Re: [PATCH v2 1/2] drm/xe/guc: Add support for w/a KLVs

[Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>]

On 3/20/2024 1:40 PM, John Harrison wrote:
> On 3/19/2024 09:08, Badal Nilawar wrote:
>> To prevent running out of bits, new w/a enable flags are being added
>> via a KLV system instead of a 32 bit flags word.
>>
>> v2: GuC version check > 70.10 is not needed as xe will not be supporting
>>      anything below < 70.19 (John Harrison)
>>
>> Cc: John Harrison <John.C.Harrison@intel.com>
>> Signed-off-by: Badal Nilawar <badal.nilawar@intel.com>
>> ---
>>   drivers/gpu/drm/xe/xe_guc_ads.c       | 61 ++++++++++++++++++++++++++-
>>   drivers/gpu/drm/xe/xe_guc_ads_types.h |  2 +
>>   drivers/gpu/drm/xe/xe_guc_fwif.h      |  5 ++-
>>   3 files changed, 65 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/xe/xe_guc_ads.c 
>> b/drivers/gpu/drm/xe/xe_guc_ads.c
>> index 6ad4c1a90a78..506f0be35763 100644
>> --- a/drivers/gpu/drm/xe/xe_guc_ads.c
>> +++ b/drivers/gpu/drm/xe/xe_guc_ads.c
>> @@ -80,6 +80,10 @@ ads_to_map(struct xe_guc_ads *ads)
>>    *      +---------------------------------------+
>>    *      | padding                               |
>>    *      +---------------------------------------+ <== 4K aligned
>> + *      | w/a KLVs                              |
>> + *      +---------------------------------------+
>> + *      | padding                               |
>> + *      +---------------------------------------+ <== 4K aligned
>>    *      | capture lists                         |
>>    *      +---------------------------------------+
>>    *      | padding                               |
>> @@ -131,6 +135,11 @@ static size_t guc_ads_golden_lrc_size(struct 
>> xe_guc_ads *ads)
>>       return PAGE_ALIGN(ads->golden_lrc_size);
>>   }
>>   +static u32 guc_ads_waklv_size(struct xe_guc_ads *ads)
>> +{
>> +    return PAGE_ALIGN(ads->ads_waklv_size);
>> +}
>> +
>>   static size_t guc_ads_capture_size(struct xe_guc_ads *ads)
>>   {
>>       /* FIXME: Allocate a proper capture list */
>> @@ -167,12 +176,22 @@ static size_t guc_ads_golden_lrc_offset(struct 
>> xe_guc_ads *ads)
>>       return PAGE_ALIGN(offset);
>>   }
>>   +static size_t guc_ads_waklv_offset(struct xe_guc_ads *ads)
>> +{
>> +    u32 offset;
>> +
>> +    offset = guc_ads_golden_lrc_offset(ads) +
>> +         guc_ads_golden_lrc_size(ads);
>> +
>> +    return PAGE_ALIGN(offset);
>> +}
>> +
>>   static size_t guc_ads_capture_offset(struct xe_guc_ads *ads)
>>   {
>>       size_t offset;
>>   -    offset = guc_ads_golden_lrc_offset(ads) +
>> -        guc_ads_golden_lrc_size(ads);
>> +    offset = guc_ads_waklv_offset(ads) +
>> +         guc_ads_waklv_size(ads);
>>         return PAGE_ALIGN(offset);
>>   }
>> @@ -260,6 +279,41 @@ static size_t calculate_golden_lrc_size(struct 
>> xe_guc_ads *ads)
>>       return total_size;
>>   }
>>   +static void guc_waklv_init(struct xe_guc_ads *ads)
>> +{
>> +    u32 addr_ggtt, offset, remain, size;
>> +
>> +    offset = guc_ads_waklv_offset(ads);
>> +    remain = guc_ads_waklv_size(ads);
>> +
>> +    /*
>> +     * Add workarounds here:
>> +     *
>> +     * if (want_wa_<name>) {
>> +     *      size = guc_waklv_<name>(guc, offset, remain);
>> +     *      offset += size;
>> +     *      remain -= size;
>> +     * }
>> +     */
>> +
>> +    size = guc_ads_waklv_size(ads) - remain;
>> +    if (!size)
>> +        return;
>> +
>> +    offset = guc_ads_waklv_offset(ads);
>> +    addr_ggtt = xe_bo_ggtt_addr(ads->bo) + offset;
>> +
>> +    ads_blob_write(ads, ads.wa_klv_addr_lo, addr_ggtt);
>> +    ads_blob_write(ads, ads.wa_klv_addr_hi, 0);
> @Daniele is pushing a change for supporting 64bit addresses (for 
> future compatibility) - https://patchwork.freedesktop.org/series/131282/
>
> Is there any value to changing this address to 64bit as well and 
> changing the hard 0 to upper_32(addr) instead?

