From b17f1222f9ea6a84e2315d9d03da793a994f3a4e Mon Sep 17 00:00:00 2001
From: Sage Weil <sage@redhat.com>
Date: Sun, 5 Nov 2017 10:44:57 -0600
Subject: [PATCH] osd: wait for pgs to catch up in handle_osd_map

Make sure we don't inject maps too far out in front of the PGs ability
to consume them.

Signed-off-by: Sage Weil <sage@redhat.com>
---
 src/common/options.cc |  5 +++++
 src/osd/OSD.cc        | 30 ++++++++++++++++++++++++++++++
 src/osd/OSD.h         | 24 ++++++++++++++++++++++++
 3 files changed, 59 insertions(+)

diff --git a/src/common/options.cc b/src/common/options.cc
index 2b311db7b93..e75f67b08b7 100644
--- a/src/common/options.cc
+++ b/src/common/options.cc
@@ -1911,6 +1911,11 @@ std::vector<Option> get_global_options() {
     .set_default(40)
     .set_description(""),
 
+    Option("osd_pg_epoch_max_lag_factor", Option::TYPE_FLOAT, Option::LEVEL_ADVANCED)
+    .set_default(2.0)
+    .set_description("Max multiple of the map cache that PGs can lag before we throttle map injest")
+    .add_see_also("osd_map_cache_size"),
+
     Option("osd_inject_bad_map_crc_probability", Option::TYPE_FLOAT, Option::LEVEL_DEV)
     .set_default(0)
     .set_description(""),
diff --git a/src/osd/OSD.cc b/src/osd/OSD.cc
index 6841a4017de..ccc11628868 100644
--- a/src/osd/OSD.cc
+++ b/src/osd/OSD.cc
@@ -1980,6 +1980,8 @@ OSD::OSD(CephContext *cct_, ObjectStore *store_,
   store_is_rotational(store->is_rotational()),
   trace_endpoint("0.0.0.0", 0, "osd"),
   asok_hook(NULL),
+  m_osd_pg_epoch_max_lag_factor(cct->_conf->get_val<double>(
+				  "osd_pg_epoch_max_lag_factor")),
   osd_compat(get_osd_compat_set()),
   osd_op_tp(cct, "OSD::osd_op_tp", "tp_osd_tp",
 	    get_num_op_threads()),
@@ -7347,6 +7349,29 @@ void OSD::handle_osd_map(MOSDMap *m)
     skip_maps = true;
   }
 
+  // wait for pgs to catch up
+  {
+    // we extend the map cache pins to accomodate pgs slow to consume maps
+    // for some period, until we hit the max_lag_factor bound, at which point
+    // we block here to stop injesting more maps than they are able to keep
+    // up with.
+    epoch_t max_lag = cct->_conf->osd_map_cache_size *
+      m_osd_pg_epoch_max_lag_factor;
+    assert(max_lag > 0);
+    if (osdmap->get_epoch() > max_lag) {
+      epoch_t min = service.get_min_pg_epoch();
+      epoch_t need = osdmap->get_epoch() - max_lag;
+      if (need > min) {
+	dout(10) << __func__ << " waiting for pgs to consume " << need
+		 << " (current min " << min
+		 << ", map cache is " << cct->_conf->osd_map_cache_size
+		 << ", max_lag_factor " << m_osd_pg_epoch_max_lag_factor
+		 << ")" << dendl;
+	service.wait_min_pg_epoch(need);
+      }
+    }
+  }
+
   ObjectStore::Transaction t;
   uint64_t txn_size = 0;
 
@@ -9314,6 +9339,7 @@ const char** OSD::get_tracked_conf_keys() const
     "osd_op_history_slow_op_threshold",
     "osd_enable_op_tracker",
     "osd_map_cache_size",
+    "osd_pg_epoch_max_lag_factor",
     "osd_pg_epoch_persisted_max_stale",
     "osd_disk_thread_ioprio_class",
     "osd_disk_thread_ioprio_priority",
@@ -9390,6 +9416,10 @@ void OSD::handle_conf_change(const struct md_config_t *conf,
       changed.count("fsid")) {
     update_log_config();
   }
+  if (changed.count("osd_pg_epoch_max_lag_factor")) {
+    m_osd_pg_epoch_max_lag_factor = conf->get_val<double>(
+      "osd_pg_epoch_max_lag_factor");
+  }
 
 #ifdef HAVE_LIBFUSE
   if (changed.count("osd_objectstore_fuse")) {
diff --git a/src/osd/OSD.h b/src/osd/OSD.h
index 02b01f6376d..cb3d8af8e43 100644
--- a/src/osd/OSD.h
+++ b/src/osd/OSD.h
@@ -384,6 +384,7 @@ public:
 private:
   // -- map epoch lower bound --
   Mutex pg_epoch_lock;
+  Cond pg_cond;
   multiset<epoch_t> pg_epochs;
   map<spg_t,epoch_t> pg_epoch;
 
@@ -394,11 +395,19 @@ public:
     assert(t == pg_epoch.end());
     pg_epoch[pgid] = epoch;
     pg_epochs.insert(epoch);
+    if (*pg_epochs.begin() == epoch) {
+      // we are the (new?) blocking epoch
+      pg_cond.Signal();
+    }
   }
   void pg_update_epoch(spg_t pgid, epoch_t epoch) {
     Mutex::Locker l(pg_epoch_lock);
     map<spg_t,epoch_t>::iterator t = pg_epoch.find(pgid);
     assert(t != pg_epoch.end());
+    if (*pg_epochs.begin() == t->second) {
+      // we were on the blocking epoch
+      pg_cond.Signal();
+    }
     pg_epochs.erase(pg_epochs.find(t->second));
     t->second = epoch;
     pg_epochs.insert(epoch);
@@ -407,6 +416,10 @@ public:
     Mutex::Locker l(pg_epoch_lock);
     map<spg_t,epoch_t>::iterator t = pg_epoch.find(pgid);
     if (t != pg_epoch.end()) {
+      if (*pg_epochs.begin() == t->second) {
+	// we were on the blocking epoch
+	pg_cond.Signal();
+      }
       pg_epochs.erase(pg_epochs.find(t->second));
       pg_epoch.erase(t);
     }
@@ -419,6 +432,14 @@ public:
       return *pg_epochs.begin();
   }
 
+  void wait_min_pg_epoch(epoch_t e) {
+    Mutex::Locker l(pg_epoch_lock);
+    while (!pg_epochs.empty() &&
+	   *pg_epochs.begin() < e) {
+      pg_cond.Wait(pg_epoch_lock);
+    }
+  }
+
 private:
   // -- superblock --
   Mutex publish_lock, pre_publish_lock; // pre-publish orders before publish
@@ -1283,6 +1304,9 @@ private:
   class C_Tick;
   class C_Tick_WithoutOSDLock;
 
+  // -- config settings --
+  float m_osd_pg_epoch_max_lag_factor;
+
   // -- superblock --
   OSDSuperblock superblock;
 
-- 
2.39.5