os/bluestore: make cache settings process-wide

do_cmake.sh

@@ -8,9 +8,13 @@ mkdir build
 cd build
 cmake $@ ..
 
+# minimal config to find plugins
 cat <<EOF > ceph.conf
 plugin dir = lib
 erasure code dir = lib
 EOF
 
+# give vstart a (hopefully) unique mon port to start with
+echo $(( RANDOM % 1000 + 40000 )) > .ceph_port
+
 echo done.

src/os/bluestore/BlueStore.cc

@@ -1695,7 +1695,7 @@ void BlueStore::ExtentMap::decode_some(bufferlist& bl)
   uint64_t prev_len = 0;
   unsigned n = 0;
   while (!p.end()) {
-    Extent *le = new Extent();
+    Extent *le = new Extent(onode->c->cache);
     uint64_t blobid;
     small_decode_varint(blobid, p);
     if ((blobid & BLOBID_FLAG_CONTIGUOUS) == 0) {
@@ -2064,7 +2064,8 @@ BlueStore::Collection::Collection(BlueStore *ns, Cache *c, coll_t cid)
     exists(true),
     // size the shared blob hash table as a ratio of the onode cache size.
     shared_blob_set(MAX(16,
-			g_conf->bluestore_onode_cache_size *
+			g_conf->bluestore_onode_cache_size /
+			store->cache_shards.size() *
 			g_conf->bluestore_shared_blob_hash_table_size_ratio)),
     onode_map(c)
 {
@@ -2190,6 +2191,13 @@ BlueStore::OnodeRef BlueStore::Collection::get_onode(
   return onode_map.add(oid, o);
 }
 
+void BlueStore::Collection::trim_cache()
+{
+  cache->trim(
+    g_conf->bluestore_onode_cache_size / store->cache_shards.size(),
+    g_conf->bluestore_buffer_cache_size / store->cache_shards.size());
+}
+
 
 
 // =======================================================
@@ -2414,7 +2422,7 @@ void BlueStore::_init_logger()
     "Sum for wal write op");
   b.add_u64(l_bluestore_wal_write_bytes, "wal_write_bytes",
     "Sum for wal write bytes");
-  b.add_u64(l_bluestore_write_penalty_read_ops, " write_penalty_read_ops",
+  b.add_u64(l_bluestore_write_penalty_read_ops, "write_penalty_read_ops",
     "Sum for write penalty read ops");
   b.add_u64(l_bluestore_allocated, "bluestore_allocated",
     "Sum for allocated bytes");
@@ -2427,10 +2435,20 @@ void BlueStore::_init_logger()
   b.add_u64(l_bluestore_compressed_original, "bluestore_compressed_original",
     "Sum for original bytes that were compressed");
 
+  b.add_u64(l_bluestore_onodes, "bluestore_onodes",
+	    "Number of onodes in cache");
   b.add_u64(l_bluestore_onode_hits, "bluestore_onode_hits",
     "Sum for onode-lookups hit in the cache");
   b.add_u64(l_bluestore_onode_misses, "bluestore_onode_misses",
     "Sum for onode-lookups missed in the cache");
+  b.add_u64(l_bluestore_extents, "bluestore_extents",
+	    "Number of extents in cache");
+  b.add_u64(l_bluestore_blobs, "bluestore_blobs",
+	    "Number of blobs in cache");
+  b.add_u64(l_bluestore_buffers, "bluestore_buffers",
+	    "Number of buffers in cache");
+  b.add_u64(l_bluestore_buffer_bytes, "bluestore_buffer_bytes",
+	    "Number of buffer bytes in cache");
   b.add_u64(l_bluestore_buffer_hit_bytes, "bluestore_buffer_hit_bytes",
     "Sum for bytes of read hit in the cache");
   b.add_u64(l_bluestore_buffer_miss_bytes, "bluestore_buffer_miss_bytes",
@@ -4508,6 +4526,24 @@ void BlueStore::_reap_collections()
   }
 }
 
+void BlueStore::_update_cache_logger()
+{
+  uint64_t num_onodes = 0;
+  uint64_t num_extents = 0;
+  uint64_t num_blobs = 0;
+  uint64_t num_buffers = 0;
+  uint64_t num_buffer_bytes = 0;
+  for (auto c : cache_shards) {
+    c->add_stats(&num_onodes, &num_extents, &num_blobs,
+		 &num_buffers, &num_buffer_bytes);
+  }
+  logger->set(l_bluestore_onodes, num_onodes);
+  logger->set(l_bluestore_extents, num_extents);
+  logger->set(l_bluestore_blobs, num_blobs);
+  logger->set(l_bluestore_buffers, num_buffers);
+  logger->set(l_bluestore_buffer_bytes, num_buffer_bytes);
+}
+
 // ---------------
 // read operations
 
