CalDet<TPCPadFlags>: Protect against invalid reads

A continuation of the CalDet<TPCFlags> saga, possibly related
to https://its.cern.ch/jira/browse/O2-4671

Tests on ARM, even after deployment of the custom streamer
in #14830, still showed
segfaults in TPC digitization.

With the relevant code isolated into a unit test in
#14850, it was possible to
do a valgrind study. This showed Invalid reads to the mData of CalArray.

Thereafter, putting assert statements showed that we often access
CalArray<PadFlags> data slightly out of bounds - irrespective of custom
streamer or not. This then either indicates a problem in the code logic
or a problem with the calibration CCDB objects. This should clearly be
fixed.

In the meantime, this commit adds a protection against invalid accesses
and returns a trivial answer as well as an error message. This is in any
case better than undefined behaviour.

In addition, this commit introduces possibility to switch off the custom
streamer for further studies.

Author: sawenzel

Detectors/TPC/base/include/TPCBase/CalArray.h

@@ -26,6 +26,11 @@
 #include <boost/format.hpp>
 #endif
 
+#ifdef NDEBUG
+#undef NDEBUG
+#include <cassert>
+#endif
+
 namespace o2
 {
 namespace tpc
@@ -93,7 +98,11 @@ class CalArray
   int getPadSubsetNumber() const { return mPadSubsetNumber; }
 
   void setValue(const size_t channel, const T& value) { mData[channel] = value; }
-  const T getValue(const size_t channel) const { return mData[channel]; }
+  const T getValue(const size_t channel) const
+  {
+    assert(channel < mData.size());
+    return mData[channel];
+  }
 
   void setValue(const size_t row, const size_t pad, const T& value);
   const T getValue(const size_t row, const size_t pad) const;

Detectors/TPC/base/include/TPCBase/CalDet.h

@@ -30,6 +30,12 @@
 #include "Rtypes.h"
 #endif
 
+#ifndef NDEBUG
+#undef NDEBUG
+// always enable assert
+#include <cassert>
+#endif
+
 namespace o2
 {
 namespace tpc
@@ -211,7 +217,26 @@ inline const T CalDet<T>::getValue(const ROC roc, const size_t row, const size_t
     }
     case PadSubset::Region: {
       const auto globalRow = roc.isOROC() ? mappedRow + mapper.getNumberOfRowsROC(ROC(0)) : mappedRow;
-      return mData[Mapper::REGION[globalRow] + roc.getSector() * Mapper::NREGIONS].getValue(Mapper::OFFSETCRUGLOBAL[globalRow] + mappedPad);
+      const auto dataRow = Mapper::REGION[globalRow] + roc.getSector() * Mapper::NREGIONS;
+      const auto index = Mapper::OFFSETCRUGLOBAL[globalRow] + mappedPad;
+      assert(dataRow < mData.size());
+      if (index >= mData[dataRow].getData().size()) {
+        // S. Wenzel: We shouldn't come here but we do. For instance for CalDet calibrations loaded from
+        // creator.loadIDCPadFlags(1731274461770);
+
+        // In this case there is an index overflow, leading to invalid reads and potentially a segfault.
+        // To increase stability, for now returning a trivial answer. This can be removed once either the algorithm
+        // or the calibration data has been fixed.
+#ifndef GPUCA_ALIGPUCODE // hide from GPU standalone compilation
+        static bool printMsg = true;
+        if (printMsg) {
+          LOG(error) << "Out of bound access in TPC CalDet ROC " << roc << " row " << row << " pad " << pad << " (no more messages printed)";
+        }
+        printMsg = false;
+#endif
+        return T{};
+      }
+      return mData[dataRow].getValue(index);
       break;
     }
   }

Detectors/TPC/base/src/TPCFlagsMemberCustomStreamer.cxx

@@ -69,7 +69,9 @@ static __attribute__((used)) int _R__dummyStreamer_3 =
   ([]() {
     auto cl = TClass::GetClass<o2::tpc::CalArray<o2::tpc::PadFlags>>();
     if (cl) {
-      cl->AdoptMemberStreamer("mData", new TMemberStreamer(MemberVectorPadFlagsStreamer));
+      if (!getenv("TPC_PADFLAGS_STREAMER_OFF")) {
+        cl->AdoptMemberStreamer("mData", new TMemberStreamer(MemberVectorPadFlagsStreamer));
+      }
     } else {
       // we should never come here ... and if we do we should assert/fail
       assert(false);