Thank you for your contribution! Sashiko AI review found 2 potential issues to consider: - [High] drm/vc4: inline flip in vc4_async_set_fence_cb() overtakes pending flips - [Medium] drm/vc4: NULL data->name and inverted drmm order in __vc4_crtc_init() --
commit 350614eb18304f31f6b7fcd5f4826c9bda20f32b Author: Maíra Canal <[email protected]> drm/vc4: Complete async page flips from an ordered workqueue This commit defers async page flip completion from the dma_fence callback to a per-CRTC ordered workqueue so that sleeping operations do not run in atomic context. > diff --git a/drivers/gpu/drm/vc4/vc4_crtc.c b/drivers/gpu/drm/vc4/vc4_crtc.c > index c558b527c7e60..623b636a03652 100644 > --- a/drivers/gpu/drm/vc4/vc4_crtc.c > +++ b/drivers/gpu/drm/vc4/vc4_crtc.c > @@ -884,15 +884,18 @@ struct vc4_async_flip_state { > struct drm_framebuffer *fb; > struct drm_framebuffer *old_fb; > struct drm_pending_vblank_event *event; > + struct dma_fence *fence; > struct dma_fence_cb cb; > + struct work_struct work; > }; > > -/* Called when the V3D execution for the BO being flipped to is done, so that > - * we can actually update the plane's address to point to it. > +/* Executed when the V3D execution for the BO being flipped to is done, so > + * that we can actually update the plane's address to point to it. > */ > -static void > -vc4_async_page_flip_complete(struct vc4_async_flip_state *flip_state) > +static void vc4_async_page_flip_work(struct work_struct *work) > { > + struct vc4_async_flip_state *flip_state = > + container_of(work, struct vc4_async_flip_state, work); > struct drm_crtc *crtc = flip_state->crtc; > struct drm_device *dev = crtc->dev; > struct vc4_dev *vc4 = to_vc4_dev(dev); > @@ -930,17 +933,19 @@ vc4_async_page_flip_complete(struct > vc4_async_flip_state *flip_state) > drm_framebuffer_put(flip_state->old_fb); > } > > + dma_fence_put(flip_state->fence); > kfree(flip_state); > } > > -static void vc4_async_page_flip_fence_complete(struct dma_fence *fence, > - struct dma_fence_cb *cb) > +static void vc4_async_page_flip_fence_cb(struct dma_fence *fence, > + struct dma_fence_cb *cb) > { > struct vc4_async_flip_state *flip_state = > container_of(cb, struct vc4_async_flip_state, cb); > + struct vc4_crtc *vc4_crtc = to_vc4_crtc(flip_state->crtc); > > - vc4_async_page_flip_complete(flip_state); > - dma_fence_put(fence); > + /* Completing the flip may sleep, so defer it to a worker. */ > + queue_work(vc4_crtc->async_flip_wq, &flip_state->work); > } > > static int vc4_async_set_fence_cb(struct drm_device *dev, > @@ -948,23 +953,26 @@ static int vc4_async_set_fence_cb(struct drm_device > *dev, > { > struct drm_framebuffer *fb = flip_state->fb; > struct drm_gem_dma_object *dma_bo = drm_fb_dma_get_gem_obj(fb, 0); > - struct dma_fence *fence; > + struct vc4_crtc *vc4_crtc = to_vc4_crtc(flip_state->crtc); > int ret; > > - ret = dma_resv_get_singleton(dma_bo->base.resv, DMA_RESV_USAGE_READ, > &fence); > + INIT_WORK(&flip_state->work, vc4_async_page_flip_work); > + > + ret = dma_resv_get_singleton(dma_bo->base.resv, DMA_RESV_USAGE_READ, > + &flip_state->fence); > if (ret) > return ret; > > - /* If there's no fence, complete the page flip immediately */ > - if (!fence) { > - vc4_async_page_flip_fence_complete(fence, &flip_state->cb); > + if (flip_state->fence && > + !dma_fence_add_callback(flip_state->fence, &flip_state->cb, > + vc4_async_page_flip_fence_cb)) > return 0; > - } > > - /* If the fence has already been completed, complete the page flip */ > - if (dma_fence_add_callback(fence, &flip_state->cb, > - vc4_async_page_flip_fence_complete)) > - vc4_async_page_flip_fence_complete(fence, &flip_state->cb); > + /* No fence or the fence has already signaled. Complete the flip > + * immediately, after earlier queued flips. > + */ > + flush_workqueue(vc4_crtc->async_flip_wq); > + vc4_async_page_flip_work(&flip_state->work); [Severity: High] Can an inline flip in vc4_async_set_fence_cb() overtake or race with an earlier flip whose fence has not yet queued its work item? flush_workqueue() only waits for work items already on async_flip_wq. If flip 1 (to fb1) is still waiting on an unsignaled fence when flip 2 (to fb2, with no fence or an already-signaled fence) runs, flip 2 completes inline first: CPU0 (ioctl) CPU1 (kworker) vc4_async_page_flip_common() [flip 1, fb1] plane->state->fb = fb1 dma_fence_add_callback(fence1, ...) vc4_async_page_flip_common() [flip 2, fb2] flip2->old_fb = fb1 plane->state->fb = fb2 flush_workqueue(async_flip_wq) vc4_async_page_flip_work(flip2) vc4_plane_async_set_fb(plane, fb2) vc4_bo_dec_usecnt(fb1) drm_framebuffer_put(fb1) <fence1 signals, queues work> vc4_async_page_flip_work(flip1) vc4_plane_async_set_fb(plane, fb1) drm_framebuffer_put(fb1) <-- HVS scans out unreferenced fb1 If fence1 signals around the time of flip 2, both vc4_async_page_flip_work() calls can also run concurrently and race in vc4_plane_async_set_fb() on vc4_state->dlist and vc4_state->hw_dlist without locking. When flip 1 finishes after flip 2, it overwrites the display list with the older fb1 and drops flip 1's reference to fb1 after flip 2 already dropped fb1's old_fb reference and BO usecnt, leaving the HVS scanning out a framebuffer whose references and usecnt have been released. [ ... ] > @@ -1387,7 +1395,13 @@ int __vc4_crtc_init(struct drm_device *drm, > vc4_crtc->data = data; > vc4_crtc->pdev = pdev; > vc4_crtc->feeds_txp = feeds_txp; > + vc4_crtc->async_flip_wq = drmm_alloc_ordered_workqueue(drm, > "%s-async-flip", > + 0, data->name); > + if (IS_ERR(vc4_crtc->async_flip_wq)) > + return PTR_ERR(vc4_crtc->async_flip_wq); > + > spin_lock_init(&vc4_crtc->irq_lock); > + > ret = drmm_crtc_init_with_planes(drm, crtc, primary_plane, NULL, > crtc_funcs, data->name); [Severity: Medium] Does allocating async_flip_wq before drmm_crtc_init_with_planes() pass a NULL data->name on BCM2712 and invert the drmm teardown order? bcm2712_pv0_data and bcm2712_pv1_data leave .base.name as NULL and rely on drmm_crtc_init_with_planes() to assign a default crtc->name, so data->name formats the workqueue name as "(null)-async-flip" on Raspberry Pi 5. In addition, drmm actions run in reverse registration order, so drmm_crtc_init_with_planes_cleanup() calls drm_crtc_cleanup() and zeroes *crtc (including crtc->dev and crtc->primary) before __drmm_workqueue_release() drains and destroys async_flip_wq. Allocating the workqueue after drmm_crtc_init_with_planes() using crtc->name avoids both issues. -- Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=2
