Thank you for your contribution! Sashiko AI review found 8 potential issues to consider: - [High] drm/imx/dcif: unclocked MMIO access when runtime PM resume fails - [Medium] drm/imx: dcif: missing vblank reference in dcif_crc_source_enable() - [Medium] drm/imx/dcif: unmasked interrupt status handling in dcif_irq_handler() - [Medium] drm/imx: dcif: unchecked runtime PM resume in dcif_hw_probe_init() - [Medium] drm/imx: dcif: unbalanced clock disable if dcif_runtime_resume() fails - [Low] drm/imx: dcif: signed overflow in dcif_crc_source_parse() ROI check - [Low] drm/imx/dcif: split __free() declaration in dcif_crtc_query_output_bus_format() - [Low] drm/imx: dcif: missing fb->pitches[0] bounds check in dcif_plane_atomic_check() --
commit b9251fe09a901a215b6a40b8b790485d134d37e2 Author: Sandor Yu <[email protected]> drm/imx: Add support for i.MX94 DCIF This commit adds DRM/KMS driver support for the NXP i.MX94 Display Control Interface (DCIF), including CRTC, primary and overlay planes, and CRC generation. > diff --git a/drivers/gpu/drm/imx/dcif/dcif-crc.c > b/drivers/gpu/drm/imx/dcif/dcif-crc.c > new file mode 100644 > index 0000000000000..132e40d3f3d45 > --- /dev/null > +++ b/drivers/gpu/drm/imx/dcif/dcif-crc.c [ ... ] > +void dcif_crc_source_enable(struct dcif_dev *dcif, enum dcif_crc_source > source, > + struct drm_rect *roi, int ncrc) > +{ > + if (ncrc >= MAX_DCIF_CRC_NUM) > + return; > + > + if (source == DCIF_CRC_SRC_NONE) > + return; > + > + if (dcif->crc_is_enabled) > + return; > + > + dcif_crc_config(dcif, roi, ncrc); > + > + regmap_set_bits(dcif->regmap, DCIF_CRC_CTRL, > + DCIF_CRC_CTRL_CRC_MODE | > DCIF_CRC_CTRL_CRC_SHADOW_LOAD_EN | > + DCIF_CRC_CTRL_CRC_TRIG); > + > + scoped_guard(spinlock_irqsave, &dcif->drm.event_lock) > + dcif->crc_is_enabled = true; > +} > + > +void dcif_crc_source_disable(struct dcif_dev *dcif, int ncrc) > +{ > + if (!dcif->crc_is_enabled) > + return; > + > + if (ncrc >= MAX_DCIF_CRC_NUM) > + return; > + > + regmap_clear_bits(dcif->regmap, DCIF_CRC_CTRL, > DCIF_CRC_CTRL_CRC_EN(ncrc)); > + > + scoped_guard(spinlock_irqsave, &dcif->drm.event_lock) > + dcif->crc_is_enabled = false; > +} [Severity: Medium] Should dcif_crc_source_enable() and dcif_crc_source_disable() acquire and release a vblank reference with drm_crtc_vblank_get() and drm_crtc_vblank_put()? dcif_irq_handler() only samples DCIF_CRC_VAL_R(0) and calls drm_crtc_add_crc_entry() when DCIF_INT0_VS_BLANK fires, and that interrupt is only unmasked while a vblank reference is held. On a steady display without continuous page flips, the vblank disable timer will mask DCIF_INT0_VS_BLANK after the atomic commit that enables CRC finishes, causing reads from debugfs crc/data to block indefinitely. [ ... ] > + while ((opt = strsep(&options, ",")) != NULL) { > + if (i > 3) > + return -EINVAL; > + > + ret = kstrtouint(opt, 10, ¶ms[i]); > + if (ret < 0) > + return ret; > + > + i++; > + } > + > + if (i != 4) > + return -EINVAL; > + > + roi->x1 = params[0]; > + roi->y1 = params[1]; > + roi->x2 = params[2]; > + roi->y2 = params[3]; > + > + if (!drm_rect_visible(roi)) > + return -EINVAL; [Severity: Low] Can