ALL;
       mds->mdcache->send_dir_updates(dir, true);
@@ -789,7 +789,7 @@ void MDBalancer::hit_recursive(CDir *dir, int type)
         dir->is_rep() &&
         dir_pop < g_conf.mds_bal_unreplicate_threshold) {
       // unreplicate
-      dout(1) << "unreplicating dir " << *dir << " pop " << dir_pop << endl;
+      dout(2) << "unreplicating dir " << *dir << " pop " << dir_pop << endl;
       
       dir->dir_rep = CDir::REP_NONE;
       mds->mdcache->send_dir_updates(dir);

diff --git a/branches/sage/cephmds2/osdc/Journaler.cc b/branches/sage/cephmds2/osdc/Journaler.cc
index cfb2828dc99f377bc17b2998e685aa0d700cb0f3..9227e5f227dd77598de889ab397d7f33afe44c42 100644 (file)
--- a/branches/sage/cephmds2/osdc/Journaler.cc
+++ b/branches/sage/cephmds2/osdc/Journaler.cc
@@ -169,7 +169,8 @@ void Journaler::write_head(Context *oncommit)
   bufferlist bl;
   bl.append((char*)&last_written, sizeof(last_written));
   filer.write(inode, 0, bl.length(), bl, 0, 
-             0, new C_WriteHead(this, last_written, oncommit));
+             0, 
+             new C_WriteHead(this, last_written, oncommit));
 }
 
 void Journaler::_finish_write_head(Header &wrote, Context *oncommit)
@@ -293,8 +294,10 @@ void Journaler::flush(Context *onsync)
   dout(10) << "flush flushing " << flush_pos << "~" << len << endl;
 
   // submit write for anything pending
+  // flush _start_ pos to _finish_flush
   filer.write(inode, flush_pos, len, write_buf, 0,
-             new C_Flush(this, flush_pos), 0);  // flush _start_ pos to _finish_flush
+             g_conf.journaler_safe ? 0:new C_Flush(this, flush_pos),  // on ACK
+             g_conf.journaler_safe ?   new C_Flush(this, flush_pos):0); // on COMMIT
   pending_flush[flush_pos] = g_clock.now();
   
   // adjust pointers