Hello,
Looking at (2.4.9's) sys_pivot_root() implementation I can see that it
moves all the processes' root/pwd references from the old root to the new
one.
However, somehow /sbin/init managed to survive the move and is keeping the
filesystem busy:
# lsof /sysroot
COMMAND PID USER FD TYPE DEVICE SIZE NODE NAME
init 1 0 txt REG 22,65 28220 11217 /sysroot/sbin/init
init 1 0 mem REG 22,65 468849 10376
/sysroot/lib/ld-2.2.2.so
init 1 0 mem REG 22,65 5636080 10370
/sysroot/lib/i686/libc-2.2.2.so
# cat /proc/1/maps
08048000-0804f000 r-xp 00000000 16:41 11217 /sysroot/sbin/init
0804f000-08050000 rw-p 00006000 16:41 11217 /sysroot/sbin/init
08050000-08054000 rwxp 00000000 00:00 0
15556000-1556c000 r-xp 00000000 16:41 10376 /sysroot/lib/ld-2.2.2.so
1556c000-1556d000 rw-p 00015000 16:41 10376 /sysroot/lib/ld-2.2.2.so
1556d000-1556e000 rw-p 00000000 00:00 0
15573000-15699000 r-xp 00000000 16:41 10370 /sysroot/lib/i686/libc-2.2.2.so
15699000-1569f000 rw-p 00125000 16:41 10370 /sysroot/lib/i686/libc-2.2.2.so
1569f000-156a3000 rw-p 00000000 00:00 0
3fffe000-40000000 rwxp fffff000 00:00 0
Any clues why and how to force it to exec a binary on the new root (or
get rid of it in any other way)?
Regards,
Tigran
Amazing that nobody knew the answer but I think I understand now the
reason, the /proc/1/cwd and /proc/1/root point to "/" and the comment
above sys_pivot_root says:
* Note:
* - we don't move root/cwd if they are not at the root (reason: if
something
* cared enough to change them, it's probably wrong to force them
elsewhere)
So, that is why the /sbin/init wasn't moved, because it's root/cwd were at
/initrd so when it got pivoted to / (and / put to /sysroot) the init was
left alone. Well, either the above policy is wrong or I need to figure out
a way to convince init to move peacebly where I want it to... :)
On Wed, 6 Mar 2002, Tigran Aivazian wrote:
> Hello,
>
> Looking at (2.4.9's) sys_pivot_root() implementation I can see that it
> moves all the processes' root/pwd references from the old root to the new
> one.
>
> However, somehow /sbin/init managed to survive the move and is keeping the
> filesystem busy:
>
> # lsof /sysroot
> COMMAND PID USER FD TYPE DEVICE SIZE NODE NAME
> init 1 0 txt REG 22,65 28220 11217 /sysroot/sbin/init
> init 1 0 mem REG 22,65 468849 10376
> /sysroot/lib/ld-2.2.2.so
> init 1 0 mem REG 22,65 5636080 10370
> /sysroot/lib/i686/libc-2.2.2.so
>
> # cat /proc/1/maps
> 08048000-0804f000 r-xp 00000000 16:41 11217 /sysroot/sbin/init
> 0804f000-08050000 rw-p 00006000 16:41 11217 /sysroot/sbin/init
> 08050000-08054000 rwxp 00000000 00:00 0
> 15556000-1556c000 r-xp 00000000 16:41 10376 /sysroot/lib/ld-2.2.2.so
> 1556c000-1556d000 rw-p 00015000 16:41 10376 /sysroot/lib/ld-2.2.2.so
> 1556d000-1556e000 rw-p 00000000 00:00 0
> 15573000-15699000 r-xp 00000000 16:41 10370 /sysroot/lib/i686/libc-2.2.2.so
> 15699000-1569f000 rw-p 00125000 16:41 10370 /sysroot/lib/i686/libc-2.2.2.so
> 1569f000-156a3000 rw-p 00000000 00:00 0
> 3fffe000-40000000 rwxp fffff000 00:00 0
>
> Any clues why and how to force it to exec a binary on the new root (or
> get rid of it in any other way)?
>
> Regards,
> Tigran
>
>
>
>
On Thu, 7 Mar 2002, Tigran Aivazian wrote:
> Amazing that nobody knew the answer but I think I understand now the
> reason, the /proc/1/cwd and /proc/1/root point to "/" and the comment
> above sys_pivot_root says:
>
> * Note:
> * - we don't move root/cwd if they are not at the root (reason: if
> something
> * cared enough to change them, it's probably wrong to force them
> elsewhere)
>
> So, that is why the /sbin/init wasn't moved, because it's root/cwd were at
~~~~~~
I often say "yes" instead of "no" (and other way around) :)
It obviously _was_ moved and the problem is that it's executable's mapping
and the libc/ld mappings are keeping /sysroot busy...
