Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
From:   Boqun Feng <boqun.feng@gmail.com>
To:     Peter Zijlstra <peterz@infradead.org>
Cc:     Ingo Molnar <mingo@redhat.com>, Will Deacon <will@kernel.org>,
        Waiman Long <longman@redhat.com>,
        Boqun Feng <boqun.feng@gmail.com>,
        linux-kernel@vger.kernel.org,
        Johannes Berg <johannes@sipsolutions.net>
Subject: [PATCH 3/4] lockding/lockdep: Avoid to find wrong lock dep path in check_irq_usage()
Date:   Sat, 19 Jun 2021 01:01:09 +0800
Message-Id: <20210618170110.3699115-4-boqun.feng@gmail.com>
In-Reply-To: <20210618170110.3699115-1-boqun.feng@gmail.com>
References: <20210618170110.3699115-1-boqun.feng@gmail.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

In the step #3 of check_irq_usage(), we seach backwards to find a lock
whose usage conflicts the usage of @target_entry1 on safe/unsafe.
However, we should only keep the irq-unsafe usage of @target_entry1 into
consideration, because it could be a case where a lock is hardirq-unsafe
but soft-safe, and in check_irq_usage() we find it because its
hardirq-unsafe could result into a hardirq-safe-unsafe deadlock, but
currently since we don't filter out the other usage bits, so we may find
a lock dependency path softirq-unsafe -> softirq-safe, which in fact
doesn't cause a deadlock. And this may cause misleading lockdep splats.

Fix this by only keeping LOCKF_ENABLED_IRQ_ALL bits when we try the
backwards search.

Reported-by: Johannes Berg <johannes@sipsolutions.net>
Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
---
 kernel/locking/lockdep.c | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

diff --git a/kernel/locking/lockdep.c b/kernel/locking/lockdep.c
index 74d084a398be..6ff1e8405a83 100644
--- a/kernel/locking/lockdep.c
+++ b/kernel/locking/lockdep.c
@@ -2768,8 +2768,18 @@ static int check_irq_usage(struct task_struct *curr, struct held_lock *prev,
 	 * Step 3: we found a bad match! Now retrieve a lock from the backward
 	 * list whose usage mask matches the exclusive usage mask from the
 	 * lock found on the forward list.
+	 *
+	 * Note, we should only keep the LOCKF_ENABLED_IRQ_ALL bits, considering
+	 * the follow case:
+	 *
+	 * When trying to add A -> B to the graph, we find that there is a
+	 * hardirq-safe L, that L -> ... -> A, and another hardirq-unsafe M,
+	 * that B -> ... -> M. However M is **softirq-safe**, if we use exact
+	 * invert bits of M's usage_mask, we will find another lock N that is
+	 * **softirq-unsafe** and N -> ... -> A, however N -> .. -> M will not
+	 * cause a inversion deadlock.
 	 */
-	backward_mask = original_mask(target_entry1->class->usage_mask);
+	backward_mask = original_mask(target_entry1->class->usage_mask & LOCKF_ENABLED_IRQ_ALL);
 
 	ret = find_usage_backwards(&this, backward_mask, &target_entry);
 	if (bfs_error(ret)) {
-- 
2.30.2