@@ -4540,10 +4576,7 @@ bool BlueStore::exists(CollectionHandle &c_, const ghobject_t& oid)
       r = false;
   }
 
-  c->cache->trim(
-    g_conf->bluestore_onode_cache_size,
-    g_conf->bluestore_buffer_cache_size);
-
+  c->trim_cache();
   return r;
 }
 
@@ -4581,9 +4614,7 @@ int BlueStore::stat(
     st->st_nlink = 1;
   }
 
-  c->cache->trim(
-    g_conf->bluestore_onode_cache_size,
-    g_conf->bluestore_buffer_cache_size);
+  c->trim_cache();
   int r = 0;
   if (_debug_mdata_eio(oid)) {
     r = -EIO;
@@ -4643,9 +4674,7 @@ int BlueStore::read(
 
  out:
   assert(allow_eio || r != -EIO);
-  c->cache->trim(
-    g_conf->bluestore_onode_cache_size,
-    g_conf->bluestore_buffer_cache_size);
+  c->trim_cache();
   if (r == 0 && _debug_data_eio(oid)) {
     r = -EIO;
     derr << __func__ << " " << c->cid << " " << oid << " INJECT EIO" << dendl;
@@ -5040,9 +5069,7 @@ int BlueStore::fiemap(
   }
 
  out:
-  c->cache->trim(
-    g_conf->bluestore_onode_cache_size,
-    g_conf->bluestore_buffer_cache_size);
+  c->trim_cache();
   ::encode(m, bl);
   dout(20) << __func__ << " 0x" << std::hex << offset << "~" << length
 	   << " size = 0x(" << m << ")" << std::dec << dendl;
@@ -5091,9 +5118,7 @@ int BlueStore::getattr(
     r = 0;
   }
  out:
-  c->cache->trim(
-    g_conf->bluestore_onode_cache_size,
-    g_conf->bluestore_buffer_cache_size);
+  c->trim_cache();
   if (r == 0 && _debug_mdata_eio(oid)) {
     r = -EIO;
     derr << __func__ << " " << c->cid << " " << oid << " INJECT EIO" << dendl;
@@ -5139,9 +5164,7 @@ int BlueStore::getattrs(
   }
 
  out:
-  c->cache->trim(
-    g_conf->bluestore_onode_cache_size,
-    g_conf->bluestore_buffer_cache_size);
+  c->trim_cache();
   if (r == 0 && _debug_mdata_eio(oid)) {
     r = -EIO;
     derr << __func__ << " " << c->cid << " " << oid << " INJECT EIO" << dendl;
@@ -5322,9 +5345,7 @@ int BlueStore::collection_list(
   }
 
  out:
-  c->cache->trim(
-    g_conf->bluestore_onode_cache_size,
-    g_conf->bluestore_buffer_cache_size);
+  c->trim_cache();
   dout(10) << __func__ << " " << c->cid
 	   << " start " << start << " end " << end << " max " << max
 	   << " = " << r << ", ls.size() = " << ls->size()
@@ -6172,9 +6193,7 @@ void BlueStore::_osr_reap_done(OpSequencer *osr)
   }
 
   if (c) {
-    c->cache->trim(
-      g_conf->bluestore_onode_cache_size,
-      g_conf->bluestore_buffer_cache_size);
+    c->trim_cache();
   }
 }
 
@@ -6357,6 +6376,8 @@ void BlueStore::_kv_sync_thread()
       // this is as good a place as any ...
       _reap_collections();
 
+      _update_cache_logger();
+
       if (bluefs) {
 	if (!bluefs_gift_extents.empty()) {
 	  _commit_bluefs_freespace(bluefs_gift_extents);

src/os/bluestore/BlueStore.h

@@ -72,8 +72,13 @@ enum {
   l_bluestore_compressed,
   l_bluestore_compressed_allocated,
   l_bluestore_compressed_original,
+  l_bluestore_onodes,
   l_bluestore_onode_hits,
   l_bluestore_onode_misses,
+  l_bluestore_extents,
+  l_bluestore_blobs,
+  l_bluestore_buffers,
+  l_bluestore_buffer_bytes,
   l_bluestore_buffer_hit_bytes,
   l_bluestore_buffer_miss_bytes,
   l_bluestore_write_big,
@@ -204,10 +209,17 @@ class BlueStore : public ObjectStore,
     Cache *cache;
     map<uint64_t, state_list_t> writing_map;
 
