On 05/06/20 14:48, Peter Xu wrote: >>> The bug is that strtoul is "impossible" to use correctly. > Could I ask why? To see see how annoying the situation is, check out utils/cutils.c in QEMU; basically, it is very hard to do error handling. From the man page: Since strtoul() can legitimately return 0 or ULONG_MAX (ULLONG_MAX for strtoull()) on both success and failure, the calling program should set errno to 0 before the call, and then determine if an error occurred by checking whether errno has a nonzero value after the call. and of course no one wants to write code for that every time they have to parse a number. In addition, if the string is empty it returns 0, and of endptr is NULL it will accept something like "123abc" and return 123. So it is not literally impossible, but it is a poorly-designed interface which is a major source of bugs. On Rusty's API design levels[1][2], I would put it at 3 if I'm feeling generous ("Read the documentation and you'll get it right"), and at -4 to -7 ("The obvious use is wrong") if it's been a bad day. Therefore it's quite common to have a wrapper like int my_strtoul(char *p, char **endptr, unsigned long *result); The wrapper will: - check that the string is not empty - always return 0 or -1 because of the by-reference output argument "result" - take care of checking that the entire input string was parsed, for example by rejecting partial parsing of the string if endptr == NULL. This version gets a solid 7 ("The obvious use is probably the correct one"); possibly even 8 ("The compiler will warn if you get it wrong") because the output argument gives you better protection against overflow. Regarding overflow, there is a strtol but not a strtoi, so you need to have a temporary long and do range checking manually. Again, you will most likely make mistakes if you use strtol, while my_strtol will merely make it annoying but it should be obvious that you're getting it wrong. Paolo [1] https://ozlabs.org/~rusty/index.cgi/tech/2008-03-30.html [2] https://ozlabs.org/~rusty/index.cgi/tech/2008-04-01.html