On Wed, Aug 14, 2019 at 06:45:03PM +0200, Patrick Steinhardt wrote: > since commit 52aa1a661 (include/closestream: avoid close more > than once, 2019-06-13), util-linux fails to build on musl libc > based systems. The culprit here is that it introduced assignments > to stderr and stdout, while the C89 standard explicitly notes > that treating stderr and stdout as valid lvalues is not a > requirement for any conforming C implementation. musl libc > implemented these streams as `extern FILE *const`, and as a > result assigning to these variables causes compiler errors. The question is if close() for stdout and stderr is the right way to go. In an ideal world it would be enough to use ferror()+fflush(), unfortunately for example NFS has never reached an ideal world and it requires fclose()... See https://lists.gnu.org/r/bug-gnulib/2019-04/msg00061.html Florian (added to CC), also suggested to use dup3() for the descriptors and then fclose() for the new handle. It sounds like a pretty elegant solution how to avoid all the issues with NULL and it's also robust enough if you accidentally call close_stream() more than once. See https://bugzilla.redhat.com/show_bug.cgi?id=1732450#c4 Maybe we can improve include/closestream.h to use dup3(), than it would be possible keep all in the header file as inline functions. Comments? > Attached is a fix for this. Instead of assigning `NULL` to the > streams, util-linux now uses a static variable `streams_closed`. I don't think we need to check if we already performed this operation as it's always called only once by atexit() and with dup3() it will be robust enough. > Unfortunately, this fix necessitated some shifting around as > closestream was previously implemented as header, only, and > implementing static variables inside of a header is not going to > work due to them being static to the single compilation unit, > only. Thus I converted the code to move the implementation into > "lib/closestream.c". Karel -- Karel Zak <kzak@xxxxxxxxxx> http://karelzak.blogspot.com
