Skip to content

Latest commit

 

History

History
194 lines (154 loc) · 8.29 KB

README.md

File metadata and controls

194 lines (154 loc) · 8.29 KB

Unthread

Unthread is an implmentation of POSIX threads designed for fuzzing and debugging of concurrent programs. Unthread is:

  • deterministic: Identical seeds will result in identical threading schedules.
  • fuzzable: Unthread can be used for instrumented fuzzing of threaded code, although the coverage-based heuristics used by most fuzzers are likely inefficient for this.
  • strict: Unthread only implementes the minimum required by the spec.
  • instant: Timed locks do not wait for wall-time to pass before a lock is acquired/times out.

With some caveats:

  • Not parallel: Unthread does not use any actual hardware threading. All threads run on a single core.
  • No preemption: Control is only transfered at "yield points", which includes most pthread functions. This means that code cannot depend on other threads making progress by eg. waiting in a loop until some condition is met. pthread_yield can be inserted in these cases to provide a yield point.
  • Non-conformance: Not being a "real" pthreads implementation that integrates with the system means some parts of the POSIX spec are difficult/impossible to implement:
    • Only pthread_cond_timedwait, pthread_cond_wait, pthread_join, and pthread_testcancel are cancellation points.
    • Process shared objects are not supported (PTHREAD_PROCESS_SHARED).
    • Signal handling is not supported.
  • Glibc only: Only Glibc is supported at the moment. PR's welcome!
  • Incomplete: Unthread is a work in progress and some pthread features have not been implemented yet:
    • Robust mutexes.
    • Semaphores.

Example

Here is a example from the Unthread test suite of a program with a race condition:

#include <stdio.h>
#include <pthread.h>
#include <stdlib.h>

static int val = 0;
static pthread_mutex_t mutex = PTHREAD_MUTEX_INITIALIZER;

void* incr(void* arg) {
    int v;
    
    pthread_mutex_lock(&mutex);
    v = val;
    pthread_mutex_unlock(&mutex);

    pthread_mutex_lock(&mutex);
    val = v + 1;
    pthread_mutex_unlock(&mutex);

    return NULL;
}

int main() {
    pthread_t a,b;
    pthread_create(&a, NULL, incr, NULL);
    pthread_create(&b, NULL, incr, NULL);
    pthread_join(a, NULL);
    pthread_join(b, NULL);
    printf("%d", val);
    
    return 0;
}

This program has no data-races as all access to val is protected behind a mutex. But it does have a scheduling-based race condition as the final value it outputs depends on the scheduling order. For this particular example and my particular machine/OS, this non-determinism does occur under standard pthreads given enough executions:

$ make bin/mutex-pthread
$ seq 1000 | xargs -i bin/mutex-pthread
2222222222222222222222222222222222222212222222222222222122222222222222222222222222222222222222222222
2222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222
2222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222
2222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222
2222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222
2222222222212222222222222222222222222222222222222222222222222222222222222122222222222222222222222222
2222222221222222222222222222222222222222222222222222222222222222222222222222222222222222222221222222
2222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222
2222222222222222222222222222221222222222222222222222222222222222221222222222222222222222222222222222
2222222222222222222222222222222222222222222222222222222222222222222222222222222122222222222222222222

This is not true in the general case. It might be that a race condition only shows under some special conditions, like high load, a certain os, or certain hardware. To make matters worse, debuggers can affect timings such that race conditions disappear. Unthread allows us to test this in a structured way instead of just hoping that we are able to trigger non-deterministic behavior.

Let us say that the expected value is 2 in the above example is 2 and the buggy value is 1. We can test this using the unthread-fuzz utility. By default, this program just executes the program with different seeds until it returns with a non-zero return code, in which case the failing seed is printed to standard out.

$ make bin/mutex
$ make bin/unthread-fuzz
$ bin/unthread-fuzz sh -c '[ $(bin/mutex) -eq "2" ]'
f503e38ee865d082f5773198fb24df71

It found a failing seed! We can execute the program with the given seed and get the exact same result each time:

$ UNTHREAD_SEED=f503e38ee865d082f5773198fb24df71 bin/mutex
1

The program can then be debugged with standard debuggers like GDB or similar.

Usage

Build Unthread by running make in this dir (or add it to your build). Then link to Unthread instead of pthread and use the Unthread include dir. Eg. instead of cc main.c -lpthread do cc main.c -Lunthread/bin/unthread.o -I unthread/include. Optionally also include the include-util dir.

The thread schedule (i.e. the sequence in which threads yield to each other) can be defined in the following ways:

  • By setting the environment variable UNTHREAD_SEED to a 32-character hexadecimal string. Unthread will use this to seed a PRNG that then picks the thread schedule.
  • By setting the environment variable UNTHREAD_FILE to point to a file that will be used as entropy source for the thread schedule. This could for example be random noise to get a random execution (much like setting UNTHREAD_SEED) or it could be the result of a instrumented fuzzing process. When using file schedule input, Unthread will exit if the entropy file is depleted.
  • By setting neither, in which case the standard input is used as an entropy source like it would be from a file.

Using Unthread with libraries that themselves link to pthread will need to be rebuilt with Unthread. Good luck. :)

Configuration

Unthread can be configured by setting the following environment variables:

Environment variable Use
UNTHREAD_SEED Seed for thread schedule
UNTHREAD_FILE Entropy file for thread schedule
UNTHREAD_VERBOSE true for verbose logging
UNTHREAD_RET_OFFSET Return code offset

Return codes

Unthread may stop the application under a number of conditions:

Retcode
40 The thread schedule entropy was exhausted
41 All threads have become deadlocked
42 An illegal operation was performed
43 Allocation failure
44 Misc. failure (usually if a system call fails)
45 Failure while reading from IO
46 An unsupported operation was performed

In case that the retcode clashes with the retcode of the application, the default retcode offset of 40 can be changed by setting the UNTHREAD_RET_OFFSET environment variable.

Testing

The Unthread test runner requires Python 3 to be installed.

Unthread is tested by the Unthread tests in the test-src dir and a set of tests from the POSIX Test Suite project in the posixtestsuite. See the POSIX Test Suite project site for more information on the POSIX Test Suite.

Unthread tests in the test-src directory are examples of non-deterministic pthread programs that are deterministic in Unthread with a given seed. The tests contains a spec containing all the possible outcomes of running the test, and a list of seeds in a seperate .seeds file. These seeds must cover all the possible outcomes or the test will fail. When emodifying Unthread, it may be necessary to generate new seeds by executing UNTHREAD_GEN=true make test.

The above will only generate the minimum number of seeds to cover the outcomes. Additional fuzzing can be done by executing UNTHREAD_ITER=N make unthread-fuzz where N is the number of iterations to fuzz.

See also

rr, a gdb debugger with deterministic replay.