[Resend. I messed up the To: line the first time]
As has been reported recently by Lennert Buytenhek, robust futexes
are broken on ARM:
>If you're also running into glibc's tst-robust1 test suite test
>locking up your ARM machine, you're probably running into the fact
>that asm-arm/futex.h includes asm-generic/futex.h, and
>asm-generic/futex.h defines futex_atomic_cmpxchg_inatomic() to
>return -ENOSYS. This causes handle_futex_death() to loop forever.
I can confirm this statement: building glibc-2.4 with NPTL on
ARM hangs the kernel when the test suite reaches tst-robust1.
The problem is that kernel/futex.c expects futex_atomic_cmpxchg_inatomic()
to return -EFAULT or the new value. It doesn't expect -ENOSYS at all, and
generally -ENOSYS causes the futex code to loop, hanging the kernel.
The higher-end archs (x86, sparc64, ppc64, etc) provide fully-functional
asm/futex.h implementations, but a number of archs (alpha, arm, arm26,
avr32, blackfin, cris, h8300, m32r, m68k, mk68knommu, sh64, sparc, um,
v850, and xtensa) use asm-generic/futex.h, which makes robust futexes
horribly broken on them. There have also been reports recently that PI
futexes are broken due to the generic futex_atomic_cmpxchg_inatomic()
just being an -ENOSYS stub.
The patch below implements the generic futex_atomic_cmpxchg_inatomic() in
terms of __copy_{from,to}_user_inatomic() and preempt_{disable,enable}().
It obviously doesn't support SMP, but UP-only support should go a long
way for users of the affected archs.
I'm using this patch now and it has allowed me to build and use glibc-2.4
with NPTL on ARM (glibc-2.4-11.src.rpm from FC5 + ARM fixes).
(Finally I can ditch LinuxThreads :->)
Comments?
/Mikael
--- linux-2.6.22/include/asm-generic/futex.h.~1~ 2007-02-04 19:44:54.000000000 +0100
+++ linux-2.6.22/include/asm-generic/futex.h 2007-08-01 19:03:43.000000000 +0200
@@ -4,6 +4,7 @@
#ifdef __KERNEL__
#include <linux/futex.h>
+#include <linux/preempt.h>
#include <asm/errno.h>
#include <asm/uaccess.h>
@@ -52,7 +53,34 @@ futex_atomic_op_inuser (int encoded_op,
static inline int
futex_atomic_cmpxchg_inatomic(int __user *uaddr, int oldval, int newval)
{
+#ifdef CONFIG_SMP
return -ENOSYS;
+#else
+ int curval, ret;
+
+ if (!access_ok(VERIFY_WRITE, uaddr, sizeof(int)))
+ return -EFAULT;
+
+ preempt_disable();
+
+ ret = -EFAULT;
+ if (__copy_from_user_inatomic(&curval, uaddr, sizeof(int)))
+ goto out;
+
+ ret = curval;
+ if (curval != oldval)
+ goto out;
+
+ ret = -EFAULT;
+ if (__copy_to_user_inatomic(uaddr, &newval, sizeof(int)))
+ goto out;
+
+ ret = newval;
+
+ out:
+ preempt_enable();
+ return ret;
+#endif
}
#endif
On Thu, Aug 02, 2007 at 01:00:21AM +0200, Mikael Pettersson wrote:
> @@ -52,7 +53,34 @@ futex_atomic_op_inuser (int encoded_op,
> static inline int
> futex_atomic_cmpxchg_inatomic(int __user *uaddr, int oldval, int newval)
> {
> +#ifdef CONFIG_SMP
> return -ENOSYS;
> +#else
Since the callers of futex_atomic_cmpxchg_inatomic() don't really
seem prepared to deal with -ENOSYS (e.g. the handle_futex_death()
infinite loop when it gets -ENOSYS), it seems better never to
return -ENOSYS from this function at all.
What if you just stick an #error in here in the SMP case?
On Thu, 2 Aug 2007 01:49:02 +0200, Lennert Buytenhek wrote:
> On Thu, Aug 02, 2007 at 01:00:21AM +0200, Mikael Pettersson wrote:
>
> > @@ -52,7 +53,34 @@ futex_atomic_op_inuser (int encoded_op,
> > static inline int
> > futex_atomic_cmpxchg_inatomic(int __user *uaddr, int oldval, int newval)
> > {
> > +#ifdef CONFIG_SMP
> > return -ENOSYS;
> > +#else
>
> Since the callers of futex_atomic_cmpxchg_inatomic() don't really
> seem prepared to deal with -ENOSYS (e.g. the handle_futex_death()
> infinite loop when it gets -ENOSYS), it seems better never to
> return -ENOSYS from this function at all.
