Message ID | 20230327-irq-affinity-upstream-v2-2-1474e518f1cb@ti.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | irqchip: irq-ti-sci-inta: Add initial IRQ affinity support | expand |
On Mon, 27 Mar 2023 16:04:27 +0100, Vignesh Raghavendra <vigneshr@ti.com> wrote: > > Certain high throughput peripherals such as ethernet etc may need > dedicated IRQ lines so the IRQ load can be balanced across cores using > IRQ affinity controls. Current driver aggregates multiple events/IRQ to > single wired IRQ line/VINT thus making it impossible to migrate events > selectively. > In order to overcome this limitation, reserve a set of VINTs as direct > interrupts and map them to events that are known to generate high IRQ > load. Use SoC specific table to determine such events. > > This allows affinity management at parent irqchip level (GIC or > ti-sci-intr + GIC). > > Signed-off-by: Vignesh Raghavendra <vigneshr@ti.com> > --- > drivers/irqchip/irq-ti-sci-inta.c | 149 +++++++++- > 1 file changed, 147 insertions(+), 2 deletions(-) > > diff --git a/drivers/irqchip/irq-ti-sci-inta.c b/drivers/irqchip/irq-ti-sci-inta.c > index a3a22edbe0f0..db098bd95edc 100644 > --- a/drivers/irqchip/irq-ti-sci-inta.c > +++ b/drivers/irqchip/irq-ti-sci-inta.c > @@ -21,6 +21,7 @@ > #include <linux/irqchip/chained_irq.h> > #include <linux/soc/ti/ti_sci_inta_msi.h> > #include <linux/soc/ti/ti_sci_protocol.h> > +#include <linux/sys_soc.h> > #include <asm-generic/msi.h> > > #define TI_SCI_DEV_ID_MASK 0xffff > @@ -40,6 +41,20 @@ > #define VINT_STATUS_OFFSET 0x18 > #define VINT_STATUS_MASKED_OFFSET 0x20 > > +#define DEV_DMASS0_INTAGGR_0 28 > +#define DEV_DMASS0_PKTDMA_0 30 > + > +/** > + * struct ti_sci_inta_soc_data - Description of SoC specific data > + * @events_list: Pointer to array of events requiring direct IRQ > + * mapping > + * @events_list_size: size of event_list array > + */ > +struct ti_sci_inta_soc_data { > + unsigned int *events_list; > + int events_list_size; > +}; > + > /** > * struct ti_sci_inta_event_desc - Description of an event coming to > * Interrupt Aggregator. This serves > @@ -64,6 +79,7 @@ struct ti_sci_inta_event_desc { > * @events: Array of event descriptors assigned to this vint. > * @parent_virq: Linux IRQ number that gets attached to parent > * @vint_id: TISCI vint ID > + * @reserved: Indicate if this VINT is reserved for direct mapping > */ > struct ti_sci_inta_vint_desc { > struct irq_domain *domain; > @@ -72,6 +88,7 @@ struct ti_sci_inta_vint_desc { > struct ti_sci_inta_event_desc events[MAX_EVENTS_PER_VINT]; > unsigned int parent_virq; > u16 vint_id; > + bool reserved; > }; > > /** > @@ -81,6 +98,8 @@ struct ti_sci_inta_vint_desc { > * @vint: TISCI resource pointer representing IA interrupts. > * @global_event: TISCI resource pointer representing global events. > * @vint_list: List of the vints active in the system > + * @resv_vint_list: List of the direct vints active in the system. > + * @resv_events: Pointer to list of direct events > * @vint_mutex: Mutex to protect vint_list > * @base: Base address of the memory mapped IO registers > * @pdev: Pointer to platform device. > @@ -97,12 +116,15 @@ struct ti_sci_inta_vint_desc { > * identifier to let sysfw know where it has to program the > * Global Event number. > * @num_vints: Number of VINT lines allocated to this IA. > + * @num_resv_vints: Number of VINT lines reserved for direct mapping. > */ > struct ti_sci_inta_irq_domain { > const struct ti_sci_handle *sci; > struct ti_sci_resource *vint; > struct ti_sci_resource *global_event; > struct list_head vint_list; > + struct list_head resv_vint_list; > + unsigned int *resv_events; > /* Mutex to protect vint list */ > struct mutex vint_mutex; > void __iomem *base; > @@ -112,6 +134,7 @@ struct ti_sci_inta_irq_domain { > int unmapped_cnt; > u16 *unmapped_dev_ids; > int num_vints; > + int num_resv_vints; > }; > > #define to_vint_desc(e, i) container_of(e, struct ti_sci_inta_vint_desc, \ > @@ -265,7 +288,11 @@ static int ti_sci_inta_alloc_parent_irqs(struct irq_domain *domain) > } > vint_desc->parent_virq = parent_virq; > > - list_add_tail(&vint_desc->list, &inta->vint_list); > + if (i < inta->num_resv_vints) > + list_add_tail(&vint_desc->list, &inta->resv_vint_list); > + else > + list_add_tail(&vint_desc->list, &inta->vint_list); > + > irq_set_chained_handler_and_data(vint_desc->parent_virq, > ti_sci_inta_irq_handler, vint_desc); > } > @@ -390,8 +417,58 @@ static void ti_sci_inta_free_irq(struct ti_sci_inta_event_desc *event_desc, > ti_sci_release_resource(inta->global_event, event_desc->global_event); > event_desc->global_event = TI_SCI_RESOURCE_NULL; > event_desc->hwirq = 0; > + vint_desc->reserved = false; > +} > + > +/** > + * ti_sci_inta_alloc_direct_irq - Allocate non-aggregated events/IRQs > + * @data: Pointer to corresponding irq_data > + * > + * Returns allocated event_desc, ERR_PTR otherwise. > + */ > +static struct ti_sci_inta_event_desc *ti_sci_inta_alloc_direct_irq(struct irq_data *data) > +{ > + struct ti_sci_inta_irq_domain *inta = data->domain->host_data; > + struct ti_sci_inta_event_desc *event_desc = NULL; > + struct ti_sci_inta_vint_desc *vint_desc; > + > + mutex_lock(&inta->vint_mutex); > + list_for_each_entry(vint_desc, &inta->resv_vint_list, list) { > + if (vint_desc->reserved) > + continue; > + > + vint_desc->reserved = true; > + event_desc = ti_sci_inta_alloc_event(vint_desc, 0, data->hwirq); > + if (IS_ERR(event_desc)) > + goto unlock; > > + data->parent_data = irq_get_irq_data(vint_desc->parent_virq); > + break; > + } > +unlock: > mutex_unlock(&inta->vint_mutex); > + > + return event_desc; > +} > + > +/** > + * ti_sci_inta_direct_irq - Allocate non-aggregated events/IRQs > + * @data: Pointer to corresponding irq_data > + * > + * Returns true if this IRQ is to be direct mapped on given platform, > + * false otherwise. > + */ > +static bool ti_sci_inta_direct_irq(struct irq_data *data) > +{ > + struct ti_sci_inta_irq_domain *inta = data->domain->host_data; > + int i; > + > + for (i = 0; i < inta->num_resv_vints; i++) { > + if (data->hwirq == inta->resv_events[i]) > + return true; > + } > + > + return false; > } > > /** > @@ -408,7 +485,11 @@ static int ti_sci_inta_request_resources(struct irq_data *data) > { > struct ti_sci_inta_event_desc *event_desc; > > - event_desc = ti_sci_inta_alloc_irq(data->domain, data->hwirq); > + if (ti_sci_inta_direct_irq(data)) > + event_desc = ti_sci_inta_alloc_direct_irq(data); > + else > + event_desc = ti_sci_inta_alloc_irq(data->domain, data->hwirq); > + > if (IS_ERR(event_desc)) > return PTR_ERR(event_desc); > > @@ -485,6 +566,15 @@ static void ti_sci_inta_ack_irq(struct irq_data *data) > static int ti_sci_inta_set_affinity(struct irq_data *d, > const struct cpumask *mask_val, bool force) > { > + int ret; > + > + if (d->parent_data) { > + ret = irq_chip_set_affinity_parent(d, mask_val, force); > + irq_data_update_effective_affinity(d, mask_val); > + > + return ret; > + } > + > return -EINVAL; > } > > @@ -627,9 +717,54 @@ static int ti_sci_inta_get_unmapped_sources(struct ti_sci_inta_irq_domain *inta) > return 0; > } > > +static unsigned int ti_sci_inta_direct_events_am62x[] = { > + /* CPSW etherenti DMA events */ > + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4627), > + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4635), > + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4643), > + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4651), > + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4659), > + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4667), > + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4675), > + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4683), > + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 5651), > +}; > + > +static struct ti_sci_inta_soc_data soc_data_am62x = { > + .events_list = ti_sci_inta_direct_events_am62x, > + .events_list_size = ARRAY_SIZE(ti_sci_inta_direct_events_am62x), > +}; I don't think these tables belong in a driver, and they are bound to grow without any obvious limits. You have firmware tables that can express these things. Surely they can be put to a good use. M.
