NFD

Can you give (steps to create) environment to reproduce an error? I don't see an obvious reason why this failure happens.

Does it happens only with the latest version of with 0.3.4 too?

Alex Afanasyev wrote:

Can you give (steps to create) environment to reproduce an error? I don't see an obvious reason why this failure happens.

Does it happens only with the latest version of with 0.3.4 too?

We are using 0.3.4, not git HEAD.

Steps to reproduce:

• Have strategy installed on a node (Let's call it nodeA)

• The node should have same routes(/prefix/test) to multiple data producers (nodeA connected to producerA and producerB)

• Kill the producer at producerA. Make sure routes for both producerA and producerB are still present at nodeA.

• Send an Interest for the prefix (/prefix/test). Sometime it takes a few tries to reproduce this crash.

In the backtrace I see some funny numbers in the use count of some shared_ptrs (very large values, negative values, ...). I suspect some kind of memory corruption is going on. Did you compile ndn-cxx and NFD yourself, both with the same compiler version?

Yes, we compiled both with same compiler version.

Are you running NFD on the same machine where it was compiled?

Yes, we are running on the same machine.

Try building and running with AddressSanitizer, see if it reports anything. Or under valgrind.

Here is the valgrind log.

This is how I ran it.

# valgrind --log-file=/tmp/den_log --track-origins=yes nfd

Ok, we found the problem.

  // FIB lookup
  shared_ptr<fib::Entry> fibEntry = m_fib.findLongestPrefixMatch(*pitEntry);
  this->dispatchToStrategy(pitEntry, bind(&Strategy::afterReceiveInterest, _1,
                                          cref(inFace), cref(interest), fibEntry, pitEntry));

A fibEntry is created by longest prefix match. If there is no match, an emptyEntry is returned. The return value isn't checked before passing it to dispatchToStrategy(). This causes a crash if fibEntry is nullptr.

Here is the code that fixes the segfault.

  // FIB lookup
  shared_ptr<fib::Entry> fibEntry = m_fib.findLongestPrefixMatch(*pitEntry);

  // XXX This assert fails
  //BOOST_ASSERT(fibEntry->m_nameTreeEntry != nullptr);
  if (fibEntry->m_nameTreeEntry == nullptr)
    {
      NFD_LOG_WARN("Received Interest " << interest.getName() << " before setting up FIB entry");
      return;
    }

  // dispatch to strategy
  this->dispatchToStrategy(pitEntry, bind(&Strategy::afterReceiveInterest, _1,
                                          cref(inFace), cref(interest), fibEntry, pitEntry));
}

Does Solution C mean that Interests with different names will share the same measurement entry if there are no matching FIB entries?

Does Solution C mean that Interests with different names will share the same measurement entry if there are no matching FIB entries?

Every solution will return the same Measurements entry, ndn:/, if the strategy invokes Measurements::get(const fib::Entry&) with the FIB entry found by longest prefix match on an empty FIB.

There's no semantic difference among the three solutions.