For consistency it'd be nice to do so, but I'm ok with just hard-coding 
a zero in the upper bits if the variable is a u32; the compiler should 
warn us if we ever switch xe_bo_ggtt_addr to return u64 so we can change 
it up at that point if needed. The things I want to avoid are either 
using upper_32_bits() on a u32 or hard-coding a 0 in the upper bits if 
the variable is a u64, and this patch does none of those things.

Daniele

>
> John.
>
>
>> +    ads_blob_write(ads, ads.wa_klv_size, size);
>> +}
>> +
>> +static int calculate_waklv_size(struct xe_guc_ads *ads)
>> +{
>> +    /* Fudge something chunky for now: */
>> +    return PAGE_SIZE;
>> +}
>> +
>>   #define MAX_GOLDEN_LRC_SIZE    (SZ_4K * 64)
>>     int xe_guc_ads_init(struct xe_guc_ads *ads)
>> @@ -271,6 +325,7 @@ int xe_guc_ads_init(struct xe_guc_ads *ads)
>>         ads->golden_lrc_size = calculate_golden_lrc_size(ads);
>>       ads->regset_size = calculate_regset_size(gt);
>> +    ads->ads_waklv_size = calculate_waklv_size(ads);
>>         bo = xe_managed_bo_create_pin_map(xe, tile, guc_ads_size(ads) 
>> + MAX_GOLDEN_LRC_SIZE,
>>                         XE_BO_CREATE_SYSTEM_BIT |
>> @@ -597,6 +652,8 @@ void xe_guc_ads_populate(struct xe_guc_ads *ads)
>>       guc_mapping_table_init(gt, &info_map);
>>       guc_capture_list_init(ads);
>>       guc_doorbell_init(ads);
>> +    /* Workaround KLV list */
>> +    guc_waklv_init(ads);
>>         if (xe->info.has_usm) {
>>           guc_um_init_params(ads);
>> diff --git a/drivers/gpu/drm/xe/xe_guc_ads_types.h 
>> b/drivers/gpu/drm/xe/xe_guc_ads_types.h
>> index 4afe44bece4b..62235b2a6fe3 100644
>> --- a/drivers/gpu/drm/xe/xe_guc_ads_types.h
>> +++ b/drivers/gpu/drm/xe/xe_guc_ads_types.h
>> @@ -20,6 +20,8 @@ struct xe_guc_ads {
>>       size_t golden_lrc_size;
>>       /** @regset_size: size of register set passed to GuC for 
>> save/restore */
>>       u32 regset_size;
>> +    /** @ads_waklv_size: waklv size */
>> +    u32 ads_waklv_size;
>>   };
>>     #endif
>> diff --git a/drivers/gpu/drm/xe/xe_guc_fwif.h 
>> b/drivers/gpu/drm/xe/xe_guc_fwif.h
>> index c281fdbfd2d6..52503719d2aa 100644
>> --- a/drivers/gpu/drm/xe/xe_guc_fwif.h
>> +++ b/drivers/gpu/drm/xe/xe_guc_fwif.h
>> @@ -207,7 +207,10 @@ struct guc_ads {
>>       u32 
>> capture_instance[GUC_CAPTURE_LIST_INDEX_MAX][GUC_MAX_ENGINE_CLASSES];
>>       u32 
>> capture_class[GUC_CAPTURE_LIST_INDEX_MAX][GUC_MAX_ENGINE_CLASSES];
>>       u32 capture_global[GUC_CAPTURE_LIST_INDEX_MAX];
>> -    u32 reserved[14];
>> +    u32 wa_klv_addr_lo;
>> +    u32 wa_klv_addr_hi;
>> +    u32 wa_klv_size;
>> +    u32 reserved[11];
>>   } __packed;
>>     /* Engine usage stats */
>