-    BufferSpace(Cache *c) : cache(c) {}
+    BufferSpace(Cache *c) : cache(c) {
+      if (cache) {
+	cache->add_blob();
+      }
+    }
     ~BufferSpace() {
       assert(buffer_map.empty());
       assert(writing_map.empty());
+      if (cache) {
+	cache->rm_blob();
+      }
     }
 
     void _add_buffer(Buffer *b, int level, Buffer *near) {
@@ -509,10 +521,24 @@ class BlueStore : public ObjectStore,
     uint8_t  blob_depth;              /// blob overlapping count
     BlobRef blob;                     ///< the blob with our data
 
-    explicit Extent() {}
-    explicit Extent(uint32_t lo) : logical_offset(lo) {}
+    /// ctor for lookup only
+    explicit Extent(uint32_t lo) : logical_offset(lo) { }
+    /// ctor for delayed intitialization (see decode_some())
+    explicit Extent(Cache *cache) {
+      cache->add_extent();
+    }
+    /// ctor for general usage
     Extent(uint32_t lo, uint32_t o, uint32_t l, uint8_t bd, BlobRef& b)
-      : logical_offset(lo), blob_offset(o), length(l), blob_depth(bd), blob(b){}
+      : logical_offset(lo), blob_offset(o), length(l), blob_depth(bd), blob(b) {
+      if (blob) {
+	blob->shared_blob->bc.cache->add_extent();
+      }
+    }
+    ~Extent() {
+      if (blob) {
+	blob->shared_blob->bc.cache->rm_extent();
+      }
+    }
 
     // comparators for intrusive_set
     friend bool operator<(const Extent &a, const Extent &b) {
@@ -699,6 +725,9 @@ class BlueStore : public ObjectStore,
     PerfCounters *logger;
     std::recursive_mutex lock;          ///< protect lru and other structures
 
+    std::atomic<uint64_t> num_extents = {0};
+    std::atomic<uint64_t> num_blobs = {0};
+
     static Cache *create(string type, PerfCounters *logger);
 
     virtual ~Cache() {}
@@ -712,8 +741,27 @@ class BlueStore : public ObjectStore,
     virtual void _adjust_buffer_size(Buffer *b, int64_t delta) = 0;
     virtual void _touch_buffer(Buffer *b) = 0;
 
+    void add_extent() {
+      ++num_extents;
+    }
+    void rm_extent() {
+      --num_extents;
+    }
+
+    void add_blob() {
+      ++num_blobs;
+    }
+    void rm_blob() {
+      --num_blobs;
+    }
+
     virtual void trim(uint64_t onode_max, uint64_t buffer_max) = 0;
 
+    virtual void add_stats(uint64_t *onodes, uint64_t *extents,
+			   uint64_t *blobs,
+			   uint64_t *buffers,
+			   uint64_t *bytes) = 0;
+
 #ifdef DEBUG_CACHE
     virtual void _audit(const char *s) = 0;
 #else
@@ -785,6 +833,18 @@ class BlueStore : public ObjectStore,
 
     void trim(uint64_t onode_max, uint64_t buffer_max) override;
 
+    void add_stats(uint64_t *onodes, uint64_t *extents,
+		   uint64_t *blobs,
+		   uint64_t *buffers,
+		   uint64_t *bytes) override {
+      std::lock_guard<std::recursive_mutex> l(lock);
+      *onodes += onode_lru.size();
+      *extents += num_extents;
+      *blobs += num_blobs;
+      *buffers += buffer_lru.size();
+      *bytes += buffer_size;
+    }
+
 #ifdef DEBUG_CACHE
     void _audit(const char *s) override;
 #endif
@@ -860,6 +920,18 @@ class BlueStore : public ObjectStore,
 
     void trim(uint64_t onode_max, uint64_t buffer_max) override;
 
+    void add_stats(uint64_t *onodes, uint64_t *extents,
+		   uint64_t *blobs,
+		   uint64_t *buffers,
+		   uint64_t *bytes) override {
+      std::lock_guard<std::recursive_mutex> l(lock);
+      *onodes += onode_lru.size();
+      *extents += num_extents;
+      *blobs += num_blobs;
+      *buffers += buffer_hot.size() + buffer_warm_in.size();
+      *bytes += buffer_bytes;
+    }
+
 #ifdef DEBUG_CACHE
     void _audit(const char *s) override;
 #endif
@@ -940,6 +1012,8 @@ class BlueStore : public ObjectStore,
       return false;
     }
 
+    void trim_cache();
+
     Collection(BlueStore *ns, Cache *ca, coll_t c);
   };
   typedef boost::intrusive_ptr<Collection> CollectionRef;
@@ -1414,6 +1488,7 @@ class BlueStore : public ObjectStore,
   CollectionRef _get_collection(const coll_t& cid);
   void _queue_reap_collection(CollectionRef& c);
   void _reap_collections();
+  void _update_cache_logger();
 
   void _assign_nid(TransContext *txc, OnodeRef o);
   uint64_t _assign_blobid(TransContext *txc);