>
> What if you just stick an #error in here in the SMP case?
The problem with #error is that it would cause compile-time
regressions. I assume that e.g. alpha supports building SMP
kernels, but #error would prevent that.
Thus I opted to fix the UP case while leaving the SMP case
unchanged. Actually I think the SMP case should be a BUG()
rather than -ENOSYS, but that's a different issue from the
UP case which I really do want to see fixed.
/Mikael
On Thu, Aug 02, 2007 at 02:06:27AM +0200, Mikael Pettersson wrote:
> > > @@ -52,7 +53,34 @@ futex_atomic_op_inuser (int encoded_op,
> > > static inline int
> > > futex_atomic_cmpxchg_inatomic(int __user *uaddr, int oldval, int newval)
> > > {
> > > +#ifdef CONFIG_SMP
> > > return -ENOSYS;
> > > +#else
> >
> > Since the callers of futex_atomic_cmpxchg_inatomic() don't really
> > seem prepared to deal with -ENOSYS (e.g. the handle_futex_death()
> > infinite loop when it gets -ENOSYS), it seems better never to
> > return -ENOSYS from this function at all.
> >
> > What if you just stick an #error in here in the SMP case?
>
> The problem with #error is that it would cause compile-time
> regressions. I assume that e.g. alpha supports building SMP
> kernels, but #error would prevent that.
>
> Thus I opted to fix the UP case while leaving the SMP case
> unchanged. Actually I think the SMP case should be a BUG()
> rather than -ENOSYS,
Probably. Or handle -ENOSYS in the callers -- but that's more
work, and would cease to be necessary once everyone implements
futex_atomic_cmpxchg_inatomic().
> but that's a different issue from the UP case which I really do
> want to see fixed.
ACK.
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
What's the state of this patch? I can confirm tst-robust1
from glibc testsuite locks a armv5 machine hard. With this patch
applied, the test succeeds.
> The higher-end archs (x86, sparc64, ppc64, etc) provide fully-functional
> asm/futex.h implementations, but a number of archs (alpha, arm, arm26,
> avr32, blackfin, cris, h8300, m32r, m68k, mk68knommu, sh64, sparc, um,
> v850, and xtensa) use asm-generic/futex.h, which makes robust futexes
> horribly broken on them. There have also been reports recently that PI
> futexes are broken due to the generic futex_atomic_cmpxchg_inatomic()
> just being an -ENOSYS stub.
This an effective local DOS bug on the affected architectures, too..
> The patch below implements the generic futex_atomic_cmpxchg_inatomic() in
> terms of __copy_{from,to}_user_inatomic() and preempt_{disable,enable}().
> It obviously doesn't support SMP, but UP-only support should go a long
> way for users of the affected archs.
>
> I'm using this patch now and it has allowed me to build and use glibc-2.4
> with NPTL on ARM (glibc-2.4-11.src.rpm from FC5 + ARM fixes).
> (Finally I can ditch LinuxThreads :->)
>
> Comments?
