diff options
| author | Nianyao Tang <tangnianyao@huawei.com> | 2025-05-15 14:53:59 +0000 |
|---|---|---|
| committer | Thomas Gleixner <tglx@linutronix.de> | 2025-05-15 16:19:22 +0200 |
| commit | f1a3fac4095c7bc0b30e2aa9921c232af8faeae0 (patch) | |
| tree | cf71eade655ea3a57df62ccd42064e7e6786707c | |
| parent | 96a8cb6d28cebd51a286fe4a8782dc8492813ac4 (diff) | |
irqchip/gic-v4.1: Use local 4_1 ITS to generate VSGI
On multi-node GICv4.1 system, VSGI senders always use one certain 4_1 ITS,
because find_4_1_its() returns the first its_node in the list, regardless of
which node the VSGI sender is on. This brings guest VSGI performance drop
when VM is not running on the same node as this returned ITS.
On a 2-socket environment, each with one ITS and 32 cpu, GICv4.1 enabled,
4U8G guest, 4 vcpu is running on same socket.
When the VM is on socket0, kvm-unit-tests ipi_hw result is 850ns.
When the VM is on socket1, it is 750ns.
The reason is that the VSGI sender always uses the last reported ITS (that
on socket1) to inject VSGI. The access from a CPU to a other-socket ITS
will cost 100ns more compared to an access to the local ITS.
Using the local ITS results in a 12% reduction in IPI latency.
Modify find_4_1_its() to return the first per-CPU local_4_1_its, which is
initialized when the VPE table is inherited from the ITS or from another
CPU. If it fails to find a local 4_1 ITS, it returns any 4_1 ITS like
before.
Suggested-by: Marc Zyngier <maz@kernel.org>
Signed-off-by: Nianyao Tang <tangnianyao@huawei.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/all/20250515145359.2795959-1-tangnianyao@huawei.com
| -rw-r--r-- | drivers/irqchip/irq-gic-v3-its.c | 6 |
1 files changed, 5 insertions, 1 deletions
diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c index 0115ad6c8259..81a44ce7619f 100644 --- a/drivers/irqchip/irq-gic-v3-its.c +++ b/drivers/irqchip/irq-gic-v3-its.c @@ -125,6 +125,8 @@ struct its_node { int vlpi_redist_offset; }; +static DEFINE_PER_CPU(struct its_node *, local_4_1_its); + #define is_v4(its) (!!((its)->typer & GITS_TYPER_VLPIS)) #define is_v4_1(its) (!!((its)->typer & GITS_TYPER_VMAPP)) #define device_ids(its) (FIELD_GET(GITS_TYPER_DEVBITS, (its)->typer) + 1) @@ -2778,6 +2780,7 @@ static u64 inherit_vpe_l1_table_from_its(void) } val |= FIELD_PREP(GICR_VPROPBASER_4_1_SIZE, GITS_BASER_NR_PAGES(baser) - 1); + *this_cpu_ptr(&local_4_1_its) = its; return val; } @@ -2815,6 +2818,7 @@ static u64 inherit_vpe_l1_table_from_rd(cpumask_t **mask) gic_data_rdist()->vpe_l1_base = gic_data_rdist_cpu(cpu)->vpe_l1_base; *mask = gic_data_rdist_cpu(cpu)->vpe_table_mask; + *this_cpu_ptr(&local_4_1_its) = *per_cpu_ptr(&local_4_1_its, cpu); return val; } @@ -4180,7 +4184,7 @@ static struct irq_chip its_vpe_irq_chip = { static struct its_node *find_4_1_its(void) { - static struct its_node *its = NULL; + struct its_node *its = *this_cpu_ptr(&local_4_1_its); if (!its) { list_for_each_entry(its, &its_nodes, entry) { |