Re: [RT] 5.9-rt14 softirq_ctrl.lock vs listening_hash[i].lock lockdep splat

[Mike Galbraith <efault@gmx.de>]

On Wed, 2020-12-09 at 11:05 +0100, Peter Zijlstra wrote:
>
> > [   47.844585]  Possible unsafe locking scenario:
> >
> > [   47.844586]        CPU0                    CPU1
> > [   47.844586]        ----                    ----
> > [   47.844587]   lock(&h->listening_hash[i].lock);
> > [   47.844588]                                lock((softirq_ctrl.lock).lock);
> > [   47.844588]                                lock(&h->listening_hash[i].lock);
> > [   47.844589]   lock((softirq_ctrl.lock).lock);
> > [   47.844590]
> >                 *** DEADLOCK ***
> >
> >
> So I've been looking at these local_lock vs lockdep splats for a bit,
> and unlike the IRQ inversions as reported here:
>
>   https://lore.kernel.org/linux-usb/20201029174348.omqiwjqy64tebg5z@linutronix.de/
>
> I think the above is an actual real problem (for RT).
>
> AFAICT the above translates to:
>
>   inet_listen()
>     lock_sock()
>       spin_lock_bh(&sk->sk_lock.slock);
> 	acquire(softirq_ctrl);
> 	acquire(&sk->sk_lock.slock);
>
>     inet_csk_listen_start()
>       sk->sk_prot->hash() := inet_hash()
> 	local_bh_disable()
> 	__inet_hash()
> 	  spin_lock(&ilb->lock);
> 	    acquire(&ilb->lock);
>
>   ----
>
>   tcp4_seq_next()
>     listening_get_next()
>       spin_lock(&ilb->lock);
> 	acquire(&ilb->lock);
>
>   tcp4_seq_show()
>     get_tcp4_sock()
>       sock_i_ino()
> 	read_lock_bh(&sk->sk_callback_lock);
> 	  acquire(softirq_ctrl)	// <---- whoops
> 	  acquire(&sk->sk_callback_lock)
>
>
> Which you can run in two tasks on the same CPU (and thus get the same
> softirq_ctrl local_lock), and deadlock.
>
> By holding softirq_ctrl we serialize against softirq-context
> (in-softirq) but that isn't relevant here, since neither context is
> that.
>
> On !RT there isn't a problem because softirq_ctrl isn't an actual lock,
> but the moment that turns into a real lock (like on RT) you're up a
> creek.
>
> In general we have the rule that as long as a lock is only ever used
> from task context (like the above ilb->lock, afaict) then it doesn't
> matter if you also take it with (soft)irqs disabled or not. But this
> softirq scheme breaks that. If you ever take a lock with BH disabled,
> you must now always take it with BH disabled, otherwise you risk
> deadlocks against the softirq_ctrl lock.
>
> Or am I missing something obvious (again) ?

Sebastian fixed this via...

From 0fe43be6c32e05d0dd692069d41a40c5453a2195 Mon Sep 17 00:00:00 2001
From: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Date: Mon, 12 Oct 2020 17:33:54 +0200
Subject: tcp: Remove superfluous BH-disable around listening_hash

Commit
   9652dc2eb9e40 ("tcp: relax listening_hash operations")

removed the need to disable bottom half while acquiring
listening_hash.lock. There are still two callers left which disable
bottom half before the lock is acquired.

Drop local_bh_disable() around __inet_hash() which acquires
listening_hash->lock, invoke inet_ehash_nolisten() with disabled BH.
inet_unhash() conditionally acquires listening_hash->lock.

Reported-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/linux-rt-users/12d6f9879a97cd56c09fb53dee343cbb14f7f1f7.camel@gmx.de/
Signed-off-by: Mike Galbraith <efault@gmx.de>
---
 net/ipv4/inet_hashtables.c  | 19 ++++++++++++-------
 net/ipv6/inet6_hashtables.c |  5 +----
 2 files changed, 13 insertions(+), 11 deletions(-)

diff --git a/net/ipv4/inet_hashtables.c b/net/ipv4/inet_hashtables.c
index 45fb450b4522..5fb95030e7c0 100644
--- a/net/ipv4/inet_hashtables.c
+++ b/net/ipv4/inet_hashtables.c
@@ -635,7 +635,9 @@ int __inet_hash(struct sock *sk, struct sock *osk)
 	int err = 0;

 	if (sk->sk_state != TCP_LISTEN) {
+		local_bh_disable();
 		inet_ehash_nolisten(sk, osk, NULL);
+		local_bh_enable();
 		return 0;
 	}
 	WARN_ON(!sk_unhashed(sk));
@@ -667,11 +669,8 @@ int inet_hash(struct sock *sk)
 {
 	int err = 0;

-	if (sk->sk_state != TCP_CLOSE) {
-		local_bh_disable();
+	if (sk->sk_state != TCP_CLOSE)
 		err = __inet_hash(sk, NULL);
-		local_bh_enable();
-	}

 	return err;
 }
@@ -682,17 +681,20 @@ void inet_unhash(struct sock *sk)
 	struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 	struct inet_listen_hashbucket *ilb = NULL;
 	spinlock_t *lock;
+	bool state_listen;

 	if (sk_unhashed(sk))
 		return;

 	if (sk->sk_state == TCP_LISTEN) {
+		state_listen = true;
 		ilb = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
-		lock = &ilb->lock;
+		spin_lock(&ilb->lock);
 	} else {
+		state_listen = false;
 		lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
+		spin_lock_bh(lock);
 	}
-	spin_lock_bh(lock);
 	if (sk_unhashed(sk))
 		goto unlock;

@@ -705,7 +707,10 @@ void inet_unhash(struct sock *sk)
 	__sk_nulls_del_node_init_rcu(sk);
 	sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 unlock:
-	spin_unlock_bh(lock);
+	if (state_listen)
+		spin_unlock(&ilb->lock);
+	else
+		spin_unlock_bh(lock);
 }
 EXPORT_SYMBOL_GPL(inet_unhash);

diff --git a/net/ipv6/inet6_hashtables.c b/net/ipv6/inet6_hashtables.c
index 55c290d55605..9bad345cba9a 100644
--- a/net/ipv6/inet6_hashtables.c
+++ b/net/ipv6/inet6_hashtables.c
@@ -333,11 +333,8 @@ int inet6_hash(struct sock *sk)
 {
 	int err = 0;

-	if (sk->sk_state != TCP_CLOSE) {
-		local_bh_disable();
+	if (sk->sk_state != TCP_CLOSE)
 		err = __inet_hash(sk, NULL);
-		local_bh_enable();
-	}

 	return err;
 }
--
2.29.2