F_DUPFD_CLOEXEC broken in 3.7.0

F_DUPFD_CLOEXEC broken in 3.7.0

[Richard W.M. Jones]

[-- Attachment #1: Type: text/plain, Size: 1432 bytes --]

Let's move this to LKML ...

On Mon, Oct 08, 2012 at 10:53:25PM +0100, Richard W.M. Jones wrote:
> On Mon, Oct 08, 2012 at 10:50:30PM +0100, Richard W.M. Jones wrote:
> [.. discussion on gnulib test-cloexec test snipped ..]
> > I'm suspicious this is a kernel bug:
> > 
> > creat("test-cloexec.tmp", 0600)         = 3
> > fcntl(3, F_GETFD)                       = 0
> > fcntl(3, F_GETFD)                       = 0
> > fcntl(3, F_SETFD, FD_CLOEXEC)           = 0
> > fcntl(3, F_GETFD)                       = 0x1 (flags FD_CLOEXEC)
> > fcntl(3, F_GETFD)                       = 0x1 (flags FD_CLOEXEC)
> > fcntl(3, F_SETFD, 0)                    = 0
> > fcntl(3, F_GETFD)                       = 0
> > fcntl(3, F_DUPFD_CLOEXEC, 0)            = 4
> > fcntl(4, F_GETFD)                       = 0
> > write(2, "test-cloexec.c:97: assertion failed\n", 36) = 36
> > 
> > It seems to me from the description in the man page that
> > F_DUPFD_CLOEXEC ought to be setting the FD_CLOEXEC flag on file
> > descriptor 4, so either it's not or else F_GETFD isn't reading the
> > flag for some reason.
> 
> Al Viro (CC'd) made some changes in this area recently ..

Attached is a self-contained test program that demonstrates the bug.

Rich.

-- 
Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones
libguestfs lets you edit virtual machines.  Supports shell scripting,
bindings from many languages.  http://libguestfs.org

[-- Attachment #2: cloexec.c --]
[-- Type: text/plain, Size: 578 bytes --]

/* Test if F_DUPFD_CLOEXEC works right.
 * by Richard W.M. Jones.
 */

#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <assert.h>

int
main ()
{
  char *file = "/tmp/cloexec.tmp";
  int fd = creat (file, 0600);
  int fd2;
  int flags;

  assert (fd >= 0);

  fd2 = fcntl (fd, F_DUPFD_CLOEXEC, 0);
  assert (fd2 >= 0);

  flags = fcntl (fd2, F_GETFD);
  assert (flags >= 0);

  if ((flags & FD_CLOEXEC) == 0) {
    fprintf (stderr, "F_DUPFD_CLOEXEC failed to set FD_CLOEXEC flag!\n");
    exit (EXIT_FAILURE);
  }

  exit (EXIT_SUCCESS);
}

Re: F_DUPFD_CLOEXEC broken in 3.7.0

[Cong Wang]

On Tue, Oct 9, 2012 at 6:05 AM, Richard W.M. Jones <rjones@redhat.com> wrote:
> Let's move this to LKML ...
>
> On Mon, Oct 08, 2012 at 10:53:25PM +0100, Richard W.M. Jones wrote:
>> On Mon, Oct 08, 2012 at 10:50:30PM +0100, Richard W.M. Jones wrote:
>> [.. discussion on gnulib test-cloexec test snipped ..]
>> > I'm suspicious this is a kernel bug:
>> >
>> > creat("test-cloexec.tmp", 0600)         = 3
>> > fcntl(3, F_GETFD)                       = 0
>> > fcntl(3, F_GETFD)                       = 0
>> > fcntl(3, F_SETFD, FD_CLOEXEC)           = 0
>> > fcntl(3, F_GETFD)                       = 0x1 (flags FD_CLOEXEC)
>> > fcntl(3, F_GETFD)                       = 0x1 (flags FD_CLOEXEC)
>> > fcntl(3, F_SETFD, 0)                    = 0
>> > fcntl(3, F_GETFD)                       = 0
>> > fcntl(3, F_DUPFD_CLOEXEC, 0)            = 4
>> > fcntl(4, F_GETFD)                       = 0
>> > write(2, "test-cloexec.c:97: assertion failed\n", 36) = 36
>> >
>> > It seems to me from the description in the man page that
>> > F_DUPFD_CLOEXEC ought to be setting the FD_CLOEXEC flag on file
>> > descriptor 4, so either it's not or else F_GETFD isn't reading the
>> > flag for some reason.
>>
>> Al Viro (CC'd) made some changes in this area recently ..
>
> Attached is a self-contained test program that demonstrates the bug.
>

Seems we passed a wrong flag to f_dupfd()...
Does the following patch help?

