Thank you for your contribution! Sashiko AI review found 16 potential issues to consider: - [High] drm/cix: unread sink DPCD caps and status race on boot, resume, and HPD - [High] drm/cix: AUX probe in cix_dp_replay_deferred_hotplug() always fails with -ENODEV - [High] drm/cix: broken VSC SDP packing, stream override, and YCbCr420 validation - [High] drm/cix: unclocked MMIO register accesses when apb_clk is gated - [High] drm/cix: NULL pointer dereference when DRM_BRIDGE_ATTACH_NO_CONNECTOR is set - [High] drm/cix: broken YCbCr 4:2:0 mode matching, format selection, and VSC SDP - [Medium] drm/cix: PHY refcount leak and state desync when cix_dp_enable() or resume fails - [Medium] drm/cix: HPD workqueue and IRQ teardown races in deinit, detach, and suspend - [Medium] drm/cix: active streams and PHY power left enabled in shutdown and remove - [Medium] drm/cix: display stays off if cix_dp_prepare() runs before cix_dp_pm_complete() - [Medium] drm/cix: missing stream TU/timing update and unchecked mode config on retrain - [Medium] drm/cix: broken BT.2020 bpc fallback in cix_dp_cal_bpc() and 6 bpc default - [Medium] drm/cix: unlocked mode state update in cix_dp_bridge_mode_set() before prepare - [Medium] drm/cix: FIFO underflow read on AUX NACK/DEFER in cix_dp_aux_cmd_submit() - [Medium] drm/cix: plain 64-bit division in cix_dp_adjust_mode() breaks 32-bit builds - [Low] drm/cix: cix_dp_aux_cmd_submit() reads reply FIFO on AUX/I2C NACK or DEFER --
commit 002cf156569e15175af86684eb3509dc08da82b1 Author: Luo Liu <[email protected]> drm/cix: add Sky1 DisplayPort transmitter driver This commit adds a DRM bridge and connector driver for the CIX Sky1 DisplayPort transmitter, driving the DisplayPort PHY via the generic PHY API and supporting link training, EDID probing, YCbCr 4:2:0, and hotplug. > diff --git a/drivers/gpu/drm/cix/dptx/cix_dptx.c > b/drivers/gpu/drm/cix/dptx/cix_dptx.c > new file mode 100644 > index 0000000000000..8915c50bc9db3 > --- /dev/null > +++ b/drivers/gpu/drm/cix/dptx/cix_dptx.c [ ... ] > +static int cix_dp_train_loop(struct cix_dp *dp) > +{ > + struct cix_dp_mode *mode = &dp->mode; > + u8 bpp = dp->connector.config.bpp; > + u8 bw_code = mode->bw_code; > + int i; > + int j; > + int ret; > + > + for (i = 0; i < DP_MAX_TRAINING_LOOP; i++) { > + for (j = 0; j <= 3; j++) { > + if (!(dp->state & DPTX_STATE_READY)) > + return 0; > + > + ret = cix_dp_train(dp); > + if (!ret) > + return 0; > + > + ret = cix_dp_mode_configure(dp, mode->pclock, bw_code, > + bpp, true); > + if (ret < 0) > + break; > + > + bw_code = dp->mode.bw_code; > + } > + > + cix_dp_mode_configure(dp, mode->pclock, 0, bpp, true); > + bw_code = dp->mode.bw_code; > + } > + > + return -EIO; > +} [Severity: Medium] When cix_dp_hpd_irq_work_func() calls cix_dp_train_loop() while a stream is enabled and link training downshifts dp->mode.bw_code or changes dp->mode.lane_cnt, does anything reprogram the active stream's transfer unit and timing registers? cix_dp_mode_set_transfer_unit() and cix_dp_mode_set_timing() depend on bw_code, link_rate, and lane_cnt, so leaving CIX_DPTX_TU_CONFIG, CIX_DPTX_USER_DATA_COUNT, CIX_DPTX_SECONDARY_DATA_WINDOW, and