Support 1 of N SPI interrupt for interrupt distribution
Subsystems (6)
*** BLURB HERE ***
The GICv3 supports 1 of N selection of SPI interrupts. When the GICD_IROUTERn.Interrupt_Routing_Mode == 1, the GIC selects the appropriate core for a SPI.
Actually, dispatch the interrupt by hardware Interrupt Control Unit is more efficient than irqbalance of software logic
and no need to implement software contoller to decide the targeted CPU on various platform.
In order to reduce interrupt latency, all interrupts are targeted to each online CPU defaultly by 1 of N selection of SPI interrupts. That is, SPI interrupts might be serviced simultaneously on different CPUs.
[default policy]
AS-IS (target to boot CPU)
CPU0 CPU1 CPU2 CPU3
29: 92 0 0 0 GICv3 141 Level ttyS0
TO-BE (enable ARM_IRQ_TARGET_ALL)
CPU0 CPU1 CPU2 CPU3
29: 23 23 24 22 GICv3 141 Level ttyS0
Hanks Chen (3):
irqchip/gic: enable irq target all
arm: disable irq on cpu shutdown flow
arm64: disable irq on cpu shutdown flow
arch/arm/kernel/smp.c | 10 ++- arch/arm64/kernel/smp.c | 9 ++- drivers/irqchip/Kconfig | 12 ++++ drivers/irqchip/irq-gic-v3.c | 107 +++++++++++++++++++++-------- include/linux/irqchip/arm-gic-v3.h | 1 + kernel/irq/cpuhotplug.c | 22 ++++++ kernel/irq/manage.c | 7 ++ 7 files changed, 138 insertions(+), 30 deletions(-)
Signature · 1 line
2.18.0
PATCH 1/3 irqchip/gic: enable irq target all
Support for interrupt distribution design for SMP system solutions.
With this feature enabled ,the SPI interrupts would be routed to all the cores rather than boot core to achieve better
load balance of interrupt handling.
That is, interrupts might be serviced simultaneously on different CPUs.
--- drivers/irqchip/Kconfig | 12 ++++ drivers/irqchip/irq-gic-v3.c | 107 +++++++++++++++++++++-------- include/linux/irqchip/arm-gic-v3.h | 1 + kernel/irq/cpuhotplug.c | 22 ++++++ kernel/irq/manage.c | 7 ++ 5 files changed, 122 insertions(+), 27 deletions(-)
drivers/irqchip/Kconfig +12 -0
@@ -597,4 +597,16 @@ config MST_IRQ help Support MStar Interrupt Controller. +config ARM_IRQ_TARGET_ALL + bool "Distribute interrupts across processors on SMP system" + depends on SMP && ARM_GIC_V3 + help + Support for interrupt distribution design for + SMP system solutions. With this feature enabled ,the + SPI interrupts would be routed to all the cores rather + than boot cpu to achieve better load balance of interrupt + handling + + If you don't know what to do here, say N. + endmenu
drivers/irqchip/irq-gic-v3.c +80 -27
137 lines not on this page · raw message
include/linux/irqchip/arm-gic-v3.h +1 -0
@@ -80,6 +80,7 @@#define GICD_CTLR_ENABLE_SS_G0 (1U << 0)#define GICD_TYPER_RSS (1U << 26) +#define GICD_TYPER_No1N (1U << 25)#define GICD_TYPER_LPIS (1U << 17)#define GICD_TYPER_MBIS (1U << 16)#define GICD_TYPER_ESPI (1U << 8)
kernel/irq/cpuhotplug.c +22 -0
@@ -87,6 +87,18 @@ static bool migrate_one_irq(struct irq_desc *desc)return false; } +#ifdef CONFIG_ARM_IRQ_TARGET_ALL +/* +* No move required, if interrupt is 1 of N IRQ. +* write current cpu_online_mask into affinity mask. +*/ +if (cpumask_weight (desc->irq_common_data.affinity) >1 ) { +cpumask_copy (desc->irq_common_data.affinity, cpu_online_mask); + +return false; + } +#endif +/* * Complete an eventually pending irq move cleanup. If this * interrupt was moved in hard irq context, then the vectors need
@@ -191,6 +203,16 @@ static void irq_restore_affinity_of_irq(struct irq_desc *desc, unsigned int cpu)struct irq_data *data =irq_desc_get_irq_data (desc);const struct cpumask *affinity =irq_data_get_affinity_mask (data); +#ifdef CONFIG_ARM_IRQ_TARGET_ALL +/* +* No restore required, if interrupt is 1 of N IRQ. +*/ +if (cpumask_weight (affinity) >1 ) { +cpumask_set_cpu (cpu,irq_data_get_affinity_mask (data)); +return ; + } +#endif +if (!irqd_affinity_is_managed (data) || !desc->action || !irq_data_get_irq_chip (data) || !cpumask_test_cpu (cpu, affinity))return ;
kernel/irq/manage.c +7 -0
@@ -270,7 +270,14 @@ int irq_do_set_affinity(struct irq_data *data, const struct cpumask *mask,switch (ret) {case IRQ_SET_MASK_OK :case IRQ_SET_MASK_OK_DONE : +#ifndef CONFIG_ARM_IRQ_TARGET_ALL cpumask_copy (desc->irq_common_data.affinity, mask); +#else +if (cpumask_weight (mask) >1 ) +cpumask_copy (desc->irq_common_data.affinity, cpu_online_mask); +else +cpumask_copy (desc->irq_common_data.affinity, mask); +#endif fallthrough;case IRQ_SET_MASK_OK_NOCOPY :irq_validate_effective_affinity (data);
Signature · 1 line
2.18.0
Marc Zyngier
maintainerRe: [PATCH v1 1/3] irqchip/gic: enable irq target all
Quoted 1 line · Hanks Chen
Support for interrupt distribution design for SMP system solutions.
As far as I know, we have been supporting interrupt distribution on ARM SMP systems pretty well for the past... what... 15 years? I'm sure Russell can dig out an ARM926 SMP system that even predates that.
Quoted 3 lines
With this feature enabled ,the SPI interrupts would be routed to all the cores rather than boot core to achieve better load balance of interrupt handling.
Please quantify this compared to the current distribution method.
Quoted 1 line
That is, interrupts might be serviced simultaneously on different CPUs.
They already can. What is new here? Or do you mean the *same* interrupt being serviced by different CPUs *at the same time*? That'd be fun.
Quoted 18 lines
Signed-off-by Hanks Chen <[email protected]>--- drivers/irqchip/Kconfig | 12 ++++ drivers/irqchip/irq-gic-v3.c | 107 +++++++++++++++++++++-------- include/linux/irqchip/arm-gic-v3.h | 1 + kernel/irq/cpuhotplug.c | 22 ++++++ kernel/irq/manage.c | 7 ++ 5 files changed, 122 insertions(+), 27 deletions(-)drivers/irqchip/Kconfig +1
-0@@ -597,4 +597,16 @@ config MST_IRQ help Support MStar Interrupt Controller. +config ARM_IRQ_TARGET_ALL
drivers/irqchip/Kconfig config MST_IRQ · line 610
Quoted 10 lines
drivers/irqchip/Kconfig +10
-0+ bool "Distribute interrupts across processors on SMP system" + depends on SMP && ARM_GIC_V3 + help + Support for interrupt distribution design for + SMP system solutions. With this feature enabled ,the + SPI interrupts would be routed to all the cores rather + than boot cpu to achieve better load balance of interrupt + handling + + If you don't know what to do here, say N.
There is no way we start introducing conditional compilation for architectural features. Either this works at all times, or it doesn't exist.
Quoted 21 lines
drivers/irqchip/Kconfig +1
-0+ endmenudrivers/irqchip/irq-gic-v3.c +6
-0@@ -381,6 +381,12 @@ static inline bool gic_supports_nmi(void)static_branch_likely (&supports_pseudo_nmis); } +static inline bool gic_supports_1n (void ) +{ +return (IS_ENABLED (CONFIG_ARM_IRQ_TARGET_ALL ) && + ~(readl_relaxed (gic_data.dist_base +GICD_TYPER ) & GICD_TYPER_No1N)); +} +static int gic_irq_set_irqchip_state (struct irq_data *d,enum irqchip_irq_state which,bool val){
drivers/irqchip/irq-gic-v3.c static void __init gic_dist_init(void) · line 725
Quoted 4 lines
drivers/irqchip/irq-gic-v3.c +1
-0@@ -716,6 +722,7 @@ static void __init gic_dist_init(void) {unsigned int i;u64 affinity; +
Spurious whitespace.
Quoted 28 lines
drivers/irqchip/irq-gic-v3.c +9
-9void __iomem *base = gic_data.dist_base;u32 val;@@ -759,16 +766,27 @@ static void __init gic_dist_init(void)/* Enable distributor with ARE, Group1 */ writel_relaxed (val, base +GICD_CTLR ); -/* -* Set all global interrupts to the boot CPU only. ARE must be -* enabled. -*/ - affinity =gic_mpidr_to_affinity (cpu_logical_map (smp_processor_id ())); -for (i =32 ; i <GIC_LINE_NR ; i++) -gic_write_irouter (affinity, base +GICD_IROUTER + i *8 ); +if (!gic_supports_1n ()) { +/* +* Set all global interrupts to the boot CPU only. ARE must be +* enabled. +*/ + affinity =gic_mpidr_to_affinity (cpu_logical_map (smp_processor_id ())); +for (i =32 ; i <GIC_LINE_NR ; i++) +gic_write_irouter (affinity, base +GICD_IROUTER + i *8 ); -for (i =0 ; i <GIC_ESPI_NR ; i++) -gic_write_irouter (affinity, base + GICD_IROUTERnE + i *8 ); +for (i =0 ; i <GIC_ESPI_NR ; i++)
drivers/irqchip/irq-gic-v3.c static void __init gic_dist_init(void) · line 790
Quoted 10 lines
drivers/irqchip/irq-gic-v3.c +10
-0+gic_write_irouter (affinity, base + GICD_IROUTERnE + i *8 ); + }else { +/* default set target all for all SPI */ +for (i =32 ; i <GIC_LINE_NR ; i++) +gic_write_irouter (GICD_IROUTER_SPI_MODE_ANY , + base +GICD_IROUTER + i *8 ); + +for (i =0 ; i <GIC_ESPI_NR ; i++) +gic_write_irouter (GICD_IROUTER_SPI_MODE_ANY , + base + GICD_IROUTERnE + i *8 );
Why should we decide to distribute interrupts to all CPUs by default when we don't even know where to route them?
Quoted 59 lines
drivers/irqchip/irq-gic-v3.c +1
-0+ } }static int gic_iterate_rdists(int (*fn)(struct
redist_region*, void__iomem*))drivers/irqchip/irq-gic-v3.c +30
-16@@ -1191,29 +1209,64 @@ static int gic_set_affinity(struct irq_data *d, const struct cpumask *mask_val, if (gic_irq_in_rdist(d)) return -EINVAL; -/* If interrupt was enabled, disable it first */ - enabled =gic_peek_irq (d,GICD_ISENABLER ); -if (enabled) -gic_mask_irq (d); +if (!gic_supports_1n ()) { +/* If interrupt was enabled, disable it first */ + enabled =gic_peek_irq (d,GICD_ISENABLER ); +if (enabled) +gic_mask_irq (d); - offset =convert_offset_index (d,GICD_IROUTER , &index); - reg =gic_dist_base (d) + offset + (index *8 ); - val =gic_mpidr_to_affinity (cpu_logical_map (cpu)); + offset =convert_offset_index (d,GICD_IROUTER , &index); + reg =gic_dist_base (d) + offset + (index *8 ); + val =gic_mpidr_to_affinity (cpu_logical_map (cpu)); -gic_write_irouter (val, reg); +gic_write_irouter (val, reg); -/* -* If the interrupt was enabled, enabled it again. Otherwise, -* just wait for the distributor to have digested our changes. -*/ -if (enabled) -gic_unmask_irq (d); -else -gic_dist_wait_for_rwp (); +/* +* If the interrupt was enabled, enabled it again. Otherwise, +* just wait for the distributor to have digested our changes. +*/ +if (enabled) +gic_unmask_irq (d); +else +gic_dist_wait_for_rwp (); + +irq_data_update_effective_affinity (d,cpumask_of (cpu)); + + }else { +/* +* no need to update when: +* input mask is equal to the current setting +*/ +if (cpumask_equal (irq_data_get_affinity_mask (d), mask_val)) +return IRQ_SET_MASK_OK_NOCOPY ; + +/* If interrupt was enabled, disable it first */ + enabled =gic_peek_irq (d,GICD_ISENABLER );
drivers/irqchip/irq-gic-v3.c static int gic_set_affinity(struct irq_data · line 1254
Quoted 10 lines
drivers/irqchip/irq-gic-v3.c +8
-1+if (enabled) +gic_mask_irq (d); + + offset =convert_offset_index (d,GICD_IROUTER , &index); + reg =gic_dist_base (d) + offset + (index *8 ); -irq_data_update_effective_affinity (d,cpumask_of (cpu)); +/* GICv3 supports target is 1 or all */ +if (cpumask_weight (mask_val) >1 ) + val =GICD_IROUTER_SPI_MODE_ANY ;
There is a massive difference between targeting more than one CPU and targeting all the CPUs. This breaks all existing drivers and userspace that need to manage interrupt affinity.
drivers/irqchip/irq-gic-v3.c static int gic_set_affinity(struct irq_data *d, const struct cpumask *mask_val, · line 1266
Quoted 13 lines
drivers/irqchip/irq-gic-v3.c +13
-0+else + val =gic_mpidr_to_affinity (cpu_logical_map (cpu)); + +gic_write_irouter (val, reg); + +/* +* If the interrupt was enabled, enabled it again. Otherwise, +* just wait for the distributor to have digested our changes. +*/ +if (enabled) +gic_unmask_irq (d); +else +gic_dist_wait_for_rwp ();
Why so much code duplication?
Quoted 25 lines
drivers/irqchip/irq-gic-v3.c +2
-1+ } - return IRQ_SET_MASK_OK_DONE; +return IRQ_SET_MASK_OK ; }#else #define
gic_set_affinityNULLinclude/linux/irqchip/arm-gic-v3.h +1
-0@@ -80,6 +80,7 @@#define GICD_CTLR_ENABLE_SS_G0 (1U << 0)#define GICD_TYPER_RSS (1U << 26) +#define GICD_TYPER_No1N (1U << 25)#define GICD_TYPER_LPIS (1U << 17)#define GICD_TYPER_MBIS (1U << 16)#define GICD_TYPER_ESPI (1U << 8)kernel/irq/cpuhotplug.c +0
-0
kernel/irq/cpuhotplug.c static bool migrate_one_irq(struct irq_desc *desc) · line 90
Quoted 4 lines
kernel/irq/cpuhotplug.c +1
-0@@ -87,6 +87,18 @@ static bool migrate_one_irq(struct irq_desc *desc)return false; } +#ifdef CONFIG_ARM_IRQ_TARGET_ALL
No way.
kernel/irq/cpuhotplug.c static bool migrate_one_irq(struct irq_desc *desc) · line 96
Quoted 6 lines
kernel/irq/cpuhotplug.c +6
-0+/* +* No move required, if interrupt is 1 of N IRQ. +* write current cpu_online_mask into affinity mask. +*/ +if (cpumask_weight (desc->irq_common_data.affinity) >1 ) { +cpumask_copy (desc->irq_common_data.affinity, cpu_online_mask);
Again, this is totally bogus.
Quoted 11 lines
kernel/irq/cpuhotplug.c +5
-0+ + return false; + } +#endif +/* * Complete an eventually pending irq move cleanup. If this * interrupt was moved in hard irq context, then the vectors need @@ -191,6 +203,16 @@ static void irq_restore_affinity_of_irq(struct irq_desc *desc, unsigned int cpu) struct irq_data *data = irq_desc_get_irq_data(desc);
kernel/irq/cpuhotplug.c static void irq_restore_affinity_of_irq(struct · line 214
Quoted 10 lines
kernel/irq/cpuhotplug.c +8
-0const struct cpumask *affinity =irq_data_get_affinity_mask (data); +#ifdef CONFIG_ARM_IRQ_TARGET_ALL +/* +* No restore required, if interrupt is 1 of N IRQ. +*/ +if (cpumask_weight (affinity) >1 ) { +cpumask_set_cpu (cpu,irq_data_get_affinity_mask (data)); +return ; + }
Same thing.
Quoted 11 lines
kernel/irq/cpuhotplug.c +2
-0+#endif +if (!irqd_affinity_is_managed (data) || !desc->action || !irq_data_get_irq_chip (data) || !cpumask_test_cpu (cpu, affinity))kernel/irq/cpuhotplug.c +0
-0return ;kernel/irq/manage.c +0
-0@@ -270,7 +270,14 @@ int irq_do_set_affinity(struct irq_data *data, const struct cpumask *mask,
kernel/irq/manage.c int irq_do_set_affinity(struct irq_data *data, · line 280
Quoted 10 lines
kernel/irq/manage.c +6
-0switch (ret) {case IRQ_SET_MASK_OK :case IRQ_SET_MASK_OK_DONE : +#ifndef CONFIG_ARM_IRQ_TARGET_ALL cpumask_copy (desc->irq_common_data.affinity, mask); +#else +if (cpumask_weight (mask) >1 ) +cpumask_copy (desc->irq_common_data.affinity, cpu_online_mask); +else +cpumask_copy (desc->irq_common_data.affinity, mask);
And again.
Quoted 4 lines
kernel/irq/manage.c +1
-0+#endif fallthrough;case IRQ_SET_MASK_OK_NOCOPY :
irq_validate_effective_affinity(data);
To sum it up:
- You claim that the current interrupt distribution model is either non functional or inefficient. I expect you to demonstrate this.
- You claim that this brings some kind of improvement: I expect you to describe workloads and numbers that demonstrate the improvements. 1:N distribution has repeatedly shown to have much worse behaviour that 1:1 distribution, so I really wonder what's new here.
- You break existing APIs where a driver or userspace can legitimately decide to route an interrupt to 2 CPUs out of 256 if it decides to. That's not acceptable.
