When we're about to run a vcpu, it is crucial that the redistributor
associated with the physical CPU is being told about the new residency.
This is abstracted by hijacking the irq_set_affinity method for the
doorbell interrupt associated with the VPE. It is expected that the
hypervisor will call this method before scheduling the VPE.
Reviewed-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
Signed-off-by: Marc Zyngier <marc.zyngier@xxxxxxx>
---
drivers/irqchip/irq-gic-v3-its.c | 96 ++++++++++++++++++++++++++++++++++++++++
1 file changed, 96 insertions(+)
diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c
index 16cdd1f60ebf..0a9aedaf6da2 100644
--- a/drivers/irqchip/irq-gic-v3-its.c
+++ b/drivers/irqchip/irq-gic-v3-its.c
@@ -148,6 +148,9 @@ static struct irq_domain *its_parent;
#define ITS_LIST_MAX 16
static unsigned long its_list_map;
+static u16 vmovp_seq_num;
+static DEFINE_RAW_SPINLOCK(vmovp_lock);
+
static DEFINE_IDA(its_vpeid_ida);
#define gic_data_rdist() (raw_cpu_ptr(gic_rdists->rdist))
@@ -238,6 +241,13 @@ struct its_cmd_desc {
u32 event_id;
bool db_enabled;
} its_vmovi_cmd;
+
+ struct {
+ struct its_vpe *vpe;
+ struct its_collection *col;
+ u16 seq_num;
+ u16 its_list;
+ } its_vmovp_cmd;
};
};
@@ -329,6 +339,16 @@ static void its_encode_db_valid(struct its_cmd_block *cmd, bool db_valid)
its_mask_encode(&cmd->raw_cmd[2], db_valid, 0, 0);
}
+static void its_encode_seq_num(struct its_cmd_block *cmd, u16 seq_num)
+{
+ its_mask_encode(&cmd->raw_cmd[0], seq_num, 47, 32);
+}
+
+static void its_encode_its_list(struct its_cmd_block *cmd, u16 its_list)
+{
+ its_mask_encode(&cmd->raw_cmd[1], its_list, 15, 0);
+}
+
static void its_encode_vpt_addr(struct its_cmd_block *cmd, u64 vpt_pa)
{
its_mask_encode(&cmd->raw_cmd[3], vpt_pa >> 16, 50, 16);
@@ -571,6 +591,20 @@ static struct its_vpe *its_build_vmovi_cmd(struct its_cmd_block *cmd,
return desc->its_vmovi_cmd.vpe;
}
+static struct its_vpe *its_build_vmovp_cmd(struct its_cmd_block *cmd,
+ struct its_cmd_desc *desc)
+{
+ its_encode_cmd(cmd, GITS_CMD_VMOVP);
+ its_encode_seq_num(cmd, desc->its_vmovp_cmd.seq_num);
+ its_encode_its_list(cmd, desc->its_vmovp_cmd.its_list);
+ its_encode_vpeid(cmd, desc->its_vmovp_cmd.vpe->vpe_id);
+ its_encode_target(cmd, desc->its_vmovp_cmd.col->target_address);
+
+ its_fixup_cmd(cmd);
+
+ return desc->its_vmovp_cmd.vpe;
+}
+
static u64 its_cmd_ptr_to_offset(struct its_node *its,
struct its_cmd_block *ptr)
{
@@ -865,6 +899,48 @@ static void its_send_vmapp(struct its_vpe *vpe, bool valid)
}
}
+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;
+ desc.its_vmovp_cmd.its_list = (u16)its_list_map;
+
+ if (!its_list_map) {
+ its = list_first_entry(&its_nodes, struct its_node, entry);
+ desc.its_vmovp_cmd.seq_num = 0;
+ desc.its_vmovp_cmd.col = &its->collections[col_id];
+ its_send_single_vcommand(its, its_build_vmovp_cmd, &desc);
+ return;
+ }
+
+ /*
+ * 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
+ * to guarantee this is to make vmovp a serialization point.
+ *
+ * Wall <-- Head.
+ */
+ raw_spin_lock_irqsave(&vmovp_lock, flags);
+
+ desc.its_vmovp_cmd.seq_num = vmovp_seq_num++;
+
+ /* Emit VMOVPs */
+ list_for_each_entry(its, &its_nodes, entry) {
+ if (!its->is_v4)
+ continue;
+
+ 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_vpe *vpe)
{
struct its_cmd_desc desc;
@@ -2148,6 +2224,25 @@ static const struct irq_domain_ops its_domain_ops = {
.deactivate = its_irq_domain_deactivate,
};
+static int its_vpe_set_affinity(struct irq_data *d,
+ const struct cpumask *mask_val,
+ bool force)
+{
+ struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
+ int cpu = cpumask_first(mask_val);
+
+ /*
+ * Changing affinity is mega expensive, so let's be as lazy as
+ * we can and only do it if we really have to.
+ */
+ if (vpe->col_idx != cpu) {
+ vpe->col_idx = cpu;
+ its_send_vmovp(vpe);
+ }
+
+ return IRQ_SET_MASK_OK_DONE;
+}
+
static void its_vpe_schedule(struct its_vpe *vpe)
{
void * __iomem vlpi_base = gic_data_rdist_vlpi_base();
@@ -2237,6 +2332,7 @@ static int its_vpe_set_vcpu_affinity(struct irq_data *d, void *vcpu_info)
static struct irq_chip its_vpe_irq_chip = {
.name = "GICv4-vpe",
+ .irq_set_affinity = its_vpe_set_affinity,
.irq_set_vcpu_affinity = its_vpe_set_vcpu_affinity,
};
--
2.11.0
