In a previous post, we looked at the /dev/fd directory.
I found out a few more things as well as FreeBSD and Linux special cases related to it.
Recently I wanted to check if standard input was connected to a pipe.
I looked at test(1) if there was something like -t,
however,
I only found -p which needs a file name,
not a file descriptor.
So I tried [ -p /dev/fd/0 ] but it didn't work.
This is, because in the FreeBSD setup without fdescfs(8), or with fdescfs(8) and default mount options, the files in /dev/fd are special character devices nodes:
% ls -l /dev/fd total 0 cr-xr-xr-x 1 root wheel 0x3 Dec 7 2024 0* cr-xr-xr-x 1 root wheel 0x4 Dec 7 2024 1* cr-xr-xr-x 1 root wheel 0x5 Dec 7 2024 2* cr-xr-xr-x 1 root wheel 0x6 Dec 7 2024 3* cr-xr-xr-x 1 root wheel 0x7 Dec 7 2024 4*
However, if you mount the fdescfs(8) with the option -o nodup, the entries correspond to the vnodes:
crw--w---- 1 root tty 0x6d Sep 24 22:23 0 crw--w---- 1 root tty 0x6d Sep 24 22:23 1 crw--w---- 1 root tty 0x6d Sep 24 22:23 2 drwxr-x--- 3 root wheel 17 Sep 24 22:22 3/ dr-xr-xr-x 2 root wheel 512 May 28 18:14 4/
Buuut, while opening file descriptors that belong to a regular file or a directory show up as a regular file or directory, pipes and FIFOs show up as... character devices.
Let's write some C to analyze this:
#include <sys/stat.h>
#include <fcntl.h>
#include <stdio.h>
int
main()
{
struct stat st;
stat("/dev/fd/0", &st);
printf("%o\n", st.st_mode);
fstat(0, &st);
printf("%o\n", st.st_mode);
int fd = open("/dev/fd/0", O_RDONLY);
fstat(fd, &st);
printf("%o\n", st.st_mode);
}
These are all the possible ways we can stat fd 0.
Now let's check:
% ./a.out </etc/passwd 100644 100644 100644 % mkfifo /tmp/myfifo % ./a.out </tmp/myfifo 20555 10644 10644 % date | ./a.out 20555 10000 10000
So indeed to get the proper st_mode we need to open the /dev/fd entry, thus test(1) cannot be used (on FreeBSD, on Linux it works).
fdescfs(5) also supports the options rdlnk and linrdlnk. These completely break opening a fifo via /dev/fd...
Working with non-blocking file handles is a conundrum that already djb complained about:
> The problem with O_NONBLOCK is that it's too heavy: > it hits an entire ofile, not just an fd.
Indeed, there's spooky action at a distance, when a different process changes your O_NONBLOCK setting:
#include <sys/wait.h>
#include <fcntl.h>
#include <unistd.h>
#include <stdio.h>
int
main()
{
int fds[2];
pipe(fds);
dprintf(1, "%o\n", fcntl(fds[1], F_GETFL));
int pid = fork();
if (pid == 0) {
fcntl(fds[1], F_SETFL, O_NONBLOCK);
dprintf(1, "in child: %d\n", fcntl(fds[1], F_GETFL));
_exit(0);
}
waitpid(pid, 0, 0);
dprintf(1, "%o\n", fcntl(fds[1], F_GETFL));
}
We run it and see:
% ./a.out 2 in child: 6 6
(That's why we need F_SETFL, the flags of F_SETFD are per file descriptor!)
As far as I can tell, there's no way in FreeBSD to get a second ofile on a pipe with different O_NONBLOCK setting.
However, in Linux it funnily works if you use /dev/fd:
#include <sys/wait.h>
#include <errno.h>
#include <fcntl.h>
#include <stdio.h>
#include <string.h>
#include <unistd.h>
int
main()
{
int fds[2];
pipe(fds);
dprintf(1, "%o\n", fcntl(fds[1], F_GETFL));
int pid = fork();
if (pid == 0) {
char *devfd;
asprintf(&devfd, "/dev/fd/%d", fds[1]);
int fd = open(devfd, O_WRONLY);
fcntl(fd, F_SETFL, O_NONBLOCK);
dprintf(1, "in child: %o\n", fcntl(fd, F_GETFL));
while (write(fd, "spam", 4) > 0)
;
dprintf(1, "error=%s\n", strerror(errno));
_exit(0);
}
waitpid(pid, 0, 0);
dprintf(1, "%o\n", fcntl(fds[1], F_GETFL));
write(fds[1], "spam", 4);
}
This program will print (on Linux 6.16):
% ./a.out 1 in child: 104001 error=Resource temporarily unavailable 1
and then hangs in the final write to prove it's indeed blocking.