(
On Thu, Dec 16, 2021 at 1:12 PM Xin Long <lucien.xin@xxxxxxxxx> wrote:
>
> On Thu, Dec 16, 2021 at 12:14 PM Lee Jones <lee.jones@xxxxxxxxxx> wrote:
> >
> > On Thu, 16 Dec 2021, Lee Jones wrote:
> >
> > > On Thu, 16 Dec 2021, Xin Long wrote:
> > >
> > > > On Thu, Dec 16, 2021 at 11:39 AM Lee Jones <lee.jones@xxxxxxxxxx> wrote:
> > > > >
> > > > > On Thu, 16 Dec 2021, Xin Long wrote:
> > > > >
> > > > > > On Wed, Dec 15, 2021 at 8:48 PM Jakub Kicinski <kuba@xxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > On Tue, 14 Dec 2021 21:57:32 +0000 Lee Jones wrote:
> > > > > > > > The cause of the resultant dump_stack() reported below is a
> > > > > > > > dereference of a freed pointer to 'struct sctp_endpoint' in
> > > > > > > > sctp_sock_dump().
> > > > > > > >
> > > > > > > > This race condition occurs when a transport is cached into its
> > > > > > > > associated hash table followed by an endpoint/sock migration to a new
> > > > > > > > association in sctp_assoc_migrate() prior to their subsequent use in
> > > > > > > > sctp_diag_dump() which uses sctp_for_each_transport() to walk the hash
> > > > > > > > table calling into sctp_sock_dump() where the dereference occurs.
> > > > >
> > > > > > in sctp_sock_dump():
> > > > > > struct sock *sk = ep->base.sk;
> > > > > > ... <--[1]
> > > > > > lock_sock(sk);
> > > > > >
> > > > > > Do you mean in [1], the sk is peeled off and gets freed elsewhere?
> > > > >
> > > > > 'ep' and 'sk' are both switched out for new ones in sctp_sock_migrate().
> > > > >
> > > > > > if that's true, it's still late to do sock_hold(sk) in your this patch.
> > > > >
> > > > > No, that's not right.
> > > > >
> > > > > The schedule happens *inside* the lock_sock() call.
> > > > Sorry, I don't follow this.
> > > > We can't expect when the schedule happens, why do you think this
> > > > can never be scheduled before the lock_sock() call?
> > >
> > > True, but I've had this running for hours and it hasn't reproduced.
> I understand, but it's a crash, we shouldn't take any risk that it
> will never happen.
> you may try to add a usleep() before the lock_sock call to reproduce it.
>
> > >
> > > Without this patch, I can reproduce this in around 2 seconds.
> > >
> > > The C-repro for this is pretty intense!
> > >
> > > If you want to be *sure* that a schedule will never happen, we can
> > > take a reference directly with:
> > >
> > > ep = sctp_endpoint_hold(tsp->asoc->ep);
> > > sk = sock_hold(ep->base.sk);
> > >
> > > Which was my original plan before I soak tested this submitted patch
> > > for hours without any sign of reproducing the issue.
> we tried to not export sctp_obj_hold/put(), that's why we had
> sctp_for_each_transport().
>
> ep itself holds a reference of sk when it's alive, so it's weird to do
> these 2 together.
>
> > >
> > > > If the sock is peeled off or is being freed, we shouldn't dump this sock,
> > > > and it's better to skip it.
> > >
> > > I guess we can do that too.
> > >
> > > Are you suggesting sctp_sock_migrate() as the call site?
> diff --git a/net/sctp/socket.c b/net/sctp/socket.c
> index 85ac2e901ffc..56ea7a0e2add 100644
> --- a/net/sctp/socket.c
> +++ b/net/sctp/socket.c
> @@ -9868,6 +9868,7 @@ static int sctp_sock_migrate(struct sock *oldsk,
> struct sock *newsk,
> inet_sk_set_state(newsk, SCTP_SS_ESTABLISHED);
> }
>
> + sock_set_flag(oldsk, SOCK_RCU_FREE);
> release_sock(newsk);
>
> return 0;
>
> SOCK_RCU_FREE is set to the previous sk, so that this sk will not
> be freed between rcu_read_lock() and rcu_read_unlock().
>
> >
> > Also, when are you planning on testing the flag?
> SOCK_RCU_FREE flag is used when freeing sk in sk_destruct(),
> and if it's set, it will be freed in the next grace period of RCU.
>
> >
> > Won't that suffer with the same issue(s)?
> diff --git a/net/sctp/diag.c b/net/sctp/diag.c
> index 7970d786c4a2..b4c4acd9e67e 100644
> --- a/net/sctp/diag.c
> +++ b/net/sctp/diag.c
> @@ -309,16 +309,21 @@ static int sctp_tsp_dump_one(struct
> sctp_transport *tsp, void *p)
>
> static int sctp_sock_dump(struct sctp_transport *tsp, void *p)
> {
> - struct sctp_endpoint *ep = tsp->asoc->ep;
> struct sctp_comm_param *commp = p;
> - struct sock *sk = ep->base.sk;
> struct sk_buff *skb = commp->skb;
> struct netlink_callback *cb = commp->cb;
> const struct inet_diag_req_v2 *r = commp->r;
> struct sctp_association *assoc;
> + struct sctp_endpoint *ep;
> + struct sock *sk;
> int err = 0;
>
> + rcu_read_lock();
> + ep = tsp->asoc->ep;
> + sk = ep->base.sk;
> lock_sock(sk);
Unfortunately, this isn't going to work, as lock_sock() may sleep,
and is not allowed to be called understand rcu_read_lock() :(
> + if (tsp->asoc->ep != ep)
> + goto release;
> list_for_each_entry(assoc, &ep->asocs, asocs) {
> if (cb->args[4] < cb->args[1])
> goto next;
> @@ -358,6 +363,7 @@ static int sctp_sock_dump(struct sctp_transport
> *tsp, void *p)
> cb->args[4] = 0;
> release:
> release_sock(sk);
> + rcu_read_unlock();
> return err;
> }
>
> rcu_read_lock() will make sure sk from tsp->asoc->ep->base.sk will not
> be freed until rcu_read_unlock().
>
> That's all I have. Do you see any other way to fix this?
>
> Thanks.
>
> >
> > --
> > Lee Jones [李琼斯]
> > Senior Technical Lead - Developer Services
> > Linaro.org │ Open source software for Arm SoCs
> > Follow Linaro: Facebook | Twitter | Blog
