[PATCH v3 1/1] tcp/dcpp: Un-pin tw_timer

[Valentin Schneider <vschneid@redhat.com>]

The TCP timewait timer is proving to be problematic for setups where scheduler
CPU isolation is achieved at runtime via cpusets (as opposed to statically via
isolcpus=domains).

What happens there is a CPU goes through tcp_time_wait(), arming the time_wait
timer, then gets isolated. TCP_TIMEWAIT_LEN later, the timer fires, causing
interference for the now-isolated CPU. This is conceptually similar to the issue
described in
  e02b93124855 ("workqueue: Unbind kworkers before sending them to exit()")

Keep softirqs disabled, but make the timer un-pinned and arm it *after* the
hashdance.

This introduces the following (non-fatal) race:

  CPU0                        CPU1
    allocates a tw
    insert it in hash table
				finds the TW and removes it
				(timer cancel does nothing)
    arms a TW timer, lasting

This partially reverts
  ed2e92394589 ("tcp/dccp: fix timewait races in timer handling")
and
  ec94c2696f0b ("tcp/dccp: avoid one atomic operation for timewait hashdance")

This also reinstores a comment from
  ec94c2696f0b ("tcp/dccp: avoid one atomic operation for timewait hashdance")
as inet_twsk_hashdance() had a "Step 1" and "Step 3" comment, but the "Step
2" had gone missing.

Link: https://lore.kernel.org/all/ZPhpfMjSiHVjQkTk@localhost.localdomain/
Signed-off-by: Valentin Schneider <vschneid@redhat.com>
---
 net/dccp/minisocks.c          | 16 +++++++---------
 net/ipv4/inet_timewait_sock.c | 20 +++++++++++++++-----
 net/ipv4/tcp_minisocks.c      | 16 +++++++---------
 3 files changed, 29 insertions(+), 23 deletions(-)

diff --git a/net/dccp/minisocks.c b/net/dccp/minisocks.c
index 64d805b27adde..2f0fad4255e36 100644
--- a/net/dccp/minisocks.c
+++ b/net/dccp/minisocks.c
@@ -53,16 +53,14 @@ void dccp_time_wait(struct sock *sk, int state, int timeo)
 		if (state == DCCP_TIME_WAIT)
 			timeo = DCCP_TIMEWAIT_LEN;
 
-		/* tw_timer is pinned, so we need to make sure BH are disabled
-		 * in following section, otherwise timer handler could run before
-		 * we complete the initialization.
-		 */
-		local_bh_disable();
-		inet_twsk_schedule(tw, timeo);
-		/* Linkage updates.
-		 * Note that access to tw after this point is illegal.
-		 */
+	       local_bh_disable();
+
+		// Linkage updates
 		inet_twsk_hashdance(tw, sk, &dccp_hashinfo);
+		inet_twsk_schedule(tw, timeo);
+		// Access to tw after this point is illegal.
+		inet_twsk_put(tw);
+
 		local_bh_enable();
 	} else {
 		/* Sorry, if we're out of memory, just CLOSE this
diff --git a/net/ipv4/inet_timewait_sock.c b/net/ipv4/inet_timewait_sock.c
index 5befa4de5b241..61a053fbd329c 100644
--- a/net/ipv4/inet_timewait_sock.c
+++ b/net/ipv4/inet_timewait_sock.c
@@ -129,6 +129,7 @@ void inet_twsk_hashdance(struct inet_timewait_sock *tw, struct sock *sk,
 
 	spin_lock(lock);
 
+	/* Step 2: Hash TW into tcp ehash chain */
 	inet_twsk_add_node_rcu(tw, &ehead->chain);
 
 	/* Step 3: Remove SK from hash chain */
@@ -137,16 +138,15 @@ void inet_twsk_hashdance(struct inet_timewait_sock *tw, struct sock *sk,
 
 	spin_unlock(lock);
 
-	/* tw_refcnt is set to 3 because we have :
+	/* tw_refcnt is set to 4 because we have :
 	 * - one reference for bhash chain.
 	 * - one reference for ehash chain.
 	 * - one reference for timer.
+	 * - one reference for ourself (our caller will release it).
 	 * We can use atomic_set() because prior spin_lock()/spin_unlock()
 	 * committed into memory all tw fields.
-	 * Also note that after this point, we lost our implicit reference
-	 * so we are not allowed to use tw anymore.
 	 */
-	refcount_set(&tw->tw_refcnt, 3);
+	refcount_set(&tw->tw_refcnt, 4);
 }
 EXPORT_SYMBOL_GPL(inet_twsk_hashdance);
 
@@ -192,7 +192,7 @@ struct inet_timewait_sock *inet_twsk_alloc(const struct sock *sk,
 		tw->tw_prot	    = sk->sk_prot_creator;
 		atomic64_set(&tw->tw_cookie, atomic64_read(&sk->sk_cookie));
 		twsk_net_set(tw, sock_net(sk));
-		timer_setup(&tw->tw_timer, tw_timer_handler, TIMER_PINNED);
+		timer_setup(&tw->tw_timer, tw_timer_handler, 0);
 		/*
 		 * Because we use RCU lookups, we should not set tw_refcnt
 		 * to a non null value before everything is setup for this
@@ -217,6 +217,16 @@ EXPORT_SYMBOL_GPL(inet_twsk_alloc);
  */
 void inet_twsk_deschedule_put(struct inet_timewait_sock *tw)
 {
+	/* This can race with tcp_time_wait() and dccp_time_wait(), as the timer
+	 * is armed /after/ adding it to the hashtables.
+	 *
+	 * If this is interleaved between inet_twsk_hashdance() and inet_twsk_put(),
+	 * then this is a no-op: the timer will still end up armed.
+	 *
+	 * Conversely, if this successfully deletes the timer, then we know we
+	 * have already gone through {tcp,dcpp}_time_wait(), and we can safely
+	 * call inet_twsk_kill().
+	 */
 	if (del_timer_sync(&tw->tw_timer))
 		inet_twsk_kill(tw);
 	inet_twsk_put(tw);
diff --git a/net/ipv4/tcp_minisocks.c b/net/ipv4/tcp_minisocks.c
index 9e85f2a0bddd4..54e025ba9b015 100644
--- a/net/ipv4/tcp_minisocks.c
+++ b/net/ipv4/tcp_minisocks.c
@@ -338,16 +338,14 @@ void tcp_time_wait(struct sock *sk, int state, int timeo)
 		if (state == TCP_TIME_WAIT)
 			timeo = TCP_TIMEWAIT_LEN;
 
-		/* tw_timer is pinned, so we need to make sure BH are disabled
-		 * in following section, otherwise timer handler could run before
-		 * we complete the initialization.
-		 */
-		local_bh_disable();
-		inet_twsk_schedule(tw, timeo);
-		/* Linkage updates.
-		 * Note that access to tw after this point is illegal.
-		 */
+	       local_bh_disable();
+
+		// Linkage updates.
 		inet_twsk_hashdance(tw, sk, net->ipv4.tcp_death_row.hashinfo);
+		inet_twsk_schedule(tw, timeo);
+		// Access to tw after this point is illegal.
+		inet_twsk_put(tw);
+
 		local_bh_enable();
 	} else {
 		/* Sorry, if we're out of memory, just CLOSE this
-- 
2.43.0