Hi Hillf, On Mon, Jul 19, 2021 at 7:16 PM Hillf Danton <hdanton@xxxxxxxx> wrote: > > On Mon, 19 Jul 2021 17:03:53 +0800 Wang ShaoBo wrote: > > > >I have tried this before, this will trigger error "underflow of refcount > >of chan" as following: > > > >[ 118.708179][ T3086] ------------[ cut here ]------------ > >[ 118.710172][ T3086] refcount_t: underflow; use-after-free. > >[ 118.713391][ T3086] WARNING: CPU: 4 PID: 3086 at lib/refcount.c:28 > >refcount_warn_saturate+0x30a/0x3c0 > >[ 118.716774][ T3086] Modules linked in: > >[ 118.718279][ T3086] CPU: 4 PID: 3086 Comm: kworker/4:2 Not tainted > >5.12.0-rc6+ #84 > >[ 118.721005][ T3086] Hardware name: QEMU Standard PC (i440FX + PIIX, > >1996), BIOS 1.13.0-1ubuntu1.1 04/01/2014 > >[ 118.722846][ T3086] Workqueue: events l2cap_chan_timeout > >[ 118.723786][ T3086] RIP: 0010:refcount_warn_saturate+0x30a/0x3c0 > >... > >[ 118.737912][ T3086] CR2: 0000000020000040 CR3: 0000000011029000 CR4: > >00000000000006e0 > >[ 118.739187][ T3086] DR0: 0000000000000000 DR1: 0000000000000000 DR2: > >0000000000000000 > >[ 118.740451][ T3086] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: > >0000000000000400 > >[ 118.741720][ T3086] Call Trace: > >[ 118.742262][ T3086] l2cap_sock_close_cb+0x165/0x170 > >[ 118.743124][ T3086] ? l2cap_sock_teardown_cb+0x560/0x560 > > > >Actually, if adding sock_hold(sk) in l2cap_sock_init(), > >l2cap_sock_kill() will continue to excute untill it found > > > >now chan's refcount is 0, this is because sock was not freed in last > >round execution of l2cap_sock_kill(). > > Well double kill cannot be walked around without adding more - add the > destroy callback to make the chan->data recorded sock survive kill. It > will be released when chan is destroyed to cut the race in reguards to > accessing sock by making chan->data stable throughout chan's lifespan. > > > +++ x/net/bluetooth/l2cap_core.c > @@ -485,7 +485,10 @@ static void l2cap_chan_destroy(struct kr > list_del(&chan->global_l); > write_unlock(&chan_list_lock); > > - kfree(chan); > + if (chan->ops && chan->ops->destroy) > + chan->ops->destroy(chan); > + else > + kfree(chan); While Im fine adding a destroy callback the kfree shall be still in l2cap_chan_destroy: if (chan->ops && chan->ops->destroy) /* Destroy chan->data */ chan->ops->destroy(chan->data); kfree(chan); > } > > void l2cap_chan_hold(struct l2cap_chan *c) > +++ x/net/bluetooth/l2cap_sock.c > @@ -1220,11 +1220,13 @@ static void l2cap_sock_kill(struct sock > > BT_DBG("sk %p state %s", sk, state_to_string(sk->sk_state)); > > + /* double kill means noop */ > + if (sock_flag(sk, SOCK_DEAD)) > + return; > /* Kill poor orphan */ > > l2cap_chan_put(l2cap_pi(sk)->chan); > sock_set_flag(sk, SOCK_DEAD); > - sock_put(sk); > } > > static int __l2cap_wait_ack(struct sock *sk, struct l2cap_chan *chan) > @@ -1504,6 +1506,14 @@ done: > return err; > } > > +static void l2cap_sock_destroy_cb(struct l2cap_chan *chan) > +{ > + struct sock *sk = chan->data; > + > + sock_put(sk); > + kfree(chan); > +} > + > static void l2cap_sock_close_cb(struct l2cap_chan *chan) > { > struct sock *sk = chan->data; > @@ -1690,6 +1700,7 @@ static const struct l2cap_ops l2cap_chan > .new_connection = l2cap_sock_new_connection_cb, > .recv = l2cap_sock_recv_cb, > .close = l2cap_sock_close_cb, > + .destroy = l2cap_sock_destroy_cb, If you do the changes above you can probably have sock_put directly set as .destroy. > .teardown = l2cap_sock_teardown_cb, > .state_change = l2cap_sock_state_change_cb, > .ready = l2cap_sock_ready_cb, -- Luiz Augusto von Dentz