The (recently established) lock ordering mandates that the per-VM
vmapp_lock is acquired before we take the per-VPE lock.

As it turns out, its_vpe_set_affinity() takes the VPE lock, and
then calls into its_send_vmovp(), which itself takes the vmapp
lock. Obviously, this isn't what we want.

As its_send_vmovp() is only called from its_vpe_set_affinity(),
hoist the vmapp locking from the former into the latter, restoring
the expected order.

Fixes: f0eb154c39471 ("irqchip/gic-v4: Substitute vmovp_lock for a per-VM lock")
Reported-by: Zhou Wang <wangzhou1@hisilicon.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
---

Notes:
    v2:
      - Dropped flags and rely on the irq_data being locked with interrupts
        disabled
    
      - Add extra documentation to explain why taking the vmapp_lock is
        dependent on the HW requiring the use of the ITS list
    
    v1: https://lore.kernel.org/r/20240723175203.3193882-1-maz@kernel.org

 drivers/irqchip/irq-gic-v3-its.c | 16 ++++++++++------
 1 file changed, 10 insertions(+), 6 deletions(-)

diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c
index 9b34596b3542..fdec478ba5e7 100644
--- a/drivers/irqchip/irq-gic-v3-its.c
+++ b/drivers/irqchip/irq-gic-v3-its.c
@@ -1329,12 +1329,6 @@ static void its_send_vmovp(struct its_vpe *vpe)
 		return;
 	}
 
-	/*
-	 * Protect against concurrent updates of the mapping state on
-	 * individual VMs.
-	 */
-	guard(raw_spinlock_irqsave)(&vpe->its_vm->vmapp_lock);
-
 	/*
 	 * Yet another marvel of the architecture. If using the
 	 * its_list "feature", we need to make sure that all ITSs
@@ -3824,7 +3818,14 @@ static int its_vpe_set_affinity(struct irq_data *d,
 	 * protect us, and that we must ensure nobody samples vpe->col_idx
 	 * during the update, hence the lock below which must also be
 	 * taken on any vLPI handling path that evaluates vpe->col_idx.
+	 *
+	 * Finally, we must protect ourselves against concurrent updates of
+	 * the mapping state on this VM should the ITS list be in use (see
+	 * the shortcut in its_send_vmovp() otherewise).
 	 */
+	if (its_list_map)
+		raw_spin_lock(&vpe->its_vm->vmapp_lock);
+
 	from = vpe_to_cpuid_lock(vpe, &flags);
 	table_mask = gic_data_rdist_cpu(from)->vpe_table_mask;
 
@@ -3854,6 +3855,9 @@ static int its_vpe_set_affinity(struct irq_data *d,
 	irq_data_update_effective_affinity(d, cpumask_of(cpu));
 	vpe_to_cpuid_unlock(vpe, flags);
 
+	if (its_list_map)
+		raw_spin_unlock(&vpe->its_vm->vmapp_lock);
+
 	return IRQ_SET_MASK_OK_DONE;
 }
 
-- 
2.39.2

The following commit has been merged into the irq/urgent branch of tip:

Commit-ID:     f97fd458763a4801d04dbb4a79d9ca6282d293ec
Gitweb:        https://git.kernel.org/tip/f97fd458763a4801d04dbb4a79d9ca6282d293ec
Author:        Marc Zyngier <maz@kernel.org>
AuthorDate:    Sun, 18 Aug 2024 18:16:25 +01:00
Committer:     Thomas Gleixner <tglx@linutronix.de>
CommitterDate: Tue, 20 Aug 2024 16:57:13 +02:00

irqchip/gic-v4: Fix ordering between vmapp and vpe locks

The recently established lock ordering mandates that the per-VM
vmapp_lock is acquired before taking the per-VPE lock.

As it turns out, its_vpe_set_affinity() takes the VPE lock, and
then calls into its_send_vmovp(), which itself takes the vmapp
lock. Obviously, this is a lock order violation.

As its_send_vmovp() is only called from its_vpe_set_affinity(),
hoist the vmapp locking from the former into the latter, restoring
the expected order.

Fixes: f0eb154c39471 ("irqchip/gic-v4: Substitute vmovp_lock for a per-VM lock")
Reported-by: Zhou Wang <wangzhou1@hisilicon.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lore.kernel.org/all/20240818171625.3030584-1-maz@kernel.org
---
 drivers/irqchip/irq-gic-v3-its.c | 16 ++++++++++------
 1 file changed, 10 insertions(+), 6 deletions(-)

diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c
index 9b34596..fdec478 100644
--- a/drivers/irqchip/irq-gic-v3-its.c
+++ b/drivers/irqchip/irq-gic-v3-its.c
@@ -1330,12 +1330,6 @@ static void its_send_vmovp(struct its_vpe *vpe)
 	}
 
 	/*
-	 * Protect against concurrent updates of the mapping state on
-	 * individual VMs.
-	 */
-	guard(raw_spinlock_irqsave)(&vpe->its_vm->vmapp_lock);
-
-	/*
 	 * Yet another marvel of the architecture. If using the
 	 * its_list "feature", we need to make sure that all ITSs
 	 * receive all VMOVP commands in the same order. The only way
@@ -3824,7 +3818,14 @@ static int its_vpe_set_affinity(struct irq_data *d,
 	 * protect us, and that we must ensure nobody samples vpe->col_idx
 	 * during the update, hence the lock below which must also be
 	 * taken on any vLPI handling path that evaluates vpe->col_idx.
+	 *
+	 * Finally, we must protect ourselves against concurrent updates of
+	 * the mapping state on this VM should the ITS list be in use (see
+	 * the shortcut in its_send_vmovp() otherewise).
 	 */
+	if (its_list_map)
+		raw_spin_lock(&vpe->its_vm->vmapp_lock);
+
 	from = vpe_to_cpuid_lock(vpe, &flags);
 	table_mask = gic_data_rdist_cpu(from)->vpe_table_mask;
 
@@ -3854,6 +3855,9 @@ out:
 	irq_data_update_effective_affinity(d, cpumask_of(cpu));
 	vpe_to_cpuid_unlock(vpe, flags);
 
+	if (its_list_map)
+		raw_spin_unlock(&vpe->its_vm->vmapp_lock);
+
 	return IRQ_SET_MASK_OK_DONE;
 }