On Friday 09 October 2015 15:41:47 Pádraig Brady wrote:
> On 09/10/15 14:38, Pino Toscano wrote:
> > On Thursday 08 October 2015 15:46:42 Pádraig Brady wrote:
> >> On 08/10/15 13:47, Pino Toscano wrote:
> >>> Pass only SOCK_* flags to accept4, as they are the only documented
> >>> ones, and passing others may trigger EINVAL.
> >>> * tests/test-accept4.c: (main): Pass SOCK_CLOEXEC instead of
> >>> O_CLOEXEC | O_BINARY to accept4.
> >>> ---
> >>> ChangeLog | 7 +++++++
> >>> tests/test-accept4.c | 4 ++--
> >>> 2 files changed, 9 insertions(+), 2 deletions(-)
> >>>
> >>> diff --git a/ChangeLog b/ChangeLog
> >>> index 02d8bf8..3601eda 100644
> >>> --- a/ChangeLog
> >>> +++ b/ChangeLog
> >>> @@ -1,3 +1,10 @@
> >>> +2015-10-08 Pino Toscano <ptosc...@redhat.com>
> >>> +
> >>> + Pass only SOCK_* flags to accept4, as they are the only documented
> >>> + ones, and passing others may trigger EINVAL.
> >>> + * tests/test-accept4.c: (main): Pass SOCK_CLOEXEC instead of
> >>> + O_CLOEXEC | O_BINARY to accept4.
> >>> +
> >>> 2015-10-06 Pavel Raiskup <prais...@redhat.com>
> >>>
> >>> gnulib-tool: fix tests of 'extensions' module
> >>> diff --git a/tests/test-accept4.c b/tests/test-accept4.c
> >>> index b24af0b..b2e6fa8 100644
> >>> --- a/tests/test-accept4.c
> >>> +++ b/tests/test-accept4.c
> >>> @@ -43,7 +43,7 @@ main (void)
> >>>
> >>> errno = 0;
> >>> ASSERT (accept4 (-1, (struct sockaddr *) &addr, &addrlen,
> >>> - O_CLOEXEC | O_BINARY)
> >>> + SOCK_CLOEXEC)
> >>> == -1);
> >>> ASSERT (errno == EBADF);
> >>> }
> >>> @@ -54,7 +54,7 @@ main (void)
> >>> close (99);
> >>> errno = 0;
> >>> ASSERT (accept4 (99, (struct sockaddr *) &addr, &addrlen,
> >>> - O_CLOEXEC | O_BINARY)
> >>> + SOCK_CLOEXEC)
> >>> == -1);
> >>> ASSERT (errno == EBADF);
> >>> }
> >>
> >> That change looks good, though it also suggests that
> >> the implementation doesn't assume the availability of SOCK_CLOEXEC etc.
> >> I think we also may need the following included in your patch,
> >> to ensure the test compiles on all platforms, and that those
> >> constants are defined appropriately on all platforms?
> >
> > The idea seems ok -- should I merge it with my patch, or can/should it
> > go as separate patch?
> >
> >> diff --git a/lib/accept4.c b/lib/accept4.c
> >> index adf0989..992dfd0 100644
> >> --- a/lib/accept4.c
> >> +++ b/lib/accept4.c
> >> @@ -24,10 +24,6 @@
> >> #include "binary-io.h"
> >> #include "msvc-nothrow.h"
> >>
> >> -#ifndef SOCK_CLOEXEC
> >> -# define SOCK_CLOEXEC 0
> >> -#endif
> >> -
> >> int
> >> accept4 (int sockfd, struct sockaddr *addr, socklen_t *addrlen, int flags)
> >> {
> >> diff --git a/lib/sys_socket.in.h b/lib/sys_socket.in.h
> >> index d29cc09..2d9df45 100644
> >> --- a/lib/sys_socket.in.h
> >> +++ b/lib/sys_socket.in.h
> >> @@ -654,10 +654,16 @@ _GL_WARN_ON_USE (shutdown, "shutdown is not always
> >> POSIX compliant - "
> >>
> >> #if @GNULIB_ACCEPT4@
> >> /* Accept a connection on a socket, with specific opening flags.