Hi, On 4/8/2023 4:10 PM, Marc Zyngier wrote: >> +static unsigned int ti_sci_inta_direct_events_am62x[] = { >> + /* CPSW etherenti DMA events */ >> + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4627), >> + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4635), >> + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4643), >> + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4651), >> + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4659), >> + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4667), >> + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4675), >> + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4683), >> + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 5651), >> +}; >> + >> +static struct ti_sci_inta_soc_data soc_data_am62x = { >> + .events_list = ti_sci_inta_direct_events_am62x, >> + .events_list_size = ARRAY_SIZE(ti_sci_inta_direct_events_am62x), >> +}; > I don't think these tables belong in a driver, and they are bound to > grow without any obvious limits. Fair point. > You have firmware tables that can express these things. Surely they can be put to a good use. By firmware tables you mean device tree? Regards Vignesh
On 2023-04-08 12:27, Raghavendra, Vignesh wrote: > Hi, > > On 4/8/2023 4:10 PM, Marc Zyngier wrote: >>> +static unsigned int ti_sci_inta_direct_events_am62x[] = { >>> + /* CPSW etherenti DMA events */ >>> + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4627), >>> + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4635), >>> + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4643), >>> + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4651), >>> + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4659), >>> + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4667), >>> + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4675), >>> + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4683), >>> + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 5651), >>> +}; >>> + >>> +static struct ti_sci_inta_soc_data soc_data_am62x = { >>> + .events_list = ti_sci_inta_direct_events_am62x, >>> + .events_list_size = ARRAY_SIZE(ti_sci_inta_direct_events_am62x), >>> +}; >> I don't think these tables belong in a driver, and they are bound to >> grow without any obvious limits. > > Fair point. > >> You have firmware tables that can express these things. Surely they >> can be put to a good use. > > By firmware tables you mean device tree? That, or any other machine-specific mean. From what I get of these systems, they already make heavy use of some runtime firmware to get things configured. That side could also provide setup information. I don't mind either way, as long as we don't end-up with forever growing in-kernel tables that are just board files in disguise. M.
diff --git a/drivers/irqchip/irq-ti-sci-inta.c b/drivers/irqchip/irq-ti-sci-inta.c index a3a22edbe0f0..db098bd95edc 100644 --- a/drivers/irqchip/irq-ti-sci-inta.c +++ b/drivers/irqchip/irq-ti-sci-inta.c @@ -21,6 +21,7 @@ #include <linux/irqchip/chained_irq.h> #include <linux/soc/ti/ti_sci_inta_msi.h> #include <linux/soc/ti/ti_sci_protocol.h> +#include <linux/sys_soc.h> #include <asm-generic/msi.h> #define TI_SCI_DEV_ID_MASK 0xffff @@ -40,6 +41,20 @@ #define VINT_STATUS_OFFSET 0x18 #define VINT_STATUS_MASKED_OFFSET 0x20 +#define DEV_DMASS0_INTAGGR_0 28 +#define DEV_DMASS0_PKTDMA_0 30 + +/** + * struct ti_sci_inta_soc_data - Description of SoC specific data + * @events_list: Pointer to array of events requiring direct IRQ + * mapping + * @events_list_size: size of event_list array + */ +struct ti_sci_inta_soc_data { + unsigned int *events_list; + int events_list_size; +}; + /** * struct ti_sci_inta_event_desc - Description of an event coming to * Interrupt Aggregator. This serves @@ -64,6 +79,7 @@ struct ti_sci_inta_event_desc { * @events: Array of event descriptors assigned to this vint. * @parent_virq: Linux IRQ number that gets attached to parent * @vint_id: TISCI vint