mon,osd,osdc: refactor snap trimming (phase 1)

qa/suites/rados/singleton/all/thrash-eio.yaml

@@ -30,6 +30,7 @@ tasks:
     - \(SLOW_OPS\)
     - \(PG_
     - \(OSD_
+    - \(OBJECT_
 - thrashosds:
     op_delay: 30
     clean_interval: 120

src/common/options.cc

@@ -1153,6 +1153,10 @@ std::vector<Option> get_global_options() {
     .set_default(true)
     .set_description("Enable POOL_APP_NOT_ENABLED health check"),
 
+    Option("mon_max_snap_prune_per_epoch", Option::TYPE_UINT, Option::LEVEL_ADVANCED)
+    .set_default(100)
+    .set_description("Max number of pruned snaps we will process in a single OSDMap epoch"),
+
     Option("mon_min_osdmap_epochs", Option::TYPE_INT, Option::LEVEL_ADVANCED)
     .set_default(500)
     .set_description(""),
@@ -2453,6 +2457,10 @@ std::vector<Option> get_global_options() {
     .set_default(.9)
     .set_description(""),
 
+    Option("osd_max_snap_prune_intervals_per_epoch", Option::TYPE_UINT, Option::LEVEL_DEV)
+    .set_default(512)
+    .set_description("Max number of snap intervals to report to mgr in pg_stat_t"),
+
     Option("osd_default_data_pool_replay_window", Option::TYPE_INT, Option::LEVEL_ADVANCED)
     .set_default(45)
     .set_description(""),

src/include/interval_set.h

@@ -76,6 +76,9 @@ class interval_set {
         T get_len() const {
                 return _iter->second;
         }
+        T get_end() const {
+                return _iter->first + _iter->second;
+        }
 
         // Set the interval length.
         void set_len(T len) {
@@ -135,6 +138,9 @@ class interval_set {
         T get_start() const {
                 return _iter->first;
         }
+        T get_end() const {
+                return _iter->first + _iter->second;
+        }
 
         // Return the interval length.
         T get_len() const {

src/include/mempool.h

@@ -24,6 +24,8 @@
 #include <mutex>
 #include <atomic>
 #include <typeinfo>
+#include <boost/container/flat_set.hpp>
+#include <boost/container/flat_map.hpp>
 
 #include <common/Formatter.h>
 #include "include/assert.h"
@@ -405,6 +407,13 @@ class pool_allocator {
     template<typename k, typename cmp = std::less<k> >			\
     using set = std::set<k,cmp,pool_allocator<k>>;			\
                                                                         \
+    template<typename k, typename cmp = std::less<k> >			\
+    using flat_set = boost::container::flat_set<k,cmp,pool_allocator<k>>; \
+									\
+    template<typename k, typename v, typename cmp = std::less<k> >	\
+    using flat_map = boost::container::flat_map<k,v,cmp,		\
+						pool_allocator<std::pair<k,v>>>; \
+                                                                        \
     template<typename v>						\
     using list = std::list<v,pool_allocator<v>>;			\
                                                                         \

src/include/rados.h

@@ -157,6 +157,7 @@ extern const char *ceph_osd_state_name(int s);
 #define CEPH_OSDMAP_REQUIRE_LUMINOUS (1<<18) /* require l for booting osds */
 #define CEPH_OSDMAP_RECOVERY_DELETES (1<<19) /* deletes performed during recovery instead of peering */
 #define CEPH_OSDMAP_PURGED_SNAPDIRS  (1<<20) /* osds have converted snapsets */
+#define CEPH_OSDMAP_NOSNAPTRIM       (1<<21) /* disable snap trimming */
 
 /* these are hidden in 'ceph status' view */
 #define CEPH_OSDMAP_SEMIHIDDEN_FLAGS (CEPH_OSDMAP_REQUIRE_JEWEL|	\

