I wrote:
> Meh --- the hazards of back-patching seem to me to be more hypothetical
> than the benefits. Still, I seem to be in the minority, so I withdraw
> the proposal to back-patch.
Actually, after digging around a bit, I'm excited about this again.
There are only a couple dozen places in our tree that pay any attention
to the result of (v)snprintf, but with the exception of psnprintf,
appendPQExpBufferVA, and one or two other places, *they're all assuming
C99 semantics*, and will fail to detect buffer overflow with the pre-C99
behavior.
Probably a lot of these are not live bugs because buffer overrun is
not ever going to occur in practice. But at least pg_upgrade and
pg_regress are constructing command strings including externally
supplied paths, so overrun doesn't seem impossible. If it happened,
they'd merrily proceed to execute a truncated command.
If we don't backpatch the snprintf change, we're morally obliged to
back-patch some other fix for these places. At least one of them,
in plperl's pport.h, is not our code and so changing it seems like
a bad idea.
Still want to argue for no backpatch?
regards, tom lane
PS: I also found a couple of places that are just wrong regardless
of semantics: they're checking overflow by "result > bufsize", not
"result >= bufsize". Will fix those in any case.