2017-03-01 16:25:03

by Thomas Gleixner

[permalink] [raw]
Subject: Re: [PATCH 4/8] x86/intel_rct/mba: Add MBA structures and initialize MBA

On Fri, 17 Feb 2017, Vikas Shivappa wrote:
> --- a/arch/x86/include/asm/intel_rdt.h
> +++ b/arch/x86/include/asm/intel_rdt.h
> @@ -11,6 +11,9 @@
> #define IA32_L3_QOS_CFG 0xc81
> #define IA32_L3_CBM_BASE 0xc90
> #define IA32_L2_CBM_BASE 0xd10
> +#define IA32_MBA_THRTL_BASE 0xd50
> +#define MAX_MBA_THRTL 100u
> +#define MBA_IS_LINEAR 0x4

I have a hard time to figure out how the latter two constants are related
to this list of registers. MBA_IS_LINEAR is used to check the CPUID bit and
MAX_MBA_THRTL is obviously a pure software constant because with a
non-linear scale the maximum value is not 100.

Just slapping defines to random places is equally bad as using hard coded
constants.

> +/*
> + * rdt_get_mb_table() - get a mapping of b/w percentage values
> + * exposed to user interface and the h/w understandable delay values.
> + *
> + * The non-linear delay values have the granularity of power of two
> + * and also the h/w does not guarantee a curve for configured delay
> + * values vs. actual b/w throttled.
> + * Hence we need a mapping that is pre caliberated for user to express
> + * the b/w in terms of any sensible number.

... calibrated so the user can express the bandwidth as a percentage value.

> +static inline int rdt_get_mb_table(struct rdt_resource *r)
> +{
> + /*
> + * There are no Intel SKUs as of now to support non-linear delay.
> + */
> + r->mb_map = NULL;

What's the point of setting this to NULL?

Also it would be helpful to emit log info here so people don't have to
start digging around.

pr_info("Bandwidth map not implemented for ....", ... model);

> +
> + return -ENODEV;

Returning -ENODEV to a function which just returns a boolean value is
pointless.

> static void rdt_get_cache_config(int idx, struct rdt_resource *r)
> {
> union cpuid_0x10_1_eax eax;
> @@ -184,9 +237,8 @@ static inline bool get_rdt_resources(void)
> ret = true;
> }
>
> - if (boot_cpu_has(X86_FEATURE_MBA)) {
> - ret = true;
> - }
> + if (boot_cpu_has(X86_FEATURE_MBA))
> + ret = rdt_get_mem_config(&rdt_resources_all[RDT_RESOURCE_MBA]);

Groan. When rdt_get_mem_config() returns false (because the map is not
implemented), then the whole function returns false and CAT is disabled.

> +static int domain_setup_ctrlval(struct rdt_resource *r, struct rdt_domain *d)
> +{
> + int i;
> +
> + d->ctrl_val = kmalloc_array(r->num_closid,
> + sizeof(*d->ctrl_val), GFP_KERNEL);
> + if (!d->ctrl_val)
> + return -ENOMEM;
> +
> + /*
> + * Initialize the Control MSRs to having no control.
> + * For Cache Allocation: Set all bits in cbm
> + * For Memory Allocation: Set b/w requested to 100
> + */
> + for (i = 0; i < r->num_closid; i++) {
> + int idx = cbm_idx(r, i);
> +
> + d->ctrl_val[i] = r->default_ctrl;
> + wrmsrl(r->msr_base + idx, d->ctrl_val[i]);
> + }

So if you use a local pointer for that, this whole mess becomes readable.

static int domain_setup_ctrlval(struct rdt_resource *r, struct rdt_domain *d)
{
u32 *p;
int i;

p = kmalloc_array(r->num_closid, sizeof(*d->ctrl_val), GFP_KERNEL);
if (!p)
return -ENOMEM;

d->ctrl_val = p;

/* Initialize the Control MSRs to the default value */
for (i = 0; i < r->num_closid; i++, p++) {
int idx = cbm_idx(r, i);

*p = r->default_ctrl;
wrmsrl(r->msr_base + idx, *p);
}
> +
> + return 0;
> +}

> static void domain_add_cpu(int cpu, struct rdt_resource *r)
> {
> - int i, id = get_cache_id(cpu, r->cache_level);
> + int id = get_cache_id(cpu, r->cache_level), ret;

Bah. If you have the same type in one line, then please move the
uninitialized variables to the front.

int ret, id = get_cache_id(cpu, r->cache_level);

But a s/i/ret/ would have been to simple and kept the code readable.

> @@ -298,19 +374,12 @@ static void domain_add_cpu(int cpu, struct rdt_resource *r)
>
> d->id = id;
>
> - d->ctrl_val = kmalloc_array(r->num_closid, sizeof(*d->ctrl_val), GFP_KERNEL);
> - if (!d->ctrl_val) {
> + ret = domain_setup_ctrlval(r, d);
> + if (ret) {
> kfree(d);
> return;
> }

What's the point of this 'ret' variable if the function is void?

if (domain_setup_ctrlval(r, d)) {
kfree(d);
return;
}

would have been to easy to read, right?

