rgw: splitting gc chains into smaller parts to prevent

OSD_WRITETOOBIG error.

fixes: https://tracker.ceph.com/issues/49823

Signed-off-by: Pritha Srivastava <prsrivas@redhat.com>

diff --git a/src/cls/rgw/cls_rgw_ops.h b/src/cls/rgw/cls_rgw_ops.h
index 42eacf00cb1dc91c6ebed3c156f320a42bc1025e..f6015eacea022d9eedfe9c35e00a4bd15d14c65f 100644 (file)
--- a/src/cls/rgw/cls_rgw_ops.h
+++ b/src/cls/rgw/cls_rgw_ops.h
@@ -838,6 +838,12 @@ struct cls_rgw_gc_set_entry_op {
 
   void dump(ceph::Formatter *f) const;
   static void generate_test_instances(std::list<cls_rgw_gc_set_entry_op*>& ls);
+
+  size_t estimate_encoded_size() const {
+    constexpr size_t start_overhead = sizeof(__u8) + sizeof(__u8) + sizeof(ceph_le32); // version and length prefix
+    constexpr size_t expr_secs_overhead = sizeof(__u32); // expiration_seconds_overhead
+    return start_overhead + expr_secs_overhead + info.estimate_encoded_size();
+  }
 };
 WRITE_CLASS_ENCODER(cls_rgw_gc_set_entry_op)
 

diff --git a/src/cls/rgw/cls_rgw_types.h b/src/cls/rgw/cls_rgw_types.h
index dac3c51353f2b0eff1709bee3faeea977fabbf79..ef76204b3941cde95b003e0317a4fc7d689f5fb7 100644 (file)
--- a/src/cls/rgw/cls_rgw_types.h
+++ b/src/cls/rgw/cls_rgw_types.h
@@ -419,6 +419,14 @@ struct cls_rgw_obj_key {
     ls.back()->name = "name";
     ls.back()->instance = "instance";
   }
+
+  size_t estimate_encoded_size() const {
+    constexpr size_t start_overhead = sizeof(__u8) + sizeof(__u8) + sizeof(ceph_le32); // version and length prefix
+    constexpr size_t string_overhead = sizeof(__u32); // strings are encoded with 32-bit length prefix
+    return start_overhead +
+        string_overhead + name.size() +
+        string_overhead + instance.size();
+  }
 };
 WRITE_CLASS_ENCODER(cls_rgw_obj_key)
 
@@ -1148,6 +1156,16 @@ struct cls_rgw_obj {
     ls.back()->key.name = "myoid";
     ls.back()->loc = "mykey";
   }
+
+  size_t estimate_encoded_size() const {
+    constexpr size_t start_overhead = sizeof(__u8) + sizeof(__u8) + sizeof(ceph_le32); // version and length prefix
+    constexpr size_t string_overhead = sizeof(__u32); // strings are encoded with 32-bit length prefix
+    return start_overhead +
+        string_overhead + pool.size() +
+        string_overhead + key.name.size() +
+        string_overhead + loc.size() +
+        key.estimate_encoded_size();
+  }
 };
 WRITE_CLASS_ENCODER(cls_rgw_obj)
 
@@ -1192,6 +1210,16 @@ struct cls_rgw_obj_chain {
   bool empty() {
     return objs.empty();
   }
+
+  size_t estimate_encoded_size() const {
+    constexpr size_t start_overhead = sizeof(__u8) + sizeof(__u8) + sizeof(ceph_le32);
+    constexpr size_t size_overhead = sizeof(__u32); // size of the chain
+    size_t chain_overhead = 0;
+    for (auto& it : objs) {
+      chain_overhead += it.estimate_encoded_size();
+    }
+    return (start_overhead + size_overhead + chain_overhead);
+  }
 };
 WRITE_CLASS_ENCODER(cls_rgw_obj_chain)
 
@@ -1233,6 +1261,14 @@ struct cls_rgw_gc_obj_info
     ceph_timespec ts{ceph_le32(21), ceph_le32(32)};
     ls.back()->time = ceph::real_clock::from_ceph_timespec(ts);
   }
+
+  size_t estimate_encoded_size() const {
+    constexpr size_t start_overhead = sizeof(__u8) + sizeof(__u8) + sizeof(ceph_le32); // version and length prefix
+    constexpr size_t string_overhead = sizeof(__u32); // strings are encoded with 32-bit length prefix
+    constexpr size_t time_overhead = 2 * sizeof(ceph_le32); // time is stored as tv_sec and tv_nsec
+    return start_overhead + string_overhead + tag.size() +
+            time_overhead + chain.estimate_encoded_size();
+  }
 };
 WRITE_CLASS_ENCODER(cls_rgw_gc_obj_info)
 

diff --git a/src/rgw/rgw_gc.cc b/src/rgw/rgw_gc.cc
index d59cc65b4b2b2117a29c42ad8cb0083a7d441d44..bd16bde1bd5dbeac47926845d1a8f16dbeac1504 100644 (file)
--- a/src/rgw/rgw_gc.cc
+++ b/src/rgw/rgw_gc.cc
@@ -65,7 +65,57 @@ int RGWGC::tag_index(const string& tag)
   return rgw_shards_mod(XXH64(tag.c_str(), tag.size(), seed), max_objs);
 }
 
