Re: [PATCH 4/9] drm/xe/pf: Use migration-friendly VRAM auto-provisioning

["Piotr Piórkowski" <piotr.piorkowski@intel.com>]

Michal Wajdeczko <michal.wajdeczko@intel.com> wrote on nie [2026-lut-15 21:33:18 +0100]:
> Instead of trying very hard to find the largest fair VRAM (aka LMEM)
> size that could be allocated for VFs on the current tile, pick some
> smaller rounded down to power-of-two value that is more likely to be
> provisioned in the same manner by the other PF instances.
> 
> In some cases, the outcome of above calculation might not be optimal,
> but it's expected that admin will do fine-tuning using sysfs files.
> 
> Signed-off-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c | 27 ++++++++++++++++++++++
>  1 file changed, 27 insertions(+)
> 
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
> index 23af49dc1bfa..43041af81518 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
> @@ -1919,6 +1919,26 @@ static u64 pf_estimate_fair_lmem(struct xe_gt *gt, unsigned int num_vfs)
>  	return fair;
>  }
>  
> +static u64 pf_profile_fair_lmem(struct xe_gt *gt, unsigned int num_vfs)
> +{
> +	struct xe_tile *tile = gt_to_tile(gt);
> +	bool admin_only_pf = xe_sriov_pf_admin_only(tile->xe);
> +	u64 usable = xe_vram_region_usable_size(tile->mem.vram);
> +	u64 shareable = ALIGN_DOWN(usable, SZ_1G);
> +	u64 alignment = pf_get_lmem_alignment(gt);
> +	u64 fair;
> +
> +	if (admin_only_pf)
> +		fair = div_u64(shareable, num_vfs);
> +	else
> +		fair = div_u64(shareable, 1 + num_vfs);
> +
> +	if (!admin_only_pf && fair)
> +		fair = rounddown_pow_of_two(fair);
> +
> +	return ALIGN_DOWN(fair, alignment);

I would add a hard guarantee that for admin_only_pf, PF will not be left without a minimum
amount of VRAM.
Currently, under certain conditions, it may run out of VRAM - although this is unlikely.
I have already discussed this with Michal offline and I am leaving it up to him to decide
what to do about it.
> +}
> +
>  /**
>   * xe_gt_sriov_pf_config_set_fair_lmem - Provision many VFs with fair LMEM.
>   * @gt: the &xe_gt (can't be media)
> @@ -1932,6 +1952,7 @@ static u64 pf_estimate_fair_lmem(struct xe_gt *gt, unsigned int num_vfs)
>  int xe_gt_sriov_pf_config_set_fair_lmem(struct xe_gt *gt, unsigned int vfid,
>  					unsigned int num_vfs)
>  {
> +	u64 profile;
>  	u64 fair;
>  
>  	xe_gt_assert(gt, vfid);
> @@ -1948,6 +1969,12 @@ int xe_gt_sriov_pf_config_set_fair_lmem(struct xe_gt *gt, unsigned int vfid,
>  	if (!fair)
>  		return -ENOSPC;
>  
> +	profile = pf_profile_fair_lmem(gt, num_vfs);
> +	fair = min(fair, profile);
> +	if (fair < profile)
> +		xe_gt_sriov_info(gt, "Using non-profile provisioning (%s %llu vs %llu)\n",
> +				 "VRAM", fair, profile);
> +
>  	return xe_gt_sriov_pf_config_bulk_set_lmem(gt, vfid, num_vfs, fair);
>  }
>  
Apart from what I mentioned earlier, the code is fine for me,
In the worst case (highly unlikely), provisioning will just fail.
So still:
Reviewed-by: Piotr Piórkowski <piotr.piorkowski@intel.com>



> -- 
> 2.47.1
> 

--