src/include/types.h

@@ -54,6 +54,8 @@ extern "C" {
 #include <string>
 #include <list>
 #include <set>
+#include <boost/container/flat_set.hpp>
+#include <boost/container/flat_map.hpp>
 #include <map>
 #include <vector>
 #include <iostream>
@@ -106,6 +108,10 @@ inline ostream& operator<<(ostream& out, const list<A,Alloc>& ilist);
 template<class A, class Comp, class Alloc>
 inline ostream& operator<<(ostream& out, const set<A, Comp, Alloc>& iset);
 template<class A, class Comp, class Alloc>
+inline ostream& operator<<(ostream& out, const boost::container::flat_set<A, Comp, Alloc>& iset);
+template<class A, class B, class Comp, class Alloc>
+inline ostream& operator<<(ostream& out, const boost::container::flat_map<A, B, Comp, Alloc>& iset);
+template<class A, class Comp, class Alloc>
 inline ostream& operator<<(ostream& out, const multiset<A,Comp,Alloc>& iset);
 template<class A, class B, class Comp, class Alloc>
 inline ostream& operator<<(ostream& out, const map<A,B,Comp,Alloc>& m);
@@ -166,6 +172,28 @@ inline ostream& operator<<(ostream& out, const set<A, Comp, Alloc>& iset) {
   return out;
 }
 
+template<class A, class Comp, class Alloc>
+inline ostream& operator<<(ostream& out, const boost::container::flat_set<A, Comp, Alloc>& iset) {
+  for (auto it = iset.begin();
+       it != iset.end();
+       ++it) {
+    if (it != iset.begin()) out << ",";
+    out << *it;
+  }
+  return out;
+}
+
+template<class A, class B, class Comp, class Alloc>
+inline ostream& operator<<(ostream& out, const boost::container::flat_map<A, B, Comp, Alloc>& m) {
+  for (auto it = m.begin();
+       it != m.end();
+       ++it) {
+    if (it != m.begin()) out << ",";
+    out << it->first << "=" << it->second;
+  }
+  return out;
+}
+
 template<class A, class Comp, class Alloc>
 inline ostream& operator<<(ostream& out, const multiset<A,Comp,Alloc>& iset) {
   for (auto it = iset.begin();

src/mds/SnapServer.cc

@@ -50,9 +50,9 @@ void SnapServer::reset_state()
 	    // needing removal, skip.
 	    continue;
 	  }
-	  if (!pi->removed_snaps.empty() &&
-	      pi->removed_snaps.range_end() > first_free)
-	    first_free = pi->removed_snaps.range_end();
+	  if (pi->snap_seq > first_free) {
+	    first_free = pi->snap_seq;
+	  }
 	}
       });
     if (first_free > last_snap)

src/messages/MOSDMap.h

@@ -22,14 +22,20 @@
 
 class MOSDMap : public Message {
 
-  static const int HEAD_VERSION = 3;
+  static const int HEAD_VERSION = 4;
+  static const int COMPAT_VERSION = 3;
 
  public:
   uuid_d fsid;
   map<epoch_t, bufferlist> maps;
   map<epoch_t, bufferlist> incremental_maps;
   epoch_t oldest_map =0, newest_map = 0;
 
+  // if we are fetching maps from the mon and have to jump a gap
+  // (client's next needed map is older than mon's oldest) we can
+  // share removed snaps from the gap here.
+  mempool::osdmap::map<int64_t,OSDMap::snap_interval_set_t> gap_removed_snaps;
+
   epoch_t get_first() const {
     epoch_t e = 0;
     map<epoch_t, bufferlist>::const_iterator i = maps.begin();
@@ -56,9 +62,9 @@ class MOSDMap : public Message {
   }
 
 
-  MOSDMap() : Message(CEPH_MSG_OSD_MAP, HEAD_VERSION) { }
+  MOSDMap() : Message(CEPH_MSG_OSD_MAP, HEAD_VERSION, COMPAT_VERSION) { }
   MOSDMap(const uuid_d &f)
-    : Message(CEPH_MSG_OSD_MAP, HEAD_VERSION),
+    : Message(CEPH_MSG_OSD_MAP, HEAD_VERSION, COMPAT_VERSION),
       fsid(f),
       oldest_map(0), newest_map(0) { }
 private:
@@ -78,9 +84,13 @@ class MOSDMap : public Message {
       oldest_map = 0;
       newest_map = 0;
     }
+    if (header.version >= 4) {
+      ::decode(gap_removed_snaps, p);
+    }
   }
   void encode_payload(uint64_t features) override {
     header.version = HEAD_VERSION;
+    header.compat_version = COMPAT_VERSION;
     ::encode(fsid, payload);
     if ((features & CEPH_FEATURE_PGID64) == 0 ||
 	(features & CEPH_FEATURE_PGPOOL3) == 0 ||
@@ -93,6 +103,7 @@ class MOSDMap : public Message {
 	header.version = 1;  // old old_client version
       else if ((features & CEPH_FEATURE_OSDENC) == 0)
 	header.version = 2;  // old pg_pool_t
+      header.compat_version = 0;
 
