out-of-bounds write access checking in Cactus

Comments (10)

1. 

   Erik Schnetter

   • removed comment

   This would be straightforward to implement in PUGH and Carpet. In Carpet, the canaries would be initialized when the memory is allocated, and would be checked regularly (e.g. during poison checking), not just before deallocating.

   I doubt that OpenMP provides its own malloc function. I expect all memory allocation (including operator new from C++) to go through libc's malloc.

   • 2013-09-02T20:34:04+00:00
2. 

   Roland Haas reporter

   • removed comment

   Sorry, I meant OpenMPI not OpenMP but could not figure out how to modify the description after having created the ticket (not enough privileges I think). See eg http://svn.open-mpi.org/svn/ompi/trunk/opal/mca/memory/linux/malloc.c though I cannot find a really good reference for this.

   • 2013-09-02T22:22:11+00:00
3. 

   Roland Haas reporter

   • removed comment

   I have a proof-of-concept implementation that implements a canary based electric fence around CarpetLib's mem objects. These are used for grid functions, scalars and arrays. The current code will check after each scheduled routine if any (write) access out-of-bounds happened in any of the variables accessible through CCTK_VarDataPtr. This is mostly ok though one can construct situations (involving Carpet's ENTER_XXX_MODE functions etc) where eg a GLOBAL routine accesses grid functions which is currently not caught.

   The functionality is activated by setting CarpetLibs::electric_fence = yes. Right now the implementation passes the test suite and there is a (designed to fail) test in CarpetExtra/CarpetEFenceTest.

   Most likely the interface will change in the future so that the check_fence function becomes a member of ggf (which is CarpetLib's representation of a grid function) rather than the data objects so that the user can just call ggf->check_fence(warning_callback) passing in a callback routine rather than iterating over all timelevels/maps/whatnot of each variable.

   The code is in a branch on bitbucket: https://bitbucket.org/rhaas80/carpet/commits/branch/fence You can click on the "Compare" button near the upper right to see the differences to (my) current Carpet master branch. Note that this branch will be rebased often.

   Comments welcome.

   • 2013-09-11T19:00:17+00:00
4. 

   Erik Schnetter

   • removed comment

   The line

   storage_ = (T*)align_up(size_t(storage_base_ + electric_fence), alignment);
   

   is wrong; electric_fence should be canary/2 instead.

   • 2013-09-17T13:04:23+00:00
5. 

   Roland Haas reporter

   • removed comment

   Right you are. Fixed now. Was a leftover from when the code only supported a fence_width of 1.

   • 2013-09-17T13:40:04+00:00
6. 

   Erik Schnetter

   • changed status to open
   • removed comment

   LGTM.

   • 2013-09-17T13:41:11+00:00
7. 

   Erik Schnetter

   • changed status to open
   • removed comment

   • 2013-09-17T13:41:17+00:00
8. 

   Roland Haas reporter

   • removed comment

   Well I was not really aiming for an actual review since I plan to make further changes. It's good to see that the concept and current implementation is approved though. Thank you.

   • 2013-09-17T15:19:08+00:00
9. 

   Roland Haas reporter

   • changed status to resolved
   • removed comment

   Committed as 0e435fca01c0ce3356565c661afb0ff9d64880a1 "CarpetLib: add some code for electric fence like functionality", d746b25592639a0f8eecab042553fe2a639be9d9 "Carpet: use electric fence provided by CarpetLib", d8cc74e37df4ff3750ccaed11b67e152cc0295f3 "CarpetEFenceTest: test electric fence in CarpetLib".

   • 2013-09-27T13:36:21+00:00
10. 

    Roland Haas reporter

    • edited description
    • changed status to closed

    • 2019-02-21T19:53:10+00:00
11. Log in to comment