Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider: - [Critical] `_intel_panel_fixed_mode` can select a fixed mode with a lower native refresh rate than requested, causing `vtotal` to be incorrectly shrunk below `vdisplay` and programmed into hardware. - [High] VRR-based seamless refresh rate fastsets fail because `is_vrr_compatible` incorrectly rejects stretched modes due to a mismatched `vtotal - vsync_start` offset. --
commit 41ed53116263d4e38c4aa61dc4fe28e6a92d7c3c Author: Ville Syrjälä <[email protected]> drm/i915/panel: Attempt VRR based refresh rate change for !allow_modeset Adjusts the panel fixed mode selection to consider VRR-compatible modes when userspace denies full modesets. This allows refresh rate changes via the fastset path by adjusting the vblank length. > diff --git a/drivers/gpu/drm/i915/display/intel_panel.c > b/drivers/gpu/drm/i915/display/intel_panel.c > index 6c539987f0a3f..a8af6a4d35bc5 100644 > --- a/drivers/gpu/drm/i915/display/intel_panel.c > +++ b/drivers/gpu/drm/i915/display/intel_panel.c > @@ -82,16 +82,37 @@ static bool is_best_fixed_mode(struct intel_connector > *connector, > abs(drm_mode_vrefresh(best_mode) - vrefresh); > } > > +static bool is_vrr_compatible(const struct drm_display_mode *mode1, > + const struct drm_display_mode *mode2) > +{ > + return drm_mode_match(mode1, mode2, > + DRM_MODE_MATCH_CLOCK | > + DRM_MODE_MATCH_TIMINGS_VRR | [Severity: High] Will this check reject valid stretched modes and prevent the fastset? drm_mode_match() with DRM_MODE_MATCH_TIMINGS_VRR requires the offset between vtotal and vsync_start to be identical between both modes. However, the VRR logic in intel_panel_compute_config() stretches vtotal without altering vsync_start or vsync_end. This means a stretched vrr_ref_mode will have a larger vtotal - vsync_start offset than its base fixed mode, causing is_vrr_compatible() to return false. If is_vrr_compatible() returns false, _intel_panel_fixed_mode() fails to find a matching mode and falls back to intel_panel_fixed_mode(). Since allow_modeset is false, this fallback causes the core DRM modeset checks to reject the commit, breaking the intended seamless switch. > + DRM_MODE_MATCH_FLAGS | > + DRM_MODE_MATCH_3D_FLAGS); > +} > + > +static const struct drm_display_mode * > +_intel_panel_fixed_mode(struct intel_connector *connector, > + const struct drm_display_mode *mode, > + const struct drm_display_mode *vrr_ref_mode) > +{ > + const struct drm_display_mode *fixed_mode, *best_mode = NULL; > + int vrefresh = drm_mode_vrefresh(mode); > + > + if (vrr_ref_mode && > + (!intel_vrr_is_in_range(connector, vrefresh) || > + !intel_vrr_is_in_range(connector, > drm_mode_vrefresh(vrr_ref_mode)))) > + return NULL; > + > list_for_each_entry(fixed_mode, &connector->panel.fixed_modes, head) { > int fixed_mode_vrefresh = drm_mode_vrefresh(fixed_mode); > > + if (vrr_ref_mode && > + (!intel_vrr_is_in_range(connector, fixed_mode_vrefresh) || > + !is_vrr_compatible(fixed_mode, vrr_ref_mode))) > + continue; > + > if (is_best_fixed_mode(connector, vrefresh, > fixed_mode_vrefresh, best_mode)) > best_mode = fixed_mode; > } [Severity: Critical] Does this loop inadvertently select a lower refresh rate mode that later causes vtotal to shrink below vdisplay? If userspace requests a transition from a lower to a higher refresh rate without a full modeset, is_vrr_compatible() filters out the high refresh rate fixed mode due to a dotclock mismatch. The loop then evaluates the lower refresh rate fixed mode. Because this is the first iteration where best_mode is NULL, is_best_fixed_mode() unconditionally returns true and accepts the mode, bypassing the requirement that fixed_mode_vrefresh >= vrefresh: intel_panel.c:is_best_fixed_mode() { if (!best_mode) return true; ... } Later in intel_panel_compute_config(), the driver shrinks vtotal to match the higher requested refresh rate using the lower dotclock: intel_panel.c:intel_panel_compute_config() { ... if (is_vrr && fixed_mode_vrefresh != vrefresh) adjusted_mode->vtotal = DIV_ROUND_CLOSEST(adjusted_mode->clock * 1000, adjusted_mode->htotal * vrefresh); ... } Since both refresh rates fall within the VRR range, update_lrr is set to true, which bypasses crtc_vtotal validation during the atomic check. Could this regression result in programming a vtotal that is smaller than vdisplay, leading to hardware hangs or pipeline underflows? -- Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=4
