On Thu, Apr 27, 2006 at 04:54:09AM +0200, Michael Kerrisk wrote:
> Justin,
>
> > Tags: upstream
>
> Any time you open a report like this you better make sure
> that whatever you say holds against the latest upstream release.
Do you mean of manpages?
> > It seems that the clone manpage is out of sync with the libc/kernel
> > reality.
> >
> > In particular, the following succeeds,
>
> The preceding is nearly meaningless, because so vague. Succeeds
> where?
I write a program that called clone() and it didn't return -1 when the
manpage suggested that it should have. If glibc passed it on to the
kernel (which it apparently did, since it succeeded), then the kernel
didn't return -1; otherwise glibc did.
> What kernel?
Linux andromeda 2.6.15-1-686 #2 Mon Mar 6 15:27:08 UTC 2006 i686 GNU/Linux
> > but the annotation says that it
> > should not:
> >
> > CLONE_NEWNS|CLONE_FS
This still seems to succeed when I do a syscall (without libc).
> > CLONE_THREAD (because CLONE_SIGHAND is not set)
> > CLONE_SIGHAND (because CLONE_VM is not set)
But these now return -1 (when using the "syscall" mechanism; see
below).
> And how did you test or verify this? If it was with a program,
> show me the complete clone() call.
Originally, I was doing it the "libc" way:
#include <sched.h>
// #include <syscall.h>
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
int f(void *v)
{
assert(v==NULL);
fprintf(stderr, "%s\n", __PRETTY_FUNCTION__);
return 0;
}
int main()
{
char buf[1<<10];
const int flags=CLONE_NEWNS|CLONE_FS;
fprintf(stderr, "%s\n", __PRETTY_FUNCTION__);
if (-1==clone(f, buf, flags, NULL)) {
perror("clone");
}
exit(EXIT_SUCCESS);
}
But I finally got the "syscall" way to work, with the above updated
results:
#include <syscall.h>
#include <asm/ldt.h>
// bits/sched.h or linux/sched.h
#include <errno.h>
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
int f(void *v)
{
assert(v==NULL);
fprintf(stderr, "%s\n", __PRETTY_FUNCTION__);
return 0;
}
_syscall5(int, clone, int, flags, void *, child_stack, int *, parent_tidptr,
struct user_desc *, newtls, int *, child_tidptr);
int main()
{
const int flags=0x0001ffff|0x00000200; /* CLONE_NEWNS|CLONE_FS; */
// 0x00000800; /* CLONE_SIGHAND */
// 0x00010000; /* CLONE_THREAD */
if (-1==clone(flags,0,0,0,0)) {
perror("clone");
}
fprintf(stderr, "%s\n", __PRETTY_FUNCTION__);
exit(EXIT_SUCCESS);
}
> > There is also mention of a "CLONE_DETACHED" flag which is not
> > otherwise mentioned; (the header files indicate that it is ignored).
>
> Mention where? Please be precise.
I'm terribly sorry but, although I had updated to manpages 2.25-3, the
manpage was already opened with 2.22 or so; CLONE_DETACHED is no
longer mentioned.
> > Also, I can't follow the description of sys_clone, mostly because the
> > prototype doesn't make sense. Instead of being a prototype of clone,
> > it looks like it is just copied from the definition (somewhere).
> > Doesn't it make more sense to include something like:
> >
> > clone(int flags, void *child_stack, int *parent_tidptr,
> > struct user_desc *newtls, int *child_tidptr);
>
> No, because it is not prototyped in header files.
> Thus either _syscall5() or syscall() is required.
Okay.
> > Instead of the stuff that some scary-looking macro in a burried header
> > file that generates the definition? :)
>
> The people who find it scary probably don't need clone().
>
> Justin -- vague bug reports like this waste my time.
Not my intent, of course.
> (I may be wrong, but this bug report sounds "speculative" -- that
> is, you don't really have much experience with clone().)
This is true, I do not.
Justin
--
To UNSUBSCRIBE, email to [EMAIL PROTECTED]
with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]
