On Sat, 2013-07-27 at 10:38 -0700, Eric Dumazet wrote:
> On Sat, 2013-07-27 at 17:55 +0100, Ben Hutchings wrote:
> > On Fri, 2013-07-26 at 13:47 -0700, Greg Kroah-Hartman wrote:
> > > 3.10-stable review patch. If anyone has any objections, please let me know.
> > >
> > > ------------------
> > >
> > > From: Eric Dumazet <edumazet@xxxxxxxxxx>
> > >
> > > [ Upstream commit 3e3aac497513c669e1c62c71e1d552ea85c1d974 ]
> > >
> > > egress_priority_map[] hash table updates are protected by rtnl,
> > > and we never remove elements until device is dismantled.
> > >
> > > We have to make sure that before inserting an new element in hash table,
> > > all its fields are committed to memory or else another cpu could
> > > find corrupt values and crash.
> > >
> > > Signed-off-by: Eric Dumazet <edumazet@xxxxxxxxxx>
> > > Cc: Patrick McHardy <kaber@xxxxxxxxx>
> > > Signed-off-by: David S. Miller <davem@xxxxxxxxxxxxx>
> > > Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
> > > ---
> > > net/8021q/vlan_dev.c | 7 +++++++
> > > 1 file changed, 7 insertions(+)
> > >
> > > --- a/net/8021q/vlan_dev.c
> > > +++ b/net/8021q/vlan_dev.c
> > > @@ -73,6 +73,8 @@ vlan_dev_get_egress_qos_mask(struct net_
> > > {
> > > struct vlan_priority_tci_mapping *mp;
> > >
> > > + smp_rmb(); /* coupled with smp_wmb() in vlan_dev_set_egress_priority() */
> > > +
> > > mp = vlan_dev_priv(dev)->egress_priority_map[(skb->priority & 0xF)];
> >
> > This barrier is the wrong type in the wrong place; there should be a
> > smp_read_barrier_depends() *after* reading the pointer from
> > egress_priority_map, and after reading each next pointer.
> >
>
> Not sure its a problem anyway, as we insert new items at the beginning
> of the chain, and we never delete any item. What you describe would make
> sense if we were deleting an item.
>
> Once a reader gets a pointer, the whole chain is stable.
For every sane architecture, data-dependency is an automatic read
barrier. But Alpha is weird: it can prefetch mp->foo using a speculated
value for mp before it sees mp. If the speculation was correct it won't
get the correct value of mp->foo. Hence smp_read_barrier_depends().
> The important part of the patch was the smp_wmb(), to make sure np->next
> was committed before "vlan->egress_priority_map[skb_prio & 0xF] = np;"
Sure. But the smp_rmb() you added doesn't synchronise with it (a
barrier is always between operations, but where is the 'before'
operation here?).
Ben.
--
Ben Hutchings
Once a job is fouled up, anything done to improve it makes it worse.
Attachment:
signature.asc
Description: This is a digitally signed message part