On 11/30/2017 08:32 PM, Tommi Rantala wrote:
>> In my opinion, the real root cause of the issue is because we too early
>> set a not-yet-initialized bearer instance to ub->bearer through
>> rcu_assign_pointer(ub->bearer, b) in tipc_udp_enable(). Instead if we
>> assign the bearer pointer at the end of tipc_udp_enable() where the
>> bearer has been completed the initialization, the issue would be avoided.
> Hi, sorry, I fail to see how that helps.
>
> bearer->tolerance is only initialized in tipc_enable_bearer() after
> calling m->enable_media() ie. tipc_udp_enable().
>
> So even if we do "rcu_assign_pointer(ub->bearer, b)" later in
> tipc_udp_enable(), bearer->tolerance will still be uninitialized, and
> the crash can happen.
Sorry, I missed the point that b->tolerance is not uninitialized when we
assign bearer pointer to ub->bearer later.
But in my view the issue happened is because we enable media too early.
So it's better to change the code as belows:
diff --git a/net/tipc/bearer.c b/net/tipc/bearer.c
index 47ec121..ec6f02a 100644
--- a/net/tipc/bearer.c
+++ b/net/tipc/bearer.c
@@ -320,19 +320,18 @@ static int tipc_enable_bearer(struct net *net,
const char *name,
strcpy(b->name, name);
b->media = m;
- res = m->enable_media(net, b, attr);
- if (res) {
- pr_warn("Bearer <%s> rejected, enable failure (%d)\n",
- name, -res);
- return -EINVAL;
- }
-
b->identity = bearer_id;
b->tolerance = m->tolerance;
b->window = m->window;
b->domain = disc_domain;
b->net_plane = bearer_id + 'A';
b->priority = priority;
+ res = m->enable_media(net, b, attr);
+ if (res) {
+ pr_warn("Bearer <%s> rejected, enable failure (%d)\n",
+ name, -res);
+ return -EINVAL;
+ }
test_and_set_bit_lock(0, &b->up);
On 01.12.2017 15:18, Ying Xue wrote:
> On 11/30/2017 08:32 PM, Tommi Rantala wrote:
>>> In my opinion, the real root cause of the issue is because we too early
>>> set a not-yet-initialized bearer instance to ub->bearer through
>>> rcu_assign_pointer(ub->bearer, b) in tipc_udp_enable(). Instead if we
>>> assign the bearer pointer at the end of tipc_udp_enable() where the
>>> bearer has been completed the initialization, the issue would be avoided.
>> Hi, sorry, I fail to see how that helps.
>>
>> bearer->tolerance is only initialized in tipc_enable_bearer() after
>> calling m->enable_media() ie. tipc_udp_enable().
>>
>> So even if we do "rcu_assign_pointer(ub->bearer, b)" later in
>> tipc_udp_enable(), bearer->tolerance will still be uninitialized, and
>> the crash can happen.
>
> Sorry, I missed the point that b->tolerance is not uninitialized when we
> assign bearer pointer to ub->bearer later.
>
> But in my view the issue happened is because we enable media too early.
> So it's better to change the code as belows:
Thanks, looks good to me!
Tested in 4.4 (which does not have b->up), and this fixes the oops also
there.
-Tommi
> diff --git a/net/tipc/bearer.c b/net/tipc/bearer.c
> index 47ec121..ec6f02a 100644
> --- a/net/tipc/bearer.c
> +++ b/net/tipc/bearer.c
> @@ -320,19 +320,18 @@ static int tipc_enable_bearer(struct net *net,
> const char *name,
>
> strcpy(b->name, name);
> b->media = m;
> - res = m->enable_media(net, b, attr);
> - if (res) {
> - pr_warn("Bearer <%s> rejected, enable failure (%d)\n",
> - name, -res);
> - return -EINVAL;
> - }
> -
> b->identity = bearer_id;
> b->tolerance = m->tolerance;
> b->window = m->window;
> b->domain = disc_domain;
> b->net_plane = bearer_id + 'A';
> b->priority = priority;
> + res = m->enable_media(net, b, attr);
> + if (res) {
> + pr_warn("Bearer <%s> rejected, enable failure (%d)\n",
> + name, -res);
> + return -EINVAL;
> + }
> test_and_set_bit_lock(0, &b->up);
>
Looks ok to me too. But I suggest we let David apply the pending patch as is to net, as that code is still wrong, and then post this improvement to net-next later.
///jon
> -----Original Message-----
> From: Tommi Rantala [mailto:[email protected]]
> Sent: Friday, December 01, 2017 09:03
> To: Ying Xue <[email protected]>; Jon Maloy
> <[email protected]>; David S. Miller <[email protected]>;
> [email protected]; [email protected]; linux-
> [email protected]
> Subject: Re: [PATCH net v2] tipc: call tipc_rcv() only if bearer is up in
> tipc_udp_recv()
>
> On 01.12.2017 15:18, Ying Xue wrote:
> > On 11/30/2017 08:32 PM, Tommi Rantala wrote:
> >>> In my opinion, the real root cause of the issue is because we too
> >>> early set a not-yet-initialized bearer instance to ub->bearer
> >>> through rcu_assign_pointer(ub->bearer, b) in tipc_udp_enable().
> >>> Instead if we assign the bearer pointer at the end of
> >>> tipc_udp_enable() where the bearer has been completed the
> initialization, the issue would be avoided.
> >> Hi, sorry, I fail to see how that helps.
> >>
> >> bearer->tolerance is only initialized in tipc_enable_bearer() after
> >> calling m->enable_media() ie. tipc_udp_enable().
> >>
> >> So even if we do "rcu_assign_pointer(ub->bearer, b)" later in
> >> tipc_udp_enable(), bearer->tolerance will still be uninitialized, and
> >> the crash can happen.
> >
> > Sorry, I missed the point that b->tolerance is not uninitialized when
> > we assign bearer pointer to ub->bearer later.
> >
> > But in my view the issue happened is because we enable media too early.
> > So it's better to change the code as belows:
>
> Thanks, looks good to me!
>
> Tested in 4.4 (which does not have b->up), and this fixes the oops also there.
>
> -Tommi
>
>
> > diff --git a/net/tipc/bearer.c b/net/tipc/bearer.c index
> > 47ec121..ec6f02a 100644
> > --- a/net/tipc/bearer.c
> > +++ b/net/tipc/bearer.c
> > @@ -320,19 +320,18 @@ static int tipc_enable_bearer(struct net *net,
> > const char *name,
> >
> > strcpy(b->name, name);
> > b->media = m;
> > - res = m->enable_media(net, b, attr);
> > - if (res) {
> > - pr_warn("Bearer <%s> rejected, enable failure (%d)\n",
> > - name, -res);
> > - return -EINVAL;
> > - }
> > -
> > b->identity = bearer_id;
> > b->tolerance = m->tolerance;
> > b->window = m->window;
> > b->domain = disc_domain;
> > b->net_plane = bearer_id + 'A';
> > b->priority = priority;
> > + res = m->enable_media(net, b, attr);
> > + if (res) {
> > + pr_warn("Bearer <%s> rejected, enable failure (%d)\n",
> > + name, -res);
> > + return -EINVAL;
> > + }
> > test_and_set_bit_lock(0, &b->up);
> >