diff --git a/fs/fcntl.c b/fs/fcntl.c
index 8f70429..71a600a 100644
--- a/fs/fcntl.c
+++ b/fs/fcntl.c
@@ -258,7 +258,7 @@ static long do_fcntl(int fd, unsigned int cmd,
unsigned long arg,
                err = f_dupfd(arg, filp, 0);
                break;
        case F_DUPFD_CLOEXEC:
-               err = f_dupfd(arg, filp, FD_CLOEXEC);
+               err = f_dupfd(arg, filp, O_CLOEXEC);
                break;
        case F_GETFD:
                err = get_close_on_exec(fd) ? FD_CLOEXEC : 0;

Re: F_DUPFD_CLOEXEC broken in 3.7.0

[Richard W.M. Jones]

On Tue, Oct 09, 2012 at 10:07:22PM +0800, Cong Wang wrote:
> On Tue, Oct 9, 2012 at 6:05 AM, Richard W.M. Jones <rjones@redhat.com> wrote:
> > Let's move this to LKML ...
> >
> > On Mon, Oct 08, 2012 at 10:53:25PM +0100, Richard W.M. Jones wrote:
> >> On Mon, Oct 08, 2012 at 10:50:30PM +0100, Richard W.M. Jones wrote:
> >> [.. discussion on gnulib test-cloexec test snipped ..]
> >> > I'm suspicious this is a kernel bug:
> >> >
> >> > creat("test-cloexec.tmp", 0600)         = 3
> >> > fcntl(3, F_GETFD)                       = 0
> >> > fcntl(3, F_GETFD)                       = 0
> >> > fcntl(3, F_SETFD, FD_CLOEXEC)           = 0
> >> > fcntl(3, F_GETFD)                       = 0x1 (flags FD_CLOEXEC)
> >> > fcntl(3, F_GETFD)                       = 0x1 (flags FD_CLOEXEC)
> >> > fcntl(3, F_SETFD, 0)                    = 0
> >> > fcntl(3, F_GETFD)                       = 0
> >> > fcntl(3, F_DUPFD_CLOEXEC, 0)            = 4
> >> > fcntl(4, F_GETFD)                       = 0
> >> > write(2, "test-cloexec.c:97: assertion failed\n", 36) = 36
> >> >
> >> > It seems to me from the description in the man page that
> >> > F_DUPFD_CLOEXEC ought to be setting the FD_CLOEXEC flag on file
> >> > descriptor 4, so either it's not or else F_GETFD isn't reading the
> >> > flag for some reason.
> >>
> >> Al Viro (CC'd) made some changes in this area recently ..
> >
> > Attached is a self-contained test program that demonstrates the bug.
> >
> 
> Seems we passed a wrong flag to f_dupfd()...
> Does the following patch help?
> 
> diff --git a/fs/fcntl.c b/fs/fcntl.c
> index 8f70429..71a600a 100644
> --- a/fs/fcntl.c
> +++ b/fs/fcntl.c
> @@ -258,7 +258,7 @@ static long do_fcntl(int fd, unsigned int cmd,
> unsigned long arg,
>                 err = f_dupfd(arg, filp, 0);
>                 break;
>         case F_DUPFD_CLOEXEC:
> -               err = f_dupfd(arg, filp, FD_CLOEXEC);
> +               err = f_dupfd(arg, filp, O_CLOEXEC);
>                 break;
>         case F_GETFD:
>                 err = get_close_on_exec(fd) ? FD_CLOEXEC : 0;

Yes ..  the same patch has already been sent upstream.

Rich.

-- 
Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones
New in Fedora 11: Fedora Windows cross-compiler. Compile Windows
programs, test, and build Windows installers. Over 70 libraries supprt'd
http://fedoraproject.org/wiki/MinGW http://www.annexia.org/fedora_mingw

[regression] F_DUPFD_CLOEXEC breakage fix

[Al Viro]

On Mon, Oct 08, 2012 at 10:53:25PM +0100, Richard W.M. Jones wrote:
> On Mon, Oct 08, 2012 at 10:50:30PM +0100, Richard W.M. Jones wrote:
> [.. discussion on gnulib test-cloexec test snipped ..]
> > I'm suspicious this is a kernel bug:
> > 
> > creat("test-cloexec.tmp", 0600)         = 3
> > fcntl(3, F_GETFD)                       = 0
> > fcntl(3, F_GETFD)                       = 0
> > fcntl(3, F_SETFD, FD_CLOEXEC)           = 0
> > fcntl(3, F_GETFD)                       = 0x1 (flags FD_CLOEXEC)
> > fcntl(3, F_GETFD)                       = 0x1 (flags FD_CLOEXEC)
> > fcntl(3, F_SETFD, 0)                    = 0
> > fcntl(3, F_GETFD)                       = 0
> > fcntl(3, F_DUPFD_CLOEXEC, 0)            = 4
> > fcntl(4, F_GETFD)                       = 0
> > write(2, "test-cloexec.c:97: assertion failed\n", 36) = 36
> > 
> > It seems to me from the description in the man page that
> > F_DUPFD_CLOEXEC ought to be setting the FD_CLOEXEC flag on file
> > descriptor 4, so either it's not or else F_GETFD isn't reading the
> > flag for some reason.

