On 12.08.2012 01:24, Peter Maydell wrote: > POSIX allows sendmsg() and recvmsg() to fail EMSGSIZE if passed a zero > msg.msg_iovlen (in particular the MacOS X implementation will do this).
> Handle the case where iov_send_recv() is passed a zero byte count > explicitly, to avoid accidentally depending on the OS to treat zero > msg_iovlen as a no-op. > Signed-off-by: Peter Maydell <peter.mayd...@linaro.org> > --- > This is what was causing 'make check' to fail on MacOS X. > The other option was to declare that a zero bytecount was illegal, I guess. I don't think sending/receiving zero bytes is a good idea in the first place. Which test were failed on MacOS? Was it failing at test-iov "random I/O"? I thought I ensured that the test does not call any i/o function with zero "count" argument. Might be I was wrong, and in that case THAT place should be fixed instead. Can you provide a bit more details please? The whole thing is actually interesting: this is indeed a system- dependent corner case which should be handled in the code to make the routine consistent. But how to fix this is an open question I think. Your approach seems to be best, but we as well may print a warning there... Thank you! /mjt > iov.c | 7 +++++++ > 1 file changed, 7 insertions(+) > > diff --git a/iov.c b/iov.c > index b333061..60705c7 100644 > --- a/iov.c > +++ b/iov.c > @@ -146,6 +146,13 @@ ssize_t iov_send_recv(int sockfd, struct iovec *iov, > unsigned iov_cnt, > { > ssize_t ret; > unsigned si, ei; /* start and end indexes */ > + if (bytes == 0) { > + /* Catch the do-nothing case early, as otherwise we will pass an > + * empty iovec to sendmsg/recvmsg(), and not all implementations > + * accept this. > + */ > + return 0; > + } > > /* Find the start position, skipping `offset' bytes: > * first, skip all full-sized vector elements, */