From bfd4091f7faf43ebc4eb81c8215e8b3ec2d4a8c4 Mon Sep 17 00:00:00 2001
From: "Adam C. Emerson" <aemerson@redhat.com>
Date: Fri, 17 Nov 2017 17:15:26 -0500
Subject: [PATCH] rgw: Expire entries in bucket info cache

To bound the degree to which an RGW instance can go out to lunch if
the watch/notify breaks down, force refresh of any cache entry over a
certain age.

Fifteen minutes by default, and expiration can be turned off entirely.

This is separate from the LRU. The LRU removes entries based on the
last time of access. This expiration patch forces refresh based on the
last time they were updated.

Signed-off-by: Adam C. Emerson <aemerson@redhat.com>
(cherry picked from commit 4489cb58a15647a31ac0546d70400af5668404cb)
Fixes: http://tracker.ceph.com/issues/22517
---
 src/common/options.cc | 20 ++++++++++++++++++++
 src/rgw/rgw_rados.h   | 18 ++++++++++++++----
 2 files changed, 34 insertions(+), 4 deletions(-)

diff --git a/src/common/options.cc b/src/common/options.cc
index 33f92052588..e190c160cbd 100644
--- a/src/common/options.cc
+++ b/src/common/options.cc
@@ -5042,6 +5042,26 @@ std::vector<Option> get_rgw_options() {
     Option("rgw_reshard_thread_interval", Option::TYPE_UINT, Option::LEVEL_ADVANCED)
     .set_default(60 * 10)
     .set_description(""),
+
+    Option("rgw_cache_expiry_interval", Option::TYPE_UINT,
+	   Option::LEVEL_ADVANCED)
+    .set_default(900)
+    .set_description("Number of seconds before entries in the bucket info "
+		     "cache are assumed stale and re-fetched. Zero is never.")
+    .add_tag("performance")
+    .add_service("rgw")
+    .set_long_description("The Rados Gateway stores metadata about buckets in "
+			  "an internal cache. This should be kept consistent "
+			  "by the OSD's relaying notify events between "
+			  "multiple watching RGW processes. In the event "
+			  "that this notification protocol fails, bounding "
+			  "the length of time that any data in the cache will "
+			  "be assumed valid will ensure that any RGW instance "
+			  "that falls out of sync will eventually recover. "
+			  "This seems to be an issue mostly for large numbers "
+			  "of RGW instances under heavy use. If you would like "
+			  "to turn off cache expiry, set this value to zero."),
+
   });
 }
 
diff --git a/src/rgw/rgw_rados.h b/src/rgw/rgw_rados.h
index 41abab17fca..24d94a7306c 100644
--- a/src/rgw/rgw_rados.h
+++ b/src/rgw/rgw_rados.h
@@ -3733,25 +3733,32 @@ public:
 
 template <class T>
 class RGWChainedCacheImpl : public RGWChainedCache {
+  ceph::timespan expiry;
   RWLock lock;
 
-  map<string, T> entries;
+  map<string, std::pair<T, ceph::coarse_mono_time>> entries;
 
 public:
   RGWChainedCacheImpl() : lock("RGWChainedCacheImpl::lock") {}
 
   void init(RGWRados *store) {
     store->register_chained_cache(this);
+    expiry = std::chrono::seconds(store->ctx()->_conf->get_val<uint64_t>(
+				    "rgw_cache_expiry_interval"));
   }
 
   bool find(const string& key, T *entry) {
     RWLock::RLocker rl(lock);
-    typename map<string, T>::iterator iter = entries.find(key);
+    auto iter = entries.find(key);
     if (iter == entries.end()) {
       return false;
     }
+    if (expiry.count() &&
+	(ceph::coarse_mono_clock::now() - iter->second.second) > expiry) {
+      return false;
+    }
 
-    *entry = iter->second;
+    *entry = iter->second.first;
     return true;
   }
 
@@ -3765,7 +3772,10 @@ public:
   void chain_cb(const string& key, void *data) override {
     T *entry = static_cast<T *>(data);
     RWLock::WLocker wl(lock);
-    entries[key] = *entry;
+    entries[key].first = *entry;
+    if (expiry.count() > 0) {
+      entries[key].second = ceph::coarse_mono_clock::now();
+    }
   }
 
   void invalidate(const string& key) override {
-- 
2.47.3