> >> - The flags are a bitmask, possibly including O_CLOEXEC (defined in
> >> <fcntl.h>)
> >> - and O_TEXT, O_BINARY (defined in "binary-io.h").
> >> + The flags are a bitmask, possibly including SOCK_NONBLOCK,
> >> + SOCK_CLOEXEC, and O_TEXT, O_BINARY (defined in "binary-io.h").
> >> See also the Linux man page at
> >>
> >> <http://www.kernel.org/doc/man-pages/online/pages/man2/accept4.2.html>. */
> >> +# ifndef SOCK_CLOEXEC
> >> +# define SOCK_CLOEXEC O_CLOEXEC
> >> +# endif
> >> +# ifndef SOCK_NONBLOCK
> >> +# define SOCK_NONBLOCK O_NONBLOCK
> >> +# endif
> >> # if @HAVE_ACCEPT4@
> >> # if !(defined __cplusplus && defined GNULIB_NAMESPACE)
> >> # define accept4 rpl_accept4
> >
> > SOCK_CLOEXEC is used only in src/accept4.c, so that seems ok.
> > OTOH, SOCK_NONBLOCK is checked in tests/test-nonblocking.c, where it
> > would enable the code passing extra flags to the socket type; defining
> > could make that check failing in case the OS does not implement accept4
> > (and thus we are providing SOCK_*). What do you think?
>
> Yes good point.
>
> accept4() makes no sense without SOCK_CLOEXEC or SOCK_NONBLOCK,
> so my change to define make sense in that respect.
> Defining them to non zero will ensure for example EINVAL is
> returned for SOCK_NONBLOCK on platforms where we replace accept4()
> (though I suppose we could simulate that also).True: if gnulib is providing accept4 to some platform which does not provide it, then SOCK_CLOEXEC and SOCK_NONBLOCK ought to be provided too, as they are part of the interface. > However SOCK_CLOEXEC and SOCK_NONBLOCK may not be defined > on platforms with select(), and we probably shouldn't define > as user space may be doing: > > #ifdef SOCK_CLOEXEC > socket(..., SOCK_CLOEXEC); > #else > socket(); > fcntl(..., FD_CLOEXEC); > #endif One workaround to that, albeit local to gnulib, would be to define something like GNULIB_DEFINED_SOCK_NONBLOCK, and adapt tests/test-nonblocking.c to # if defined(SOCK_NONBLOCK) && !defined(GNULIB_DEFINED_SOCK_NONBLOCK) or, even better, add some configure check to test whether socket(.., .. | SOCK_NONBLOCK) works. However, the above would not work with 3rd party code doing that check; one could think that such code may be broken already, as the underlying kernel could not support those extra flags for socket. > So let's forget my adjustment and instead I suggest > merging this into your change: > > diff --git a/doc/glibc-functions/accept4.texi > b/doc/glibc-functions/accept4.texi > index 20386e9..b4114db 100644 > --- a/doc/glibc-functions/accept4.texi > +++ b/doc/glibc-functions/accept4.texi > @@ -16,4 +16,7 @@ programs that spawn child processes. > > Portability problems not fixed by Gnulib: > @itemize > +@item > +SOCK_CLOEXEC and SOCK_NONBLOCK may not be defined > +as they're also significant to the socket() function. > @end itemize > diff --git a/tests/test-accept4.c b/tests/test-accept4.c > index b24af0b..5a29680 100644 > --- a/tests/test-accept4.c > +++ b/tests/test-accept4.c > @@ -31,6 +31,10 @@ SIGNATURE_CHECK (accept4, int, (int, struct sockaddr *, > socklen_t *, int)); > > #include "macros.h" > > +#ifndef SOCK_CLOEXEC > +# define SOCK_CLOEXEC 0 > +#endif > + I was not unsure about this, but then I noticed we do the same also in lib/accept4.c itself, so it will not make things worse (platforms without accept4 and SOCK_* flags already have an accept4 implementation which basically works like socket). Should I merge the above bit in my patch? What about authorship of it? Thanks, -- Pino Toscano
signature.asc
Description: This is a digitally signed message part.
