Skip to content

Latest commit

 

History

History
48 lines (35 loc) · 1.87 KB

known_races.md

File metadata and controls

48 lines (35 loc) · 1.87 KB
Raw

For those interested in the attempt to get rid of concurrency bugs, this is the current state.

Things already tried

  • Search for all static and global variables across TestU01 and consider them for thread safety, adapting as necessary (part of the conversion process). Wherever possible things are converted to static const.

  • Look for any "escapes" via file IO or operating system functions where files are written then read from.

  • Run many instances of the benchmarks against each other in different threads with different source RNGs. Compare the output of each run against the previous run for exact pvalue outputs. Leave it running for hours with 32-threads on a 16 core machine.

  • Use valgrind with drd analyser on the above soak program.

  • Use clang with -fsanitise=threads on the above soak program.

Known data-races

  • The posix version of lgamma writes to a shared-variable signgam (both valgrind and -fsantise=threads pick this up). This is benign, because no part of TestU01 actually reads from it.

  • There are some potentially racing accesses to global variables related to output by printing in TestU01. These are benign as long as nobody flips them during output, which is not how they are used (as far as I know they will only be modified by the user).

  • If you enable printing of logging to stdout using swrite_Basic and friends, then output will potentially be interleaved. This can't really be helped due to the deep assumption of stdout - getting rid of it would require a lot of work, but actually is not that much benefit.

Potential weak points

I'm still not absolutely sure about Crush, as it is the most complex use-case, and takes so long to run. However I have no reason to think it is broken, unless some kind of global variable or side-effect was missed.