On Wed, Jul 06, 2022 at 04:50:04PM +0200, Florian Westphal wrote:
> Kajetan Puchalski reports crash on ARM, with backtrace of:
>
> __nf_ct_delete_from_lists
> nf_ct_delete
> early_drop
> __nf_conntrack_alloc
>
> Unlike atomic_inc_not_zero, refcount_inc_not_zero is not a full barrier.
> conntrack uses SLAB_TYPESAFE_BY_RCU, i.e. it is possible that a 'newly'
> allocated object is still in use on another CPU:
>
> CPU1 CPU2
> encounter 'ct' during hlist walk
> delete_from_lists
> refcount drops to 0
> kmem_cache_free(ct);
> __nf_conntrack_alloc() // returns same object
> refcount_inc_not_zero(ct); /* might fail */
>
> /* If set, ct is public/in the hash table */
> test_bit(IPS_CONFIRMED_BIT, &ct->status);
>
> In case CPU1 already set refcount back to 1, refcount_inc_not_zero()
> will succeed.
>
> The expected possibilities for a CPU that obtained the object 'ct'
> (but no reference so far) are:
>
> 1. refcount_inc_not_zero() fails. CPU2 ignores the object and moves to
> the next entry in the list. This happens for objects that are about
> to be free'd, that have been free'd, or that have been reallocated
> by __nf_conntrack_alloc(), but where the refcount has not been
> increased back to 1 yet.
>
> 2. refcount_inc_not_zero() succeeds. CPU2 checks the CONFIRMED bit
> in ct->status. If set, the object is public/in the table.
>
> If not, the object must be skipped; CPU2 calls nf_ct_put() to
> un-do the refcount increment and moves to the next object.
>
> Parallel deletion from the hlists is prevented by a
> 'test_and_set_bit(IPS_DYING_BIT, &ct->status);' check, i.e. only one
> cpu will do the unlink, the other one will only drop its reference count.
>
> Because refcount_inc_not_zero is not a full barrier, CPU2 may try to
> delete an object that is not on any list:
>
> 1. refcount_inc_not_zero() successful (refcount inited to 1 on other CPU)
> 2. CONFIRMED test also successful (load was reordered or zeroing
> of ct->status not yet visible)
> 3. delete_from_lists unlinks entry not on the hlist, because
> IPS_DYING_BIT is 0 (already cleared).
>
> 2) is already wrong: CPU2 will handle a partially initited object
> that is supposed to be private to CPU1.
>
> Add needed barriers when refcount_inc_not_zero() is successful.
>
> It also inserts a smp_wmb() before the refcount is set to 1 during
> allocation.
>
> Because other CPU might still 'see' the object, refcount_set(1)
> "resurrects" the object, so we need to make sure that other CPUs will
> also observe the right contents. In particular, the CONFIRMED bit test
> must only pass once the object is fully initialised and either in the
> hash or about to be inserted (with locks held to delay possible unlink from
> early_drop or gc worker).
>
> I did not change flow_offload_alloc(), as far as I can see it should call
> refcount_inc(), not refcount_inc_not_zero(): the ct object is attached to
> the skb so its refcount should be >= 1 in all cases.
>
> v2: prefer smp_acquire__after_ctrl_dep to smp_rmb (Will Deacon).
> v3: keep smp_acquire__after_ctrl_dep close to refcount_inc_not_zero call
> add comment in nf_conntrack_netlink, no control dependency there
> due to locks.
>
> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> Reported-by: Kajetan Puchalski <kajetan.puchalski@xxxxxxx>
> Diagnosed-by: Will Deacon <will@xxxxxxxxxx>
> Fixes: 719774377622 ("netfilter: conntrack: convert to refcount_t api")
> Signed-off-by: Florian Westphal <fw@xxxxxxxxx>
> ---
> net/netfilter/nf_conntrack_core.c | 22 ++++++++++++++++++++++
> net/netfilter/nf_conntrack_netlink.c | 1 +
> net/netfilter/nf_conntrack_standalone.c | 3 +++
> 3 files changed, 26 insertions(+)
Acked-by: Will Deacon <will@xxxxxxxxxx>
Will
