Re: Bash printf should diagnose integer overflow

[Paul Eggert]

On 3/18/24 12:41, Chet Ramey wrote:
> I'm not sure what you're using, but that was not my experience on
> macOS.

I am using Fedora 39 (the current version) on x86-64. That could explain 
our differing experiences.

I see several diagnostics (mostly diff output) with "make check" on 
Fedora 39. The diagnostics can vary from run to run, i.e., they aren't 
necessarily reproducible. I assumed these were OK because I got 'em 
before any patches. From my point of view the diff output was sort of 
random, so I eyeballed it and guessed which outputs mattered and which 
didn't. Apparently I guessed incorrectly with fw.

At some point I suppose it'd be nice if 'make check' succeeded (exit 
status 0) or failed (nonzero exit status) so that it is easy for 
non-experts to tell which diagnostics matter; that's what many other 
packages do. For now I'd rather focus on the integer overflow issues in 
Bash, while they're fresh in my mind.


> First, the patched version doesn't build on macOS because your patches
> don't include <stdbool.h>. Once you get past that, printf goes into an
> infinite loop on
>
> printf -v s "%b" ""
>
> in printstr because `fw' is used unititialized (and randomly set to some
> ridiculously large value). That and the old test's incorrect expectation
> that a field width of 9223372036854775825 would always overflow to -1 (a
> left-adjusted field width of 1) instead of being flagged as overflow are
> the UB I was talking about.

Yes, I see now. I didn't get that behavior on Fedora, perhaps because 
the junk in fw was benign there.

Perhaps at some point we could enable more of GCC's static checking to 
catch silly mistakes like that. Again, a task for another time.


> It should be mostly there in the changes I pushed today, once I made it
> through the above.

Thanks, I checked the devel branch against what I submitted, found a few 
errors, and while reviewing all this found and fixed a few other 
integer-overflow issues in Bash. Proposed patches attached, in "git 
format-patch" format so you can use "git am" on them. If there's some 
reason a patch shouldn't be applied please let me know so that I can 
stop worrying about that subissue.

0001-Improve-use-of-HAVE_C_BOOL.patch
Description: Text Data

0002-Minor-mkseq-clarification-tuning.patch
Description: Text Data

0003-Simplify-mkseq-control.patch
Description: Text Data

0004-printf-M.Nq-should-respect-N-even-if-M-overflows.patch
Description: Text Data

0005-Fix-bug-with-printf-.NQ-when-N-is-large.patch
Description: Text Data

0006-Omit-unnecessary-INT_MAX-checks-in-printf_builtin.patch
Description: Text Data

0007-Omit-unused-typemax.h-macros.patch
Description: Text Data

0008-Omit-unused-stringlist-functions.patch
Description: Text Data

0009-Define-NULL-offsetof-in-bashansi.h.patch
Description: Text Data

0010-Do-not-crash-if-NMEMB-SIZE-overflows-in-alloc.patch
Description: Text Data