[IMPALA-1046] Race condition in IoMgr. Blocked ranges enqueued after cancel.

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: Impala 1.4
• Fix Version/s: Impala 1.4
• Assignee: Nong Li
• Reporter: Nong Li
• Created: 13/Jun/14 00:02
• Updated: 14/Jun/14 00:17
• Resolved: 14/Jun/14 00:17

Description

if (!enough_memory) {
    RequestContext::PerDiskState& state = reader->disk_states_[disk_queue->disk_id];
    unique_lock<mutex> reader_lock(reader->lock_);
    if (!range->ready_buffers_.empty()) {
      // We have memory pressure and this range doesn't need another buffer
      // (it already has one queued). Skip this range and pick it up later.
      range->blocked_on_queue_ = true;
      reader->blocked_ranges_.Enqueue(range);                                                <-- This is bad and breaks the invariant that blocked ranges are not queued anymore if the reader is cancelled. Not sure if this is a good invariant or if it should be fixed here.
      state.DecrementRequestThread();
      return;
    } else {
      // We need to get a buffer anyway since there are none queued. The query
      // is likely to fail due to mem limits but there's nothing we can do about that
      // now.
    }
  }