Intrepid gcc -O2 breaks string appending with sprintf(), due to fortify source patch

Anders Kaseorg noticed that the use of _FORTIFY_SOURCE breaks a specific use of
sprintf (see attached):

$ gcc -O0 -o foo foo.c && ./foo
not fail
$ gcc -O2 -o foo foo.c && ./foo
not fail
$ gcc -O2 -D_FORTIFY_SOURCE=2 -o foo foo.c && ./foo
fail

The original report was filed in Ubuntu, where -D_FORTIFY_SOURCE=2 is enabled by
default: https://bugs.launchpad.net/ubuntu/+source/glibc/+bug/305901

C99 states:
The sprintf function is equivalent to fprintf, except that the output is written
into an array (specified by the argument s) rather than to a stream. A null
character is written at the end of the characters written; it is not counted as
part of the returned value. If copying takes place between objects that overlap,
the behavior is undefined.

The man page does not mention this limitation, and prior to the use of
__sprintf_chk, this style of call worked as expected. As such, a large volume
of source code uses this style of call:
http://web.mit.edu/andersk/Public/sprintf-results

It seems that it would make sense to fix __sprintf_chk, or very loudly mention
the C99-described overlap-is-undefined behavior in sprintf documentation.

Created attachment 3095
test case

sprintf(buf, "%sfoo", buf) is UNDEFINED.

Thanks for the clarification. However, I think it is still a bug that the
limitation is not mentioned in the manpage.

Then contact whoever wrote it.

man 3p sprintf certainly documents it:
"If copying takes place between objects that overlap as a result of a call
to sprintf() or snprintf(), the results are undefined."

I have submitted a patch for linux-manpages:
http://thread.gmane.org/gmane.linux.man/639

(In reply to comment #6)
> I have submitted a patch for linux-manpages:
> http://thread.gmane.org/gmane.linux.man/639

I've applied the following patch for man-pages-3.16.

--- a/man3/printf.3
+++ b/man3/printf.3
@@ -133,6 +133,17 @@ string that specifies how subsequent arguments (or
arguments accessed via
 the variable-length argument facilities of
 .BR stdarg (3))
 are converted for output.
+
+C99 and POSIX.1-2001 specify that the results are undefined if a call to
+.BR sprintf (),
+.BR snprintf (),
+.BR vsprintf (),
+or
+.BR vsnprintf ()
+would cause to copying to take place between objects that overlap
+(e.g., if the target string array and one of the supplied input arguments
+refer to the same buffer).
+See NOTES.
 .SS "Return value"
 Upon successful return, these functions return the number of characters
 printed (not including the
@@ -851,6 +862,26 @@ and conversion characters \fBa\fP and \fBA\fP.
 glibc 2.2 adds the conversion character \fBF\fP with C99 semantics,
 and the flag character \fBI\fP.
 .SH NOTES
+Some programs imprudently rely on code such as the following
+
+ sprintf(buf, "%s some further text", buf);
+
+to append text to
+.IR buf .
+However, the standards explicitly note that the results are undefined
+if source and destination buffers overlap when calling
+.BR sprintf (),
+.BR snprintf (),
+.BR vsprintf (),
+and
+.BR vsnprintf ().
+.\" http://sourceware.org/bugzilla/show_bug.cgi?id=7075
+Depending on the version of
+.BR gcc (1)
+used, and the compiler options employed, calls such as the above will
+.B not
+produce the expected results.
+
 The glibc implementation of the functions
 .BR snprintf ()
 and

Created attachment 3625
work-around pre-trunc behavior

This patch restores the prior sprintf behavior. Looking through
_IO_str_init_static_internal seems to indicate that nothing actually requires
"s" to lead with a NULL. Is there anything wrong with this work-around, which
could be used until the number of affected upstream sources is not quite so
large?

*** Bug 260998 has been marked as a duplicate of this bug. ***
Seen from the domain http://volichat.com
Page where seen: http://volichat.com/adult-chat-rooms
Marked for reference. Resolved as fixed @bugzilla.

I'd still like to have this patch applied -- while we can claim the behavior is "undefined", it is not, in fact, undefined. It behaves one way without -D_FORTIFY_SOURCE=2, and differently with it. And that difference doesn't need to exist. Ubuntu carried this patch for quite a while.

The point of _FORTIFY_SOURCE is to expose undefined behaviour.

It's not defined in POSIX, but it has worked a certain way in glibc for decades. There's no _reason_ to break it for _FORTIFY_SOURCE. Pre-truncating just silently breaks programs and does weird stuff. If you want to expose it with _FORITFY_SOURCE then have vsprintf notice that the target and first format argument are the same variable, and refuse to build.

Either pretruncation should be eliminated, or the undefined behavior should be explicitly detected and dealt with. Just having programs lose data while running with no indication of the cause seems like a terrible user experience.

It might be a good idea to take this discussion to the libc-alpha mailing list.

So I'd like to bring this back up and reiterate the issue: there is no benefit to the early truncation, and it actively breaks lots of existing software (which is why Debian and Ubuntu have had this fix for 10 years now).

What is the _benefit_ of early truncation that justifies breaking so many existing cases?

Can glibc please take this patch? http://paste.ubuntu.com/p/CbrxmSfKD4/

There was a pretty lengthy discussion on this late last year:

https://sourceware.org/ml/libc-alpha/2018-12/msg00838.html

where the behaviour breakage was introduced in the non-fortified path and then reverted. It might be a good idea to resume that discussion for the fortified case as well.

(In reply to Kees Cook from comment #14)
> So I'd like to bring this back up and reiterate the issue: there is no
> benefit to the early truncation, and it actively breaks lots of existing
> software (which is why Debian and Ubuntu have had this fix for 10 years now).
>
> What is the _benefit_ of early truncation that justifies breaking so many
> existing cases?

ideally sprintf, snprintf and sprintf_chk would be able to share code and have identical behaviour (currently there is a lot of duplicated logic in glibc with a potential for inconsistent behaviour).

that said, i think _FORTIFY_SOURCE should detect undefined behaviour if possible since it's a bug that breaks portability.

note that it does not matter what guarantees a library documents: there are plenty of precedents for compiler optimizations to break code based on ub in library calls, a compiler can remove all code paths leading to a sprintf(s, "%s", s), trying to make such code work in glibc is just hiding a ticking time bomb.

(In reply to Szabolcs Nagy from comment #16)
> (In reply to Kees Cook from comment #14)
> > So I'd like to bring this back up and reiterate the issue: there is no
> > benefit to the early truncation, and it actively breaks lots of existing
> > software (which is why Debian and Ubuntu have had this fix for 10 years now).
> >
> > What is the _benefit_ of early truncation that justifies breaking so many
> > existing cases?

I wonder if the early truncation was introduced to avoid cases where aliasing can be used to avoid fortify length checks. But then again, truncation might not effectively prevent that after all. And we do not seem to use strlen followed by strcpy in vfprintf.

I haven't looked at this in detail, though.

> ideally sprintf, snprintf and sprintf_chk would be able to share code and
> have identical behaviour (currently there is a lot of duplicated logic in
> glibc with a potential for inconsistent behaviour).

Not sure what you mean by this. The core vfprintf engine is shared, of course.