>
> /Mikael
>
> --- linux-2.6.22/include/asm-generic/futex.h.~1~ 2007-02-04 19:44:54.000000000 +0100
> +++ linux-2.6.22/include/asm-generic/futex.h 2007-08-01 19:03:43.000000000 +0200
> @@ -4,6 +4,7 @@
> #ifdef __KERNEL__
>
> #include <linux/futex.h>
> +#include <linux/preempt.h>
> #include <asm/errno.h>
> #include <asm/uaccess.h>
>
> @@ -52,7 +53,34 @@ futex_atomic_op_inuser (int encoded_op,
> static inline int
> futex_atomic_cmpxchg_inatomic(int __user *uaddr, int oldval, int newval)
> {
> +#ifdef CONFIG_SMP
> return -ENOSYS;
> +#else
> + int curval, ret;
> +
> + if (!access_ok(VERIFY_WRITE, uaddr, sizeof(int)))
> + return -EFAULT;
> +
> + preempt_disable();
> +
> + ret = -EFAULT;
> + if (__copy_from_user_inatomic(&curval, uaddr, sizeof(int)))
> + goto out;
> +
> + ret = curval;
> + if (curval != oldval)
> + goto out;
> +
> + ret = -EFAULT;
> + if (__copy_to_user_inatomic(uaddr, &newval, sizeof(int)))
> + goto out;
> +
> + ret = newval;
> +
> + out:
> + preempt_enable();
> + return ret;
> +#endif
> }
>
> #endif
> -
- --
"rm -rf" only sounds scary if you don't have backups
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.6 (GNU/Linux)
iD8DBQFHA+6aibPvMsrqrwMRAlMuAKCBf7qpD2dETgU+RgnDG4ArVvFp3gCgrewq
4KvOt40U1MAPM7g4F/Ps5jk=
=68gr
-----END PGP SIGNATURE-----
On Wed, 3 Oct 2007 22:33:46 +0300, Riku Voipio wrote:
> What's the state of this patch? I can confirm tst-robust1
> from glibc testsuite locks a armv5 machine hard. With this patch
> applied, the test succeeds.
There were no comments from any Linux arch or futex maintainer.
Because of that I intend to submit an ARM-only patch when 2.6.23
has been released.
/Mikael
> > The higher-end archs (x86, sparc64, ppc64, etc) provide fully-functional
> > asm/futex.h implementations, but a number of archs (alpha, arm, arm26,
> > avr32, blackfin, cris, h8300, m32r, m68k, mk68knommu, sh64, sparc, um,
> > v850, and xtensa) use asm-generic/futex.h, which makes robust futexes
> > horribly broken on them. There have also been reports recently that PI
> > futexes are broken due to the generic futex_atomic_cmpxchg_inatomic()
> > just being an -ENOSYS stub.
>
> This an effective local DOS bug on the affected architectures, too..
>
> > The patch below implements the generic futex_atomic_cmpxchg_inatomic() in
> > terms of __copy_{from,to}_user_inatomic() and preempt_{disable,enable}().
> > It obviously doesn't support SMP, but UP-only support should go a long
> > way for users of the affected archs.
> >
> > I'm using this patch now and it has allowed me to build and use glibc-2.4
> > with NPTL on ARM (glibc-2.4-11.src.rpm from FC5 + ARM fixes).
> > (Finally I can ditch LinuxThreads :->)
> >
> > Comments?
> >
> > /Mikael
> >
> > --- linux-2.6.22/include/asm-generic/futex.h.~1~ 2007-02-04 19:44:54.000000000 +0100
> > +++ linux-2.6.22/include/asm-generic/futex.h 2007-08-01 19:03:43.000000000 +0200
> > @@ -4,6 +4,7 @@
> > #ifdef __KERNEL__
> >
> > #include <linux/futex.h>
> > +#include <linux/preempt.h>
> > #include <asm/errno.h>
> > #include <asm/uaccess.h>
> >
> > @@ -52,7 +53,34 @@ futex_atomic_op_inuser (int encoded_op,
> > static inline int
> > futex_atomic_cmpxchg_inatomic(int __user *uaddr, int oldval, int newval)
> > {
> > +#ifdef CONFIG_SMP
> > return -ENOSYS;
> > +#else
> > + int curval, ret;
> > +
> > + if (!access_ok(VERIFY_WRITE, uaddr, sizeof(int)))
> > + return -EFAULT;
> > +
> > + preempt_disable();
> > +
> > + ret = -EFAULT;
> > + if (__copy_from_user_inatomic(&curval, uaddr, sizeof(int)))
> > + goto out;
> > +
> > + ret = curval;
> > + if (curval != oldval)
> > + goto out;
> > +
> > + ret = -EFAULT;
> > + if (__copy_to_user_inatomic(uaddr, &newval, sizeof(int)))
> > + goto out;
> > +
> > + ret = newval;
> > +
> > + out:
> > + preempt_enable();
> > + return ret;
> > +#endif
> > }
> >
> > #endif
> > -
>
> - --
> "rm -rf" only sounds scary if you don't have backups
> -----BEGIN PGP SIGNATURE-----
> Version: GnuPG v1.4.6 (GNU/Linux)
>
> iD8DBQFHA+6aibPvMsrqrwMRAlMuAKCBf7qpD2dETgU+RgnDG4ArVvFp3gCgrewq
> 4KvOt40U1MAPM7g4F/Ps5jk=
> =68gr
> -----END PGP SIGNATURE-----
>
On Thu, 4 Oct 2007, Mikael Pettersson wrote:
> On Wed, 3 Oct 2007 22:33:46 +0300, Riku Voipio wrote:
> > What's the state of this patch? I can confirm tst-robust1
> > from glibc testsuite locks a armv5 machine hard. With this patch
> > applied, the test succeeds.
