detecting and escaping infinite loops with jolt michael carbin with sasa misailovic, michael kling,...
TRANSCRIPT
Detecting and Escaping Infinite Loops with Jolt
Michael Carbin with Sasa Misailovic, Michael Kling, and Martin Rinard
Massachusetts Institute of Technology
Exceptions?
Our Research Context
Researcher
Developer
User
Motivation
• You run a program – Edit a document– Search for a keyword– Indent source code files
• Program infinite loops– You get no output– You lose your work
Logo courtesy J.D. Rhoades
Our Solution:
J lt
Automatically detect and escape infinite loops
Program produces output User doesn’t lose work
Basic Approach
1. Compile program with Jolt compiler. (adds instrumentation to mark loop boundaries)
2. Execute program.3. Program stops responding.4. Launch Jolt monitor.
– Monitor takes snapshots at each loop head.
– If two snapshots are same, infinite loop!
5. Jolt jumps to instruction after loop.
Does this work?
1. Are real infinite loops this simple?2. Does Jolt produce a safe execution?3. Does Jolt produce better output than
Ctrl-C?4. Does Jolt match the developer’s fix?5. Can this be implemented efficiently?
stdout
Infinite Loop in Grep
v2.5
25th ECOOP
conf.in
v2.5.1
25th ECOOP
grep –E “[0-9]*” --color=always conf.in
Infinite Loop
Infinite Loop in Grep 2.5
1: void prline(regex_t regex, char *beg) { 3: while (true) { 4: int size = 0; 5: int off = match(regex, beg, &size); 6: if (off == -1) { 7: break; 8: } 9: char const *b = beg + off;10: fwrite(beg, 1, off, stdout);11: enable_color();10: fwrite (b, 1, size, stdout);11: disable_color();11: beg = b + size;12: }14: }
regex
“[0-9]*”
*beg“25th
ECOOP”
regex
“[0-9]*”
*beg“25th
ECOOP”
size 2
off 0rege
x“[0-9]*”
*beg“25th
ECOOP”
size 2
off 0
*b“25th
ECOOP”
regex
“[0-9]*”
*beg “th ECOOP”
size 2
off 6
*b“25th
ECOOP”regex
“[0-9]*”
*beg “th ECOOP”
size 0
off 0
*b“25th
ECOOP”
regex
“[0-9]*”
*beg“th
ECOOP”
size 0
off 0
*b“th
ECOOP”
Applying Jolt to Grep
Step 1: Compile Grep
JoltCompile
r
Adds instructions that mark loop boundaries.
Jolt Compiles Grep 1: void prline(regex_t regex, char *beg) { 2: … 3: 4: while (true) { 5: 6: int size = 0; 7: int offset = match(regex, beg, &size); 8: if (offset == -1) { 9:10: break;11: }12: char const *b = beg + offset;13: fwrite (b, sizeof (char), size, stdout);14: beg = b + size;15: }16:17: }18: …
Jolt Compiles Grep 1: void prline(regex_t regex, char *beg) { 2: … 3: jolt_loop_entry(ID); 4: while (true) { 5: jolt_loop_body(ID, &jolt_escape_dest); 6: int size = 0; 7: int offset = match(regex, beg, &size); 8: if (offset == -1) { 9: jolt_loop_exit(ID);10: break;11: }12: char const *b = beg + offset;13: fwrite (b, sizeof (char), size, stdout);14: beg = b + size;15: }16: jolt_escape_dest:17: …18: }
• Grep runs as normal.• Control instrumentation overhead
~2.5%.
Step 2: Run Grep
• Grep enters infinite loop, does not respond!
Step 3: Grep stops Responding
• Jolt dynamically attaches to grep• Monitors execution
– Computes snapshot at each loop head– If two snapshots are the same, infinite
loop!
Step 4: Launch Jolt Monitor
Jolt
How to Compute a Snapshot Efficiently
1. Log all registers and memory addresses that each loop iteration accesses.
2. Record the contents of these registers and addresses at the head of the loop.
Comparing Two Snapshots
• First, check if snapshots have the same accessed– Registers and– Memory addresses
• Next check if contents of– Accessed registers and– Accessed memory
addresses
are the same
regex “[0-9]*”
beg 0xdeadbeef
*beg “th ECOOP”
size 0
off 0
b 0xdeadbeef
*b “th ECOOP”
regex “[0-9]*”
beg 0xdeadbeef
*beg “th ECOOP”
size 0
off 0
b 0xdeadbeef
*b “th ECOOP”
Snapshot 1
Snapshot 2
1: void prline(regex_t regex, char *beg) { 2: jolt_loop_entry(ID); 3: while (true) { 4: jolt_loop_body(ID, &jolt_escape_dest); 5: int size = 0; 6: int offset = match(regex, beg, &size); 7: if (offset == -1) { 8: jolt_loop_exit(ID); 9: break;10: }11: char const *b = beg + offset;12: fwrite (b, sizeof (char), size, stdout);13: beg = b + size;14: }15: jolt_escape_dest:16: }
Step 5: Jolt Exits the Loop
Output
Ctrl-C
Jolt
Does this work for Grep?
