On Thu, Aug 12, 2021 at 03:43:56PM -0500, Eric Blake wrote:
$ ./nbdkit -f memory 1 --filter=delay delay-read=1oopsms
Oops - our use of sscanf didn't check for trailing garbage, and this is behaving as delay-read=1ms.
$ ./nbdkit -fv memory 1 --filter=delay delay-read=999999999999999999999ms
Using gdb, I see that in glibc this results in the same as delay-read=4294967295, but that behavior is unspecified by POSIX and may result in other values on other platforms. Better would be detecting overflow, but sscanf() cannot detect numeric overflow.
Detecting trailing garbage could be done with sscanf(value, "%ums%n", r, &n) == 1 followed by checking that n consumed strlen(value) bytes, but detecting overflow really needs strtol() rather than sscanf.
We have other filters and plugins that use sscanf. As long as their inputs come from stable sources (such as scanning kernel /proc files) or don't parse numbers, that is safe; but in general, use of sscanf to parse user-provided data is risky.
I'm wondering if we want to add int nbdkit_parse_int_suffix (const char *what, const char *str, char **suff, int *r); and friends, which sets *suff to the first unparsed byte in str on a successful integer prefix parse. -- Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3266 Virtualization: qemu.org | libvirt.org