- You pollute the core code with hacks that should never be there. If the current behaviour is a problem, please state what the problem is.
- You don't factor in the requirements of the pseudo-NMI masking, meaning the IRI doesn't know about the current PMR and will not be able to efficiently dispatch interrupts.
- I don't see anything that configures the participating nodes to the 1:N distribution, meaning it currently works by chance.
- How does it work with CPU hotplug, where we actively move the interrupts away from a CPU going down? With this scheme, the interrupts keep
being
delivered to the wrong CPU, resulting in lost interrupts if they were edge triggered.
As it stands, I can't imagine this code making it in the kernel.
Thanks,
M.
Signature · 1 line
Jazz is not dead. It just smells funny...
Catalin Marinas
maintainerRe: [PATCH v1 1/3] irqchip/gic: enable irq target all
Quoted 12 lines · Marc Zyngier
Quoted 1 line · Hanks Chen
Support for interrupt distribution design for SMP system solutions.
As far as I know, we have been supporting interrupt distribution on ARM SMP systems pretty well for the past... what... 15 years? I'm sure Russell can dig out an ARM926 SMP system that even predates that.
Quoted 3 lines
With this feature enabled ,the SPI interrupts would be routed to all the cores rather than boot core to achieve better load balance of interrupt handling.
Quoted 6 lines
Please quantify this compared to the current distribution method.
Quoted 1 line
That is, interrupts might be serviced simultaneously on different CPUs.
They already can. What is new here? Or do you mean the *same* interrupt being serviced by different CPUs *at the same time*? That'd be fun.
IIRC (it's been many years since I looked at the GIC), more than one CPU is woken and if they all read the INTACK, only one of them gets the actual IRQ number, the others see a spurious interrupt. I thought we decided that's not an efficient way to handle interrupts, so a software irqbalancer is better.
Has anything changed since then?
I'm also concerned that in a big.LITTLE system, you may see the big CPUs taking the interrupts all the time, which is nice for energy efficiency.
Signature · 6 lines
Catalin _______________________________________________ Linux-mediatek mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-mediatek
Marc Zyngier
maintainerRe: [PATCH v1 1/3] irqchip/gic: enable irq target all
Quoted 20 lines · Catalin Marinas
Quoted 12 lines · Marc Zyngier
Quoted 1 line · Hanks Chen
Support for interrupt distribution design for SMP system solutions.
As far as I know, we have been supporting interrupt distribution on ARM SMP systems pretty well for the past... what... 15 years? I'm sure Russell can dig out an ARM926 SMP system that even predates that.
Quoted 3 lines
With this feature enabled ,the SPI interrupts would be routed to all the cores rather than boot core to achieve better load balance of interrupt handling.
Quoted 6 lines
Please quantify this compared to the current distribution method.
Quoted 1 line
That is, interrupts might be serviced simultaneously on different CPUs.
They already can. What is new here? Or do you mean the *same* interrupt being serviced by different CPUs *at the same time*? That'd be fun.
Quoted 7 lines
IIRC (it's been many years since I looked at the GIC), more than one CPU is woken and if they all read the INTACK, only one of them gets the actual IRQ number, the others see a spurious interrupt. I thought we decided that's not an efficient way to handle interrupts, so a software irqbalancer is better.
Has anything changed since then?
What you describe is mostly the GICv1/v2 behaviour.
GICv3 *can* be slightly better in that respect, as long as you force the synchronization from the CPU interface back to the redistributor via PHME and a DSB SY on each priority change.
Which means what whatever you gain from not interrupting the other CPUs, you waste it by forcing synchronization system wide.
Quoted 2 lines
I'm also concerned that in a big.LITTLE system, you may see the big CPUs taking the interrupts all the time, which is nice for energy efficiency.
Yes, this is bound to happen. It means you cannot place interrupts on a small cluster unless you completely hotplug the big cores off, thanks to the "more than one means all" nonsense.
And since these patches seem to also break hotplug, that's... fun.
M.
Signature · 6 lines
Jazz is not dead. It just smells funny... _______________________________________________ Linux-mediatek mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-mediatek
Re: [PATCH v1 1/3] irqchip/gic: enable irq target all
In reply to Marc ZyngierHi Marc,
Sorry for the late reply.
Quoted 8 lines · Marc Zyngier
Quoted 1 line · Hanks Chen
Support for interrupt distribution design for SMP system solutions.
As far as I know, we have been supporting interrupt distribution on ARM SMP systems pretty well for the past... what... 15 years? I'm sure Russell can dig out an ARM926 SMP system that even predates that.
Quoted 6 lines
Quoted 3 lines
With this feature enabled ,the SPI interrupts would be routed to all the cores rather than boot core to achieve better load balance of interrupt handling.
Please quantify this compared to the current distribution method.
Got it, but I need some time to prepare for the demonstration. I'll show it soon.
Btw, it is good on IO bandwidth-sensitive applications, we could also use androbench to recognize it.
Quoted 5 lines
Quoted 1 line
That is, interrupts might be serviced simultaneously on different CPUs.
They already can. What is new here? Or do you mean the *same* interrupt being serviced by different CPUs *at the same time*? That'd be fun.
It is my fault, I used the imprecise word to describe it. Not mean 'same interrupt being serviced by different CPUs'. You are right, we can use affinity API to achieve 'different CPUs' However, I focus on HW behavior.
If we don't change the setting of affinity(target to boot CPU), SPI IRQs might still be serviced on different CPUs after applying target all.
Quoted 17 lines
Quoted 17 lines
Signed-off-by Hanks Chen <[email protected]>--- drivers/irqchip/Kconfig | 12 ++++ drivers/irqchip/irq-gic-v3.c | 107 +++++++++++++++++++++-------- include/linux/irqchip/arm-gic-v3.h | 1 + kernel/irq/cpuhotplug.c | 22 ++++++ kernel/irq/manage.c | 7 ++ 5 files changed, 122 insertions(+), 27 deletions(-)drivers/irqchip/Kconfig +0
-0@@ -597,4 +597,16 @@ config MST_IRQ help Support MStar Interrupt Controller.
