Exercises

There are 10 programs under directory /root/workspace/razor/benchmarks/core-utilities. For each program, run Razor with/without applying heuristics to get the results: (1) code reduction rate (CRR); (2) number of failed testing inputs.

Let's take bzip2 as an example.

• bzip2-1.0.5.c.orig.c : the source code.
• bzip2.orig : the binary.
• train : the directory that contains training inputs for profiling.
• test : the directory that contains testing inputs for evaluating the debloated binary.
• run_razor.py : the helper script for running Razor, it takes commands:
  • train : Trace the original binary with the inputs under train, get the traces → ./logs.
  • debloat :
    • Merge the logs → %program-trace.log
    • Dump the executed instructions → instr.s
    • Instrument the executed instructions → %program.s
    • Rewrite the binary → %program.orig_temp/%program.orig.debloated
  • test : Run the debloated binary with the inputs under test.
  • extend_debloat [1|2|3|4] : Use heuristics (1 → zCode, 2 → zCall, 3 → zLib, 4 → zFunc) to extend the merged log (i.e., %program-trace.log) and debloat the binary (i.e., dumping the executed instruction, instrumentation, rewriting the binary). The newly debloated binary would overwrite %program.orig_temp/%program.orig.debloated.
  • clean : Clean the outputs under current directory.

The following are an example to get the code reduction rates (CRRs) and the numbers of failed testing inputs for bzip2.

1. Run and trace the binary with inputs under train : python run_razor.py train.
2. Debloat the binary without applying heuristics : python run_razor.py debloat.
3. Run the debloated binary with inputs under test : python run_razor.py test.
   • To focus on the failed cases with Floating point exception : python run_razor.py test | grep "Floating point exception"
   • As we can see, there are 6 failed testing inputs without applying any heuristics.
4. Get CRR:
   • Get the code size of the original binary : python ../../../tools/get_code_size.py ./bzip2.orig
   • Get the code size of the debloated binary : python ../../../tools/get_code_size.py ./bzip2.orig_temp/bzip2.orig.debloated
   • Calculate the CRR : python -c "print (old_size-new_size)/float(old_size)"
   • As we can see, the CRR without applying any heuristics is 62.4% .
5. Apply heuristic zCode:
   • python run_razor.py extend_debloat 1
   • Get failed cases when running the new debloated binary: python run_razor.py test | grep "Floating point"
   • As we can see, there are 5 failed testing inputs with applying heuristic zCode.
   • Get the code size of the original binary : python ../../../tools/get_code_size.py ./bzip2.orig
   • Get the code size of the debloated binary : python ../../../tools/get_code_size.py ./bzip2.orig_temp/bzip2.orig.debloated
   • Calculate the CRR : python -c "print (old_size-new_size)/float(old_size)"
   • As we can see, the CRR with applying heuristics zCode is 62.4% .
6. Apply heuristic zCall:
   • python run_razor.py extend_debloat 2
   • Get failed cases when running the new debloated binary: python run_razor.py test | grep "Floating point"
   • As we can see, there are 1 failed testing inputs with applying heuristic zCall.
   • Get the code size of the original binary : python ../../../tools/get_code_size.py ./bzip2.orig
   • Get the code size of the debloated binary : python ../../../tools/get_code_size.py ./bzip2.orig_temp/bzip2.orig.debloated
   • Calculate the CRR : python -c "print (old_size-new_size)/float(old_size)"
   • As we can see, the CRR with applying heuristics zCall is 59.9% .
7. Apply heuristic zLib:
   • python run_razor.py extend_debloat 3
   • Get failed cases when running the new debloated binary: python run_razor.py test | grep "Floating point"
   • As we can see, there are 0 failed testing inputs with applying heuristic zLib.
   • Get the code size of the original binary : python ../../../tools/get_code_size.py ./bzip2.orig
   • Get the code size of the debloated binary : python ../../../tools/get_code_size.py ./bzip2.orig_temp/bzip2.orig.debloated
   • Calculate the CRR : python -c "print (old_size-new_size)/float(old_size)"
   • As we can see, the CRR with applying heuristics zLib is 24.8% .
   • Since there are no failed cases, you might stop here.
8. Apply heuristic zFunc:
   • python run_razor.py extend_debloat 4
   • Get failed cases when running the new debloated binary: python run_razor.py test | grep "Floating point"
   • As we can see, there are 0 failed testing inputs with applying heuristic zFunc.
   • Get the code size of the original binary : python ../../../tools/get_code_size.py ./bzip2.orig
   • Get the code size of the debloated binary : python ../../../tools/get_code_size.py ./bzip2.orig_temp/bzip2.orig.debloated
   • Calculate the CRR : python -c "print (old_size-new_size)/float(old_size)"
   • As we can see, the CRR with applying heuristics zFunc is 24.8% .
9. So the CRRs and the numbers of failed testing inputs for bzip2 are: