[AMD Official Use Only - AMD Internal Distribution Only]
> -----Original Message-----
> From: amd-gfx <amd-gfx-bounces@xxxxxxxxxxxxxxxxxxxxx> On Behalf Of Hawking
> Zhang
> Sent: Tuesday, May 28, 2024 10:21 AM
> To: amd-gfx@xxxxxxxxxxxxxxxxxxxxx
> Cc: Zhou1, Tao <Tao.Zhou1@xxxxxxx>; Kuehling, Felix
> <Felix.Kuehling@xxxxxxx>; Kasiviswanathan, Harish
> <Harish.Kasiviswanathan@xxxxxxx>; Zhang, Hawking
> <Hawking.Zhang@xxxxxxx>
> Subject: [PATCH] drm/amdgpu: Estimate RAS reservation when report capacity
>
> Add estimate of how much vram we need to reserve for RAS when caculating the
> total available vram.
>
> Signed-off-by: Hawking Zhang <Hawking.Zhang@xxxxxxx>
> ---
> .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 9 +++++++--
> drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 18 ++++++++++++++++++
> drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h | 2 ++
> 3 files changed, 27 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> index e98927529f61..ad813772f8a1 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> @@ -173,6 +173,8 @@ int amdgpu_amdkfd_reserve_mem_limit(struct
> amdgpu_device *adev, {
> uint64_t reserved_for_pt =
> ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
> + struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
> + uint64_t reserved_for_ras = (con ? con->reserved_pages_in_bytes : 0);
> size_t system_mem_needed, ttm_mem_needed, vram_needed;
> int ret = 0;
> uint64_t vram_size = 0;
> @@ -221,7 +223,7 @@ int amdgpu_amdkfd_reserve_mem_limit(struct
> amdgpu_device *adev,
> (kfd_mem_limit.ttm_mem_used + ttm_mem_needed >
> kfd_mem_limit.max_ttm_mem_limit) ||
> (adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] +
> vram_needed >
> - vram_size - reserved_for_pt - atomic64_read(&adev->vram_pin_size)
> +
> + vram_size - reserved_for_pt - reserved_for_ras -
> +atomic64_read(&adev->vram_pin_size) +
> atomic64_read(&adev->kfd.vram_pinned))) {
> ret = -ENOMEM;
> goto release;
> @@ -1694,6 +1696,8 @@ size_t amdgpu_amdkfd_get_available_memory(struct
> amdgpu_device *adev, {
> uint64_t reserved_for_pt =
> ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
> + struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
> + uint64_t reserved_for_ras = (con ? con->reserved_pages_in_bytes : 0);
> ssize_t available;
> uint64_t vram_available, system_mem_available, ttm_mem_available;
>
> @@ -1702,7 +1706,8 @@ size_t amdgpu_amdkfd_get_available_memory(struct
> amdgpu_device *adev,
> - adev->kfd.vram_used_aligned[xcp_id]
> - atomic64_read(&adev->vram_pin_size)
> + atomic64_read(&adev->kfd.vram_pinned)
> - - reserved_for_pt;
> + - reserved_for_pt
> + - reserved_for_ras;
>
> if (adev->gmc.is_app_apu || adev->flags & AMD_IS_APU) {
> system_mem_available = no_system_mem_limit ?
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c
> index ecce022c657b..a6334e0e62dc 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c
> @@ -3317,6 +3317,22 @@ static void amdgpu_ras_event_mgr_init(struct
> amdgpu_device *adev)
> amdgpu_put_xgmi_hive(hive);
> }
>
> +static void amdgpu_ras_init_reserved_vram_size(struct amdgpu_device
> +*adev) {
> + struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
> +
> + if (!con || (adev->flags & AMD_IS_APU))
> + return;
> +
> + switch (amdgpu_ip_version(adev, MP0_HWIP, 0)) {
> + case IP_VERSION(13, 0, 6):
[Tao] can we apply the change for all ASICs which support RAS?
> + con->reserved_pages_in_bytes =
> AMDGPU_RAS_RESERVED_VRAM_SIZE;
> + break;
> + default:
> + break;
> + }
> +}
> +
> int amdgpu_ras_init(struct amdgpu_device *adev) {
> struct amdgpu_ras *con = amdgpu_ras_get_context(adev); @@ -3422,6
> +3438,8 @@ int amdgpu_ras_init(struct amdgpu_device *adev)
> /* Get RAS schema for particular SOC */
> con->schema = amdgpu_get_ras_schema(adev);
>
> + amdgpu_ras_init_reserved_vram_size(adev);
> +
> if (amdgpu_ras_fs_init(adev)) {
> r = -EINVAL;
> goto release_con;
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h
> index 6a8c7b1609df..bee622c4268a 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h
> @@ -64,6 +64,7 @@ struct amdgpu_iv_entry; #define
> AMDGPU_RAS_FEATURES_SOCKETID_SHIFT 29 #define
> AMDGPU_RAS_FEATURES_SOCKETID_MASK 0xe0000000
>
> +#define AMDGPU_RAS_RESERVED_VRAM_SIZE (16ULL << 20)
[Tao] it's better to add comment here to explain why the value is 16MB.
> /* The high three bits indicates socketid */ #define
> AMDGPU_RAS_GET_FEATURES(val) ((val) &
> ~AMDGPU_RAS_FEATURES_SOCKETID_MASK)
>
> @@ -541,6 +542,7 @@ struct amdgpu_ras {
> struct ras_event_manager __event_mgr;
> struct ras_event_manager *event_mgr;
>
> + uint64_t reserved_pages_in_bytes;
> };
>
> struct ras_fs_data {
> --
> 2.17.1
