Hi, Simon Josefsson <si...@josefsson.org> writes:
> l...@gnu.org (Ludovic Courtès) writes: > >> Simon Josefsson <si...@josefsson.org> writes: >> >>> l...@gnu.org (Ludovic Courtès) writes: >>> >>>> What I meant to say is that it should return an error because resolving >>>> the ‘does-not-exist’ service fails. >>> >>> Does it only fail (well, falsely succeed) for unknown services? >> >> When asked for a host name that does not resolve, it returns an error, >> as expected. > > For unknown services? Interesting. What about known services, does it > resolve them correctly? I don’t know, but I’d expect it to work since otherwise the function is really useless. > If the problem is that unknown services (for existing names) results in > port 0, we could work around that by wrapping the system getaddrinfo in > a check whether port is 0 and then return EAI_SERVICE. Possibly, unless PORT = 0 can have a special meaning, but that doesn’t appear to be the case. [...] >>>>>> If it is indeed a bug, should we try to work around it in Gnulib? >>>>>> Any idea how this could be done? >>>>> >>>>> Given that this is a bug (which isn't clear to me yet), we could work >>>>> around it. >>>> >>>> I agree in principle, of course, but I can’t think of a way to work >>>> around that since we’ll always have to rely on libc at one point or >>>> another. >>> >>> It should be possible to look up services in other ways (getservbyname), >>> right? >> >> Yes, but what if getservbyname is similarly broken? That wouldn’t be >> surprising if these functions are factorized. > > We can replace broken getservbyname too. Yes, but sooner or later, we’ll have to build on a libc function that works. :-) >>> If you use the getaddrinfo replacement in gnulib on that system, >>> doesn't it produce correct results? >> >> It’s not used by default. How can I force it to be used? > > Try building using './configure gl_cv_func_getaddrinfo=no'. OK, long story short: Guile has new ‘getaddrinfo’ bindings, with the same parameters as the C version. The tests are here: http://git.savannah.gnu.org/gitweb/?p=guile.git;a=blob;f=test-suite/tests/net-db.test When run with the Gnulib replacement on Darwin 9.2.2, we get: --8<---------------cut here---------------start------------->8--- Running net-db.test ERROR: net-db.test: getaddrinfo: 127.0.0.1, any service - arguments: ((getaddrinfo-error 3)) ERROR: net-db.test: getaddrinfo: 127.0.0.1:80 - arguments: ((getaddrinfo-error 3)) ERROR: net-db.test: getaddrinfo: port 80 - arguments: ((getaddrinfo-error 3)) ERROR: net-db.test: getaddrinfo: port 80 with family and socket type - arguments: ((getaddrinfo-error 3)) FAIL: net-db.test: getaddrinfo: no name FAIL: net-db.test: getaddrinfo: wrong service name --8<---------------cut here---------------end--------------->8--- Error code 3 is ‘EAI_BADFLAGS’. When run without the Gnulib replacement, the only test that fails is “wrong service name”. IOW the Gnulib replacement seems to be bogus. Sorry, I can’t really provide better info. >>>>> But it is also not clear to me that the test-case above indicate a >>>>> useful way of using getaddrinfo, what kind of application uses >>>>> parameters like that, and what result did you expect? >>>> >>>> An error, because the service name doesn’t resolve. >>> >>> What if the system is configured (or programmed) to resolve all services >>> names? Isn't that a valid POSIX compliant getaddrinfo implementation? >>> It's not particular useful one, I admit. >> >> But then a given program wouldn’t be able to do anything useful with it. >> For instance, getaddrinfo ("host", "http") would be useless if the >> resulting port were not 80, because clients elsewhere may still get 80, >> because that’s what IANA has always said. > > Right. I'm not sure how gnulib should behave here. > > First, there is nothing in POSIX (I think?) that says "http" must always > resolve to port 80, and that "unknown-non-existing-service" must result > in an error. To the contrary, getaddrinfo/getservbyname typically reads > /etc/services, which allows administrators to modify the "http" port to > something else. So "http" is likely not even intended to be guaranteed > to translate into 80 by POSIX. Right. > But anything but 80 is useless for portable applications. Should > gnulib to replace a system getaddrinfo that doesn't translate "http" > to 80? No, I think this is getting too far. My initial request was: make sure getaddrinfo(3) returns an error when asked for non-existent service. The only OS that fails at this test is Darwin 9.2.2. Thanks, Ludo’.
