On Wed, Oct 15, 2025 at 03:52:40PM +0530, Ankit Nautiyal wrote:
> Add a check during atomic crtc check phase to ensure the programmed
> guardband is sufficient to cover latencies introduced by enabled features
> such as DSC, PSR/PR, scalers, and DP SDPs.
>
> Signed-off-by: Ankit Nautiyal <[email protected]>
> ---
> drivers/gpu/drm/i915/display/intel_display.c | 56 ++++++++++++++++++++
> 1 file changed, 56 insertions(+)
>
> diff --git a/drivers/gpu/drm/i915/display/intel_display.c
> b/drivers/gpu/drm/i915/display/intel_display.c
> index 4367ecfab2b3..4e3f08a8cd9c 100644
> --- a/drivers/gpu/drm/i915/display/intel_display.c
> +++ b/drivers/gpu/drm/i915/display/intel_display.c
> @@ -126,6 +126,7 @@
> #include "intel_vga.h"
> #include "intel_vrr.h"
> #include "intel_wm.h"
> +#include "skl_prefill.h"
> #include "skl_scaler.h"
> #include "skl_universal_plane.h"
> #include "skl_watermark.h"
> @@ -4191,6 +4192,57 @@ static int hsw_compute_linetime_wm(struct
> intel_atomic_state *state,
> return 0;
> }
>
> +static int intel_crtc_guardband_atomic_check(struct intel_atomic_state
> *state,
> + struct intel_crtc *crtc)
Not sure why you're adding this. We already have the
compute_guardband().
> +{
> + struct intel_display *display = to_intel_display(crtc);
> + struct intel_crtc_state *crtc_state =
> + intel_atomic_get_new_crtc_state(state, crtc);
> + const struct drm_display_mode *adjusted_mode =
> + &crtc_state->hw.adjusted_mode;
> + struct skl_prefill_ctx prefill_ctx;
> + int prefill_framestart_delay = 1;
> + int prefill_min_guardband;
> + int prefill_latency_us;
> + int prefill_wm0_lines;
> + int prefill_sagv_us;
> + int psr_latency = 0;
> + int sdp_latency = 0;
> + int min_guardband;
> + int guardband;
> +
> + skl_prefill_init(&prefill_ctx, crtc_state);
> + prefill_wm0_lines = skl_wm0_prefill_lines(crtc_state);
> + prefill_sagv_us = display->sagv.block_time_us;
> + prefill_latency_us = prefill_sagv_us +
> + intel_scanlines_to_usecs(adjusted_mode,
> + prefill_framestart_delay +
> + prefill_wm0_lines);
All of that should pretty much just be skl_prefill_init_worst()
> + prefill_min_guardband =
> + skl_prefill_min_guardband(&prefill_ctx,
> + crtc_state,
> + prefill_latency_us);
The only question really is what use as the latency here.
I think we want it to be:
max(sagv_block_time, skl_watermark_max_latency(1))
which should guarantee that we get the max power savings.
> +
> + if (intel_crtc_has_dp_encoder(crtc_state)) {
> + psr_latency = intel_psr_max_link_wake_latency(crtc_state);
> + sdp_latency = intel_dp_compute_sdp_latency(crtc_state);
> + }
> +
> + min_guardband = max(sdp_latency, psr_latency);
> +
> + min_guardband = max(min_guardband, prefill_min_guardband);
> +
> + guardband = intel_crtc_vblank_length(crtc_state);
> +
> + if (guardband < min_guardband) {
> + drm_dbg_kms(display->drm, "actual guardband: %d shorter than
> min guardband: %d\n",
> + guardband, min_guardband);
> + return -EINVAL;
> + }
I don't think we want to do any checks here. This whole thing
should just be something like:
guardband = prefill_min_guardband()
guardband = max(guardband, psr_min_guardband())
guardband = max(guardband, sdp_min_guardband())
crtc_state->vrr.guardband = min(guardband, intel_vrr_max_guardband())
And then we need to check the final value against
sdp_min_guardband() in .compute_config_late() to make sure
we got enough for the SDPs. So quite similar to PSR, except
we just want .compute_config_late() to fail if we don't have
enough for the SDPs.
I think that should be good enough for now. It may force a modeset
if the SDPs change though, so later we might want to think about
using a better worst case estimate here, eg. assume HDR metadata may
get enabled later, which we'd like to do without changing the guardband.
> +
> + return 0;
> +}
> +
> static int intel_crtc_atomic_check(struct intel_atomic_state *state,
> struct intel_crtc *crtc)
> {
> @@ -4253,6 +4305,10 @@ static int intel_crtc_atomic_check(struct
> intel_atomic_state *state,
> if (ret)
> return ret;
>
> + ret = intel_crtc_guardband_atomic_check(state, crtc);
> + if (ret)
> + return ret;
> +
> return 0;
> }
>
> --
> 2.45.2
--
Ville Syrjälä
Intel