Re: [PATCH v2 3/3] fuzz: add virtio-9p configurations for fuzzing

[Darren Kenny <darren.kenny@oracle.com>]

On Monday, 2021-01-18 at 10:30:33 -05, Alexander Bulekov wrote:
> On 210118 1334, Christian Schoenebeck wrote:
>> On Montag, 18. Januar 2021 00:09:24 CET Alexander Bulekov wrote:
>> > virtio-9p devices are often used to expose a virtual-filesystem to the
>> > guest. There have been some bugs reported in this device, such as
>> > CVE-2018-19364, and CVE-2021-20181. We should fuzz this device
>> > 
>> > This patch adds two virtio-9p configurations:
>> >  * One with the widely used -fsdev local driver. This driver leaks some
>> >    state in the form of files/directories created in the shared dir.
>> >  * One with the synth driver. While it is not used in the real world, this
>> >    driver won't leak leak state between fuzz inputs.
>> > 
>> > Signed-off-by: Alexander Bulekov <alxndr@bu.edu>
>> > ---
>> > CC: Christian Schoenebeck <qemu_oss@crudebyte.com>
>> > CC: Greg Kurz <groug@kaod.org>
>> > 
>> > I considered adding an atexit handler to remove the temp directory,
>> > however I am worried that there might be some error that results in a
>> > call to exit(), rather than abort(), which will cause problems for
>> > future fork()-ed fuzzers. I don't think there are such calls in the 9p
>> > code, however there might be something in the APIs used by 9p. As this
>> > code is primarily for ephemeral OSS-Fuzz conainers, this shouldn't be
>> > too much of an issue.
>> 
>> Yes, dealing with signal handlers for that is probably a bit intransparent and 
>> would leave a questionable feeling about its reliability.
>> 
>> What about __attribute__((destructor)) to auto delete the fuzzer directory, 
>> like virtio-9p-test.c does for the same task?
>> 
>
> My worry is that we will end up deleting it while it is still in use.
> The scenario I am worried about:
> [parent process ] set up temp directory for virtio-9p
> [parent process ] initialize QEMU 
> [parent process ] fork() and wait()
> [child process 1] Run the fuzzing input.
> [child process 1] Once the input has been executed, call _Exit(). This
> should skip the atexit()/__attribute((destructor)) handlers. For reasons
> related to libfuzzer, we need to do this so that libfuzzer doesn't treat
> each child exit()-ing as a crash.
> [parent process ] wait() returns.
> [parent process ] generate a new input.. fork() and wait()
> [child process 2] Run the fuzzing input.
> [child process 2] Somewhere we hit an abort(). libfuzzer hooks the abort
> and dumps the crashing input and stack trace. Since abort() doesn't call
> exit handlers, it will skip over atexit()/__attribute((destructor)) handlers
> [parent process ] wait() returns.
> [parent process ] generate a new input.. fork() and wait()
> [child process 3] Run the fuzzing input.
> [child process 3] Somewhere we hit an exit(). This will dump the
> input/stacktrace and it will run the exit handlers (removing the shared
> 9p directory)
> [parent process ] wait() returns. generate a new input.. fork() and wait()
> [child process 4] ...

OK, that answer's my question :)

> Now all the subsequent forked children will refer to a shared directory
> that we already removed. Ideally, we would run the cleanup handler only
> after the parent process exit()s. I think there are some ways to do
> this, by placing the atexit() call in a place only reachable by the
> parent. However, on OSS-Fuzz this shouldn't be a problem as everything
> is cleaned up automatically anyway..

Yep, agreed.

> I am more worried about the fact that files/directories/links that are
> created by 9p in the child processes, persist across inputs. I think
> Thomas suggested a way to work-around this for PATCH 1/3. We could have
> a function that runs in the parent after each wait() returns, that would
> remove all the files in the temp directory and scrub the extended
> attributes applied by 9p to the shared dir.

Hmm, that sounds like something to consider, but it may also end up
slowing down the execution during the turn-around - guess it depends on
how much noise is being generated.

Thanks,

Darren.

> -Alex
>
>
>> >  tests/qtest/fuzz/generic_fuzz_configs.h | 20 ++++++++++++++++++++
>> >  1 file changed, 20 insertions(+)
>> > 
>> > diff --git a/tests/qtest/fuzz/generic_fuzz_configs.h
>> > b/tests/qtest/fuzz/generic_fuzz_configs.h index 1a133655ee..f99657cdbc
>> > 100644
>> > --- a/tests/qtest/fuzz/generic_fuzz_configs.h
>> > +++ b/tests/qtest/fuzz/generic_fuzz_configs.h
>> > @@ -19,6 +19,16 @@ typedef struct generic_fuzz_config {
>> >      gchar* (*argfunc)(void); /* Result must be freeable by g_free() */
>> >  } generic_fuzz_config;
>> > 
>> > +static inline gchar *generic_fuzzer_virtio_9p_args(void){
>> > +    char tmpdir[] = "/tmp/qemu-fuzz.XXXXXX";
>> > +    g_assert_nonnull(mkdtemp(tmpdir));
>> > +
>> > +    return g_strdup_printf("-machine q35 -nodefaults "
>> > +    "-device virtio-9p,fsdev=hshare,mount_tag=hshare "
>> > +    "-fsdev local,id=hshare,path=%s,security_model=mapped-xattr,"
>> > +    "writeout=immediate,fmode=0600,dmode=0700", tmpdir);
>> > +}
>> > +
>> >  const generic_fuzz_config predefined_configs[] = {
>> >      {
>> >          .name = "virtio-net-pci-slirp",
>> > @@ -60,6 +70,16 @@ const generic_fuzz_config predefined_configs[] = {
>> >          .name = "virtio-mouse",
>> >          .args = "-machine q35 -nodefaults -device virtio-mouse",
>> >          .objects = "virtio*",
>> > +    },{
>> > +        .name = "virtio-9p",
>> > +        .argfunc = generic_fuzzer_virtio_9p_args,
>> > +        .objects = "virtio*",
>> > +    },{
>> > +        .name = "virtio-9p-synth",
>> > +        .args = "-machine q35 -nodefaults "
>> > +        "-device virtio-9p,fsdev=hshare,mount_tag=hshare "
>> > +        "-fsdev synth,id=hshare",
>> > +        .objects = "virtio*",
>> >      },{
>> >          .name = "e1000",
>> >          .args = "-M q35 -nodefaults "
>> 
>> 
>>