mds: allow disabling batch ops

To address a bug and future ones where batching lookup/getattr does not help
"kick" the MDS in switching state more quickly (e.g. flushing the MDS journal).

Signed-off-by: Patrick Donnelly <pdonnell@ibm.com>

diff --git a/src/common/options/mds.yaml.in b/src/common/options/mds.yaml.in
index 06c12ea671dbe1b37d06e43b10120b23113abc2f..80be6be4d6d5a7b375cbc04e806879fd95459e18 100644 (file)
--- a/src/common/options/mds.yaml.in
+++ b/src/common/options/mds.yaml.in
@@ -793,6 +793,21 @@ options:
   services:
   - mds
   with_legacy: true
+- name: mds_allow_batched_ops
+  type: bool
+  level: advanced
+  desc: allow MDS to batch lookup/getattr RPCs
+  long_desc: >
+    The MDS will batch a lookup or getattr RPC on the same inode when
+    possible to avoid repetitive locks on metadata and to bypass other
+    requests acquiring write locks. Generally, this should only
+    improve performance but this switch exists to provide a means to
+    turn this behavior off for comparison.
+  default: true
+  services:
+  - mds
+  flags:
+  - runtime
 # multiple of size_max that triggers immediate split
 - name: mds_bal_fragment_fast_factor
   type: float

diff --git a/src/mds/MDSRank.cc b/src/mds/MDSRank.cc
index 6f4adb81be491a289242ce2b6552864a6d18b511..59f695a8c0422b3416c4206f3d7ffa87b314ca2c 100644 (file)
--- a/src/mds/MDSRank.cc
+++ b/src/mds/MDSRank.cc
@@ -4070,6 +4070,7 @@ std::vector<std::string> MDSRankDispatcher::get_tracked_keys()
     "fsid",
     "host",
     "mds_allow_async_dirops",
+    "mds_allow_batched_ops",
     "mds_alternate_name_max",
     "mds_bal_export_pin",
     "mds_bal_fragment_dirs",

diff --git a/src/mds/Server.cc b/src/mds/Server.cc
index 65033c052af788cf195f730bac984681f5a11d11..00e04b23bf0b37ade07b56059a9e2767fd0277b5 100644 (file)
--- a/src/mds/Server.cc
+++ b/src/mds/Server.cc
@@ -288,6 +288,7 @@ Server::Server(MDSRank *m, MetricsHandler *metrics_handler) :
 {
   forward_all_requests_to_auth = g_conf().get_val<bool>("mds_forward_all_requests_to_auth");
   replay_unsafe_with_closed_session = g_conf().get_val<bool>("mds_replay_unsafe_with_closed_session");
+  allow_batched_ops = g_conf().get_val<bool>("mds_allow_batched_ops");
   cap_revoke_eviction_timeout = g_conf().get_val<double>("mds_cap_revoke_eviction_timeout");
   max_snaps_per_dir = g_conf().get_val<uint64_t>("mds_max_snaps_per_dir");
   delegate_inos_pct = g_conf().get_val<uint64_t>("mds_client_delegate_inos_pct");
@@ -1386,6 +1387,9 @@ void Server::handle_conf_change(const std::set<std::string>& changed) {
   if (changed.count("mds_forward_all_requests_to_auth")){
     forward_all_requests_to_auth = g_conf().get_val<bool>("mds_forward_all_requests_to_auth");
   }
+  if (changed.count("mds_allow_batched_ops")) {
+    allow_batched_ops = g_conf().get_val<bool>("mds_allow_batched_ops");
+  }
   if (changed.count("mds_cap_revoke_eviction_timeout")) {
     cap_revoke_eviction_timeout = g_conf().get_val<double>("mds_cap_revoke_eviction_timeout");
     dout(20) << __func__ << " cap revoke eviction timeout changed to "
@@ -4220,7 +4224,7 @@ void Server::handle_client_getattr(const MDRequestRef& mdr, bool is_lookup)
   if (mask & CEPH_STAT_RSTAT)
     want_auth = true; // set want_auth for CEPH_STAT_RSTAT mask
 
-  if (!mdr->is_batch_head() && mdr->can_batch()) {
+  if (!mdr->is_batch_head() && allow_batched_ops && mdr->can_batch()) {
     CF_MDS_RetryRequestFactory cf(mdcache, mdr, false);
     int r = mdcache->path_traverse(mdr, cf, mdr->get_filepath(),
                                   (want_auth ? MDS_TRAVERSE_WANT_AUTH : 0),

diff --git a/src/mds/Server.h b/src/mds/Server.h
index 41f73523b32a05a8e441054bc97f1904bd313f05..136794262a8802f543026f61331d969fab19c30b 100644 (file)
--- a/src/mds/Server.h
+++ b/src/mds/Server.h
@@ -607,6 +607,7 @@ private:
   unsigned delegate_inos_pct = 0;
   uint64_t dir_max_entries = 0;
   int64_t bal_fragment_size_max = 0;
+  bool allow_batched_ops = true;
 
   double inject_rename_corrupt_dentry_first = 0.0;