unsigned values in params[] wrap to negative ints in struct drm_rect and bypass the ROI bounds checks? For example, "roi:10,0,2147483648,10" sets roi->x1 = 10 and roi->x2 = INT_MIN, which wraps in drm_rect_visible() to a positive width and then passes dcif_crtc_check_crc_roi() because roi->x1 >= 0 and roi->x2 <= m->hdisplay, programming truncated 12-bit coordinates into DCIF_CRC_POS_R and DCIF_CRC_SIZE_R. > diff --git a/drivers/gpu/drm/imx/dcif/dcif-crtc.c > b/drivers/gpu/drm/imx/dcif/dcif-crtc.c > new file mode 100644 > index 0000000000000..d5285105fa4df > --- /dev/null > +++ b/drivers/gpu/drm/imx/dcif/dcif-crtc.c [ ... ] > +static int dcif_crtc_query_output_bus_format(struct drm_crtc *crtc, > + struct drm_crtc_state *crtc_state) > +{ > + struct dcif_crtc_state *dcif_state = to_dcif_crtc_state(crtc_state); > + struct drm_bridge *bridge __free(drm_bridge_put) = NULL; > + struct drm_bridge_state *bridge_state; > + u32 bus_format; > + > + dcif_state->bus_format = MEDIA_BUS_FMT_RGB888_1X24; > + dcif_state->bus_flags = 0; > + > + bridge = dcif_crtc_get_bridge(crtc, crtc_state); [Severity: Low] This isn't a bug, but per Cleanup API conventions, bridge should be defined and initialized in a single statement rather than declared with __free(drm_bridge_put) = NULL at the top of the function and assigned later. [ ... ] > +static void dcif_crtc_atomic_enable(struct drm_crtc *crtc, > + struct drm_atomic_commit *state) > +{ > + struct drm_crtc_state *crtc_state = > drm_atomic_get_new_crtc_state(state, crtc); > + struct dcif_crtc_state *dcif_crtc_state = > to_dcif_crtc_state(crtc_state); > + struct drm_display_mode *adj = &crtc_state->adjusted_mode; > + struct dcif_dev *dcif = crtc_to_dcif_dev(crtc); > + struct drm_device *drm = crtc->dev; > + int ret; > + > + dev_dbg(drm->dev, "mode " DRM_MODE_FMT "\n", DRM_MODE_ARG(adj)); > + > + /* enable power when we start to set mode for CRTC */ > + ret = pm_runtime_resume_and_get(drm->dev); > + if (ret < 0) { > + drm_err(drm, "failed to resume DCIF, ret = %d\n", ret); > + return; > + } > + dcif->crtc_pm_enabled = true; [ ... ] > +} > + > +static void dcif_crtc_atomic_disable(struct drm_crtc *crtc, > + struct drm_atomic_commit *state) > +{ > + struct drm_crtc_state *crtc_state = > drm_atomic_get_new_crtc_state(state, crtc); > + struct dcif_crtc_state *dcif_crtc_state = > to_dcif_crtc_state(crtc_state); > + struct dcif_dev *dcif = crtc_to_dcif_dev(crtc); > + struct drm_device *drm = crtc->dev; > + > + if (dcif->has_crc && dcif_crtc_state->crc.source != DCIF_CRC_SRC_NONE) > + dcif_crc_source_disable(dcif, 0); > + > + dcif_disable_controller(dcif); > + dcif_disable_plane_panic(dcif); > + > + drm_crtc_vblank_off(crtc); > + > + if (dcif->crtc_pm_enabled) { > + dcif->crtc_pm_enabled = false; > + pm_runtime_put_sync(drm->dev); > + } [Severity: High] If pm_runtime_resume_and_get() fails in dcif_crtc_atomic_enable(), won't subsequent callbacks still access DCIF MMIO registers while the clocks and power domain are off? After dcif_crtc_atomic_enable() returns early with dcif->crtc_pm_enabled set to false, drm_atomic_helper_commit_tail_rpm() still runs