In this class project I will explore a preliminary implementation of the tag policy language.

Writeup

• Projects:Mike_Dalton:Writeup

Progress/Milestones

Current status:

• Language specification complete (ran it by Michael McMartin as well), supports all hardware features

Milestone by end of week

• Initial lexer/parser finished, port a few small analyses
• Compile and run on FPGA

Milestone by end of next week

• Add support for flexible specification of check handlers and basic modularity/composability
• Implement more complex analyses

Update - Weekend of 5/21

• Updated language specification, now happy with final version
  • Added support for check handlers
  • Modularity will be provided by allowing policies to specify their own configuration using DSLs
    • Languages for various policies are too different to try to unify (syscall tainting resembles AppArmor or conventional RBAC/SELinux, while breakpoints/watchpoints might just be symbol lists, etc)
• Implementation delayed by new issue - role of multiplexing register windows in Linux kernel
  • See internal Kozyrakis wiki for a much longer writeup (ask me if you want it reposted here)
  • _Serious_ threat to reliability - without modifications, Linux would allow stale tags from one process's register windows to show up in another process, potentially causing false positives/negatives
    • Artifact of the shared register window format of the SPARC register file
  • Devised solution that prevents tags from one process' register window showing up in another process (with no appreciable performance impact)
    • Thankfully Linux guarantees isolation of the _data_ in register windows already
  • Laid out plan to ensure all register window tags are zero at execve() time
  • Laid out plan to prevent OS tag policies from interfering with user tag policies
    • Two solutions, one trivial but wasteful, one complete but more time-intensive
• Beginning implementation Monday/Tuesday now that register window problems are sorted out
  • Did not encounter these register window issues before because we weren't concurrently running processes which used the same tag bit for different tag policies
  • Thankfully issue turned out to be quite resolveable