Hi Bruno,
When I build a coreutils snapshot with -D_FORTIFY_SOURCE=2 on a
relatively recent fedora-based system, seq always aborts like this:
$ ./seq 1
*** %n in writable segment detected ***
1zsh: abort ./seq 1
[Exit 134 (ABRT)]
That is due to the fact that vasnprintf writes %n into a format
string that is subsequently used by snprintf.
And why is that done? For portability to crufty systems on which
s*printf doesn't even return a valid count.
But the particular bug that code is working around doesn't
even affect glibc's snprintf function.
This is why gnulib should be written to rely on posix- (or c99-)
compliant functions whenever possible: so that conforming systems
aren't penalized. I haven't looked closely enough, but perhaps
vasnprintf itself could require snprintf-posix? It looks like
that might induce a dependency loop, though, since the snprintf-posix
module uses vasnprintf.c.
Here's a minimal change to avoid using %n in this specific case.
It works because there is already code to handle the case in which
there is no %n directive, i.e., when fbp[1] == '\0'.
Another approach is to avoid using %n with any glibc-based system.
diff --git a/lib/vasnprintf.c b/lib/vasnprintf.c
index f563823..9692562 100644
--- a/lib/vasnprintf.c
+++ b/lib/vasnprintf.c
@@ -3384,7 +3384,7 @@ VASNPRINTF (DCHAR_T *resultbuf, size_t *lengthp,
else
#endif
*fbp = dp->conversion;
-#if USE_SNPRINTF
+#if USE_SNPRINTF && _FORTIFY_SOURCE != 2
fbp[1] = '%';
fbp[2] = 'n';
fbp[3] = '\0';