>
> There were no comments from any Linux arch or futex maintainer.
Probably because it wasn't posted to linux-arch?
> Because of that I intend to submit an ARM-only patch when 2.6.23
> has been released.
>
> /Mikael
>
> > > The higher-end archs (x86, sparc64, ppc64, etc) provide fully-functional
> > > asm/futex.h implementations, but a number of archs (alpha, arm, arm26,
> > > avr32, blackfin, cris, h8300, m32r, m68k, mk68knommu, sh64, sparc, um,
> > > v850, and xtensa) use asm-generic/futex.h, which makes robust futexes
> > > horribly broken on them. There have also been reports recently that PI
> > > futexes are broken due to the generic futex_atomic_cmpxchg_inatomic()
> > > just being an -ENOSYS stub.
> >
> > This an effective local DOS bug on the affected architectures, too..
> >
> > > The patch below implements the generic futex_atomic_cmpxchg_inatomic() in
> > > terms of __copy_{from,to}_user_inatomic() and preempt_{disable,enable}().
> > > It obviously doesn't support SMP, but UP-only support should go a long
> > > way for users of the affected archs.
> > >
> > > I'm using this patch now and it has allowed me to build and use glibc-2.4
> > > with NPTL on ARM (glibc-2.4-11.src.rpm from FC5 + ARM fixes).
> > > (Finally I can ditch LinuxThreads :->)
> > >
> > > Comments?
> > >
> > > /Mikael
> > >
> > > --- linux-2.6.22/include/asm-generic/futex.h.~1~ 2007-02-04 19:44:54.000000000 +0100
> > > +++ linux-2.6.22/include/asm-generic/futex.h 2007-08-01 19:03:43.000000000 +0200
> > > @@ -4,6 +4,7 @@
> > > #ifdef __KERNEL__
> > >
> > > #include <linux/futex.h>
> > > +#include <linux/preempt.h>
> > > #include <asm/errno.h>
> > > #include <asm/uaccess.h>
> > >
> > > @@ -52,7 +53,34 @@ futex_atomic_op_inuser (int encoded_op,
> > > static inline int
> > > futex_atomic_cmpxchg_inatomic(int __user *uaddr, int oldval, int newval)
> > > {
> > > +#ifdef CONFIG_SMP
> > > return -ENOSYS;
> > > +#else
> > > + int curval, ret;
> > > +
> > > + if (!access_ok(VERIFY_WRITE, uaddr, sizeof(int)))
> > > + return -EFAULT;
> > > +
> > > + preempt_disable();
> > > +
> > > + ret = -EFAULT;
> > > + if (__copy_from_user_inatomic(&curval, uaddr, sizeof(int)))
> > > + goto out;
> > > +
> > > + ret = curval;
> > > + if (curval != oldval)
> > > + goto out;
> > > +
> > > + ret = -EFAULT;
> > > + if (__copy_to_user_inatomic(uaddr, &newval, sizeof(int)))
> > > + goto out;
> > > +
> > > + ret = newval;
> > > +
> > > + out:
> > > + preempt_enable();
> > > + return ret;
> > > +#endif
> > > }
> > >
> > > #endif
> > > -
> >
> > - --
> > "rm -rf" only sounds scary if you don't have backups
> > -----BEGIN PGP SIGNATURE-----
> > Version: GnuPG v1.4.6 (GNU/Linux)
> >
> > iD8DBQFHA+6aibPvMsrqrwMRAlMuAKCBf7qpD2dETgU+RgnDG4ArVvFp3gCgrewq
> > 4KvOt40U1MAPM7g4F/Ps5jk=
> > =68gr
> > -----END PGP SIGNATURE-----
> >
> -
> 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/
>
Gr{oetje,eeting}s,
Geert
--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected]
In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds