Re: [PATCH] irqchip/gicv3: Add support for Range Selector (RS) feature

[Shanker Donthineni]

Hi Marc,

On 09/16/2017 01:14 PM, Marc Zyngier wrote:
> On Fri, Sep 15 2017 at 10:08:56 am BST, Shanker Donthineni <shankerd@xxxxxxxxxxxxxx> wrote:
> 
> Hi Shanker,
> 
>> A new feature Range Selector (RS) has been added to GIC specification
>> in order to support more than 16 CPUs at affinity level 0. New fields
>> are introduced in SGI system registers (ICC_SGI0R_EL1, ICC_SGI1R_EL1
>> and ICC_ASGI1R_EL1) to relax an artificial limit of 16 at level 0.
>>
>> - A new RSS field in ICC_CTLR_EL3, ICC_CTLR_EL1 and ICV_CTLR_EL1:
>>   [18] - Range Selector Support (RSS)
>>   0b0 = Targeted SGIs with affinity level 0 values of 0-15 are supported.
>>   0b1 = Targeted SGIs with affinity level 0 values of 0-255 are supported.
>>
>> - A new RS field in ICC_SGI0R_EL1, ICC_SGI1R_EL1 and ICC_ASGI1R_EL1:
>>   [47:44] - RangeSelector (RS) which group of 16 TargetList[n] field
>>             TargetList[n] represents aff0 value ((RS*16)+n)
>>             When ICC_CTLR_EL3.RSS==0 or ICC_CTLR_EL1.RSS==0, RS is RES0.
>>
>> - A new RSS field in GICD_TYPER:
>>   [26] - Range Selector Support (RSS)
>>   0b0 = Targeted SGIs with affinity level 0 values of 0-15 are supported.
>>   0b1 = Targeted SGIs with affinity level 0 values of 0-255 are supported.
>>
>> Signed-off-by: Shanker Donthineni <shankerd@xxxxxxxxxxxxxx>
>> ---
>>  drivers/irqchip/irq-gic-v3.c       | 79 ++++++++++++++++++++------------------
>>  include/linux/irqchip/arm-gic-v3.h |  4 ++
>>  2 files changed, 46 insertions(+), 37 deletions(-)
>>
>> diff --git a/drivers/irqchip/irq-gic-v3.c b/drivers/irqchip/irq-gic-v3.c
>> index 984c3ec..ba98c94 100644
>> --- a/drivers/irqchip/irq-gic-v3.c
>> +++ b/drivers/irqchip/irq-gic-v3.c
>> @@ -56,6 +56,7 @@ struct gic_chip_data {
>>  	u64			redist_stride;
>>  	u32			nr_redist_regions;
>>  	unsigned int		irq_nr;
>> +	bool			has_rss;
>>  	struct partition_desc	*ppi_descs[16];
>>  };
>>  
>> @@ -483,6 +484,9 @@ static int gic_populate_rdist(void)
>>  
>>  static void gic_cpu_sys_reg_init(void)
>>  {
>> +	int cpu = smp_processor_id();
>> +	bool rss;
>> +
>>  	/*
>>  	 * Need to check that the SRE bit has actually been set. If
>>  	 * not, it means that SRE is disabled at EL2. We're going to
>> @@ -514,6 +518,15 @@ static void gic_cpu_sys_reg_init(void)
>>  
>>  	/* ... and let's hit the road... */
>>  	gic_write_grpen1(1);
>> +
>> +	/* GIC CPU interface has RSS capability? */
>> +	rss = !!(read_sysreg_s(SYS_ICC_CTLR_EL1) & ICC_CTLR_EL1_RSS);
>> +
>> +	if (gic_data.has_rss != rss)
>> +		pr_info("Broken hardware, mismatch RSS, SGIs may not work\n");
> 
> I disagree with this statement. The architecture allows for a non-RSS
> GIC to be matched with RSS-aware CPUs, and vice-versa. This doesn't mean
> the system is broken.
> 

Yes, I agree with you, overlooked the spec. I'll fix it in v2 patch.


>> +	else if ((cpu_logical_map(cpu) & 0xf0) && (!rss))
>> +		pr_crit("MPIDR bits[7..4] must be zero, cpu=%d\n", cpu);
> 
> That's the real "Gotcha". Finding an MPIDR_EL1.Aff0 >= 16 on any CPU if
> any of the CPUs or the distributor doesn't support RSS is an integration
> blunder. Anything else is perfectly acceptable.
> 
> Unfortunately, this is not something you can decide by looking at a
> single CPU, so you'll need to accumulate some state somehow.
> 
 
Sure, I'll enhance the validation code to check CPU AFF0 value and RSS capability 
of each online CPU. Something like this.

 static void gic_cpu_sys_reg_init(void)
 {
 ....

        /* Keep the RSS capability status in per_cpu variable */
        per_cpu(has_rss, cpu) = !!(gic_read_ctlr() & ICC_CTLR_EL1_RSS);

        /* Does it requires RSS support to send SGIs to this CPU? */
        if (!MPIDR_RS(cpu_logical_map(cpu)))
                return;

        /* Check all the other CPUs have capable of sending SGIs to this CPU */
        for_each_online_cpu(i) {
                if (i == cpu)
                        continue;

                if (!per_cpu(has_rss, i)) {
                        pr_crit("CPU%d doesn't support RSS, Can't send SGIs "
                                "to CPU%d\n", i, cpu);
                        continue;
                }

                /**
                 * GIC spec says, When ICC_CTLR_EL1.RSS==1 and GICD_TYPER.RSS==0,
                 * writing ICC_ASGI1R_EL1 register with RS != 0 is a CONSTRAINED
                 * UNPREDICTABLE choice of :
                 *   - The write is ignored.
                 *   - The RS field is treated as 0.
                 */
                if (!gic_data.has_rss)
                        pr_crit("CPU%d has RSS support but not GIC Distributor,"
                                " Can't send SGIs to CPU%d", i, cpu);
        }
 }

>> +
>>  }
>>  
>>  static int gic_dist_supports_lpis(void)
>> @@ -554,43 +567,13 @@ static int gic_starting_cpu(unsigned int cpu)
>>  	return 0;
>>  }
>>  
>> -static u16 gic_compute_target_list(int *base_cpu, const struct cpumask *mask,
>> -				   unsigned long cluster_id)
>> -{
>> -	int next_cpu, cpu = *base_cpu;
>> -	unsigned long mpidr = cpu_logical_map(cpu);
>> -	u16 tlist = 0;
>> -
>> -	while (cpu < nr_cpu_ids) {
>> -		/*
>> -		 * If we ever get a cluster of more than 16 CPUs, just
>> -		 * scream and skip that CPU.
>> -		 */
>> -		if (WARN_ON((mpidr & 0xff) >= 16))
>> -			goto out;
>> -
>> -		tlist |= 1 << (mpidr & 0xf);
>> -
>> -		next_cpu = cpumask_next(cpu, mask);
>> -		if (next_cpu >= nr_cpu_ids)
>> -			goto out;
>> -		cpu = next_cpu;
>> -
>> -		mpidr = cpu_logical_map(cpu);
>> -
>> -		if (cluster_id != (mpidr & ~0xffUL)) {
>> -			cpu--;
>> -			goto out;
>> -		}
>> -	}
>> -out:
>> -	*base_cpu = cpu;
>> -	return tlist;
>> -}
>> -
>>  #define MPIDR_TO_SGI_AFFINITY(cluster_id, level) \
>>  	(MPIDR_AFFINITY_LEVEL(cluster_id, level) \
>>  		<< ICC_SGI1R_AFFINITY_## level ##_SHIFT)
>> +#define MPIDR_RS(mpidr)			(((mpidr) & 0xF0) >> 4)
>> +#define MPIDR_TO_SGI_TARGETLIST(mpidr)	(1 << ((mpidr) & 0xF))
>> +#define MPIDR_TO_SGI_CLUSTER_ID(mpidr)	(mpidr & ~0xF)
>> +#define MPIDR_TO_SGI_RS(mpidr)		(MPIDR_RS(mpidr) << ICC_SGI1R_RS_SHIFT)
>>  
>>  static void gic_send_sgi(u64 cluster_id, u16 tlist, unsigned int irq)
>>  {
>> @@ -600,6 +583,7 @@ static void gic_send_sgi(u64 cluster_id, u16 tlist, unsigned int irq)
>>  	       MPIDR_TO_SGI_AFFINITY(cluster_id, 2)	|
>>  	       irq << ICC_SGI1R_SGI_ID_SHIFT		|
>>  	       MPIDR_TO_SGI_AFFINITY(cluster_id, 1)	|
>> +	       MPIDR_TO_SGI_RS(cluster_id)		|
>>  	       tlist << ICC_SGI1R_TARGET_LIST_SHIFT);
>>  
>>  	pr_debug("CPU%d: ICC_SGI1R_EL1 %llx\n", smp_processor_id(), val);
>> @@ -620,10 +604,27 @@ static void gic_raise_softirq(const struct cpumask *mask, unsigned int irq)
>>  	smp_wmb();
>>  
>>  	for_each_cpu(cpu, mask) {
>> -		unsigned long cluster_id = cpu_logical_map(cpu) & ~0xffUL;
>> -		u16 tlist;
>> +		u64 mpidr = cpu_logical_map(cpu);
>> +		u64 cluster_id = MPIDR_TO_SGI_CLUSTER_ID(mpidr);
>> +		u16 tlist = 0;
>> +		int next_cpu;
>> +
>> +		if (WARN_ON((!gic_data.has_rss) && MPIDR_RS(mpidr)))
>> +			continue;
>> +
>> +		/* Prepare TARGETLIST which have the same cluster_id */
>> +		do {
>> +			tlist |= MPIDR_TO_SGI_TARGETLIST(mpidr);
>> +			next_cpu = cpumask_next(cpu, mask);
>> +			if (next_cpu >= nr_cpu_ids)
>> +				break;
>> +
>> +			mpidr = cpu_logical_map(next_cpu);
>> +			if (cluster_id != MPIDR_TO_SGI_CLUSTER_ID(mpidr))
>> +				break;
>> +			cpu = next_cpu;
>> +		} while (true);
> 
> I really dislike this. Why can't you just adapt the function we already
> have instead of making the patch pointlessly hard to review? How about
> something like this (completely untested and probably buggy, but you'll
> get the idea):
> 
> diff --git a/drivers/irqchip/irq-gic-v3.c b/drivers/irqchip/irq-gic-v3.c
> index c4bec908ccea..d0539155ebae 100644
> --- a/drivers/irqchip/irq-gic-v3.c
> +++ b/drivers/irqchip/irq-gic-v3.c
> @@ -603,13 +603,6 @@ static u16 gic_compute_target_list(int *base_cpu, const struct cpumask *mask,
>  	u16 tlist = 0;
>  
>  	while (cpu < nr_cpu_ids) {
> -		/*
> -		 * If we ever get a cluster of more than 16 CPUs, just
> -		 * scream and skip that CPU.
> -		 */
> -		if (WARN_ON((mpidr & 0xff) >= 16))
> -			goto out;
> -
>  		tlist |= 1 << (mpidr & 0xf);
>  
>  		next_cpu = cpumask_next(cpu, mask);
> @@ -633,7 +626,7 @@ static u16 gic_compute_target_list(int *base_cpu, const struct cpumask *mask,
>  	(MPIDR_AFFINITY_LEVEL(cluster_id, level) \
>  		<< ICC_SGI1R_AFFINITY_## level ##_SHIFT)
>  
> -static void gic_send_sgi(u64 cluster_id, u16 tlist, unsigned int irq)
> +static void gic_send_sgi(u64 cluster_id, u64 rs, u16 tlist, unsigned int irq)
>  {
>  	u64 val;
>  
> @@ -641,6 +634,7 @@ static void gic_send_sgi(u64 cluster_id, u16 tlist, unsigned int irq)
>  	       MPIDR_TO_SGI_AFFINITY(cluster_id, 2)	|
>  	       irq << ICC_SGI1R_SGI_ID_SHIFT		|
>  	       MPIDR_TO_SGI_AFFINITY(cluster_id, 1)	|
> +	       rs << ICC_SGI1R_RS_SHIFT		|
>  	       tlist << ICC_SGI1R_TARGET_LIST_SHIFT);
>  
>  	pr_debug("CPU%d: ICC_SGI1R_EL1 %llx\n", smp_processor_id(), val);
> @@ -662,10 +656,11 @@ static void gic_raise_softirq(const struct cpumask *mask, unsigned int irq)
>  
>  	for_each_cpu(cpu, mask) {
>  		unsigned long cluster_id = cpu_logical_map(cpu) & ~0xffUL;
> +		u64 rs = (cpu_logical_map(cpu) >> 4) & 0xfUL;
>  		u16 tlist;
>  
>  		tlist = gic_compute_target_list(&cpu, mask, cluster_id);
> -		gic_send_sgi(cluster_id, tlist, irq);
> +		gic_send_sgi(cluster_id, rs, tlist, irq);
>  	}
>  
>  	/* Force the above writes to ICC_SGI1R_EL1 to be executed */
> 
> It'd definitely look much better and would be easier to review.
> 

