[FFmpeg-devel] [PATCH] avcodec/mjpegdec: Fixes runtime error: signed integer overflow: -24543 * 2031616 cannot be represented in type 'int'
wm4
nfxjfg at googlemail.com
Fri Apr 7 14:30:56 EEST 2017
On Fri, 7 Apr 2017 13:07:23 +0200
Michael Niedermayer <michael at niedermayer.cc> wrote:
> On Fri, Apr 07, 2017 at 12:35:37PM +0200, wm4 wrote:
> > On Fri, 7 Apr 2017 12:22:07 +0200
> > Michael Niedermayer <michael at niedermayer.cc> wrote:
> >
> > > On Fri, Apr 07, 2017 at 08:30:50AM +0200, wm4 wrote:
> > > > On Fri, 7 Apr 2017 02:17:37 +0200
> > > > Michael Niedermayer <michael at niedermayer.cc> wrote:
> > > >
> > > > > On Wed, Mar 29, 2017 at 11:47:31AM +0200, wm4 wrote:
> > > > > > On Mon, 27 Mar 2017 00:41:05 +0200
> > > > > > Michael Niedermayer <michael at niedermayer.cc> wrote:
> > > > > >
> > > > > > > On Sun, Mar 26, 2017 at 05:30:05PM -0400, Ronald S. Bultje wrote:
> > > > > > > > Hi,
> > > > > > > >
> > > > > > > > On Sun, Mar 26, 2017 at 3:31 PM, Michael Niedermayer <michael at niedermayer.cc
> > > > > > > > > wrote:
> > > > > > > >
> > > > > > > > > On Sun, Mar 26, 2017 at 07:41:58PM +0200, wm4 wrote:
> > > > > > > > > > On Sun, 26 Mar 2017 19:16:26 +0200
> > > > > > > > > > Michael Niedermayer <michael at niedermayer.cc> wrote:
> > > > > > > > > >
> > > > > > > > > > > On Sun, Mar 26, 2017 at 06:51:11PM +0200, wm4 wrote:
> > > > > > > > > > > > On Sun, 26 Mar 2017 18:11:01 +0200
> > > > > > > > > > > > Michael Niedermayer <michael at niedermayer.cc> wrote:
> > > > > > > > > > > >
> > > > > > > > > > > > > Fixes: 943/clusterfuzz-testcase-5114865297391616
> > > > > > > > > > > > >
> > > > > > > > > > > > > Found-by: continuous fuzzing process
> > > > > > > > > https://github.com/google/oss-fuzz/tree/master/targets/ffmpeg
> > > > > > > > > > > > > Signed-off-by: Michael Niedermayer <michael at niedermayer.cc>
> > > > > > > > > > > > > ---
> > > > > > > > > > > > > libavcodec/mjpegdec.c | 3 ++-
> > > > > > > > > > > > > 1 file changed, 2 insertions(+), 1 deletion(-)
> > > > > > > > > > > > >
> > > > > > > > > > > > > diff --git a/libavcodec/mjpegdec.c b/libavcodec/mjpegdec.c
> > > > > > > > > > > > > index f26e8a3f9a..e08b045fe7 100644
> > > > > > > > > > > > > --- a/libavcodec/mjpegdec.c
> > > > > > > > > > > > > +++ b/libavcodec/mjpegdec.c
> > > > > > > > > > > > > @@ -757,7 +757,8 @@ static int decode_block_progressive(MJpegDecodeContext
> > > > > > > > > *s, int16_t *block,
> > > > > > > > > > > > > uint16_t *quant_matrix,
> > > > > > > > > > > > > int ss, int se, int Al, int
> > > > > > > > > *EOBRUN)
> > > > > > > > > > > > > {
> > > > > > > > > > > > > - int code, i, j, level, val, run;
> > > > > > > > > > > > > + int code, i, j, val, run;
> > > > > > > > > > > > > + SUINT level;
> > > > > > > > > > > > >
> > > > > > > > > > > > > if (*EOBRUN) {
> > > > > > > > > > > > > (*EOBRUN)--;
> > > > > > > > > > > >
> > > > > > > > > > > > Please make the type either signed or unsigned. Making it both
> > > > > > > > > > > > (depending on the debug level) just to make the fuzzer happy (or
> > > > > > > > > > > > something more complicated than that?) isn't a good idea. You
> > > > > > > > > probably
> > > > > > > > > > > > want to make it always unsigned?
> > > > > > > > > > >
> > > > > > > > > > > No, i want to make it SUINT
> > > > > > > > > > >
> > > > > > > > > > > If it is always unsigned then its not possible to detect overflows
> > > > > > > > > > > without explicitly checking for overflows.
> > > > > > > > > > > If it is SUINT then ubsan can be used to detect overflows, this is
> > > > > > > > > > > usefull to test files showing artifacts but no decode errors.
> > > > > > > > > > >
> > > > > > > > > >
> > > > > > > > > > The point of these tools (static analyzers, sanitizers, fuzzers) is to
> > > > > > > > > > improve the correctness of the code.
> > > > > > > > >
> > > > > > > > > > SUINT is still defined to "int" if
> > > > > > > > > > CHECKED is not defined
> > > > > > > > >
> > > > > > > > > no
> > > > > > > > >
> > > > > > > > > internal.h:
> > > > > > > > > #ifdef CHECKED
> > > > > > > > > #define SUINT int
> > > > > > > > > #define SUINT32 int32_t
> > > > > > > > > #else
> > > > > > > > > #define SUINT unsigned
> > > > > > > > > #define SUINT32 uint32_t
> > > > > > > > > #endif
> > > > > > > > >
> > > > > > > > > I belive the rest of your mail assumes this condition is backward to
> > > > > > > > > how it is
> > > > > > > > >
> > > > > > > > > SUINT is not there to make any tools happy its there to allow finding
> > > > > > > > > overflows in debug more while having valid c code in normal builds
> > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > > > > Why don't we want to detect overflows in debug mode?