       // reencode maps using old format
       //
@@ -138,13 +149,18 @@ class MOSDMap : public Message {
       ::encode(oldest_map, payload);
       ::encode(newest_map, payload);
     }
+    if (header.version >= 4) {
+      ::encode(gap_removed_snaps, payload);
+    }
   }
 
   const char *get_type_name() const override { return "osdmap"; }
   void print(ostream& out) const override {
     out << "osd_map(" << get_first() << ".." << get_last();
     if (oldest_map || newest_map)
       out << " src has " << oldest_map << ".." << newest_map;
+    if (!gap_removed_snaps.empty())
+      out << " +gap_removed_snaps";
     out << ")";
   }
 };

src/mgr/ClusterState.h

@@ -103,6 +103,14 @@ class ClusterState
     return std::forward<Callback>(cb)(pg_map, std::forward<Args>(args)...);
   }
 
+  template<typename Callback, typename...Args>
+  auto with_mutable_pgmap(Callback&& cb, Args&&...args) ->
+    decltype(cb(pg_map, std::forward<Args>(args)...))
+  {
+    Mutex::Locker l(lock);
+    return std::forward<Callback>(cb)(pg_map, std::forward<Args>(args)...);
+  }
+
   template<typename... Args>
   void with_monmap(Args &&... args) const
   {

src/mgr/DaemonServer.cc

@@ -1422,7 +1422,7 @@ void DaemonServer::send_report()
   auto m = new MMonMgrReport();
   py_modules.get_health_checks(&m->health_checks);
 
-  cluster_state.with_pgmap([&](const PGMap& pg_map) {
+  cluster_state.with_mutable_pgmap([&](PGMap& pg_map) {
       cluster_state.update_delta_stats();
 
       if (pending_service_map.epoch) {

src/mon/MgrStatMonitor.cc

@@ -188,6 +188,11 @@ bool MgrStatMonitor::prepare_report(MonOpRequestRef op)
   }
   dout(10) << __func__ << " " << pending_digest << ", "
 	   << pending_health_checks.checks.size() << " health checks" << dendl;
+  dout(20) << "pending_digest:\n";
+  JSONFormatter jf(true);
+  pending_digest.dump(&jf);
+  jf.flush(*_dout);
+  *_dout << dendl;
   return true;
 }
 

src/mon/MgrStatMonitor.h

@@ -73,6 +73,10 @@ class MgrStatMonitor : public PaxosService {
     return nullptr;
   }
 
+  const PGMapDigest& get_digest() {
+    return digest;
+  }
+
   ceph_statfs get_statfs(OSDMap& osdmap,
 			 boost::optional<int64_t> data_pool) const {
     return digest.get_statfs(osdmap, data_pool);

src/mon/MonCommands.h

@@ -728,11 +728,11 @@ COMMAND("osd erasure-code-profile ls", \
 	"list all erasure code profiles", \
 	"osd", "r", "cli,rest")
 COMMAND("osd set " \
-	"name=key,type=CephChoices,strings=full|pause|noup|nodown|noout|noin|nobackfill|norebalance|norecover|noscrub|nodeep-scrub|notieragent|sortbitwise|recovery_deletes|require_jewel_osds|require_kraken_osds " \
+	"name=key,type=CephChoices,strings=full|pause|noup|nodown|noout|noin|nobackfill|norebalance|norecover|noscrub|nodeep-scrub|notieragent|nosnaptrim|sortbitwise|recovery_deletes|require_jewel_osds|require_kraken_osds " \
 	"name=sure,type=CephChoices,strings=--yes-i-really-mean-it,req=false", \
 	"set <key>", "osd", "rw", "cli,rest")
 COMMAND("osd unset " \
-	"name=key,type=CephChoices,strings=full|pause|noup|nodown|noout|noin|nobackfill|norebalance|norecover|noscrub|nodeep-scrub|notieragent", \
+	"name=key,type=CephChoices,strings=full|pause|noup|nodown|noout|noin|nobackfill|norebalance|norecover|noscrub|nodeep-scrub|notieragent|nosnaptrim", \
 	"unset <key>", "osd", "rw", "cli,rest")
 COMMAND("osd require-osd-release "\
 	"name=release,type=CephChoices,strings=luminous|mimic " \