[14681] Fixing thread sanitizer issue 4169 <master>

[MiguelBarro]

Description

Review and testing against intermediate branch #2704

This pull request addresses the following family of deadlocks A => B => C => A where:

• A RecursiveTimedMutex& Endpoint::getMutex() Endpoint is EDP DataWriter in Participant 2.
• B Participant 1. std::recursive_mutex RTPSParticipantImpl::getParticipantMutex()
• C Participant 2. std::recursive_mutex RTPSParticipantImpl::getParticipantMutex()

It involves a test case where two participants are destroyed from the Main Thread.
The execution sequence that leads to deadlocks:

• A=>B Flow controller thread for EDP DataWriter in Participant 2.
  EDP DataWriter sending messages through its flow controller, that is, flow controller's asynchronous thread
  (FlowControllerImpl::run()). The asynchronous thread takes the DataWriter mutex (A).
  Later it delivers the sample via intraprocess to EDPDataReader in participant 2 which takes the participant
  mutex (B) to traverse endpoint collections and do matching.

• B=>C Main Thread.
  On test destruction comes participant 1 removal.
  Participant 1 calls RTPSParticipantImpl::disable() which takes mutex B to traverse it's endpoint list for
  removal. Here the EDP demise notifications go via intraprocess and ends up in EDP::unpairWriterProxy()
  call that requires C (participant 2 mutex) for unmatching.

• C=>A Main Thread
  On test destruction comes participant 2 removal.
  Participant 2 calls RTPSParticipantImpl::disable() which takes mutex C to traverse it's endpoint list for
  removal. In order to send DATA[U]R message it requires mutex (A) to allocate a new CacheChange (single mutex for
  history and endpoint).

In order to solve the deadlock, the participant's endpoint collections will no longer be protected by the general
participant mutex but by a devoted one (this was partially done before but the refactor extends it to all collections).

    mutable shared_mutex endpoints_list_mutex;

    std::vector<RTPSWriter*> m_allWriterList;
    std::vector<RTPSReader*> m_allReaderList;
    std::vector<RTPSWriter*> m_userWriterList;
    std::vector<RTPSReader*> m_userReaderList;

Other changes added in this pull request:

• In order to prevent deadlocks with the temp_data_lock mutex (devoted to protect EDP temporary proxies access) a new proxy pool is created in EDP. The proxies are then synchronized via ownership instead of mutual exclusion.
• A thread local storage variable is added to the RTPSParticipantImpl class in order to prevent reentrancy in the endpoints_list_mutex (this situation takes place very often in intraprocess calls).

Contributor Checklist

• Commit messages follow the project guidelines.
• The code follows the style guidelines of this project.
• NA Tests that thoroughly check the new feature have been added/Regression tests checking the bug and its fix have been added.
• Any new/modified methods have been properly documented using Doxygen.
• Fast DDS test suite has been run locally.
• Changes are ABI compatible.
• Changes are API compatible.
• NA Documentation builds and tests pass locally.
• NA New feature has been added to the versions.md file (if applicable).
• NA New feature has been documented/Current behavior is correctly described in the documentation.

Reviewer Checklist

• Check contributor checklist is correct.
• Check CI results: changes do not issue any warning.
• Check CI results: failing tests are unrelated with the changes.

[jsan-rt]

@richiprosima please test mac

[MiguelCompany]

@Mergifyio backport 2.6.x

[mergify]

backport 2.6.x

✅ Backports have been created

• #2774 [14681] Fixing thread sanitizer issue 4169 (backport #2717) has been created for branch 2.6.x