-int RGWGC::send_chain(cls_rgw_obj_chain& chain, const string& tag)
+std::tuple<int, std::optional<cls_rgw_obj_chain>> RGWGC::send_split_chain(const cls_rgw_obj_chain& chain, const std::string& tag)
+{
+  ldpp_dout(this, 20) << "RGWGC::send_split_chain - tag is: " << tag << dendl;
+
+  if (cct->_conf->rgw_max_chunk_size) {
+    cls_rgw_obj_chain broken_chain;
+    ldpp_dout(this, 20) << "RGWGC::send_split_chain - rgw_max_chunk_size is: " << cct->_conf->rgw_max_chunk_size << dendl;
+
+    for (auto it = chain.objs.begin(); it != chain.objs.end(); it++) {
+      ldpp_dout(this, 20) << "RGWGC::send_split_chain - adding obj with name: " << it->key << dendl;
+      broken_chain.objs.emplace_back(*it);
+      cls_rgw_gc_obj_info info;
+      info.tag = tag;
+      info.chain = broken_chain;
+      cls_rgw_gc_set_entry_op op;
+      op.info = info;
+      size_t total_encoded_size = op.estimate_encoded_size();
+      ldpp_dout(this, 20) << "RGWGC::send_split_chain - total_encoded_size is: " << total_encoded_size << dendl;
+
+      if (total_encoded_size > cct->_conf->rgw_max_chunk_size) { //dont add to chain, and send to gc
+        broken_chain.objs.pop_back();
+        --it;
+        ldpp_dout(this, 20) << "RGWGC::send_split_chain - more than, dont add to broken chain and send chain" << dendl;
+        auto ret = send_chain(broken_chain, tag);
+        if (ret < 0) {
+          broken_chain.objs.insert(broken_chain.objs.end(), it, chain.objs.end()); // add all the remainder objs to the list to be deleted inline
+          ldpp_dout(this, 0) << "RGWGC::send_split_chain - send chain returned error: " << ret << dendl;
+          return {ret, {broken_chain}};
+        }
+        broken_chain.objs.clear();
+      }
+    }
+    if (!broken_chain.objs.empty()) { //when the chain is smaller than or equal to rgw_max_chunk_size
+      ldpp_dout(this, 20) << "RGWGC::send_split_chain - sending leftover objects" << dendl;
+      auto ret = send_chain(broken_chain, tag);
+      if (ret < 0) {
+        ldpp_dout(this, 0) << "RGWGC::send_split_chain - send chain returned error: " << ret << dendl;
+        return {ret, {broken_chain}};
+      }
+    }
+  } else {
+    auto ret = send_chain(chain, tag);
+    if (ret < 0) {
+      ldpp_dout(this, 0) << "RGWGC::send_split_chain - send chain returned error: " << ret << dendl;
+      return {ret, {std::move(chain)}};
+    }
+  }
+  return {0, {}};
+}
+
+int RGWGC::send_chain(const cls_rgw_obj_chain& chain, const string& tag)
 {
   ObjectWriteOperation op;
   cls_rgw_gc_obj_info info;

diff --git a/src/rgw/rgw_gc.h b/src/rgw/rgw_gc.h
index f66b0be89c8482dd65e512f5cae5f0bba1e15479..196f2802c1649a2c4d84232695e31147f172bb16 100644 (file)
--- a/src/rgw/rgw_gc.h
+++ b/src/rgw/rgw_gc.h
@@ -29,6 +29,7 @@ class RGWGC : public DoutPrefixProvider {
   static constexpr uint64_t seed = 8675309;
 
   int tag_index(const std::string& tag);
+  int send_chain(const cls_rgw_obj_chain& chain, const std::string& tag);
 
   class GCWorker : public Thread {
     const DoutPrefixProvider *dpp;
@@ -51,7 +52,7 @@ public:
     finalize();
   }
   std::vector<bool> transitioned_objects_cache;
-  int send_chain(cls_rgw_obj_chain& chain, const std::string& tag);
+  std::tuple<int, std::optional<cls_rgw_obj_chain>> send_split_chain(const cls_rgw_obj_chain& chain, const std::string& tag);
 
