irqchip/gic-v4: Substitute vmovp_lock for a per-VM lock

vmovp_lock is abused in a number of cases to serialise updates
to vlpi_count[] and deal with map/unmap of a VM to ITSs.

Instead, provide a per-VM lock and revisit the use of vlpi_count[]
so that it is always wrapped in this per-VM vmapp_lock.

This reduces the potential contention on a concurrent VMOVP command,
and paves the way for subsequent VPE locking that holding vmovp_lock
actively prevents due to the lock ordering.

Signed-off-by: Marc Zyngier <[email protected]>
Signed-off-by: Thomas Gleixner <[email protected]>
Tested-by: Nianyao Tang <[email protected]>
Link: https://lore.kernel.org/r/[email protected]

Author: Marc Zyngier

drivers/irqchip/irq-gic-v3-its.c

@@ -1317,7 +1317,6 @@ static void its_send_vmovp(struct its_vpe *vpe)
 {
 	struct its_cmd_desc desc = {};
 	struct its_node *its;
-	unsigned long flags;
 	int col_id = vpe->col_idx;
 
 	desc.its_vmovp_cmd.vpe = vpe;
@@ -1329,6 +1328,12 @@ 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
@@ -1337,8 +1342,7 @@ static void its_send_vmovp(struct its_vpe *vpe)
 	 *
 	 * Wall <-- Head.
 	 */
-	raw_spin_lock_irqsave(&vmovp_lock, flags);
-
+	guard(raw_spinlock)(&vmovp_lock);
 	desc.its_vmovp_cmd.seq_num = vmovp_seq_num++;
 	desc.its_vmovp_cmd.its_list = get_its_list(vpe->its_vm);
 
@@ -1353,8 +1357,6 @@ static void its_send_vmovp(struct its_vpe *vpe)
 		desc.its_vmovp_cmd.col = &its->collections[col_id];
 		its_send_single_vcommand(its, its_build_vmovp_cmd, &desc);
 	}
-
-	raw_spin_unlock_irqrestore(&vmovp_lock, flags);
 }
 
 static void its_send_vinvall(struct its_node *its, struct its_vpe *vpe)
@@ -1791,12 +1793,10 @@ static bool gic_requires_eager_mapping(void)
 
 static void its_map_vm(struct its_node *its, struct its_vm *vm)
 {
-	unsigned long flags;
-
 	if (gic_requires_eager_mapping())
 		return;
 
-	raw_spin_lock_irqsave(&vmovp_lock, flags);
+	guard(raw_spinlock_irqsave)(&vm->vmapp_lock);
 
 	/*
 	 * If the VM wasn't mapped yet, iterate over the vpes and get
@@ -1814,28 +1814,22 @@ static void its_map_vm(struct its_node *its, struct its_vm *vm)
 			its_send_vinvall(its, vpe);
 		}
 	}
-
-	raw_spin_unlock_irqrestore(&vmovp_lock, flags);
 }
 
 static void its_unmap_vm(struct its_node *its, struct its_vm *vm)
 {
-	unsigned long flags;
-
 	/* Not using the ITS list? Everything is always mapped. */
 	if (gic_requires_eager_mapping())
 		return;
 
-	raw_spin_lock_irqsave(&vmovp_lock, flags);
+	guard(raw_spinlock_irqsave)(&vm->vmapp_lock);
 
 	if (!--vm->vlpi_count[its->list_nr]) {
 		int i;
 
 		for (i = 0; i < vm->nr_vpes; i++)
 			its_send_vmapp(its, vm->vpes[i], false);
 	}
-
-	raw_spin_unlock_irqrestore(&vmovp_lock, flags);
 }
 
 static int its_vlpi_map(struct irq_data *d, struct its_cmd_info *info)
@@ -3942,6 +3936,8 @@ static void its_vpe_invall(struct its_vpe *vpe)
 {
 	struct its_node *its;
 
+	guard(raw_spinlock_irqsave)(&vpe->its_vm->vmapp_lock);
+
 	list_for_each_entry(its, &its_nodes, entry) {
 		if (!is_v4(its))
 			continue;
@@ -4547,6 +4543,7 @@ static int its_vpe_irq_domain_alloc(struct irq_domain *domain, unsigned int virq
 	vm->db_lpi_base = base;
 	vm->nr_db_lpis = nr_ids;
 	vm->vprop_page = vprop_page;
+	raw_spin_lock_init(&vm->vmapp_lock);
 
 	if (gic_rdists->has_rvpeid)
 		irqchip = &its_vpe_4_1_irq_chip;

include/linux/irqchip/arm-gic-v4.h

@@ -25,6 +25,14 @@ struct its_vm {
 	irq_hw_number_t		db_lpi_base;
 	unsigned long		*db_bitmap;
 	int			nr_db_lpis;
+	/*
+	 * Ensures mutual exclusion between updates to vlpi_count[]
+	 * and map/unmap when using the ITSList mechanism.
+	 *
+	 * The lock order for any sequence involving the ITSList is
+	 * vmapp_lock -> vpe_lock ->vmovp_lock.
+	 */
+	raw_spinlock_t		vmapp_lock;
 	u32			vlpi_count[GICv4_ITS_LIST_MAX];
 };