> /initrd so when it got pivoted to / (and / put to /sysroot) the init was
> left alone. Well, either the above policy is wrong or I need to figure out
> a way to convince init to move peacebly where I want it to... :)
>
> On Wed, 6 Mar 2002, Tigran Aivazian wrote:
>
> > Hello,
> >
> > Looking at (2.4.9's) sys_pivot_root() implementation I can see that it
> > moves all the processes' root/pwd references from the old root to the new
> > one.
> >
> > However, somehow /sbin/init managed to survive the move and is keeping the
> > filesystem busy:
> >
> > # lsof /sysroot
> > COMMAND PID USER FD TYPE DEVICE SIZE NODE NAME
> > init 1 0 txt REG 22,65 28220 11217 /sysroot/sbin/init
> > init 1 0 mem REG 22,65 468849 10376
> > /sysroot/lib/ld-2.2.2.so
> > init 1 0 mem REG 22,65 5636080 10370
> > /sysroot/lib/i686/libc-2.2.2.so
> >
> > # cat /proc/1/maps
> > 08048000-0804f000 r-xp 00000000 16:41 11217 /sysroot/sbin/init
> > 0804f000-08050000 rw-p 00006000 16:41 11217 /sysroot/sbin/init
> > 08050000-08054000 rwxp 00000000 00:00 0
> > 15556000-1556c000 r-xp 00000000 16:41 10376 /sysroot/lib/ld-2.2.2.so
> > 1556c000-1556d000 rw-p 00015000 16:41 10376 /sysroot/lib/ld-2.2.2.so
> > 1556d000-1556e000 rw-p 00000000 00:00 0
> > 15573000-15699000 r-xp 00000000 16:41 10370 /sysroot/lib/i686/libc-2.2.2.so
> > 15699000-1569f000 rw-p 00125000 16:41 10370 /sysroot/lib/i686/libc-2.2.2.so
> > 1569f000-156a3000 rw-p 00000000 00:00 0
> > 3fffe000-40000000 rwxp fffff000 00:00 0
> >
> > Any clues why and how to force it to exec a binary on the new root (or
> > get rid of it in any other way)?
> >
> > Regards,
> > Tigran
> >
> >
> >
> >
>
>
just for the record (and for the curious) -- the solution to this problem
is to re-exec /sbin/init. I wonder if this sort of scenario is what Al
Viro had in mind when he added the re-exec code to init in 1998...
On Thu, 7 Mar 2002, Tigran Aivazian wrote:
> On Thu, 7 Mar 2002, Tigran Aivazian wrote:
>
> > Amazing that nobody knew the answer but I think I understand now the
> > reason, the /proc/1/cwd and /proc/1/root point to "/" and the comment
> > above sys_pivot_root says:
> >
> > * Note:
> > * - we don't move root/cwd if they are not at the root (reason: if
> > something
> > * cared enough to change them, it's probably wrong to force them
> > elsewhere)
> >
> > So, that is why the /sbin/init wasn't moved, because it's root/cwd were at
> ~~~~~~
>
> I often say "yes" instead of "no" (and other way around) :)
>
> It obviously _was_ moved and the problem is that it's executable's mapping
> and the libc/ld mappings are keeping /sysroot busy...
>
> > /initrd so when it got pivoted to / (and / put to /sysroot) the init was
> > left alone. Well, either the above policy is wrong or I need to figure out
> > a way to convince init to move peacebly where I want it to... :)
> >
> > On Wed, 6 Mar 2002, Tigran Aivazian wrote:
> >
> > > Hello,
> > >
> > > Looking at (2.4.9's) sys_pivot_root() implementation I can see that it
> > > moves all the processes' root/pwd references from the old root to the new
> > > one.
> > >
> > > However, somehow /sbin/init managed to survive the move and is keeping the
> > > filesystem busy:
> > >
> > > # lsof /sysroot
> > > COMMAND PID USER FD TYPE DEVICE SIZE NODE NAME
> > > init 1 0 txt REG 22,65 28220 11217 /sysroot/sbin/init
> > > init 1 0 mem REG 22,65 468849 10376
> > > /sysroot/lib/ld-2.2.2.so
> > > init 1 0 mem REG 22,65 5636080 10370
> > > /sysroot/lib/i686/libc-2.2.2.so
> > >
> > > # cat /proc/1/maps
> > > 08048000-0804f000 r-xp 00000000 16:41 11217 /sysroot/sbin/init
> > > 0804f000-08050000 rw-p 00006000 16:41 11217 /sysroot/sbin/init
> > > 08050000-08054000 rwxp 00000000 00:00 0
> > > 15556000-1556c000 r-xp 00000000 16:41 10376 /sysroot/lib/ld-2.2.2.so
> > > 1556c000-1556d000 rw-p 00015000 16:41 10376 /sysroot/lib/ld-2.2.2.so
> > > 1556d000-1556e000 rw-p 00000000 00:00 0
> > > 15573000-15699000 r-xp 00000000 16:41 10370 /sysroot/lib/i686/libc-2.2.2.so
> > > 15699000-1569f000 rw-p 00125000 16:41 10370 /sysroot/lib/i686/libc-2.2.2.so
> > > 1569f000-156a3000 rw-p 00000000 00:00 0
> > > 3fffe000-40000000 rwxp fffff000 00:00 0
> > >
> > > Any clues why and how to force it to exec a binary on the new root (or
> > > get rid of it in any other way)?