Interesting...  Oh, crap.  OK, that's easily fixed: in fs/fcntl.c
                err = f_dupfd(arg, filp, FD_CLOEXEC);
should get s/FD_/O_/.  Linus, could you apply the following?

Fix a braino in F_DUPFD_CLOEXEC; f_dupfd() expects flags for
alloc_fd()/get_unused_fd()/etc. and there clone-on-exec if
O_CLOEXEC, not FD_CLOEXEC.

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
---
diff --git a/fs/fcntl.c b/fs/fcntl.c
index 8f70429..71a600a 100644
--- a/fs/fcntl.c
+++ b/fs/fcntl.c
@@ -258,7 +258,7 @@ static long do_fcntl(int fd, unsigned int cmd, unsigned long arg,
 		err = f_dupfd(arg, filp, 0);
 		break;
 	case F_DUPFD_CLOEXEC:
-		err = f_dupfd(arg, filp, FD_CLOEXEC);
+		err = f_dupfd(arg, filp, O_CLOEXEC);
 		break;
 	case F_GETFD:
 		err = get_close_on_exec(fd) ? FD_CLOEXEC : 0;

Re: [regression] F_DUPFD_CLOEXEC breakage fix

[Richard W.M. Jones]

On Mon, Oct 08, 2012 at 11:21:58PM +0100, Al Viro wrote:
> On Mon, Oct 08, 2012 at 10:53:25PM +0100, Richard W.M. Jones wrote:
> > On Mon, Oct 08, 2012 at 10:50:30PM +0100, Richard W.M. Jones wrote:
> > [.. discussion on gnulib test-cloexec test snipped ..]
> > > I'm suspicious this is a kernel bug:
> > > 
> > > creat("test-cloexec.tmp", 0600)         = 3
> > > fcntl(3, F_GETFD)                       = 0
> > > fcntl(3, F_GETFD)                       = 0
> > > fcntl(3, F_SETFD, FD_CLOEXEC)           = 0
> > > fcntl(3, F_GETFD)                       = 0x1 (flags FD_CLOEXEC)
> > > fcntl(3, F_GETFD)                       = 0x1 (flags FD_CLOEXEC)
> > > fcntl(3, F_SETFD, 0)                    = 0
> > > fcntl(3, F_GETFD)                       = 0
> > > fcntl(3, F_DUPFD_CLOEXEC, 0)            = 4
> > > fcntl(4, F_GETFD)                       = 0
> > > write(2, "test-cloexec.c:97: assertion failed\n", 36) = 36
> > > 
> > > It seems to me from the description in the man page that
> > > F_DUPFD_CLOEXEC ought to be setting the FD_CLOEXEC flag on file
> > > descriptor 4, so either it's not or else F_GETFD isn't reading the
> > > flag for some reason.
> 
> Interesting...  Oh, crap.  OK, that's easily fixed: in fs/fcntl.c
>                 err = f_dupfd(arg, filp, FD_CLOEXEC);
> should get s/FD_/O_/.  Linus, could you apply the following?
> 
> Fix a braino in F_DUPFD_CLOEXEC; f_dupfd() expects flags for
> alloc_fd()/get_unused_fd()/etc. and there clone-on-exec if
> O_CLOEXEC, not FD_CLOEXEC.
> 
> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>

I can confirm that this patch fixed the problem for me.

Tested-by: Richard W.M. Jones <rjones@redhat.com>

> ---
> diff --git a/fs/fcntl.c b/fs/fcntl.c
> index 8f70429..71a600a 100644
> --- a/fs/fcntl.c
> +++ b/fs/fcntl.c
> @@ -258,7 +258,7 @@ static long do_fcntl(int fd, unsigned int cmd, unsigned long arg,
>  		err = f_dupfd(arg, filp, 0);
>  		break;
>  	case F_DUPFD_CLOEXEC:
> -		err = f_dupfd(arg, filp, FD_CLOEXEC);
> +		err = f_dupfd(arg, filp, O_CLOEXEC);
>  		break;
>  	case F_GETFD:
>  		err = get_close_on_exec(fd) ? FD_CLOEXEC : 0;

Rich.

-- 
Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones
virt-df lists disk usage of guests without needing to install any
software inside the virtual machine.  Supports Linux and Windows.
http://et.redhat.com/~rjones/virt-df/