kern/181127: [PATCH] set{domain, host}name doesn't permit NUL terminated strings that are MAXHOSTNAMELEN long

[Bruce Evans]

On Thu, 8 Aug 2013, Garrett Cooper wrote:

> On Aug 8, 2013, at 4:35 AM, Bruce Evans wrote:
>
>> On Thu, 8 Aug 2013, Garrett Cooper wrote:
>>
>>>> Synopsis:       [PATCH] set{domain,host}name doesn't permit NUL terminated strings that are MAXHOSTNAMELEN long
>>> ...
>>>> Description:
>>> The noted link/patch fixes POSIX and generic requirement compliance for set{domain,host}name per the manpages by accounting for the fact that the string
>>> must be NUL terminated.
>>
>> The bugs seem to be mainly in the tests, so the proposed fix enlarges them.
>> MAXHOSTNAMELEN is already 1 larger than the POSIX limit {HOST_NAME_MAX}
>> (see the sysconf(3) sources).
>
> So the fix is bogus. Ok, missed that MAXHOSTNAMELEN was '\0' inclusive.
>
>>> Found with the NetBSD t_set{domain,host}name testcases:
>>>
>>> Before:
>>>
>>> $ pwd
>>> /usr/tests/lib/libc/gen
>>> $ sudo atf-run t_setdomainname | atf-report
>>> t_setdomainname (1/1): 3 test cases
>>>   setdomainname_basic: [0.019497s] Failed: /usr/src/lib/libc/tests/gen/t_setdomainname.c:66: setdomainname(domains[i],sizeof(domains[i])) == 0 not met
>>>   setdomainname_limit: [0.004173s] Passed.
>>>   setdomainname_perm: [0.005297s] Passed.
>>> [0.029872s]
>>
>> I'm not sure what these do, but according to the Synopsis,
>> set{domain,host}name correctly doesn't permit NUL terminated strings that
>> are MAXHOSTNAMELEN long (not counting space for the NUL).  MAXHOSTNAMELEN
>> counts space for the NUL and is 1 larger than {HOST_NAME_MAX}.
>
> Yes. It's kind of odd why NetBSD passes here, but this should work on FreeBSD as well as they aren't doing anything going out-of-bounds in the testcases (see https://github.com/yaneurabeya/freebsd/blob/master/lib/libc/tests/gen/t_setdomainname.c , https://github.com/yaneurabeya/freebsd/blob/master/lib/libc/tests/gen/t_sethostname.c if you're curious).

It uses MAXHOSTNAMELEN + 1 in (only) 1 place, and then seems to check that
setdomainname() on a null name with "length" MAXHOSTNAMELEN + 1 fails.
It doesn't seem to test any strings of nearly length MAXHOSTNAMELEN.

> ...
>
>>> @@ -314,11 +314,11 @@ sysctl_hostname(SYSCTL_HANDLER_ARGS)
>>>
>>> SYSCTL_PROC(_kern, KERN_HOSTNAME, hostname,
>>>    CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_PRISON | CTLFLAG_MPSAFE,
>>> -    (void *)(offsetof(struct prison, pr_hostname)), MAXHOSTNAMELEN,
>>> +    (void *)(offsetof(struct prison, pr_hostname)), MAXHOSTNAMELEN+1,
>>>    sysctl_hostname, "A", "Hostname");
>>> SYSCTL_PROC(_kern, KERN_NISDOMAINNAME, domainname,
>>>    CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_PRISON | CTLFLAG_MPSAFE,
>>> -    (void *)(offsetof(struct prison, pr_domainname)), MAXHOSTNAMELEN,
>>> +    (void *)(offsetof(struct prison, pr_domainname)), MAXHOSTNAMELEN+1,
>>>    sysctl_hostname, "A", "Name of the current YP/NIS domain");
>>> SYSCTL_PROC(_kern, KERN_HOSTUUID, hostuuid,
>>>    CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_PRISON | CTLFLAG_MPSAFE,
>>
>> The sysctls were originally simple SYSCTL_STRING()s and I think they
>> worked then.  Now they are quite complicated, to  support jails, etc.,
>> but they still use sysctl_handle_string() so I think they handle
>> (non)strings and (non)termination the same.  Note that
>> sysctl_handle_string() doesn't actually return strings unless the
>> buffer is large enough to hold the NUL terminator.  It just truncates.
>> This is reflected in the gethostname(3) API.  The name length for
>> gethostname() must be 1 larger than {HOST_NAME_MAX} to ensure
>> getting a string.  OTOH, the name length for sethostname(3) should
>> not include space for the NUL, so it must not be larger than
>> {HOST_NAME_MAX}.  If it is larger than {HOST_NAME_MAX}, then the
>> syscall will just fail.  If it is larger than the string length
>> (to include the NUL and possibly more) but not larger than
>> {HOST_NAME_MAX}, then the syscall will succeed and the string will
>> just be terminated more than once.  (It would be safer to write NULs
>> from the end of the string until the end of the buffer in all cases.)
>
> So translation is: is there's a bug in the sysctl handler after jail support was added and there's no reasonable way to fix it without reverting things back to their sane forms?

No.  I suspected a bug in the jail support, but couldn't see any.  You will
have to check with name and string lengths of nearly MAXHOSTNAMELEN + 1 on
(or whatever the kernel buffer size is for plain SYSCTL_STRING()) to see if
the jail support gives any differences.

Bruce