[PATCH v5 09/19] arch_topology: Use the last level cache information from the cacheinfo

Conor.Dooley at microchip.com Conor.Dooley at microchip.com
Fri Jul 1 07:47:16 PDT 2022


On 01/07/2022 12:11, Sudeep Holla wrote:
> EXTERNAL EMAIL: Do not click links or open attachments unless you know the content is safe
> 
> On Thu, Jun 30, 2022 at 10:07:49PM +0000, Conor.Dooley at microchip.com wrote:
>>
>>
>> On 30/06/2022 21:21, Sudeep Holla wrote:
>>> On Thu, Jun 30, 2022 at 08:13:55PM +0000, Conor.Dooley at microchip.com wrote:
>>>>
>>>> I didn't have the time to go digging into things, but the following
>>>> macro looked odd:
>>>> #define per_cpu_cacheinfo_idx(cpu, idx)            \
>>>>                             (per_cpu_cacheinfo(cpu) + (idx))
>>>> Maybe it is just badly named, but is this getting the per_cpu_cacheinfo
>>>> and then incrementing intentionally, or is it meant to get the
>>>> per_cpu_cacheinfo of cpu + idx?
>>>
>>> OK, basically per_cpu_cacheinfo(cpu) get the information for a cpu
>>> while per_cpu_cacheinfo_idx(cpu, idx) will fetch the information for a
>>> given cpu and given index within the cpu. So we are incrementing the
>>> pointer by the index. These work just fine on arm64 platform.
>>
>> Right, that's what I figured but wanted to be sure.
>>
> 
> OK
> 
>>>
>>> Not sure if compiler is optimising something as I still can't understand
>>> how we can end up with valid llc but fw_token as NULL.
>> See idk about that. The following fails to boot.
>> index 167abfa6f37d..9d45c37fb004 100644
>> --- a/drivers/base/cacheinfo.c
>> +++ b/drivers/base/cacheinfo.c
>> @@ -36,6 +36,8 @@ struct cpu_cacheinfo *get_cpu_cacheinfo(unsigned int cpu)
>>   static inline bool cache_leaves_are_shared(struct cacheinfo *this_leaf,
>>                                             struct cacheinfo *sib_leaf)
>>   {
>> +       if (!this_leaf || !sib_leaf)
>> +               return false;
> 
> Did you hit this ?

Ah I forgot to remove this, I had added it (since I knew a return value
of false was correct) but it was still failing to boot. It was my step
prior to adding the if(!llc...

> 
>>          /*
>>           * For non DT/ACPI systems, assume unique level 1 caches,
>>           * system-wide shared caches for all other levels. This will be used
>> @@ -74,8 +76,12 @@ bool last_level_cache_is_shared(unsigned int cpu_x, unsigned int cpu_y)
>>
>>          llc_x = per_cpu_cacheinfo_idx(cpu_x, cache_leaves(cpu_x) - 1);
>>          llc_y = per_cpu_cacheinfo_idx(cpu_y, cache_leaves(cpu_y) - 1);
>> +       if (!llc_x || !llc_y){
>> +               printk("llc was null\n");
> 
> Or this ?

This printk never prints out.

> 
>> +               return false;
>> +       }
>>
>> -       return cache_leaves_are_shared(llc_x, llc_y);
>> +       return false; //cache_leaves_are_shared(llc_x, llc_y);
> 
> Even the above change fails to boot ? Coz you are always returning false here
> too.

Correct, fails to boot.

> 
>>   }
>>
>>   #ifdef CONFIG_OF
>>
>> and this boots:
>>
>> diff --git a/drivers/base/cacheinfo.c b/drivers/base/cacheinfo.c
>> index 167abfa6f37d..01900908fe31 100644
>> --- a/drivers/base/cacheinfo.c
>> +++ b/drivers/base/cacheinfo.c
>> @@ -36,6 +36,8 @@ struct cpu_cacheinfo *get_cpu_cacheinfo(unsigned int cpu)
>>   static inline bool cache_leaves_are_shared(struct cacheinfo *this_leaf,
>>                                             struct cacheinfo *sib_leaf)
>>   {
>> +       if (!this_leaf || !sib_leaf)
>> +               return false;
>>          /*
>>           * For non DT/ACPI systems, assume unique level 1 caches,
>>           * system-wide shared caches for all other levels. This will be used
>> @@ -75,7 +77,7 @@ bool last_level_cache_is_shared(unsigned int cpu_x, unsigned int cpu_y)
>>          llc_x = per_cpu_cacheinfo_idx(cpu_x, cache_leaves(cpu_x) - 1);
>>          llc_y = per_cpu_cacheinfo_idx(cpu_y, cache_leaves(cpu_y) - 1);
>>
> 
> You are just missing the checks for llc_x and llc_y and it works which
> means llc_x and llc_y is where things are going wrong.
> 

Sounds about right to me.



More information about the linux-riscv mailing list