ID + * @reserved: Indicate if this VINT is reserved for direct mapping */ struct ti_sci_inta_vint_desc { struct irq_domain *domain; @@ -72,6 +88,7 @@ struct ti_sci_inta_vint_desc { struct ti_sci_inta_event_desc events[MAX_EVENTS_PER_VINT]; unsigned int parent_virq; u16 vint_id; + bool reserved; }; /** @@ -81,6 +98,8 @@ struct ti_sci_inta_vint_desc { * @vint: TISCI resource pointer representing IA interrupts. * @global_event: TISCI resource pointer representing global events. * @vint_list: List of the vints active in the system + * @resv_vint_list: List of the direct vints active in the system. + * @resv_events: Pointer to list of direct events * @vint_mutex: Mutex to protect vint_list * @base: Base address of the memory mapped IO registers * @pdev: Pointer to platform device. @@ -97,12 +116,15 @@ struct ti_sci_inta_vint_desc { * identifier to let sysfw know where it has to program the * Global Event number. * @num_vints: Number of VINT lines allocated to this IA. + * @num_resv_vints: Number of VINT lines reserved for direct mapping. */ struct ti_sci_inta_irq_domain { const struct ti_sci_handle *sci; struct ti_sci_resource *vint; struct ti_sci_resource *global_event; struct list_head vint_list; + struct list_head resv_vint_list; + unsigned int *resv_events; /* Mutex to protect vint list */ struct mutex vint_mutex; void __iomem *base; @@ -112,6 +134,7 @@ struct ti_sci_inta_irq_domain { int unmapped_cnt; u16 *unmapped_dev_ids; int num_vints; + int num_resv_vints; }; #define to_vint_desc(e, i) container_of(e, struct ti_sci_inta_vint_desc, \ @@ -265,7 +288,11 @@ static int ti_sci_inta_alloc_parent_irqs(struct irq_domain *domain) } vint_desc->parent_virq = parent_virq; - list_add_tail(&vint_desc->list, &inta->vint_list); + if (i < inta->num_resv_vints) + list_add_tail(&vint_desc->list, &inta->resv_vint_list); + else + list_add_tail(&vint_desc->list, &inta->vint_list); + irq_set_chained_handler_and_data(vint_desc->parent_virq, ti_sci_inta_irq_handler, vint_desc); } @@ -390,8 +417,58 @@ static void ti_sci_inta_free_irq(struct ti_sci_inta_event_desc *event_desc, ti_sci_release_resource(inta->global_event, event_desc->global_event); event_desc->global_event = TI_SCI_RESOURCE_NULL; event_desc->hwirq = 0; + vint_desc->reserved = false; +} + +/** + * ti_sci_inta_alloc_direct_irq - Allocate non-aggregated events/IRQs + * @data: Pointer to corresponding irq_data + * + * Returns allocated event_desc, ERR_PTR otherwise. + */ +static struct ti_sci_inta_event_desc *ti_sci_inta_alloc_direct_irq(struct irq_data *data) +{ + struct ti_sci_inta_irq_domain *inta = data->domain->host_data; + struct ti_sci_inta_event_desc *event_desc = NULL; + struct ti_sci_inta_vint_desc *vint_desc; + + mutex_lock(&inta->vint_mutex); + list_for_each_entry(vint_desc, &inta->resv_vint_list, list) { + if (vint_desc->reserved) + continue; + + vint_desc->reserved = true; + event_desc = ti_sci_inta_alloc_event(vint_desc, 0, data->hwirq); + if (IS_ERR(event_desc)) + goto unlock; + data->parent_data = irq_get_irq_data(vint_desc->parent_virq); + break; + } +unlock: mutex_unlock(&inta->vint_mutex); + + return event_desc; +} + +/** + * ti_sci_inta_direct_irq - Allocate non-aggregated events/IRQs + * @data: Pointer to corresponding irq_data + * + * Returns true if this IRQ is to be direct mapped on given platform, + * false otherwise. + */ +static bool ti_sci_inta_direct_irq(struct irq_data *data) +{ + struct ti_sci_inta_irq_domain *inta = data->domain->host_data; + int i; + + for (i = 0; i < inta->num_resv_vints; i++) { + if (data->hwirq == inta->resv_events[i]) + return true; + } + + return false; } /** @@ -408,7 +485,11 @@ static int ti_sci_inta_request_resources(struct irq_data *data) { struct ti_sci_inta_event_desc *event_desc; - event_desc = ti_sci_inta_alloc_irq(data->domain, data->hwirq); + if (ti_sci_inta_direct_irq(data)) + event_desc = ti_sci_inta_alloc_direct_irq(data); + else + event_desc = ti_sci_inta_alloc_irq(data->domain, data->hwirq); + if (IS_ERR(event_desc)) return PTR_ERR(event_desc); @@ -485,6 +566,15 @@ static void ti_sci_inta_ack_irq(struct irq_data *data) static int ti_sci_inta_set_affinity(struct irq_data *d, const struct cpumask *mask_val, bool force) { + int ret; + + if (d->parent_data) { + ret = irq_chip_set_affinity_parent(d, mask_val, force); + irq_data_update_effective_affinity(d, mask_val); + + return ret; + } + return -EINVAL; } @@ -627,9 +717,54 @@ static int ti_sci_inta_get_unmapped_sources(struct ti_sci_inta_irq_domain *inta) return 0; } +static unsigned int ti_sci_inta_direct_events_am62x[] = { + /* CPSW etherenti DMA events */ + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4627), + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4635), + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4643), + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4651), + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4659), + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4667), + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4675), + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4683), + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 5651), +}; + +static struct ti_sci_inta_soc_data soc_data_am62x = { + .events_list = ti_sci_inta_direct_events_am62x, + .events_list_size = ARRAY_SIZE(ti_sci_inta_direct_events_am62x), +}; + +static unsigned int ti_sci_inta_direct_events_am64x[] = { + /* CPSW etherenti DMA events */ + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4624), + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4632), + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4640), + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4648), + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4656), + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4664), + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4672), + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 4680), + TO_HWIRQ(DEV_DMASS0_PKTDMA_0, 5648), +}; + +static struct ti_sci_inta_soc_data soc_data_am64x = { + .events_list = ti_sci_inta_direct_events_am64x, + .events_list_size = ARRAY_SIZE(ti_sci_inta_direct_events_am64x), +}; + +static const struct soc_device_attribute k3_soc_devices[] = { + { .family = "AM64X", .data = &soc_data_am64x }, + { .family = "AM62X", .data = &soc_data_am62x }, + { .family = "AM62AX", .data = &soc_data_am62x }, + { /* sentinel */ } +}; + static int ti_sci_inta_irq_domain_probe(struct platform_device *pdev) { struct irq_domain *parent_domain, *domain, *msi_domain; + const struct ti_sci_inta_soc_data *soc_data; + const struct soc_device_attribute *soc; struct device_node *parent_node, *node; struct ti_sci_inta_irq_domain *inta; struct device *dev = &pdev->dev; @@ -703,8 +838,18 @@ static int ti_sci_inta_irq_domain_probe(struct platform_device *pdev) } INIT_LIST_HEAD(&inta->vint_list); + INIT_LIST_HEAD(&inta->resv_vint_list); mutex_init(&inta->vint_mutex); + soc = soc_device_match(k3_soc_devices); + if (soc) { + soc_data = soc->data; + if (inta->num_vints > soc_data->events_list_size) { + inta->num_resv_vints = soc_data->events_list_size; + inta->resv_events = soc_data->events_list; + } + } + ret = ti_sci_inta_alloc_parent_irqs(domain); if (ret) { irq_domain_remove(msi_domain);
Certain high throughput peripherals such as ethernet etc may need dedicated IRQ lines so the IRQ load can be balanced across cores using IRQ affinity controls. Current driver aggregates multiple events/IRQ to single wired IRQ line/VINT thus making it impossible to migrate events selectively. In order to overcome this limitation, reserve a set of VINTs as direct interrupts and map them to events that are known to generate high IRQ load. Use SoC specific table to determine such events. This allows affinity management at parent irqchip level (GIC or ti-sci-intr + GIC). Signed-off-by: Vignesh Raghavendra <vigneshr@ti.com> --- drivers/irqchip/irq-ti-sci-inta.c | 149 +++++++++- 1 file changed, 147 insertions(+), 2 deletions(-)