Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider: - [High] drm/crtc-helper: unsynchronized crtc_state mutation in VRR helpers - [Medium] drm/crtc-helper: vrr->dynamic set and left true when VRR is disabled --
commit 2e8aace704ca21ab8efb4aacd3b771407706ec16 Author: Nicolas Frattaroli <[email protected]> drm/crtc-helper: Add VRR helper functions This commit adds drm_crtc_helper_vrr_flush() and drm_crtc_helper_vrr_vsync() to update the internal VRR state in struct drm_crtc_state when a new frame is ready or when a vertical blank ends, returning whether cur_vtotal changed. > diff --git a/drivers/gpu/drm/drm_crtc_helper.c > b/drivers/gpu/drm/drm_crtc_helper.c > index 4a8cfe046e4ab..4ffb3b0b78780 100644 > --- a/drivers/gpu/drm/drm_crtc_helper.c > +++ b/drivers/gpu/drm/drm_crtc_helper.c > @@ -838,6 +838,69 @@ int drm_crtc_helper_set_config(struct drm_mode_set *set, > } > EXPORT_SYMBOL(drm_crtc_helper_set_config); > > +/** > + * drm_crtc_helper_vrr_flush - Update VRR state on frame ready > + * @crtc_state: pointer to &struct drm_crtc_state of CRTC that has a new > frame ready > + * > + * Update the VRR state by communicating that a frame is ready on > + * @crtc_state.crtc. > + * > + * Should be called from the @crtc_state.crtc > &drm_crtc_helper_funcs.atomic_flush > + * hook. > + * > + * Returns: > + * - %true if @crtc_state.vrr_state.cur_vtotal changed > + * - %false otherwise > + */ > +bool drm_crtc_helper_vrr_flush(struct drm_crtc_state *crtc_state) > +{ > + struct drm_crtc_vrr_state *vrr = &crtc_state->vrr_state; > + > + if (!crtc_state->vrr_enabled) { > + if (vrr->cur_vtotal) { > + vrr->cur_vtotal = 0; > + return true; > + } > + return false; > + } > + > + if (vrr->cur_vtotal == vrr->base_vtotal) > + return false; > + > + vrr->cur_vtotal = vrr->base_vtotal; > + > + return true; > +} > +EXPORT_SYMBOL(drm_crtc_helper_vrr_flush); [Severity: High] As reported by Sashiko on RFC [1], is it safe for drm_crtc_helper_vrr_vsync() and drm_crtc_helper_vrr_flush() to mutate crtc_state->vrr_state in place across commits and from the vblank interrupt handler? Consider a sequence of two atomic commits when VRR is enabled: 1. Commit 1 runs atomic_flush and calls drm_crtc_helper_vrr_flush(state_1), which sets state_1->vrr_state.cur_vtotal = base_vtotal and programs the hardware vtotal to base_vtotal. 2. Commit 2 starts and calls drm_atomic_get_crtc_state() -> __drm_atomic_helper_crtc_duplicate_state(), which copies crtc->state (state_1) into state_2 via memcpy(), snapshotting state_2->vrr_state.cur_vtotal = base_vtotal. 3. Before commit 2 calls drm_atomic_helper_swap_state() (or while waiting on fences / hw_done), a vblank interrupt fires and the driver's vblank handler calls drm_crtc_helper_vrr_vsync(crtc->state) on state_1, setting state_1->vrr_state.cur_vtotal = max_vtotal and programming the hardware vtotal to max_vtotal. 4. Commit 2 runs drm_atomic_helper_swap_state(), replacing crtc->state with state_2, which still holds the stale vrr_state.cur_vtotal == base_vtotal copied in step 2. 5. When commit 2 reaches atomic_flush and calls drm_crtc_helper_vrr_flush(state_2), vrr->cur_vtotal == vrr->base_vtotal evaluates to true and the helper returns false, causing the driver to skip reprogramming the hardware vtotal back from max_vtotal to base_vtotal for the new frame. In addition, drm_crtc_helper_vrr_vsync() mutates crtc->state->vrr_state from hardirq context without synchronization against __drm_atomic_helper_crtc_duplicate_state(), drm_atomic_helper_swap_state(), drm_crtc_helper_vrr_flush(), or state destruction after drm_atomic_helper_commit_hw_done(). Should the mutable runtime VRR state live on struct drm_crtc protected by an irq-safe spinlock rather than in struct drm_crtc_state? > + > +/** > + * drm_crtc_helper_vrr_vsync - Update VRR state on vblank end > + * @crtc_state: pointer to &struct drm_crtc_state of CRTC the vblank > occurred on > + * > + * Update the VRR state by communicating that a vertical blank has ended and > a > + * new frame is about to start. > + * > + * Returns: > + * - %true if @crtc_state.vrr_state.cur_vtotal changed > + * - %false otherwise > + */ > +bool drm_crtc_helper_vrr_vsync(struct drm_crtc_state *crtc_state) > +{ > + struct drm_crtc_vrr_state *vrr = &crtc_state->vrr_state; > + u16 old_vtotal = vrr->cur_vtotal; > + > + if (!crtc_state->vrr_enabled) > + vrr->cur_vtotal = 0; > + else > + vrr->cur_vtotal = vrr->max_vtotal; > + > + if (abs_diff(vrr->cur_vtotal, old_vtotal) > 1) > + vrr->dynamic = true; > + > + return old_vtotal != vrr->cur_vtotal; > +} [Severity: Medium] When crtc_state->vrr_enabled is false, drm_crtc_helper_vrr_vsync() sets vrr->cur_vtotal = 0 and then checks: if (abs_diff(vrr->cur_vtotal, old_vtotal) > 1) vrr->dynamic = true; If VRR was previously enabled so old_vtotal > 1 (including fixed-rate VRR where vrr->dynamic was false), doesn't abs_diff(0, old_vtotal) > 1 evaluate to true and set vrr->dynamic = true when VRR is disabled? Neither drm_crtc_helper_vrr_vsync() nor drm_crtc_helper_vrr_flush() resets vrr->dynamic to false in the !crtc_state->vrr_enabled path either, leaving vrr->dynamic stuck at true after VRR is turned off. Should both helpers clear vrr->dynamic = false when !crtc_state->vrr_enabled? [1] https://lore.kernel.org/r/[email protected] -- Sashiko AI review ยท https://sashiko.dev/#/patchset/[email protected]?part=6