> > > > > > >
> > > > > > > like wm4 you read "#if A" as "#if not A", all your mail and questions
> > > > > > > seem based on reading the condition for SUINT flipped around
> > > > > > >
> > > > > > > in DEBUG mode CHECKED is enabled, SUINT is int and overflows are
> > > > > > > undefined behaviour which can be detected easily with ubsan.
> > > > > > >
> > > > > > > This allows us to debug samples producing artifacts but no decode
> > > > > > > errors due to overflows.
> > > > > > >
> > > > > > >
> > > > > > > in normal mode CHECKED is disabled, SUINT is unsigned and overflows
> > > > > > > are defined behavior. There must be no undefined behavior in releases
> > > > > > >
> > > > > > > maybe you belive everyone is using debug mode and the fuzzers run in
> > > > > > > debug mode. Maybe this is why everyone belives the condition is
> > > > > > > backward
> > > > > >
> > > > > > How would we know this? Maybe I've been assuming too much in order to
> > > > > > try to make sense out of it.
> > > > > >
> > > > > > > I might be wrong but unless you manually pass -DDEBUG you dont use
> > > > > > > debug mode, adding -DDEBUG is how our debug mode fate client tests that
> > > > > > > mode
> > > > > >
> > > > >
> > > > > > But why do you want to detect overflows in debug mode, if in release
> > > > >
> > > > > To debug.
> > > > > Like i want compiler warnings to "debug"
> > > > > or coverity warnings to find and correct bugs. == "debug"
> > > > >
> > > > > If a undamaged file triggers an overflow thats very likely a bug and
> > > > > a easy to fix one if one knows about the overflow.
> > > > > If one doesnt know of that overflow it can be very hard to find
> > > >
> > >
> > > > Then it's not a bug by definition, and you can ignore it.
> > >
> > > overflows have been bugs in the past
> > > warnings have pointed to bugs in the past
> > > ubsan has pointed to bugs in the past
> >
> > What I'm saying is:
> >
> > Either an overflow is a bug. Then you should not apply hacks like
> > making all types unsigned just to hide these bugs not show up on
> > sanitizers or static analyzers. And you shouldn't add another hack just
> > to make them show up again in debug mode. You could argue that this is
> > some kind of "hardening" technique (with the second hack still allowing
> > debugging it), but how about let's not? Just fix overflows when you
> > find them.
> >
> > Or the overflow is not a bug. Then the type should always be unsigned,
> > because overflows would be relatively meaningless. Why make the source
> > code harder to follow and add noise in debug mode just because of a
> > vague idea that unknown overflows could occur and which could be bugs?
> > If overflows are a problem, make the type always signed, and always fix
> > overflows as you encounter them.
>
> The problem is what we have is a useless fuzzed file triggering this
> the fuzzed file shows that the overflow can be triggered it tells us
> nothing about if its a real bug
Look at the code and determine if it could be a bug or not?
Changing a type to unsigned is a nice way to get rid of undefined
behavior, but do it only if the result of an overflow does not matter.
> can it be triggered also by a file we want to support?
> if so its a real bug
And how does SUINT help you make the decision?
> in some cases this can be awnsered by reading some specification
> still in reality what really matters if such files exist even if we
> have a spec and even if it does say what the range is.
> Some specs are non public, some lack detailed limits on intermediates
> of computations.
>
> We can spend a long time to determine if something is needed by a
> real file and thus a bug or not.
> This is what i generally do but often theres just no final clear yes
> or no. Theres a probably yes / probably no / maybe
> and for that SUINT comes in handy
> we fix the undefined behavior but if a user submits a file that doesnt
> decode correctly we can still flip a switch and use ubsan to see if
> it involves any overflows
Well there are 3 reasons why you'd change signed to unsigned:
1. unsigned overflow actually gives the intended result
2. avoiding undefined behavior when the result does not matter
3. shutting up a fuzzer without fixing the actual issue
None of these fix themselves when changing the type back to signed,
except for operations which needed to be signed in the first place. I'm
not sure how changing back to signed is supposed to help anyone with
actual overflows, since these would have been buggy in the first place.
Sure, it helps you better to spot a bug you could have spotted easier
with the original code when all was still signed.
But with 1. and 2. you'd have confidence in the change to unsigned. We
don't add arbitrary ifdefs and typedefs to other code either, just
because it'd help us to easily "revert" a change. And for 3. checking
the overflow explicitly would have been better.
So, you can fix the overflows by actually fixing the overflows, instead
of switching the type to unsigned and hoping it doesn't break anything.
And even if what you said about "flipping a switch" is true, you could
just edit the source code at the questionable location, which in fact
would be quicker than changing SUINT declared in a central location or
rebuilding the entire FFmpeg to change the debug level.
Maybe I'm indeed missing something here.
>
> This is simply minimizing the time needed to debug such issues
> * spend more and more time to make really really sure nothing valid
> can trigger an overflow, no unexpected odd files exist, ...
> * vs just run files reported by users under ubsan
>
> You could also say iam trying to use our userbases time to help
> double check if what i think is likely not a bug really is not a bug
I don't see how any of this helps users. They don't run fuzzers or
debug tools, and they are not going to try to change the definition of
SUINT.
If something is _really_ questionable we still could print an error
message if a calculation overflows. That would _actually_ help finding
the issue. (But please, let's not do this with every overflow check.)
More information about the ffmpeg-devel
mailing list