[PATCH] syslogd +realpath

[Denis Vlasenko]

On Thursday 08 February 2007 20:39, Marc Leeman wrote:
> > I tried. /dev/log is created by bind() of AF_UNIX.
> 
> doh, on a RO fs?

It is indeed created with bind(), but on RO fs it fails.

#include <sys/types.h>
#include <sys/socket.h>
#include <sys/un.h>
int main(int argc, char **argv)
{
        struct sockaddr_un sunx;
        socklen_t addr_len;
        int sock_fd;
//      unlink(argv[1]);
        memset(&sunx, 0, sizeof(sunx));
        sunx.sun_family = AF_UNIX;
        strncpy(sunx.sun_path, argv[1], sizeof(sunx.sun_path));
        sock_fd = socket(AF_UNIX, SOCK_DGRAM, 0);
        if (sock_fd < 0) return 1;
        addr_len = sizeof(sunx.sun_family) + strlen(sunx.sun_path);
        bind(sock_fd, (struct sockaddr *) &sunx, addr_len);
        return 0;
}

Testing by stracing "testprog /dev/log2"

...
socket(PF_FILE, SOCK_DGRAM, 0)          = 4
bind(4, {sa_family=AF_FILE, path="/dev/log2"}, 11) = 0
exit_group(0)

and also now I see /dev/log2 socket.

Running that again:

...
socket(PF_FILE, SOCK_DGRAM, 0)          = 4
bind(4, {sa_family=AF_FILE, path="/dev/log2"}, 11) = -1 EADDRINUSE (Address already in use)
exit_group(0)


On RO fs (my / in RO, so I just do "testprog /bin/log2")

socket(PF_FILE, SOCK_DGRAM, 0)          = 4
bind(4, {sa_family=AF_FILE, path="/bin/log2"}, 11) = -1 EROFS (Read-only file system)
exit_group(0)  


> > If it exists, bind() will fail, even if no other process did bind().
> > That's why removing /dev/log is really needed.
> 
> No, /dev/ is not normally mounted on a rw filesystem (tmpfs). That's
> probably why a symlink was used in the firsts place (to /tmp/log)

Of sourse you may want to symlink /dev/log, and that's exactly why syslogd
tries to resolve the symlink - it needs to remove the socket, or else
it won't be able to bind to it.
--
vda