linux-kernel - Re: [PATCH v3] skbuff: fix a data race in skb_queue

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Thu, 6 Feb 2020 20:29:06 +0100
From:   Marco Elver <elver@...gle.com>
To:     "Jason A. Donenfeld" <Jason@...c4.com>
Cc:     Eric Dumazet <eric.dumazet@...il.com>, Qian Cai <cai@....pw>,
        Netdev <netdev@...r.kernel.org>,
        LKML <linux-kernel@...r.kernel.org>,
        Dmitry Vyukov <dvyukov@...gle.com>
Subject: Re: [PATCH v3] skbuff: fix a data race in skb_queue_len()

On Thu, 6 Feb 2020 at 19:43, Jason A. Donenfeld <Jason@...c4.com> wrote:
>
> On Thu, Feb 06, 2020 at 10:22:02AM -0800, Eric Dumazet wrote:
> > On 2/6/20 10:12 AM, Jason A. Donenfeld wrote:
> > > On Thu, Feb 6, 2020 at 6:10 PM Eric Dumazet <eric.dumazet@...il.com> wrote:
> > >> Unfortunately we do not have ADD_ONCE() or something like that.
> > >
> > > I guess normally this is called "atomic_add", unless you're thinking
> > > instead about something like this, which generates the same
> > > inefficient code as WRITE_ONCE:
> > >
> > > #define ADD_ONCE(d, s) *(volatile typeof(d) *)&(d) += (s)
> > >
> >
> > Dmitry Vyukov had a nice suggestion few months back how to implement this.
> >
> > https://lkml.org/lkml/2019/10/5/6
>
> That trick appears to work well in clang but not gcc:
>
> #define ADD_ONCE(d, i) ({ \
>        typeof(d) *__p = &(d); \
>        __atomic_store_n(__p, (i) + __atomic_load_n(__p, __ATOMIC_RELAXED), __ATOMIC_RELAXED); \
> })
>
> gcc 9.2 gives:
>
>   0:   8b 47 10                mov    0x10(%rdi),%eax
>   3:   83 e8 01                sub    $0x1,%eax
>   6:   89 47 10                mov    %eax,0x10(%rdi)
>
> clang 9.0.1 gives:
>
>    0:   81 47 10 ff ff ff ff    addl   $0xffffffff,0x10(%rdi)
>
> But actually, clang does equally as well with:
>
> #define ADD_ONCE(d, i) *(volatile typeof(d) *)&(d) += (i)

I feel that ADD_ONCE conveys that it adds actually once (atomically),
that is, if there are concurrent ADD_ONCE, all of them will succeed.
This is not the case with the above variants and the 'ONCE' can turn
into a 'MAYBE', and since we probably want to avoid making this more
expensive on e.g. x86 that would need a LOCK-prefix.

In the case here, what we actually want is something that safely
increments/decrements if there are only concurrent readers (concurrent
writers disallowed). So 'add_exclusive(var, val)' (all-caps or not)
might be more appropriate. [As an aside, recent changes to KCSAN would
also allow us to assert for something like 'add_exclusive()' that
there are in fact no other writers but only concurrent readers, even
if all accesses are marked.]

If the single-writer constraint isn't wanted, but should still not be
atomic, maybe 'add_lossy()'?

Thanks,
-- Marco

> And testing further back, it generates the same code with your original
> WRITE_ONCE.
>
> If clang's optimization here is technically correct, maybe we should go
> talk to the gcc people about catching this case?