In order to avoid closing standard streams multiple times, commit
52aa1a661 (include/closestream: avoid close more than once, 2019-06-13)
introduced code to set the standard output and error streams to `NULL`.
As musl libc defines standard streams as constant pointers, the change
causes compiler errors on systems with that libc. According to ISO C89,
being able to assign to the standard text streams is not a requirement
for any C implementation, see footnote 238 in chapter ยง7.19.5.6:
The primary use of the freopen function is to change the file
associated with a standard text stream (stderr, stdin, or stdout),
as those identifiers need not be modifiable lvalues to which the
value returned by the fopen function may be assigned.
This commit implements a new function `flush_standard_stream` that tries
to reliably flush standard streams without actually closing them. By not
calling fclose(3P), we can neatly avoid the issue of accessing standard
streams in an unspecified state and thus remove the infringing `NULL`
assignments.
Properly flushing standard streams without fclose(3P) proves to be more
intricate than one may expect, though, as some filesystems like NFS may
defer flushing until they see a close(3P) of the underlying descriptor.
One may call fsync(3P) to remedy that, but this may incur a heavy
performance penalty in some scenarios. To work around the issue and
still get proper errors, we duplicate the stream's file descriptor and
close that one instead, which is sufficient to cause a flush.
Note that both `close_stdout` and `close_stdout_atexit` are misnamed
after this change, as we do not actually close the streams now. In order
to avoid unnecessary code churn, we still retain their current names.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
return 0;
}
+static inline int
+flush_standard_stream(FILE *stream)
+{
+ int fd;
+
+ errno = 0;
+
+ if (ferror(stream) != 0 || fflush(stream) != 0)
+ goto error;
+
+ /*
+ * Calling fflush is not sufficient on some filesystems
+ * like e.g. NFS, which may defer the actual flush until
+ * close. Calling fsync would help solve this, but would
+ * probably result in a performance hit. Thus, we work
+ * around this issue by calling close on a dup'd file
+ * descriptor from the stream.
+ */
+ if ((fd = fileno(stream)) < 0 || (fd = dup(fd)) < 0 || close(fd) != 0)
+ goto error;
+
+ return 0;
+error:
+ return (errno == EBADF) ? 0 : EOF;
+}
+
/* Meant to be used atexit(close_stdout); */
static inline void
close_stdout(void)
{
- if (stdout && close_stream(stdout) != 0 && !(errno == EPIPE)) {
+ if (flush_standard_stream(stdout) != 0 && !(errno == EPIPE)) {
if (errno)
warn(_("write error"));
else
_exit(CLOSE_EXIT_CODE);
}
- if (stderr && close_stream(stderr) != 0)
+ if (flush_standard_stream(stderr) != 0)
_exit(CLOSE_EXIT_CODE);
-
- stdout = NULL;
- stderr = NULL;
}
static inline void