Thanks,

tglx


2017-03-10 21:50:57

by Shivappa Vikas

[permalink] [raw]
Subject: Re: [PATCH 4/8] x86/intel_rct/mba: Add MBA structures and initialize MBA



On Wed, 1 Mar 2017, Thomas Gleixner wrote:

> On Fri, 17 Feb 2017, Vikas Shivappa wrote:
>> --- a/arch/x86/include/asm/intel_rdt.h
>> +++ b/arch/x86/include/asm/intel_rdt.h
>> @@ -11,6 +11,9 @@
>> #define IA32_L3_QOS_CFG 0xc81
>> #define IA32_L3_CBM_BASE 0xc90
>> #define IA32_L2_CBM_BASE 0xd10
>> +#define IA32_MBA_THRTL_BASE 0xd50
>> +#define MAX_MBA_THRTL 100u
>> +#define MBA_IS_LINEAR 0x4
>
> I have a hard time to figure out how the latter two constants are related
> to this list of registers. MBA_IS_LINEAR is used to check the CPUID bit and
> MAX_MBA_THRTL is obviously a pure software constant because with a
> non-linear scale the maximum value is not 100.
>
> Just slapping defines to random places is equally bad as using hard coded
> constants.
>
>> +/*
>> + * rdt_get_mb_table() - get a mapping of b/w percentage values
>> + * exposed to user interface and the h/w understandable delay values.
>> + *
>> + * The non-linear delay values have the granularity of power of two
>> + * and also the h/w does not guarantee a curve for configured delay
>> + * values vs. actual b/w throttled.
>> + * Hence we need a mapping that is pre caliberated for user to express
>> + * the b/w in terms of any sensible number.
>
> ... calibrated so the user can express the bandwidth as a percentage value.
>
>> +static inline int rdt_get_mb_table(struct rdt_resource *r)
>> +{
>> + /*
>> + * There are no Intel SKUs as of now to support non-linear delay.
>> + */
>> + r->mb_map = NULL;
>
> What's the point of setting this to NULL?
>
> Also it would be helpful to emit log info here so people don't have to
> start digging around.
>
> pr_info("Bandwidth map not implemented for ....", ... model);
>
>> +
>> + return -ENODEV;
>
> Returning -ENODEV to a function which just returns a boolean value is
> pointless.
>
>> static void rdt_get_cache_config(int idx, struct rdt_resource *r)
>> {
>> union cpuid_0x10_1_eax eax;
>> @@ -184,9 +237,8 @@ static inline bool get_rdt_resources(void)
>> ret = true;
>> }
>>
>> - if (boot_cpu_has(X86_FEATURE_MBA)) {
>> - ret = true;
>> - }
>> + if (boot_cpu_has(X86_FEATURE_MBA))
>> + ret = rdt_get_mem_config(&rdt_resources_all[RDT_RESOURCE_MBA]);
>
> Groan. When rdt_get_mem_config() returns false (because the map is not
> implemented), then the whole function returns false and CAT is disabled.
>
>> +static int domain_setup_ctrlval(struct rdt_resource *r, struct rdt_domain *d)
>> +{
>> + int i;
>> +
>> + d->ctrl_val = kmalloc_array(r->num_closid,
>> + sizeof(*d->ctrl_val), GFP_KERNEL);
>> + if (!d->ctrl_val)
>> + return -ENOMEM;
>> +
>> + /*
>> + * Initialize the Control MSRs to having no control.
>> + * For Cache Allocation: Set all bits in cbm
>> + * For Memory Allocation: Set b/w requested to 100
>> + */
>> + for (i = 0; i < r->num_closid; i++) {
>> + int idx = cbm_idx(r, i);
>> +
>> + d->ctrl_val[i] = r->default_ctrl;
>> + wrmsrl(r->msr_base + idx, d->ctrl_val[i]);
>> + }
>
> So if you use a local pointer for that, this whole mess becomes readable.
>
> static int domain_setup_ctrlval(struct rdt_resource *r, struct rdt_domain *d)
> {
> u32 *p;
> int i;
>
> p = kmalloc_array(r->num_closid, sizeof(*d->ctrl_val), GFP_KERNEL);
> if (!p)
> return -ENOMEM;
>
> d->ctrl_val = p;
>
> /* Initialize the Control MSRs to the default value */
> for (i = 0; i < r->num_closid; i++, p++) {
> int idx = cbm_idx(r, i);
>
> *p = r->default_ctrl;
> wrmsrl(r->msr_base + idx, *p);
> }
>> +
>> + return 0;
>> +}
>
>> static void domain_add_cpu(int cpu, struct rdt_resource *r)
>> {
>> - int i, id = get_cache_id(cpu, r->cache_level);
>> + int id = get_cache_id(cpu, r->cache_level), ret;
>
> Bah. If you have the same type in one line, then please move the
> uninitialized variables to the front.
>
> int ret, id = get_cache_id(cpu, r->cache_level);
>
> But a s/i/ret/ would have been to simple and kept the code readable.
>
>> @@ -298,19 +374,12 @@ static void domain_add_cpu(int cpu, struct rdt_resource *r)
>>
>> d->id = id;
>>
>> - d->ctrl_val = kmalloc_array(r->num_closid, sizeof(*d->ctrl_val), GFP_KERNEL);
>> - if (!d->ctrl_val) {
>> + ret = domain_setup_ctrlval(r, d);
>> + if (ret) {
>> kfree(d);
>> return;
>> }
>
> What's the point of this 'ret' variable if the function is void?
>
> if (domain_setup_ctrlval(r, d)) {
> kfree(d);
> return;
> }
>
> would have been to easy to read, right?

Will fix all the issues pointed. Thanks for pointing out.

>
> Thanks,
>
> tglx
>