• Is it safe?• Yes, no side-effects (except on output files).
• Is it better than Ctrl-C?• Yes, continues to match additional lines and
files.
• Is it equivalent to the correct fix?• Yes, corrected by break if size is 0 (v2.5.1). • 3 years later no, v2.5.3 skips all size 0
matches.
25th ECOOP ‘11
Does this work in general?
5 Applications and 8 Infinite Loops
1. ctags : line numbers of functions in code.– v5.5 : one loop in fortran module.– v5.7b : one loop in python module.
2. grep (v2.5): matches regexp against files (3 loops).
3. ping (v20100214): icmp utility.
4. indent (v1.1-svr 4): indents source code.
5. look (v1.9.1): matches a word against dictionary file.
Question #1
Are infinite loops this simple?
Benchmark
Detected
ctags-f Yesctags-p Yes
grep Yesping Yeslook Yes
indent No
7 of 8
Question #2
Does escape produce a safe execution?
• Methodology– Validated execution with Valgrind and by hand.– Tested over available infinite loop triggering
inputs.
• Results– Yes, side effects often localized = consistent
state. – Or, simple correctness invariants.
Question #3
Does Jolt produce a better output than Ctrl-C?
– Methodology• Defined output abstraction, and compared outputs.
– Results• Yes, errors often isolated to single output unit (e.g.,
file).
– Example• indent: correct indention resumes on next file.• Terminating indent deletes your source code!!!• Development hint: don’t update files in-place.
Question #4
Does escape model the developers’ fix?
– Methodology• Manually inspected a later version of each application.
– Results• ctags: no, output semantically different on some inputs.• grep: jolt matches fix for two of three loops (3 years).• ping, indent, look: yes, in all cases.
– Example• ping: developer used continue instead of break.
Can this be implemented efficiently?
Question #5
Benchmark
Overhead (%)
Detection(s)
ctags-f 7.3 0.319
ctags-p 5.2 0.334
grep 2.5 0.551*
ping 1.6 0.287
look 0.0 0.296
indent 8.4 x*Average over the three loops
Observations
• Infinite loops can (and often do) frustrate users.
• Infinite loops can be (and often are) simple.
• Errors are localized in some cases.
• Jolt’s results can be (and often are) better than no results at all.
• Applying Jolt can (and often does) model the developer’s fix.
Did We Just Solve the Halting Problem?
No
Limitations
• Jolt doesn’t detect infinite loops that always change state.– Though still possible to exit loop if user
desires.
• We did not consider busy-wait loops.– Though a good question is what does it
mean when a user thinks a program is waiting too long?
Future Work
• Off-the-shelf binaries - pure dynamic analysis
• Safety guarantees – shepherded execution
• Larger classes of loops - symbolic non-termination analysis
Related Work
• Bounding Loop Length– Detecting and Eliminating Memory Leaks Using Cyclic
Memory Allocation (Nguyen and Rinard, ISMM ‘07)
• Non-termination Provers– TNT (Gupta, et al.) using Invariant Generation (POPL
‘08)– Looper (Burnim, et al.) using Symbolic Execution (ASE
‘09)
• Termination Provers– Terminator (Cook, et al.) (PLDI ‘06)
Takeaway
Jolt gives users another option to deal with programs that infinite loop.
Another hammer in the toolbox to help users deal with otherwise fatal errors.
From: "Armando Solar-Lezama" <[email protected]> To: "Martin Rinard" <[email protected]> Subject: Thanks
I was writing a document in Word this morning, and after about an hour of unsaved work, Word went into an infinite loop that made the application completely frozen.
So, having listened to your talks too many times, I got my debugger, paused the program, changed the program counter to a point a few instructions past the end of the loop, and let it keep running from there.
Word went back to working as if nothing had ever happened. I was able to finish my document, save it, and close Word without problems.
So thanks, Armando.
Thank You/Questions?
Supplementary Slides
Bug Reports
grep2.5 -- 13 Mar 2002
2.5.1 -- 29 Oct 2004
2.5.3 -- 02 Aug 2007
indent1994
ctags5.5 – 20035.7b –2007
Instrumentation Overhead
Bench. Average Lowest Highest
ctags-f 1.073 1.068 1.08
ctags-p 1.052 1.035 1.057
grep 1.025 1.014 1.028
ping 1.016 1.005 1.024
look 1 1 1
indent 1.084 1.082 1.086
Detection Time
Bench. Time (s)Footprint
(b)Length
ctags-f 0.319 240 256
ctags-p 0.334 312 992
grep-c 0.585 992 4030
grep-cc 0.579 992 4036
grep-m 0.490 846 2506
ping 0.287 192 54
look 0.296 300 378