I tried to avoid the function call overhead by changing to inline and fix one
coding bug. We're writing SGI1R register with tlist=0 when function 
gic_send_sgi() fails to return valid tlist.
  
Anyway, I'll keep the original function and add RS support in v2 patch.

>>  
>> -		tlist = gic_compute_target_list(&cpu, mask, cluster_id);
>>  		gic_send_sgi(cluster_id, tlist, irq);
>>  	}
>>  
>> @@ -959,6 +960,10 @@ static int __init gic_init_bases(void __iomem *dist_base,
>>  		goto out_free;
>>  	}
>>  
>> +	gic_data.has_rss = !!(typer & GICD_TYPER_RSS);
>> +	pr_info("Distributor has %s Range Selector support\n",
>> +		gic_data.has_rss ? "" : "no");
>> +
>>  	set_handle_irq(gic_handle_irq);
>>  
>>  	if (IS_ENABLED(CONFIG_ARM_GIC_V3_ITS) && gic_dist_supports_lpis())
>> diff --git a/include/linux/irqchip/arm-gic-v3.h b/include/linux/irqchip/arm-gic-v3.h
>> index 6a1f87f..eb9ff9b 100644
>> --- a/include/linux/irqchip/arm-gic-v3.h
>> +++ b/include/linux/irqchip/arm-gic-v3.h
>> @@ -68,6 +68,7 @@
>>  #define GICD_CTLR_ENABLE_SS_G1		(1U << 1)
>>  #define GICD_CTLR_ENABLE_SS_G0		(1U << 0)
>>  
>> +#define GICD_TYPER_RSS			(1U << 26)
>>  #define GICD_TYPER_LPIS			(1U << 17)
>>  #define GICD_TYPER_MBIS			(1U << 16)
>>  
>> @@ -377,6 +378,7 @@
>>  #define ICC_CTLR_EL1_SEIS_MASK		(0x1 << ICC_CTLR_EL1_SEIS_SHIFT)
>>  #define ICC_CTLR_EL1_A3V_SHIFT		15
>>  #define ICC_CTLR_EL1_A3V_MASK		(0x1 << ICC_CTLR_EL1_A3V_SHIFT)
>> +#define ICC_CTLR_EL1_RSS		(0x1 << 18)
>>  #define ICC_PMR_EL1_SHIFT		0
>>  #define ICC_PMR_EL1_MASK		(0xff << ICC_PMR_EL1_SHIFT)
>>  #define ICC_BPR0_EL1_SHIFT		0
>> @@ -465,6 +467,8 @@
>>  #define ICC_SGI1R_AFFINITY_2_SHIFT	32
>>  #define ICC_SGI1R_AFFINITY_2_MASK	(0xffULL << ICC_SGI1R_AFFINITY_2_SHIFT)
>>  #define ICC_SGI1R_IRQ_ROUTING_MODE_BIT	40
>> +#define ICC_SGI1R_RS_SHIFT		44
>> +#define ICC_SGI1R_RS_MASK		(0xfULL << ICC_SGI1R_RS_SHIFT)
>>  #define ICC_SGI1R_AFFINITY_3_SHIFT	48
>>  #define ICC_SGI1R_AFFINITY_3_MASK	(0xffULL << ICC_SGI1R_AFFINITY_3_SHIFT)
> 
> Thanks,
> 
> 	M.
> 

-- 
Shanker Donthineni
Qualcomm Datacenter Technologies, Inc. as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project.