drm_atomic_helper_commit_planes() because crtc_state->active is true, invoking dcif_plane_atomic_update() and dcif_crtc_atomic_flush() without checking dcif->crtc_pm_enabled. Later, dcif_crtc_atomic_disable() also calls dcif_crc_source_disable(), dcif_disable_controller(), and dcif_disable_plane_panic() before checking dcif->crtc_pm_enabled, causing unclocked MMIO accesses and an external abort on ARM64. [ ... ] > +irqreturn_t dcif_irq_handler(int irq, void *data) > +{ > + struct drm_device *drm = data; > + struct dcif_dev *dcif = to_dcif_dev(drm); > + int domain = dcif->cpu_domain; > + u32 stat0, stat1, crc; > + > + regmap_read(dcif->regmap, DCIF_IS0(domain), &stat0); > + regmap_read(dcif->regmap, DCIF_IS1(domain), &stat1); > + regmap_write(dcif->regmap, DCIF_IS0(domain), stat0); > + regmap_write(dcif->regmap, DCIF_IS1(domain), stat1); > + > + if (stat0 & DCIF_INT0_VS_BLANK) { > + drm_crtc_handle_vblank(&dcif->crtc); > + > + scoped_guard(spinlock_irqsave, &drm->event_lock) { > + if (dcif->crc_is_enabled) { > + regmap_read(dcif->regmap, DCIF_CRC_VAL_R(0), > &crc); > + drm_crtc_add_crc_entry(&dcif->crtc, false, 0, > &crc); > + dev_dbg(drm->dev, "crc=0x%x\n", crc); > + } > + } > + } > + > + if (stat1 & (DCIF_INT1_FIFO_PANIC0 | DCIF_INT1_FIFO_PANIC1)) { > + u32 panic = stat1 & (DCIF_INT1_FIFO_PANIC0 | > DCIF_INT1_FIFO_PANIC1); > + > + dev_dbg_ratelimited(drm->dev, "FIFO panic on %s\n", > + panic == (DCIF_INT1_FIFO_PANIC0 | > DCIF_INT1_FIFO_PANIC1) ? > + "layers 0 & 1" : panic == > DCIF_INT1_FIFO_PANIC0 ? "layer 0" : > + "layer 1"); > + } > + > + return IRQ_HANDLED; > +} [Severity: Medium] Should dcif_irq_handler() mask stat0 and stat1 with DCIF_IE0(domain) and DCIF_IE1(domain), and return IRQ_NONE when no enabled interrupt is pending? DCIF_IS0(domain) latches DCIF_INT0_VS_BLANK on every vertical blank even when it is masked in DCIF_IE0(domain), which is why dcif_crtc_enable_vblank() clears DCIF_IS0(domain) before enabling it. Because FIFO panic interrupts remain enabled in DCIF_IE1(domain) whenever the CRTC is active, a FIFO panic interrupt firing mid-frame while vblanks are disabled will see the latched DCIF_INT0_VS_BLANK bit in stat0 and spuriously call drm_crtc_handle_vblank() and drm_crtc_add_crc_entry(). > diff --git a/drivers/gpu/drm/imx/dcif/dcif-drv.c > b/drivers/gpu/drm/imx/dcif/dcif-drv.c > new file mode 100644 > index 0000000000000..3b1b9dbf0c6b0 > --- /dev/null > +++ b/drivers/gpu/drm/imx/dcif/dcif-drv.c [ ... ] > +static void dcif_hw_probe_init(struct dcif_dev *dcif) > +{ > + struct drm_device *drm = &dcif->drm; > + u32 val, vmin, vmaj; > + int ret; > + > + ret = pm_runtime_resume_and_get(drm->dev); > + if (ret < 0) { > + drm_err(drm, "failed to resume DCIF: %d\n", ret); > + return; > + } [ ... ] > +static int dcif_probe(struct platform_device *pdev) > +{ [ ... ] > + dcif_hw_probe_init(dcif); > + > + ret = dcif_kms_prepare(dcif); > + if (ret) > + return ret; > + > + ret = devm_request_irq(drm->dev, dcif->irq[dcif->cpu_domain], > + dcif_irq_handler, 0, drm->driver->name, drm); [Severity: Medium] Should dcif_hw_probe_init() return an error code so dcif_probe() can abort when