Quoted 16 lines
drivers/irqchip/Kconfig config MST_IRQ · line 610
Quoted 11 lines
drivers/irqchip/Kconfig +11
-0+config ARM_IRQ_TARGET_ALL + bool "Distribute interrupts across processors on SMP system" + depends on SMP && ARM_GIC_V3 + help + Support for interrupt distribution design for + SMP system solutions. With this feature enabled ,the + SPI interrupts would be routed to all the cores rather + than boot cpu to achieve better load balance of interrupt + handling + + If you don't know what to do here, say N.There is no way we start introducing conditional compilation for architectural features. Either this works at all times, or it doesn't exist.
I have no idea how to solve it.
Indeed, it is not good to enable the feature on every ARM-based platform.
maybe I could use device tree to determine whether or not the feature should be enabled.
e.g.
gic: interrupt-controller {
compatible = "arm,gic-v3";
...
irq-target-all-enable = <1>;
};Would this be better?
Quoted 20 lines
Quoted 20 lines
drivers/irqchip/Kconfig +1
-0+ endmenudrivers/irqchip/irq-gic-v3.c +6
-0@@ -381,6 +381,12 @@ static inline bool gic_supports_nmi(void)static_branch_likely (&supports_pseudo_nmis); } +static inline bool gic_supports_1n (void ) +{ +return (IS_ENABLED (CONFIG_ARM_IRQ_TARGET_ALL ) && + ~(readl_relaxed (gic_data.dist_base +GICD_TYPER ) & GICD_TYPER_No1N)); +} +static int gic_irq_set_irqchip_state (struct irq_data *d,enum irqchip_irq_state which,bool val)
Quoted 8 lines
drivers/irqchip/irq-gic-v3.c static void __init gic_dist_init(void) · line 725
Quoted 6 lines
{
drivers/irqchip/irq-gic-v3.c +1
-0@@ -716,6 +722,7 @@ static void __init gic_dist_init(void) {unsigned int i;u64 affinity; +Spurious whitespace.
Got it, I'll fix it.
Quoted 35 lines
Quoted 35 lines
drivers/irqchip/irq-gic-v3.c +15
-9void __iomem *base = gic_data.dist_base;u32 val;@@ -759,16 +766,27 @@ static void __init gic_dist_init(void)/* Enable distributor with ARE, Group1 */ writel_relaxed (val, base +GICD_CTLR ); -/* -* Set all global interrupts to the boot CPU only. ARE must be -* enabled. -*/ - affinity =gic_mpidr_to_affinity (cpu_logical_map (smp_processor_id ())); -for (i =32 ; i <GIC_LINE_NR ; i++) -gic_write_irouter (affinity, base +GICD_IROUTER + i *8 ); +if (!gic_supports_1n ()) { +/* +* Set all global interrupts to the boot CPU only. ARE must be +* enabled. +*/ + affinity =gic_mpidr_to_affinity (cpu_logical_map (smp_processor_id ())); +for (i =32 ; i <GIC_LINE_NR ; i++) +gic_write_irouter (affinity, base +GICD_IROUTER + i *8 ); -for (i =0 ; i <GIC_ESPI_NR ; i++) -gic_write_irouter (affinity, base + GICD_IROUTERnE + i *8 ); +for (i =0 ; i <GIC_ESPI_NR ; i++) +gic_write_irouter (affinity, base + GICD_IROUTERnE + i *8 ); + }else { +/* default set target all for all SPI */ +for (i =32 ; i <GIC_LINE_NR ; i++) +gic_write_irouter (GICD_IROUTER_SPI_MODE_ANY , + base +GICD_IROUTER + i *8 );
Quoted 9 lines
drivers/irqchip/irq-gic-v3.c static void __init gic_dist_init(void) · line 788
Quoted 4 lines
drivers/irqchip/irq-gic-v3.c +4
-0+ +for (i =0 ; i <GIC_ESPI_NR ; i++) +gic_write_irouter (GICD_IROUTER_SPI_MODE_ANY , + base + GICD_IROUTERnE + i *8 );Why should we decide to distribute interrupts to all CPUs by default when
we don't even know where to route them?
It makes sense to dispatch SPI interrupts defaultly by GICD when we transfer the control to GIC HW.
It's more real time to track the IRQ balance by HW controller and the policy is based on CPU state or low power purpose.
btw, we also want to introduce the interrupt class to our big-little system.It is an implementation-defined feature that the GIC600 provides for the 1 of N SPIs.
Quoted 65 lines
Quoted 65 lines
drivers/irqchip/irq-gic-v3.c +1
-0+ } }static int gic_iterate_rdists(int (*fn)(struct
redist_region*, void__iomem*))drivers/irqchip/irq-gic-v3.c +35
-16@@ -1191,29 +1209,64 @@ static int gic_set_affinity(struct irq_data *d, const struct cpumask *mask_val, if (gic_irq_in_rdist(d)) return -EINVAL; -/* If interrupt was enabled, disable it first */ - enabled =gic_peek_irq (d,GICD_ISENABLER ); -if (enabled) -gic_mask_irq (d); +if (!gic_supports_1n ()) { +/* If interrupt was enabled, disable it first */ + enabled =gic_peek_irq (d,GICD_ISENABLER ); +if (enabled) +gic_mask_irq (d); - offset =convert_offset_index (d,GICD_IROUTER , &index); - reg =gic_dist_base (d) + offset + (index *8 ); - val =gic_mpidr_to_affinity (cpu_logical_map (cpu)); + offset =convert_offset_index (d,GICD_IROUTER , &index); + reg =gic_dist_base (d) + offset + (index *8 ); + val =gic_mpidr_to_affinity (cpu_logical_map (cpu)); -gic_write_irouter (val, reg); +gic_write_irouter (val, reg); -/* -* If the interrupt was enabled, enabled it again. Otherwise, -* just wait for the distributor to have digested our changes. -*/ -if (enabled) -gic_unmask_irq (d); -else -gic_dist_wait_for_rwp (); +/* +* If the interrupt was enabled, enabled it again. Otherwise, +* just wait for the distributor to have digested our changes. +*/ +if (enabled) +gic_unmask_irq (d); +else +gic_dist_wait_for_rwp (); + +irq_data_update_effective_affinity (d,cpumask_of (cpu)); + + }else { +/* +* no need to update when: +* input mask is equal to the current setting +*/ +if (cpumask_equal (irq_data_get_affinity_mask (d), mask_val)) +return IRQ_SET_MASK_OK_NOCOPY ; + +/* If interrupt was enabled, disable it first */ + enabled =gic_peek_irq (d,GICD_ISENABLER ); +if (enabled) +gic_mask_irq (d); + + offset =convert_offset_index (d,GICD_IROUTER , &index); + reg =gic_dist_base (d) + offset + (index *8 );
Quoted 9 lines
drivers/irqchip/irq-gic-v3.c static int gic_set_affinity(struct irq_data *d, const struct cpumask *mask_val, · line 1253
Quoted 4 lines
drivers/irqchip/irq-gic-v3.c +3
-1-irq_data_update_effective_affinity (d,cpumask_of (cpu)); +/* GICv3 supports target is 1 or all */ +if (cpumask_weight (mask_val) >1 ) + val =GICD_IROUTER_SPI_MODE_ANY ;There is a massive difference between targeting more than one CPU and targeting all the CPUs. This breaks all existing drivers and userspace that need to manage interrupt affinity.