   // asynchronously defer garbage collection on an object that's still being read
   int async_defer_chain(const std::string& tag, const cls_rgw_obj_chain& info);

diff --git a/src/rgw/rgw_rados.cc b/src/rgw/rgw_rados.cc
index e51cba0d4d21e5526e09f0a7ff75aa04f2b55839..b57cbc188e84aecf20f0437dc558bc7f87502b49 100644 (file)
--- a/src/rgw/rgw_rados.cc
+++ b/src/rgw/rgw_rados.cc
@@ -4927,10 +4927,10 @@ int RGWRados::Object::complete_atomic_modification(const DoutPrefixProvider *dpp
     //Delete objects inline just in case gc hasn't been initialised, prevents crashes
     store->delete_objs_inline(dpp, chain, tag);
   } else {
-    auto ret = store->gc->send_chain(chain, tag); // do it synchronously
-    if (ret < 0) {
+    auto [ret, leftover_chain] = store->gc->send_split_chain(chain, tag); // do it synchronously
+    if (ret < 0 && leftover_chain) {
       //Delete objects inline if send chain to gc fails
-      store->delete_objs_inline(dpp, chain, tag);
+      store->delete_objs_inline(dpp, *leftover_chain, tag);
     }
   }
   return 0;
@@ -4950,13 +4950,13 @@ void RGWRados::update_gc_chain(const DoutPrefixProvider *dpp, rgw_obj head_obj,
   }
 }
 
-int RGWRados::send_chain_to_gc(cls_rgw_obj_chain& chain, const string& tag)
+std::tuple<int, std::optional<cls_rgw_obj_chain>> RGWRados::send_chain_to_gc(cls_rgw_obj_chain& chain, const string& tag)
 {
   if (chain.empty()) {
-    return 0;
+    return {0, std::nullopt};
   }
 
-  return gc->send_chain(chain, tag);
+  return gc->send_split_chain(chain, tag);
 }
 
 void RGWRados::delete_objs_inline(const DoutPrefixProvider *dpp, cls_rgw_obj_chain& chain, const string& tag)

diff --git a/src/rgw/rgw_rados.h b/src/rgw/rgw_rados.h
index 0bac6c607b1816049af75ade5841ad3c15cefc86..243963711b8ea33b12cbf7cbf5d0a93d56e24691 100644 (file)
--- a/src/rgw/rgw_rados.h
+++ b/src/rgw/rgw_rados.h
@@ -1469,7 +1469,7 @@ public:
   int unlock(const rgw_pool& pool, const std::string& oid, rgw_zone_id& zone_id, std::string& owner_id);
 
   void update_gc_chain(const DoutPrefixProvider *dpp, rgw_obj head_obj, RGWObjManifest& manifest, cls_rgw_obj_chain *chain);
-  int send_chain_to_gc(cls_rgw_obj_chain& chain, const std::string& tag);
+  std::tuple<int, std::optional<cls_rgw_obj_chain>> send_chain_to_gc(cls_rgw_obj_chain& chain, const std::string& tag);
   void delete_objs_inline(const DoutPrefixProvider *dpp, cls_rgw_obj_chain& chain, const std::string& tag);
   int gc_operate(const DoutPrefixProvider *dpp, std::string& oid, librados::ObjectWriteOperation *op);
   int gc_aio_operate(const std::string& oid, librados::AioCompletion *c,

diff --git a/src/rgw/rgw_sal_rados.cc b/src/rgw/rgw_sal_rados.cc
index be2ecf9048195280af6db9ebf8d76421d6072ff9..6ca6703c484fc9efb9ad54daa6d03daf1076cad7 100644 (file)
--- a/src/rgw/rgw_sal_rados.cc
+++ b/src/rgw/rgw_sal_rados.cc
@@ -2003,14 +2003,14 @@ int RadosMultipartUpload::abort(const DoutPrefixProvider *dpp, CephContext *cct)
     store->getRados()->delete_objs_inline(dpp, chain, mp_obj.get_upload_id());
   } else {
     /* use upload id as tag and do it synchronously */
-    ret = store->getRados()->send_chain_to_gc(chain, mp_obj.get_upload_id());
-    if (ret < 0) {
+    auto [ret, leftover_chain] = store->getRados()->send_chain_to_gc(chain, mp_obj.get_upload_id());
+    if (ret < 0 && leftover_chain) {
       ldpp_dout(dpp, 5) << __func__ << ": gc->send_chain() returned " << ret << dendl;
       if (ret == -ENOENT) {
         return -ERR_NO_SUCH_UPLOAD;
       }
       //Delete objects inline if send chain to gc fails
-      store->getRados()->delete_objs_inline(dpp, chain, mp_obj.get_upload_id());
+      store->getRados()->delete_objs_inline(dpp, *leftover_chain, mp_obj.get_upload_id());
     }
   }