> > >
> > > Regards,
> > > Tigran
> > >
> > >
> > >
> > >
> >
> >
>
> -
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
>
even more curiously, swapoff is the _final_ answer.
who on Earth would have imagined that swapping-on increments the mnt_count
of the root filesystem? Very very curious.
So, initially, the mnt_count was too high but most references were dropped
by re-execing /sbin/init (I did have to hack /sbin/init to allow
re-execing in runlevel 0 though).
Then, I still had a problem because the refcount was surprizingly 2
instead of 1. Now I discovered that the extra reference was due to
swapping.
Regards,
Tigran
On Thu, 7 Mar 2002, Tigran Aivazian wrote:
> just for the record (and for the curious) -- the solution to this problem
> is to re-exec /sbin/init. I wonder if this sort of scenario is what Al
> Viro had in mind when he added the re-exec code to init in 1998...
>
> On Thu, 7 Mar 2002, Tigran Aivazian wrote:
>
> > On Thu, 7 Mar 2002, Tigran Aivazian wrote:
> >
> > > Amazing that nobody knew the answer but I think I understand now the
> > > reason, the /proc/1/cwd and /proc/1/root point to "/" and the comment
> > > above sys_pivot_root says:
> > >
> > > * Note:
> > > * - we don't move root/cwd if they are not at the root (reason: if
> > > something
> > > * cared enough to change them, it's probably wrong to force them
> > > elsewhere)
> > >
> > > So, that is why the /sbin/init wasn't moved, because it's root/cwd were at
> > ~~~~~~
> >
> > I often say "yes" instead of "no" (and other way around) :)
> >
> > It obviously _was_ moved and the problem is that it's executable's mapping
> > and the libc/ld mappings are keeping /sysroot busy...
> >
> > > /initrd so when it got pivoted to / (and / put to /sysroot) the init was
> > > left alone. Well, either the above policy is wrong or I need to figure out
> > > a way to convince init to move peacebly where I want it to... :)
> > >
> > > On Wed, 6 Mar 2002, Tigran Aivazian wrote:
> > >
> > > > Hello,
> > > >
> > > > Looking at (2.4.9's) sys_pivot_root() implementation I can see that it
> > > > moves all the processes' root/pwd references from the old root to the new
> > > > one.
> > > >
> > > > However, somehow /sbin/init managed to survive the move and is keeping the
> > > > filesystem busy:
> > > >
> > > > # lsof /sysroot
> > > > COMMAND PID USER FD TYPE DEVICE SIZE NODE NAME
> > > > init 1 0 txt REG 22,65 28220 11217 /sysroot/sbin/init
> > > > init 1 0 mem REG 22,65 468849 10376
> > > > /sysroot/lib/ld-2.2.2.so
> > > > init 1 0 mem REG 22,65 5636080 10370
> > > > /sysroot/lib/i686/libc-2.2.2.so
> > > >
> > > > # cat /proc/1/maps
> > > > 08048000-0804f000 r-xp 00000000 16:41 11217 /sysroot/sbin/init
> > > > 0804f000-08050000 rw-p 00006000 16:41 11217 /sysroot/sbin/init
> > > > 08050000-08054000 rwxp 00000000 00:00 0
> > > > 15556000-1556c000 r-xp 00000000 16:41 10376 /sysroot/lib/ld-2.2.2.so
> > > > 1556c000-1556d000 rw-p 00015000 16:41 10376 /sysroot/lib/ld-2.2.2.so
> > > > 1556d000-1556e000 rw-p 00000000 00:00 0
> > > > 15573000-15699000 r-xp 00000000 16:41 10370 /sysroot/lib/i686/libc-2.2.2.so
> > > > 15699000-1569f000 rw-p 00125000 16:41 10370 /sysroot/lib/i686/libc-2.2.2.so
> > > > 1569f000-156a3000 rw-p 00000000 00:00 0
> > > > 3fffe000-40000000 rwxp fffff000 00:00 0
> > > >
> > > > Any clues why and how to force it to exec a binary on the new root (or
> > > > get rid of it in any other way)?
> > > >
> > > > Regards,
> > > > Tigran
> > > >
> > > >
> > > >
> > > >
> > >
> > >
> >
> > -
> > To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> > the body of a message to [email protected]
> > More majordomo info at http://vger.kernel.org/majordomo-info.html
> > Please read the FAQ at http://www.tux.org/lkml/
> >
>
>