Yes, it is not a good implementation that cause the user-space confused. Should I modify the code as follows?
If (bitmap_equal(mask_val, online_mask, nr_cpumask_bits))
val = GICD_IROUTER_SPI_MODE_ANY;
redefine the coverage
Quoted 7 lines
Quoted 7 lines
drivers/irqchip/irq-gic-v3.c +7
-0+else + val =gic_mpidr_to_affinity (cpu_logical_map (cpu)); + +gic_write_irouter (val, reg); + +/* +* If the interrupt was enabled, enabled it again. Otherwise,
Quoted 8 lines
drivers/irqchip/irq-gic-v3.c static int gic_set_affinity(struct irq_data *d, const struct cpumask *mask_val, · line 1231
Quoted 6 lines
drivers/irqchip/irq-gic-v3.c +6
-0+* just wait for the distributor to have digested our changes. +*/ +if (enabled) +gic_unmask_irq (d); +else +gic_dist_wait_for_rwp ();Why so much code duplication?
Got it, I'll remove the duplicated code.
Quoted 33 lines
Quoted 26 lines
drivers/irqchip/irq-gic-v3.c +2
-1+ } - return IRQ_SET_MASK_OK_DONE; +return IRQ_SET_MASK_OK ; }#else #define
gic_set_affinityNULLinclude/linux/irqchip/arm-gic-v3.h +1
-0@@ -80,6 +80,7 @@#define GICD_CTLR_ENABLE_SS_G0 (1U << 0)#define GICD_TYPER_RSS (1U << 26) +#define GICD_TYPER_No1N (1U << 25)#define GICD_TYPER_LPIS (1U << 17)#define GICD_TYPER_MBIS (1U << 16)#define GICD_TYPER_ESPI (1U << 8)kernel/irq/cpuhotplug.c +0
-0kernel/irq/cpuhotplug.c static bool migrate_one_irq(struct irq_desc *desc) · line 90
Quoted 4 lines
kernel/irq/cpuhotplug.c +1
-0@@ -87,6 +87,18 @@ static bool migrate_one_irq(struct irq_desc *desc)return false; } +#ifdef CONFIG_ARM_IRQ_TARGET_ALL No way.
Quoted 9 lines
kernel/irq/cpuhotplug.c static bool migrate_one_irq(struct irq_desc *desc) · line 96
Quoted 6 lines
kernel/irq/cpuhotplug.c +6
-0+/* +* No move required, if interrupt is 1 of N IRQ. +* write current cpu_online_mask into affinity mask. +*/ +if (cpumask_weight (desc->irq_common_data.affinity) >1 ) { +cpumask_copy (desc->irq_common_data.affinity, cpu_online_mask);Again, this is totally bogus.
If I add the target all flag into irq_common_data stucture to distinguish it. Would this be better?
remove #ifdef CONFIG_ARM_IRQ_TARGET_ALL
and
if (desc->irq_common_data.irq_target_all)
cpumask_copy(desc->irq_common_data.affinity, cpu_online_mask);
Quoted 13 lines
Quoted 13 lines
kernel/irq/cpuhotplug.c +5
-0+ + return false; + } +#endif +/* * Complete an eventually pending irq move cleanup. If this * interrupt was moved in hard irq context, then the vectors need @@ -191,6 +203,16 @@ static void irq_restore_affinity_of_irq(struct irq_desc *desc, unsigned int cpu) struct irq_data *data = irq_desc_get_irq_data(desc);const struct cpumask *affinity =irq_data_get_affinity_mask (data);
Quoted 11 lines
kernel/irq/cpuhotplug.c static void irq_restore_affinity_of_irq(struct irq_desc *desc, unsigned int cpu) · line 213
Quoted 8 lines
kernel/irq/cpuhotplug.c +8
-0+#ifdef CONFIG_ARM_IRQ_TARGET_ALL +/* +* No restore required, if interrupt is 1 of N IRQ. +*/ +if (cpumask_weight (affinity) >1 ) { +cpumask_set_cpu (cpu,irq_data_get_affinity_mask (data)); +return ; + }Same thing.
as above
remove #ifdef CONFIG_ARM_IRQ_TARGET_ALL
and
if (desc->irq_common_data.irq_target_all) {
cpumask_set_cpu(cpu, irq_data_get_affinity_mask(data));
return;
}Quoted 15 lines
Quoted 15 lines
kernel/irq/cpuhotplug.c +2
-0+#endif +if (!irqd_affinity_is_managed (data) || !desc->action || !irq_data_get_irq_chip (data) || !cpumask_test_cpu (cpu, affinity))kernel/irq/cpuhotplug.c +0
-0return ;kernel/irq/manage.c +1
-0@@ -270,7 +270,14 @@ int irq_do_set_affinity(struct irq_data *data, const struct cpumask *mask, switch (ret) {case IRQ_SET_MASK_OK :case IRQ_SET_MASK_OK_DONE : +#ifndef CONFIG_ARM_IRQ_TARGET_ALL
Quoted 9 lines
kernel/irq/manage.c int irq_do_set_affinity(struct irq_data *data, const struct cpumask *mask, · line 279
Quoted 6 lines
kernel/irq/manage.c +5
-0cpumask_copy (desc->irq_common_data.affinity, mask); +#else +if (cpumask_weight (mask) >1 ) +cpumask_copy (desc->irq_common_data.affinity, cpu_online_mask); +else +cpumask_copy (desc->irq_common_data.affinity, mask);And again.
