Merge pull request ceph#61214 from aclamk/wip-aclamk-bs-fragmentation-health

os/bluestore: Add health warning for bluestore fragmentation

Author: aclamk

src/common/options/global.yaml.in

@@ -5490,6 +5490,29 @@ options:
   desc: Enable health indication when spurious read errors are observed by OSD
   default: true
   with_legacy: true
+- name: bluestore_warn_on_free_fragmentation
+  type: float
+  level: basic
+  desc: Level at which disk free fragmentation causes health warning. Set "1" to disable.
+    This is same value as admin command "bluestore allocator score block".
+  default: 0.8
+  with_legacy: false
+  flags:
+  - runtime
+  see_also:
+  - bluestore_fragmentation_check_period
+- name: bluestore_fragmentation_check_period
+  type: uint
+  level: basic
+  desc: The period to perform bluestore free fragmentation check.
+    Checking fragmentation is usually almost immediate. For highly fragmented storage,
+    it can take several miliseconds. It can cause a stall to a write operation.
+  default: 3600
+  with_legacy: false
+  flags:
+  - runtime
+  see_also:
+  - bluestore_warn_on_free_fragmentation
 - name: bluestore_slow_ops_warn_lifetime
   type: uint
   level: advanced

src/os/CMakeLists.txt

@@ -43,7 +43,8 @@ endif()
 add_library(os STATIC ${libos_srcs})
 target_link_libraries(os
   legacy-option-headers
-  blk)
+  blk
+  ${FMT_LIB})
 
 target_link_libraries(os heap_profiler kv)
 

src/os/bluestore/BlueStore.cc

@@ -5278,7 +5278,7 @@ void BlueStore::Collection::split_cache(
 // MempoolThread
 
 #undef dout_prefix
-#define dout_prefix *_dout << "bluestore.MempoolThread(" << this << ") "
+#define dout_prefix *_dout << "bluestore.MempoolThread "
 #undef dout_context
 #define dout_context store->cct
 
@@ -5405,6 +5405,23 @@ void *BlueStore::MempoolThread::entry()
     interval_stats_trim = false;
 
     store->refresh_perf_counters();
+    uint64_t period = store->cct->_conf.get_val<uint64_t>("bluestore_fragmentation_check_period");
+    if (period != 0 && store->alloc) {
+      auto now = mono_clock::now();
+      timespan elapsed = now - last_fragmentation_check;
+      if (elapsed > make_timespan(period)) {
+        last_fragmentation_check = now;
+        double score;
+        score = store->alloc->get_fragmentation_score();
+        store->logger->set(l_bluestore_fragmentation, score * 1e6);
+        now = mono_clock::now();
+        elapsed = now - last_fragmentation_check;
+        auto seconds = elapsed.count() * 1e-9;
+        dout(0) << std::fixed << std::setprecision(6)
+          << "fragmentation_score=" << score << " took=" << seconds << "s" << dendl;
+      }
+    }
+
     auto wait = ceph::make_timespan(
       store->cct->_conf->bluestore_cache_trim_interval);
     cond.wait_for(l, wait);
@@ -15110,9 +15127,6 @@ void BlueStore::_kv_finalize_thread()
 
       // this is as good a place as any ...
       _reap_collections();
-      logger->set(l_bluestore_fragmentation,
-	(uint64_t)(alloc ? alloc->get_fragmentation() * 1000 : 0));
-
       log_latency("kv_final",
 	l_bluestore_kv_final_lat,
 	mono_clock::now() - start,
@@ -19106,6 +19120,11 @@ void BlueStore::_log_alerts(osd_alert_list_t& alerts)
       "BLUESTORE_NO_COMPRESSION",
       s0);
   }
+  if (logger->get(l_bluestore_fragmentation) >
+    cct->_conf.get_val<double>("bluestore_warn_on_free_fragmentation") * 1e6) {
+    alerts.emplace("BLUESTORE_FREE_FRAGMENTATION",
+      fmt::format("{0:.6f}", logger->get(l_bluestore_fragmentation) * 1e-6));
+  }
 }
 
 void BlueStore::_collect_allocation_stats(uint64_t need, uint32_t alloc_size,

src/os/bluestore/BlueStore.h

@@ -2772,6 +2772,8 @@ class BlueStore : public ObjectStore,
   private:
     void _update_cache_settings();
     void _resize_shards(bool interval_stats);
+
+    mono_clock::time_point last_fragmentation_check;
   } mempool_thread;
 
 #ifdef WITH_BLKIN