SIGFAULT while running select * in a 3 node cluster on AWS about scylladb

[slivne]

We do not support it at this point - doing range_queries in read in cluster - closing

[gleb-cloudius]

On Wed, Jul 15, 2015 at 09:31:05AM -0700, slivne wrote:

We do not support it at this point - doing range_queries in read in cluster - closing

The crash has nothing to do with range_queries read in cluster though.
It crashes in deserializer because it uses assign operator instead of
placement new. This patch should fix the crash:

diff --git a/mutation_query.cc b/mutation_query.cc
index 50b6f27..539dd1e 100644
--- a/mutation_query.cc
+++ b/mutation_query.cc
@@ -87,7 +87,7 @@ void db::serializer<reconcilable_result>::read(reconcilable_result& v, input& in
auto fm = db::frozen_mutation_serializer::read(in);
partitions.emplace_back(partition(p_row_count, std::move(fm)));
}

• v = reconcilable_result(row_count, std::move(partitions));
• new (&v) reconcilable_result(row_count, std::move(partitions));
  }

future<reconcilable_result>

    Gleb.

[tgrabiec]

2015-07-15 18:49 GMT+02:00 Gleb Natapov [email protected]:

On Wed, Jul 15, 2015 at 09:31:05AM -0700, slivne wrote:

We do not support it at this point - doing range_queries in read in
cluster - closing

The crash has nothing to do with range_queries read in cluster though.
It crashes in deserializer because it uses assign operator instead of
placement new. This patch should fix the crash:

diff --git a/mutation_query.cc b/mutation_query.cc
index 50b6f27..539dd1e 100644
--- a/mutation_query.cc
+++ b/mutation_query.cc
@@ -87,7 +87,7 @@ void
db::serializer<reconcilable_result>::read(reconcilable_result& v, input& in
auto fm = db::frozen_mutation_serializer::read(in);
partitions.emplace_back(partition(p_row_count, std::move(fm)));
}

• v = reconcilable_result(row_count, std::move(partitions));
• new (&v) reconcilable_result(row_count, std::move(partitions));

The output of db::serializer::read() is not said to be uninitialized,
there's this for example:

template<>
void db::serializer::read(sstring& s, input& in) {
s = in.read();
}

The fix could be in net::serializer::read_serializable(), but it assumes
that values are movable:

diff --git a/message/messaging_service.hh b/message/messaging_service.hh
index 2a3033d..897969b 100644
--- a/message/messaging_service.hh
+++ b/message/messaging_service.hh
@@ -136,7 +136,7 @@ struct serializer {
}
bytes_view bv(reinterpret_cast<const int8_t*>(buf.get()),
sz);
data_input in(bv);

•            db::serializer<Serializable>::read(v, in);
  

•            new (&v) (db::serializer<Serializable>::read(in));
       });
   });
  

  }

Or we could align db::serializer<>::read() so that they're like rpc
deserializers wrt to output parameter.

[gleb-cloudius]

On Wed, Jul 15, 2015 at 10:09:07AM -0700, Tomasz Grabiec wrote:

2015-07-15 18:49 GMT+02:00 Gleb Natapov [email protected]:

On Wed, Jul 15, 2015 at 09:31:05AM -0700, slivne wrote:

We do not support it at this point - doing range_queries in read in
cluster - closing

The crash has nothing to do with range_queries read in cluster though.
It crashes in deserializer because it uses assign operator instead of
placement new. This patch should fix the crash:

diff --git a/mutation_query.cc b/mutation_query.cc
index 50b6f27..539dd1e 100644
--- a/mutation_query.cc
+++ b/mutation_query.cc
@@ -87,7 +87,7 @@ void
db::serializer<reconcilable_result>::read(reconcilable_result& v, input& in
auto fm = db::frozen_mutation_serializer::read(in);
partitions.emplace_back(partition(p_row_count, std::move(fm)));
}

• v = reconcilable_result(row_count, std::move(partitions));
• new (&v) reconcilable_result(row_count, std::move(partitions));

The output of db::serializer::read() is not said to be uninitialized,
there's this for example:

template<>
void db::serializer::read(sstring& s, input& in) {
s = in.read();
}

The fix could be in net::serializer::read_serializable(), but it assumes
that values are movable:

diff --git a/message/messaging_service.hh b/message/messaging_service.hh
index 2a3033d..897969b 100644
--- a/message/messaging_service.hh
+++ b/message/messaging_service.hh
@@ -136,7 +136,7 @@ struct serializer {
}
bytes_view bv(reinterpret_cast<const int8_t*>(buf.get()),
sz);
data_input in(bv);

•            db::serializer<Serializable>::read(v, in);
  

•            new (&v) (db::serializer<Serializable>::read(in));
       });
   });
  

  }

Or we could align db::serializer<>::read() so that they're like rpc
deserializers wrt to output parameter.

I think your fix is good. Serializable class should be easily movable.

        Gleb.