CIX_DPTX_DATA_CONTROL at their old values corrupts the running video stream. Also, the fallback cix_dp_mode_configure() call at the end of the outer loop ignores its return value. > + > +static int cix_dp_aux_cmd_submit(struct cix_dp *dp, u32 cmd, u32 addr, > + u8 *buf, u32 bytes, u8 *reply) > +{ [ ... ] > + value = cix_dp_read(dp, CIX_DPTX_AUX_REPLY_CODE); > + if (reply) > + *reply = value; > + > + if (is_read) { > + if (value == CIX_DPTX_AUX_REPLY_CODE_AUX_ACK || > + value == CIX_DPTX_AUX_REPLY_CODE_I2C_ACK) { > + value = cix_dp_read(dp, CIX_DPTX_AUX_REPLY_DATA_COUNT); > + if ((value & CIX_DPTX_REPLY_DATA_COUNT_MASK) != bytes) > + return -EIO; > + } > + for (i = 0; i < bytes; i++) > + buf[i] = cix_dp_read(dp, CIX_DPTX_AUX_REPLY_DATA); > + } > + > + return 0; > +} [Severity: Low] Should the CIX_DPTX_AUX_REPLY_DATA read loop be inside the ACK check? When a sink responds to a read with NACK or DEFER, the loop still runs and overwrites buf with stale or empty FIFO reads before drm_dp_aux retries. > + > +static ssize_t cix_dp_aux_transfer(struct drm_dp_aux *aux, > + struct drm_dp_aux_msg *msg) > +{ > + struct cix_dp *dp = container_of(aux, struct cix_dp, aux); > + size_t to_access = 0; > + size_t accessed_bytes = 0; > + int ret; > + > + if (!cix_dp_get_hpd_state(dp)) > + return -ENODEV; > + > + do { > + to_access = min_t(size_t, DP_AUX_MAX_PAYLOAD_BYTES, > + msg->size - accessed_bytes); > + ret = cix_dp_aux_cmd_submit(dp, msg->request, > + msg->address + accessed_bytes, > + msg->buffer + accessed_bytes, > + to_access, &msg->reply); > + if (ret) > + return accessed_bytes ? (ssize_t)accessed_bytes : ret; > + > + accessed_bytes += to_access; > + } while (accessed_bytes < msg->size); > + > + return msg->size; > +} [Severity: Medium] In addition to cix_dp_aux_cmd_submit() reading from the empty CIX_DPTX_AUX_REPLY_DATA FIFO when the sink replies with NACK or DEFER, does cix_dp_aux_transfer() keep submitting later 16-byte chunks after a non-ACK reply? cix_dp_aux_cmd_submit() returns 0 on NACK and DEFER, so a multi-chunk transfer continues looping and overwrites msg->reply instead of stopping and returning to the caller. [ ... ] > +static int cix_dp_set_misc(struct cix_dp *dp, > + struct cix_dp_panel *dp_panel) > +{ > + struct cix_dp_config *config; > + > + if (!dp_panel || !dp_panel->connector) > + return -EINVAL; > + > + config = &dp_panel->connector->config; > + > + config->misc0 &= ~CIX_DPTX_STREAM_MISC0_SYNC_LOCK; > + config->misc1 &= ~CIX_DPTX_VSC_COLORIMETRY_EN; > + config->override = 0; > + > + if (dp->caps.vsc_supported && > + cix_dp_needs_vsc_sdp(config->format, config->colorspace)) { > + config->misc1 |= CIX_DPTX_VSC_COLORIMETRY_EN; > + > + switch (config->bpc) { > + case 6: > + config->override = CIX_DPTX_STREAM_OVERRIDE_BPC_6; > + break; [ ... ] > + default: > + config->override = CIX_DPTX_STREAM_OVERRIDE_BPC_8; > + break; > + } > + > + if (config->format == CIX_DP_FORMAT_YCBCR420) > + config->override |= CIX_DPTX_STREAM_OVERRIDE_YCbCR_420; > + > + config->override |= CIX_DPTX_STREAM_OVERRIDE_ENABLE; > + config->dynamic_range = DP_DYNAMIC_RANGE_CTA; > + config->content_type = DP_CONTENT_TYPE_NOT_DEFINED; > + > + goto end; > + } [ ... ] > +static int dptx_setup_vsc_sdp(struct dp_sdp *sdp, > + struct cix_dp_config *config) > +{ > + sdp->sdp_header.HB0 = 0; > + sdp->sdp_header.HB1 = 0x7; > + sdp->sdp_header.HB2 = 0x5; > + sdp->sdp_header.HB3 = 0x13; > + > + switch (config->format) { > + case CIX_DP_FORMAT_YCBCR444: > + sdp->db[16] = 0x1 << 4; > + break; > + case CIX_DP_FORMAT_YCBCR422: > + sdp->db[16] = 0x1 << 4; > + break; > + case CIX_DP_FORMAT_YCBCR420: > + sdp->db[16] = 0x3 << 4; > + break; > + default: > + break; > + } [Severity: High] Does dptx_setup_vsc_sdp() program the wrong VSC SDP payload for YCbCr 4:2:2, RGB, and 6 bpc? CIX_DP_FORMAT_YCBCR422 sets sdp->db[16] to 0x1 << 4 (YCbCr 4:4:4) instead of 0x2 << 4 (YCbCr 4:2:2). In the default (RGB) branch, sdp->db[16] is not cleared before the colorspace switch ORs in the colorimetry bits, leaving stale format and colorimetry bits from earlier modesets in conn->sdp[CIX_VSC_SDP]. Further down in dptx_setup_vsc_sdp(), switch (config->bpc) has no case 6, so 6 bpc RGB falls into default and encodes 0x1 (8 bpc) instead of 0x0 in sdp->db[17]. In cix_dp_set_misc() above, the VSC path only sets the pixel format bits in config->override for CIX_DP_FORMAT_YCBCR420 (leaving YCbCr 4:2:2 and 4:4:4 as 0/RGB in the override register) and overwrites config->content_type with DP_CONTENT_TYPE_NOT_DEFINED instead of using conn_state->content_type. Also, cix_dp_pick_format_bpc() and cix_dp_connector_mode_valid() select DRM_OUTPUT_COLOR_FORMAT_YCBCR420 without checking dp->caps.vsc_supported; when vsc_supported is false, cix_dp_set_misc() falls through and returns -EINVAL (which cix_dp_panel_hw_cfg() ignores), leaving MISC0, MISC1, OVERRIDE, and the VSC SDP unconfigured while the RCSU adapter is set to YCbCr 4:2:0. [ ... ] > +static int cix_dp_panel_setup_colorimetry_sdp(struct cix_dp *dp, > + struct cix_dp_panel *dp_panel) > +{ > + struct cix_connector *conn = dp_panel->connector; > + > + if (conn->config.override & CIX_DPTX_STREAM_OVERRIDE_ENABLE) { > + dptx_setup_vsc_sdp(&conn->sdp[CIX_VSC_SDP], &conn->config); > + infoframe_write_packet(dp, dp_panel->source_id, CIX_VSC_SDP, > + &conn->sdp[CIX_VSC_SDP], 1); > + } > + > + return 0; > +} [ ... ] > +static int cix_dp_panel_hw_cfg(struct cix_dp *dp, int stream_id) > +{ > + struct cix_dp_panel *dp_panel = dp->active_panels[stream_id]; > + struct drm_display_mode *mode = cix_dp_get_mode(dp, stream_id); > + struct cix_connector *conn; > + > + if (!mode) > + return -EINVAL; > + > + conn = dp_panel->connector; > + > + cix_dp_cfg_adapter(dp, stream_id); > + > + cix_dp_set_misc(dp, dp_panel); > + cix_dp_panel_setup_colorimetry_sdp(dp, dp_panel); > + > + cix_dp_mode_set_transfer_unit(dp, dp_panel->source_id, mode, > + conn->config.bpp); > + cix_dp_mode_set_timing(dp, dp_panel->source_id, mode, conn->config.bpp); > + return 0; > +} [Severity: High] Does enabling the bridge when attached with DRM_BRIDGE_ATTACH_NO_CONNECTOR cause a NULL pointer dereference on dp_panel->connector? When DRM_BRIDGE_ATTACH_NO_CONNECTOR is passed to cix_dp_bridge_attach(), it returns 0 without calling cix_dp_connector_create(), leaving dp->dp_panel.connector NULL. On modeset enable, cix_dp_panel_hw_cfg() ignores the -EINVAL return from cix_dp_set_misc() and calls cix_dp_panel_setup_colorimetry_sdp(), which dereferences conn->config.override, followed by cix_dp_panel_hw_cfg() dereferencing conn->config.bpp. In addition, cix_dp_hpd_event_work_func() exits early when !dp->dp_panel.connector and cix_dp_bridge_funcs does not implement the bridge detect, EDID, or HPD callbacks needed without an internal connector. [ ... ] > +/* > + * Pure probe: report the current HPD level without touching the session > state > + * or the hardware. This runs on every DRM probe, poll and hotplug, so > + * DPTX_STATE_CONNECTED is advanced from the HPD event work and the encoder > + * prepare path instead. > + * > + * The host is initialized by probe, resume and cix_dp_prepare(). Bail out > + * loudly if we are ever called before that: cix_dp_read() is a plain readl() > + * and would hang on a gated APB clock. > + */ > +static int cix_dp_detect(struct cix_dp *dp) > +{ > + int ret = 0; > + > + mutex_lock(&dp->session_lock); > + > + if (dp->state & DPTX_STATE_SUSPENDED) { > + ret = -EBUSY; > + } else if (!(dp->state & DPTX_STATE_INITIALIZED)) { > + dev_warn_once(dp->dev, "detect with uninitialized host\n"); > + ret = -ENODEV; > + } > + > + mutex_unlock(&dp->session_lock); > + > + if (ret) > + return ret; > + > + return cix_dp_get_hpd_state(dp) ? 0 : -ENODEV; > +} [Severity: High] Can MMIO registers be read while dp->apb_clk is gated, hanging the bus? As noted in the comment above cix_dp_detect(), reading DPTX registers while the host is uninitialized and apb_clk is gated hangs the bus: - cix_dp_aux_transfer() calls cix_dp_get_hpd_state() and cix_dp_aux_cmd_submit() without acquiring dp->session_lock or checking DPTX_STATE_INITIALIZED / DPTX_STATE_SUSPENDED, so /dev/drm_dp_aux* or I2C transfers during suspend read unclocked MMIO registers. - cix_dp_hpd_irq_work_func() and cix_dp_hpd_event_work_func() call cix_dp_get_hpd_state() before acquiring dp->session_lock and checking DPTX_STATE_INITIALIZED. - cix_dp_detect() drops dp->session_lock before calling cix_dp_get_hpd_state(), and cix_dp_prepare() drops and re-acquires dp->session_lock around cix_dp_handle_connect() without rechecking DPTX_STATE_INITIALIZED, allowing a concurrent cix_dp_pm_prepare() to gate apb_clk in between. [ ... ] > +static void cix_dp_replay_deferred_hotplug(struct cix_dp *dp) > +{ > + struct phy *phy = dp->phy; > + union phy_configure_opts opts = { 0 }; > + u8 dpcd_rev; > + > + if (!(dp->state & DPTX_STATE_INITIALIZED)) > + return; > + if (dp->force_hpd != connector_status_unknown) > + return; > + if (cix_dp_get_hpd_state(dp)) > + return; > + > + phy_validate(phy, PHY_MODE_DP, 0, &opts); > + if (phy_get_mode(phy) != PHY_MODE_DP) > + return; > + > + /* > + * The combo PHY is already muxed to DP, so the sink may have been > + * plugged in before the connector existed. Probe the AUX channel to > + * make sure one is really there before forcing a hotplug event. > + */ > + if (drm_dp_dpcd_readb(&dp->aux, DP_DPCD_REV, &dpcd_rev) < 