Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
From:   Yuyang Du <duyuyang@gmail.com>
To:     peterz@infradead.org, will.deacon@arm.com, mingo@kernel.org
Cc:     bvanassche@acm.org, ming.lei@redhat.com, frederic@kernel.org,
        tglx@linutronix.de, linux-kernel@vger.kernel.org,
        longman@redhat.com, paulmck@linux.vnet.ibm.com,
        boqun.feng@gmail.com, Yuyang Du <duyuyang@gmail.com>
Subject: [PATCH v4 11/30] locking/lockdep: Remove irq-safe to irq-unsafe read check
Date:   Thu, 29 Aug 2019 16:31:13 +0800
Message-Id: <20190829083132.22394-12-duyuyang@gmail.com>
In-Reply-To: <20190829083132.22394-1-duyuyang@gmail.com>
References: <20190829083132.22394-1-duyuyang@gmail.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

We have a lockdep warning:

  ========================================================
  WARNING: possible irq lock inversion dependency detected
  5.1.0-rc7+ #141 Not tainted
  --------------------------------------------------------
  kworker/8:2/328 just changed the state of lock:
  0000000007f1a95b (&(&host->lock)->rlock){-...}, at: ata_bmdma_interrupt+0x27/0x1c0 [libata]
  but this lock took another, HARDIRQ-READ-unsafe lock in the past:
   (&trig->leddev_list_lock){.+.?}

and interrupts could create inverse lock ordering between them.

other info that might help us debug this:
   Possible interrupt unsafe locking scenario:

         CPU0                    CPU1
         ----                    ----
    lock(&trig->leddev_list_lock);
                                 local_irq_disable();
                                 lock(&(&host->lock)->rlock);
                                 lock(&trig->leddev_list_lock);
    <Interrupt>
      lock(&(&host->lock)->rlock);

 *** DEADLOCK ***

This splat is a false positive, which is enabled by the addition of
recursive read locks in the graph. Specifically, trig->leddev_list_lock is a
rwlock_t type, which was not in the graph before recursive read lock support
was added in lockdep.

This false positve is caused by a "false-positive" check in IRQ usage check.

In mark_lock_irq(), the following checks are currently performed:

   ----------------------------------
  |   ->      | unsafe | read unsafe |
  |----------------------------------|
  | safe      |  F  B  |    F* B*    |
  |----------------------------------|
  | read safe |  F* B* |      -      |
   ----------------------------------

Where:
F: check_usage_forwards
B: check_usage_backwards
*: check enabled by STRICT_READ_CHECKS

But actually the safe -> unsafe read dependency does not create a deadlock
scenario.

Fix this by simply removing those two checks, and since safe read -> unsafe
is indeed a problem, these checks are not actually strict per se, so remove
the macro STRICT_READ_CHECKS, and we have the following checks:

   ----------------------------------
  |   ->      | unsafe | read unsafe |
  |----------------------------------|
  | safe      |  F  B  |      -      |
  |----------------------------------|
  | read safe |  F  B  |      -      |
   ----------------------------------

Signed-off-by: Yuyang Du <duyuyang@gmail.com>
---
 kernel/locking/lockdep.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/kernel/locking/lockdep.c b/kernel/locking/lockdep.c
index acbd538..1dda9de 100644
--- a/kernel/locking/lockdep.c
+++ b/kernel/locking/lockdep.c
@@ -3221,8 +3221,6 @@ static int SOFTIRQ_verbose(struct lock_class *class)
 	return 0;
 }
 
-#define STRICT_READ_CHECKS	1
-
 static int (*state_verbose_f[])(struct lock_class *class) = {
 #define LOCKDEP_STATE(__STATE) \
 	__STATE##_verbose,
@@ -3268,7 +3266,7 @@ typedef int (*check_usage_f)(struct task_struct *, struct held_lock *,
 	 * Validate that the lock dependencies don't have conflicting usage
 	 * states.
 	 */
-	if ((!read || STRICT_READ_CHECKS) &&
+	if ((!read || !dir) &&
 			!usage(curr, this, excl_bit, state_name(new_bit & ~LOCK_USAGE_READ_MASK)))
 		return 0;
 
@@ -3279,7 +3277,7 @@ typedef int (*check_usage_f)(struct task_struct *, struct held_lock *,
 		if (!valid_state(curr, this, new_bit, excl_bit + LOCK_USAGE_READ_MASK))
 			return 0;
 
-		if (STRICT_READ_CHECKS &&
+		if (dir &&
 			!usage(curr, this, excl_bit + LOCK_USAGE_READ_MASK,
 				state_name(new_bit + LOCK_USAGE_READ_MASK)))
 			return 0;
-- 
1.8.3.1