[Intel-gfx] [PATCH 18/22] drm/i915/tgl: Define MOCS entries for Tigerlake

Thu Jul 25 18:26:26 UTC 2019

On 2019-07-25 00:32, Lucas De Marchi wrote:
> On Thu, Jul 18, 2019 at 10:09:27AM -0700, Daniele Ceraolo Spurio wrote:
>>
>>
>> On 7/18/19 6:08 AM, Ville Syrjälä wrote:
>>> On Fri, Jul 12, 2019 at 06:09:36PM -0700, Lucas De Marchi wrote:
>>>> From: Tomasz Lis <tomasz.lis at intel.com>
>>>>
>>>> The MOCS table is published as part of bspec, and versioned. Entries
>>>> are supposed to never be modified, but new ones can be added. Adding
>>>> entries increases table version. The patch includes version 1 entries.
>>>>
>>>> Two of the 3 legacy entries used for gen9 are no longer expected to 
>>>> work.
>>>> Although we are changing the gen11 table, those changes are 
>>>> supposed to
>>>> be backward compatible since we are only touching previously undefined
>>>> entries.
>>>>
>>>> Cc: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
>>>> Cc: Mika Kuoppala <mika.kuoppala at intel.com>
>>>> Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio at intel.com>
>>>> Signed-off-by: Tomasz Lis <tomasz.lis at intel.com>
>>>> Signed-off-by: Lucas De Marchi <lucas.demarchi at intel.com>
>>>> ---
>>>>  drivers/gpu/drm/i915/gt/intel_mocs.c | 25 ++++++++++++++++++++++---
>>>>  1 file changed, 22 insertions(+), 3 deletions(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/i915/gt/intel_mocs.c 
>>>> b/drivers/gpu/drm/i915/gt/intel_mocs.c
>>>> index 290a5e9b90b9..259e7bec0a63 100644
>>>> --- a/drivers/gpu/drm/i915/gt/intel_mocs.c
>>>> +++ b/drivers/gpu/drm/i915/gt/intel_mocs.c
>>>> @@ -62,6 +62,10 @@ struct drm_i915_mocs_table {
>>>>  #define GEN11_NUM_MOCS_ENTRIES    64  /* 63-64 are reserved, but 
>>>> configured. */
>>>>  /* (e)LLC caching options */
>>>> +/*
>>>> + * Note: LE_0_PAGETABLE works only up to Gen11; for newer gens it 
>>>> means
>>>> + * the same as LE_UC
>>>> + */
>>>>  #define LE_0_PAGETABLE        _LE_CACHEABILITY(0)
>>>>  #define LE_1_UC            _LE_CACHEABILITY(1)
>>>>  #define LE_2_WT            _LE_CACHEABILITY(2)
>>>> @@ -100,8 +104,9 @@ struct drm_i915_mocs_table {
>>>>   * of bspec.
>>>>   *
>>>>   * Entries not part of the following tables are undefined as far as
>>>> - * userspace is concerned and shouldn't be relied upon. For the time
>>>> - * being they will be initialized to PTE.
>>>> + * userspace is concerned and shouldn't be relied upon. For Gen < 12
>>>> + * they will be initialized to PTE. Gen >= 12 onwards don't have a 
>>>> setting for
>>>> + * PTE. We use the same value, but that actually means Uncached.
>>>>   *
>>>>   * The last two entries are reserved by the hardware. For ICL+ they
>>>>   * should be initialized according to bspec and never used, for older
>>>> @@ -137,11 +142,13 @@ static const struct drm_i915_mocs_entry 
>>>> broxton_mocs_table[] = {
>>>>  };
>>>>  #define GEN11_MOCS_ENTRIES \
>>>> -    /* Base - Uncached (Deprecated) */ \
>>>> +    /* Gen11: Base - Uncached (Deprecated) */ \
>>>> +    /* Gen12+: Base - Error (Reserved for Non-Use) */ \
>>>>      MOCS_ENTRY(I915_MOCS_UNCACHED, \
>>>>             LE_1_UC | LE_TC_1_LLC, \
>>>>             L3_1_UC), \
>>>>      /* Base - L3 + LeCC:PAT (Deprecated) */ \
>>>> +    /* Gen12+: Base - Reserved */ \
>>>>      MOCS_ENTRY(I915_MOCS_PTE, \
>>>>             LE_0_PAGETABLE | LE_TC_1_LLC, \
>>>>             L3_3_WB), \
>>>> @@ -233,6 +240,18 @@ static const struct drm_i915_mocs_entry 
>>>> broxton_mocs_table[] = {
>>>>      MOCS_ENTRY(23, \
>>>>             LE_3_WB | LE_TC_1_LLC | LE_LRUM(3) | LE_RSC(1) | 
>>>> LE_SCC(7), \
>>>>             L3_3_WB), \
>>>> +    /* Gen12+: HW Reserved - HDC:L1 + L3 + LLC */ \
>>>
>>> Why is this marked as reserved? From the looks of things 48-61 should
>>> just be normal entries that userspace can select to get HDC L1$. And
>>> looks like icl already has that stuff. So someone should probably 
>>> figure
>>> out if Mesa/etc. can make use of the HDC L1$, and if so we should add
>>> the relevant MOCS entries for icl as well.
>>
>> Here the reserved terminology is indeed misleading. The 48-59 range 
>> is a "special" range where L1 usage is implicitly enabled by the HW, 
>> as there is no explicit L1 toggle in the MOCS registers. The reserved 
>> here means that the range shouldn't be used for "normal" MOCS 
>> settings, but SW can freely use these entries as needed. Similarly, 
>> MOCS 60 and 61 are reserved for other special purposes, but are still 
>> usable by SW. The only entries SW shouldn't touch are 62 and 63.
>>
>> Regarding ICL, Gen11 HW doesn't have the capability so no new entries 
>> are required there.
It might be a good idea to find a better word for it than "reserved". 
Not only for this patch, but to be used everywhere, especially the spec.
But that exceeds the scope of this review.
>>
>>>
>>>> +    MOCS_ENTRY(48, \
>>>> +           LE_3_WB | LE_TC_1_LLC | LE_LRUM(3), \
>>>> +           L3_3_WB), \
>>>> +    /* Gen12+: HW Reserved - HW Special Case (CCS) */ \
>>
>> The specs have MOCS 49-51 defined as well.
>
> humn... it seems they got added later.
>
> I'm not sure anymore if we should update igt so it doesn't expect those
> entries to be set to PTE or if we should stop reusing the same table for
> ICL and TGL. Spec doesn't mention the compatibility of this table with
> gen 11 anymore. Thoughts?
>
>
It doesn't sound right to change the implementation decision to keep them
together, only because this allows to keep tests intact.

I don't believe there's a reason to verify undefined entries in IGT.
Sometimes we do want to verify our specific implementation instead of
pure specs compliance; but I don't see a reason for this should be the 
case here.

The existing MOCS entries are supposed to be unchangeable (in the boundary
of specific platform). So the chances of having to split the tables in 
the future are
low (both tables would have to define V2 entries at same indexes, and define
them differently).

-Tomasz

> Lucas De Marchi
>
>>
>> Daniele
>>
>>>> +    MOCS_ENTRY(60, \
>>>> +           LE_3_WB | LE_TC_1_LLC | LE_LRUM(3), \
>>>> +           L3_1_UC), \
>>>> +    /* Gen12+: HW Reserved - HW Special Case (Displayable) */ \
>>>> +    MOCS_ENTRY(61, \
>>>> +           LE_1_UC | LE_TC_1_LLC | LE_SCF(1), \
>>>> +           L3_3_WB), \
>>>>      /* HW Reserved - SW program but never use */ \
>>>>      MOCS_ENTRY(62, \
>>>>             LE_3_WB | LE_TC_1_LLC | LE_LRUM(3), \
>>>> -- 
>>>> 2.21.0
>>>