Re: [PATCH v2 1/3] set_dumpable: fix the theoretical race with itself

[Andrew Morton <akpm@linux-foundation.org>]

On Tue, 19 Nov 2013 15:43:15 +0100 Oleg Nesterov <oleg@redhat.com> wrote:

> set_dumpable() updates MMF_DUMPABLE_MASK in a non-trivial way to
> ensure that get_dumpable() can't observe the intermediate state,
> but this all can't help if multiple threads call set_dumpable()
> at the same time.
> 
> And in theory commit_creds()->set_dumpable(SUID_DUMP_ROOT) racing
> with sys_prctl()->set_dumpable(SUID_DUMP_DISABLE) can result in
> SUID_DUMP_USER.
> 
> Change this code to update both bits atomically via cmpxchg().
> 
> Note: this assumes that it is safe to mix bitops and cmpxchg. IOW,
> if, say, an architecture implements cmpxchg() using the locking
> (like arch/parisc/lib/bitops.c does), then it should use the same
> locks for set_bit/etc.

I suppose that is reasonable - atomic operations on memory should be
atomic wrt other classes of atomic operations on the same memory.  So
set_bit() has to "know" about concurrent cmpxchg().  It *has* to be
this way to fully emulate the hardware-based RMW operations.

parisc got that right.  I wonder if all other architectures did, and
how on earth we can communicate this to future arch developers..