[v2,net-next] af_unix: fix a fatal race with bit fields

On Wed, 2013-05-01 at 03:36 -0400, David Miller wrote:
> From: Benjamin Herrenschmidt <benh@kernel.crashing.org>
> Date: Wed, 01 May 2013 11:39:53 +1000
> 
> > I'm not even completely certain bytes are safe to be honest, though
> > probably more than bitfields. I'll poke our compiler people.
> 
> Older Alpha only has 32-bit and 64-bit loads and stores, so byte sized
> accesses are not atomic, and therefore use racey read-modify-write
> sequences.

Right, so what about the following more general fix ?

Thanks !

[PATCH v2] af_unix: fix a fatal race with bit fields

Using bit fields is dangerous on ppc64/sparc64, as the compiler [1]
uses 64bit instructions to manipulate them.
If the 64bit word includes any atomic_t or spinlock_t, we can lose
critical concurrent changes.

This is happening in af_unix, where unix_sk(sk)->gc_candidate/
gc_maybe_cycle/lock share the same 64bit word.

This leads to fatal deadlock, as one/several cpus spin forever
on a spinlock that will never be available again.

A safer way would be to use a long to store flags.
This way we are sure compiler/arch wont do bad things.

As we own unix_gc_lock spinlock when clearing or setting bits,
we can use the non atomic __set_bit()/__clear_bit().

recursion_level can share the same 64bit location with the spinlock,
as it is set only with this spinlock held.

[1] bug fixed in gcc-4.8.0 :
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52080

Reported-by: Ambrose Feinstein <ambrose@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
---
 include/net/af_unix.h |    5 +++--
 net/unix/garbage.c    |   12 ++++++------
 2 files changed, 9 insertions(+), 8 deletions(-)

> diff --git a/include/net/af_unix.h b/include/net/af_unix.h
> index a8836e8..dbdfd2b 100644
> --- a/include/net/af_unix.h
> +++ b/include/net/af_unix.h
> @@ -57,9 +57,10 @@ struct unix_sock {
>  	struct list_head	link;
>  	atomic_long_t		inflight;
>  	spinlock_t		lock;
> -	unsigned int		gc_candidate : 1;
> -	unsigned int		gc_maybe_cycle : 1;
>  	unsigned char		recursion_level;
> +	unsigned long		gc_flags;
> +#define UNIX_GC_CANDIDATE	0
> +#define UNIX_GC_MAYBE_CYCLE	1
>  	struct socket_wq	peer_wq;
>  };

Why not just change gc_candidate and gc_maybe_cycle to
unsigned char?
It might reduce the number of pad bytes somewhat.

	David

On Wed, 2013-05-01 at 16:53 +0100, David Laight wrote:
> Why not just change gc_candidate and gc_maybe_cycle to
> unsigned char?
> It might reduce the number of pad bytes somewhat.

You didn't quite follow the discussion.

I used bytes on V1, and we are not 100% sure its safe on all arches.

unsigned long is guaranteed to be safe. We absolutely rely on this.

Better use more bytes on a socket (with no impact on real memory use),
than spending hours to debug some strange issues.

From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Wed, 01 May 2013 08:24:03 -0700

> [PATCH v2] af_unix: fix a fatal race with bit fields
> 
> Using bit fields is dangerous on ppc64/sparc64, as the compiler [1]
> uses 64bit instructions to manipulate them.
> If the 64bit word includes any atomic_t or spinlock_t, we can lose
> critical concurrent changes.
> 
> This is happening in af_unix, where unix_sk(sk)->gc_candidate/
> gc_maybe_cycle/lock share the same 64bit word.
> 
> This leads to fatal deadlock, as one/several cpus spin forever
> on a spinlock that will never be available again.
> 
> A safer way would be to use a long to store flags.
> This way we are sure compiler/arch wont do bad things.
> 
> As we own unix_gc_lock spinlock when clearing or setting bits,
> we can use the non atomic __set_bit()/__clear_bit().
> 
> recursion_level can share the same 64bit location with the spinlock,
> as it is set only with this spinlock held.
> 
> [1] bug fixed in gcc-4.8.0 :
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52080
> 
> Reported-by: Ambrose Feinstein <ambrose@google.com>
> Signed-off-by: Eric Dumazet <edumazet@google.com>

Applied and queued up for -stable, thanks Eric.