as above
if (!desc->irq_common_data.irq_target_all)
cpumask_copy(desc->irq_common_data.affinity, mask); else cpumask_copy(desc->irq_common_data.affinity, cpu_online_mask);
Quoted 7 lines
Quoted 4 lines
kernel/irq/manage.c +1
-0+#endif fallthrough;case IRQ_SET_MASK_OK_NOCOPY :
irq_validate_effective_affinity(data);To sum it up:
Quoted 8 lines
- You claim that the current interrupt distribution model is either non functional or inefficient. I expect you to demonstrate this.
- You claim that this brings some kind of improvement: I expect you to describe workloads and numbers that demonstrate the improvements. 1:N distribution has repeatedly shown to have much worse behaviour that 1:1 distribution, so I really wonder what's new here.
I'll show the demonstration to answer first two questions.
Quoted 4 lines
- You break existing APIs where a driver or userspace can legitimately decide to route an interrupt to 2 CPUs out of 256 if it decides to. That's not acceptable.
I redefine the coverage as above
Quoted 4 lines
- You pollute the core code with hacks that should never be there. If the
current behaviour is a problem, please state what the problem is.
We found the RCU warn when IRQs target to a offline CPU without I bit masked in CPU hot-plug flow
I'll reproduce the issue again and show the log analysis for it.
Quoted 5 lines
- You don't factor in the requirements of the pseudo-NMI masking, meaning
the IRI doesn't know about the current PMR and will not be able to efficiently dispatch interrupts.
Yes. mask interrupts using PMR becomes very expensive. (require a dsb) We don't use the pseduo-NMI in our platform.
Thank you for pointing that out, I need some time to mull it over.
Quoted 4 lines
- I don't see anything that configures the participating nodes to the 1:N
distribution, meaning it currently works by chance.
1 of N SPI is dispatched by GICD with round robin policy by default. However, GIC support the interface to control the policy of dispatching SPI interrupt.
In MTK platform, GICD could dispatch SPI interrupts to suitable target CPU by CPU state or low power purpose.
Quoted 7 lines
- How does it work with CPU hotplug, where we actively move the interrupts
away from a CPU going down? With this scheme, the interrupts keep
being
delivered to the wrong CPU, resulting in lost interrupts if they were edge triggered.
Hot-plug off CPU would assert the GICR_WAKER.ProcessorSleep and GICR_CTLR.DPG in trusted firmware cortex-a (TF-a)
1) When ProcessorSleep is 1
All interrupts that arrive at the re-distributor are not communicated to the CPU interface
2) When DPG is 1
Any interrupts that have not reached a core at the time of the change are recalled and reprioritized by
the GIC
hot-plug off sequence. (PSCI in TF-a)
1. Mask interrupts on the core
2. Clear the CPU interface enables
3. GICR_WAKER.ProcessorSleep = 1
4. GICR_CTLR.DPG = 1 (we need to add it when 1 of N SPIs feature is enabled)
According to the above, I think it would keep the edge-triggered interrupt in GIC.
Quoted 5 lines
As it stands, I can't imagine this code making it in the kernel.
Thanks,
M.
In MTK platform, We change the dispatch policy of 1:N SPIs by interface of interrupt routing infrastructure
and it would improve our system performance.
GICR_CTLR.DPG, cpu_active and interrupt class could also change the policy of 1:N SPIs.
So we want to introduce the interface for 1 of N SPIs to provide more choices.
But I need to carefully consider the various scenario on all arm64 system and don't break the established API.
I need some time to mull it over.
Thank you for your review
Hanks
_______________________________________________
Linux-mediatek mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-mediatek
Re: [PATCH v1 1/3] irqchip/gic: enable irq target all
Hanks,
Quoted 15 lines · Hanks Chen
Quoted 10 lines · Marc Zyngier
kernel/irq/cpuhotplug.c static bool migrate_one_irq(struct irq_desc *desc) · line 96
Quoted 6 lines · Hanks Chen
kernel/irq/cpuhotplug.c +6
-0+/* +* No move required, if interrupt is 1 of N IRQ. +* write current cpu_online_mask into affinity mask. +*/ +if (cpumask_weight (desc->irq_common_data.affinity) >1 ) { +cpumask_copy (desc->irq_common_data.affinity, cpu_online_mask);Again, this is totally bogus.
If I add the target all flag into
irq_common_datastucture to distinguish it. Would this be better?
Quoted 6 lines
remove #ifdef
CONFIG_ARM_IRQ_TARGET_ALLand
if (
desc->irq_common_data.irq_target_all)cpumask_copy(desc->irq_common_data.affinity, cpu_online_mask);
No. We are not adding random members to irq_common_data just to support this. If at all this is a flag of the interrupt chip.
Also this copy is wrong to begin with. The affinity mask is only updated on cpu hot-unplug when the outgoing CPU is the last online CPU in the mask. Then we break the user/kernel supplied affinity mask and we only do that when the interrupt is not affinity managed.
We do not remove offline CPUs from the affinity mask. The affinity code of the irq chip has to ensure that only online CPUs can be targeted. That's what ends up in effective_affinity.
Quoted 8 lines
Quoted 8 lines
kernel/irq/cpuhotplug.c static void irq_restore_affinity_of_irq(struct irq_desc *desc, unsigned int cpu) · line 213
Quoted 8 lines
kernel/irq/cpuhotplug.c +8
-0+#ifdef CONFIG_ARM_IRQ_TARGET_ALL +/* +* No restore required, if interrupt is 1 of N IRQ. +*/ +if (cpumask_weight (affinity) >1 ) { +cpumask_set_cpu (cpu,irq_data_get_affinity_mask (data)); +return ; + }
Heck no. This breaks managed interrupts and some more. Fiddling with the irq affinity mask is wrong to begin with. Don't touch it ever.