pm_runtime_resume_and_get() fails? If runtime resume fails here, dcif_hw_probe_init() returns early without reading DCIF_VER, masking bootloader interrupts, or stopping the controller, and dcif_probe() continues on to request the IRQ and register the DRM device with uninitialized hardware state. Also, calling dcif_hw_probe_init() before dcif_kms_prepare() powers on and disables the hardware before devm_drm_of_get_bridge() checks for probe deferral and before dcif_kms_link_suppliers() sets up the supplier device links. [ ... ] > +static int dcif_suspend(struct device *dev) > +{ > + struct dcif_dev *dcif = dev_get_drvdata(dev); > + int ret; > + > + ret = drm_mode_config_helper_suspend(&dcif->drm); > + if (ret < 0) > + return ret; > + > + if (pm_runtime_suspended(dev)) > + return 0; > + > + return dcif_runtime_suspend(dev); > +} > + > +static int dcif_resume(struct device *dev) > +{ > + struct dcif_dev *dcif = dev_get_drvdata(dev); > + int ret; > + > + if (!pm_runtime_suspended(dev)) { > + ret = dcif_runtime_resume(dev); > + if (ret < 0) > + return ret; > + } > + > + return drm_mode_config_helper_resume(&dcif->drm); > +} [Severity: Medium] If dcif_runtime_resume() fails in dcif_resume(), won't the runtime PM core call dcif_runtime_suspend() a second time on already disabled clocks? When the CRTC is active before system suspend, device_prepare() increments the runtime PM usage count so drm_mode_config_helper_suspend() -> pm_runtime_put_sync() leaves the device RPM_ACTIVE, and dcif_suspend() then calls dcif_runtime_suspend() directly without updating the runtime PM status. If dcif_runtime_resume() fails during system resume, the clocks remain disabled while runtime_status is still RPM_ACTIVE, so device_complete() -> pm_runtime_put() will invoke dcif_runtime_suspend() -> clk_bulk_disable_unprepare() again. Using pm_runtime_force_suspend() and pm_runtime_force_resume() would keep the runtime PM state in sync with the clocks. > diff --git a/drivers/gpu/drm/imx/dcif/dcif-plane.c > b/drivers/gpu/drm/imx/dcif/dcif-plane.c > new file mode 100644 > index 0000000000000..9f54f1b2d8d17 > --- /dev/null > +++ b/drivers/gpu/drm/imx/dcif/dcif-plane.c [ ... ] > +static int dcif_plane_atomic_check(struct drm_plane *plane, struct > drm_atomic_commit *state) > +{ > + struct drm_plane_state *new_plane_state = > drm_atomic_get_new_plane_state(state, plane); > + struct dcif_dev *dcif = plane_to_dcif_dev(plane); > + struct drm_framebuffer *fb = new_plane_state->fb; > + struct drm_crtc_state *crtc_state; > + int ret; > + > + if (!fb) > + return 0; [ ... ] > + if (new_plane_state->fb->format->has_alpha && > + new_plane_state->pixel_blend_mode == DRM_MODE_BLEND_COVERAGE && > + new_plane_state->alpha != DRM_BLEND_ALPHA_OPAQUE) > + return -EINVAL; > + > + return 0; > +} [Severity: Low] Should dcif_plane_atomic_check() reject framebuffers with fb->pitches[0] > 0xffff? DCIF_CTRLDESC3_PITCH() is a 16-bit field (GENMASK(15, 0)), so a framebuffer stride >= 65536 bytes is silently truncated in dcif_plane_atomic_update(), corrupting scanout. -- Sashiko AI review ยท https://sashiko.dev/#/patchset/[email protected]?part=4