0) > + return; > + > + dp->force_hpd = connector_status_connected; > + dp->status = connector_status_unknown; > + schedule_delayed_work(&dp->hpd_event_work, 0); > +} [Severity: High] Won't drm_dp_dpcd_readb() in cix_dp_replay_deferred_hotplug() always fail with -ENODEV? cix_dp_replay_deferred_hotplug() only reaches drm_dp_dpcd_readb() when cix_dp_get_hpd_state(dp) is false, and it does not set dp->force_hpd = connector_status_connected until after the read succeeds. However, drm_dp_dpcd_readb() calls cix_dp_aux_transfer(), which immediately checks if (!cix_dp_get_hpd_state(dp)) and returns -ENODEV, so a USB-C display plugged in before bridge attach is never detected. > + > +static void cix_dp_hpd_event_work_func(struct work_struct *work) > +{ > + struct cix_dp *dp = container_of(work, struct cix_dp, > + hpd_event_work.work); > + struct phy *phy = dp->phy; > + union phy_configure_opts opts = { 0 }; > + enum drm_connector_status old_status; > + enum phy_mode mode; > + int i; > + > + if (!dp->dp_panel.connector) > + return; > + > + /* > + * On a USB-C DP alt-mode PHY the hotplug only becomes valid once the > + * PHY has been switched to DP mode, so wait for the mode change. > + * A native DP PHY has no ->set_mode() operation, which leaves the > + * mode at PHY_MODE_INVALID forever; there is nothing to wait for. > + */ > + if (cix_dp_get_hpd_state(dp) && > + phy_get_mode(phy) != PHY_MODE_INVALID) { > + for (i = 0; i < DP_HPD_MAX_TRIES; i++) { > + phy_validate(phy, PHY_MODE_DP, 0, &opts); > + mode = phy_get_mode(phy); > + if (mode == PHY_MODE_DP) > + break; > + msleep(100); > + } > + } > + > + old_status = dp->status; > + drm_helper_probe_detect(&dp->connector.base, NULL, false); > + > + if (dp->status == connector_status_connected) > + cix_dp_handle_connect(dp); > + else > + cix_dp_handle_disconnect(dp); > + > + if (old_status != dp->status && dp->drm) > + drm_helper_hpd_irq_event(dp->drm); > +} [Severity: High] Can sink DPCD capabilities remain unread across boot, resume, and hotplug, and can concurrent connector probes suppress hotplug events? If a native DP/eDP sink is already connected at boot, the HPD interrupt fires in cix_dp_init() before cix_dp_bridge_attach() sets dp->dp_panel.connector, so cix_dp_hpd_event_work_func() returns early; later, cix_dp_replay_deferred_hotplug() also returns early because cix_dp_get_hpd_state(dp) is true. Likewise, cix_dp_pm_prepare() clears DPTX_STATE_CONNECTED on suspend without cix_dp_pm_complete() re-probing DPCD on resume. Because cix_dp_connector_detect() only checks HPD level without calling cix_dp_handle_connect(), dp->dpcd, dp->downstream_ports, dp->caps, and dp->link_config remain unpopulated during mode validation, atomic_check, and cix_dp_bridge_mode_set() (which runs before encoder_enable() -> cix_dp_prepare() calls cix_dp_handle_connect()). Even when cix_dp_hpd_event_work_func() runs, it calls drm_helper_probe_detect() before cix_dp_handle_connect(), so cix_dp_connector_detect() passes stale dp->dpcd and dp->downstream_ports to drm_dp_set_subconnector_property(), and any concurrent DRM connector probe updates dp->status before old_status = dp->status is read: CPU0 (kworker) CPU1 (userspace probe) cix_dp_hpd_event_work_func() msleep(100) waiting for PHY cix_dp_connector_detect() dp->status = connected old_status = dp->status (connected) drm_helper_probe_detect() if (old_status != dp->status) <-- false, HPD event dropped Finally, cix_dp_handle_connect() only clears DPTX_STATE_CONNECTED when cix_dp_panel_read_sink_caps() returns -ETIMEDOUT or -ENOTCONN; if it fails with -EIO, -EPROTO, -EBUSY, or -ENODEV, DPTX_STATE_CONNECTED stays set and prevents later cix_dp_handle_connect() calls from reading the DPCD. [ ... ] > +static int cix_dp_prepare(struct cix_dp *dp) > +{ > + int ret; > + > + mutex_lock(&dp->session_lock); > + > + if (dp->state & (DPTX_STATE_ENABLED)) { > + ret = 0; > + goto end; > + } > + > + if (!(dp->state & DPTX_STATE_INITIALIZED)) { > + ret = cix_dp_host_init(dp); > + if (ret) > + goto end; > + } else if (dp->state & DPTX_STATE_INIT_TRAIN) { > + reset_dp_and_reinit(dp); > + usleep_range(100, 200); > + } > + > + if (cix_dp_get_hpd_state(dp) && > + !(dp->state & DPTX_STATE_CONNECTED)) { > + mutex_unlock(&dp->session_lock); > + cix_dp_handle_connect(dp); > + mutex_lock(&dp->session_lock); > + } > + > + if (!(cix_dp_get_hpd_state(dp) && (dp->state & DPTX_STATE_CONNECTED))) { > + ret = 0; > + goto end; > + } > + > + ret = cix_dp_core_on(dp, true); > + > +end: > + mutex_unlock(&dp->session_lock); > + return ret; > +} [Severity: Medium] If the display controller resumes and calls cix_dp_bridge_atomic_enable() -> cix_dp_prepare() before cix_dptx_pm_resume() -> cix_dp_pm_complete() clears DPTX_STATE_SUSPENDED, will the display stay blank? cix_dp_prepare() handles !(dp->state & DPTX_STATE_INITIALIZED) by calling cix_dp_host_init(), and then calls cix_dp_handle_connect() to restore DPTX_STATE_CONNECTED (which cix_dp_pm_prepare() cleared). However, cix_dp_handle_connect() bails out immediately when DPTX_STATE_SUSPENDED is still set, leaving DPTX_STATE_CONNECTED clear so cix_dp_prepare() skips cix_dp_core_on() and cix_dp_enable() skips enabling the stream. [ ... ] > +static int cix_dp_encoder_enable(struct cix_dp *dp, int stream_id, int > enable) > +{ > + int ret; > + > + if (enable) { > + ret = cix_dp_prepare(dp); > + if (ret) > + return ret; > + > + ret = cix_dp_enable(dp, stream_id); > + } else { > + if (!(dp->state & DPTX_STATE_INITIALIZED)) > + return 0; > + > + ret = cix_dp_disable(dp, stream_id); > + if (ret) > + return ret; > + > + ret = cix_dp_unprepare(dp); > + } > + > + return ret; > +} [Severity: Medium] Does a failure in cix_dp_enable() leak PHY init/power references and leave dp->state inconsistent? cix_dp_prepare() -> cix_dp_core_on() calls phy_init() and phy_power_on() and sets DPTX_STATE_READY | DPTX_STATE_INIT_TRAIN. If cix_dp_enable() then fails, DPTX_STATE_ENABLED is not set, cix_dp_encoder_enable() does not roll back via cix_dp_core_off(), and a later disable also skips cix_dp_core_off() because cix_dp_unprepare() returns early when !