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
kernel/irq/manage.c +0
-0@@ -270,7 +270,14 @@ int irq_do_set_affinity(struct irq_data *data, const struct cpumask *mask, switch (ret) {case IRQ_SET_MASK_OK :
Quoted 8 lines
Quoted 8 lines
kernel/irq/manage.c int irq_do_set_affinity(struct irq_data *data, const struct cpumask *mask, · line 279
Quoted 8 lines
kernel/irq/manage.c +6
-0case IRQ_SET_MASK_OK_DONE : +#ifndef CONFIG_ARM_IRQ_TARGET_ALL cpumask_copy (desc->irq_common_data.affinity, mask); +#else +if (cpumask_weight (mask) >1 ) +cpumask_copy (desc->irq_common_data.affinity, cpu_online_mask); +else +cpumask_copy (desc->irq_common_data.affinity, mask);
Again, no. Touching the affinity mask is a NONO. The affinity mask is handed in from either the kernel or from user space. This code has no business to override that.
The only case where the kernel touches it is on CPU hot-unplug when the last cpu in the affinity mask goes offline and if the interrupt is not affinity managed.
Quoted 7 lines
Quoted 4 lines
- You pollute the core code with hacks that should never be there. If the
current behaviour is a problem, please state what the problem is.We found the RCU warn when IRQs target to a offline CPU without I bit masked in CPU hot-plug flow
I'll reproduce the issue again and show the log analysis for it.
That has absolutely nothing to do with any of the functions where you fiddle with the affinity mask.
Thanks,
tglx
_______________________________________________
Linux-mediatek mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-mediatek
PATCH 2/3 arm: disable irq on cpu shutdown flow
In reply to Hanks ChenDisable irq on cpu shutdown flow to ensure interrupts did not bother this cpu after status as offline.
To avoid suspicious RCU usage
[0:swapper/0]RCU used illegally from offline CPU! ... [0:swapper/0]lockdep: [name:lockdep&]cpu_id = 0, cpu_is_offline = 1
--- arch/arm/kernel/smp.c | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-)
arch/arm/kernel/smp.c +8 -2
@@ -262,6 +262,12 @@ int __cpu_disable(void)remove_cpu_topology (cpu);#endif +/* +* we disable irq here to ensure interrupts +* did not bother this cpu after status as offline. +*/ +local_irq_disable (); +/* * Take this CPU offline. Once we clear this, we can't return, * and we must not schedule until we're ready to give up the cpu.
@@ -600,11 +606,11 @@ static void ipi_cpu_stop(unsigned int cpu)raw_spin_unlock (&stop_lock); } -set_cpu_online (cpu, false); -local_fiq_disable ();local_irq_disable (); +set_cpu_online (cpu, false); +while (1 ) {cpu_relax ();wfe ();
Signature · 1 line
2.18.0
PATCH 3/3 arm64: disable irq on cpu shutdown flow
In reply to Hanks ChenDisable irq on cpu shutdown flow to ensure interrupts did not bother this cpu after status as offline.
To avoid suspicious RCU usage
(0)[0:swapper/0]RCU used illegally from offline CPU! ... (0)[0:swapper/0]lockdep: [name:lockdep&]cpu_id = 0, cpu_is_offline = 1
--- arch/arm64/kernel/smp.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)
arch/arm64/kernel/smp.c +8 -1
@@ -308,6 +308,12 @@ int __cpu_disable(void)remove_cpu_topology (cpu);numa_remove_cpu (cpu); +/* +* we disable irq here to ensure interrupts +* did not bother this cpu after status as offline. +*/ +local_irq_disable (); +/* * Take this CPU offline. Once we clear this, we can't return, * and we must not schedule until we're ready to give up the cpu.
@@ -842,9 +848,10 @@ void arch_irq_work_raise(void)static void local_cpu_stop (void ) { +local_daif_mask (); +set_cpu_online (smp_processor_id (), false); -local_daif_mask ();sdei_mask_local_cpu ();cpu_park_loop (); }
Signature · 1 line
2.18.0
Marc Zyngier
maintainerRe: [PATCH v1 3/3] arm64: disable irq on cpu shutdown flow
Quoted 6 lines · Hanks Chen
Disable irq on cpu shutdown flow to ensure interrupts did not bother this cpu after status as offline.
To avoid suspicious RCU usage
(0)[0:swapper/0]RCU used illegally from offline CPU! ... (0)[0:swapper/0]lockdep: [name:lockdep&]cpu_id = 0,cpu_is_offline= 1
This needs to be explained *a lot* more .
My hunch is that because a CPU going offline can still receive interrupts thanks to your interrupt broadcast hack, you break some the core expectations, and RCU shouts at you.
If that's indeed the case, I don't think the architecture code needs fixing (or at least, not for that).
Quoted 13 lines
Signed-off-by Hanks Chen <[email protected]>--- arch/arm64/kernel/smp.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)arch/arm64/kernel/smp.c +0
-0@@ -308,6 +308,12 @@ int __cpu_disable(void)remove_cpu_topology (cpu);numa_remove_cpu (cpu);
arch/arm64/kernel/smp.c int __cpu_disable(void) · line 319
Quoted 10 lines
arch/arm64/kernel/smp.c +6
-0+/* +* we disable irq here to ensure interrupts +* did not bother this cpu after status as offline. +*/ +local_irq_disable (); +/* * Take this CPU offline. Once we clear this, we can't return, * and we must not schedule until we're ready to give up the cpu.
Conveniently, the code that takes care of migrating the interrupts is just below this comment. Which strongly suggests that the interrupt migration is broken by your earlier patch.
arch/arm64/kernel/smp.c void arch_irq_work_raise(void) · line 857
Quoted 12 lines
arch/arm64/kernel/smp.c +2
-1@@ -842,9 +848,10 @@ void arch_irq_work_raise(void)static void local_cpu_stop (void ) { +local_daif_mask (); +set_cpu_online (smp_processor_id (), false); -local_daif_mask ();sdei_mask_local_cpu ();cpu_park_loop (); }
What problem are you addressing here?
Thanks,
M.
Signature · 6 lines
Jazz is not dead. It just smells funny... _______________________________________________ Linux-mediatek mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-mediatek