(dp->state & DPTX_STATE_ENABLED). On the next cix_dp_prepare() call, the else if (dp->state & DPTX_STATE_INIT_TRAIN) branch calls reset_dp_and_reinit(), which resets the controller and calls phy_init() again without a matching phy_exit(), while cix_dp_core_on() sees DPTX_STATE_READY still set and returns 0 without retraining the link. Similarly, in cix_dp_pm_complete(), if cix_dp_host_init() fails on resume, enable_irq(dp->irq) is skipped and DPTX_STATE_SUSPENDED stays set, while the next cix_dp_pm_prepare() or cix_dp_deinit() calls disable_irq(dp->irq) unconditionally. [ ... ] > +static void cix_dp_deinit(struct cix_dp *dp) > +{ > + cancel_delayed_work_sync(&dp->sw_hpd_irq_work); > + cancel_delayed_work_sync(&dp->hpd_irq_work); > + cancel_delayed_work_sync(&dp->hpd_event_work); > + disable_irq(dp->irq); > + > + mutex_lock(&dp->session_lock); > + cix_dp_host_deinit(dp); > + dp->state &= ~(DPTX_STATE_CONFIGURED); > + mutex_unlock(&dp->session_lock); > + > + mutex_destroy(&dp->session_lock); > +} [Severity: Medium] Can HPD delayed works be re-queued while cix_dp_deinit(), cix_dp_bridge_detach(), or cix_dp_pm_prepare() is tearing down the driver? In cix_dp_deinit(), cancel_delayed_work_sync() runs before disable_irq(dp->irq): CPU0 CPU1 (irq/kworker) cix_dp_deinit() cancel_delayed_work_sync(...) cix_dp_irq_handler() queue_delayed_work(..., &dp->hpd_event_work, 0) disable_irq(dp->irq) cix_dp_host_deinit() mutex_destroy(&dp->session_lock) (devm frees dp) cix_dp_hpd_event_work_func() <-- use-after-free / gated APB read Two related races exist in the other teardown paths: - cix_dp_bridge_detach() cancels hpd_irq_work before sw_hpd_irq_work (even though cix_dp_sw_hpd_irq_work_func() queues hpd_irq_work) and leaves dp->irq enabled and dp->dp_panel.connector non-NULL while calling drm_connector_cleanup(&dp->connector.base). - cix_dp_pm_prepare() cancels the delayed works before setting DPTX_STATE_SUSPENDED, while cix_dp_oob_hotplug_event() checks DPTX_STATE_SUSPENDED unlocked and can schedule hpd_event_work and sw_hpd_irq_work after the cancel calls. [ ... ] > +static void cix_dptx_remove(struct platform_device *pdev) > +{ > + struct cix_dp *dp = dev_get_drvdata(&pdev->dev); > + > + if (!dp) > + return; > + > + drm_bridge_remove(&dp->bridge); > + cix_dp_deinit(dp); > +} [ ... ] > +static void cix_dptx_shutdown(struct platform_device *pdev) > +{ > + struct cix_dp *dp = platform_get_drvdata(pdev); > + > + if (dp) { > + mutex_lock(&dp->session_lock); > + dp->active_stream_cnt = 0; > + mutex_unlock(&dp->session_lock); > + cix_dp_pm_prepare(dp); > + } > +} [Severity: Medium] What happens to active streams and PHY power if cix_dptx_shutdown() or cix_dptx_remove() runs while a display stream is enabled? In cix_dptx_shutdown(), zeroing dp->active_stream_cnt without calling cix_dp_ctrl_stream_off() or cix_dp_core_off() causes cix_dp_pm_prepare() -> cix_dp_host_deinit() to call phy_exit(dp->phy) while the PHY is still powered on (and with the second phy_init() reference from cix_dp_core_on() and stream_clk still active). In cix_dptx_remove(), dp->active_stream_cnt is not cleared and streams are not stopped, so cix_dp_host_deinit() returns early at if (dp->active_stream_cnt) and leaves the transmitter, clocks, and PHY enabled. > diff --git a/drivers/gpu/drm/cix/dptx/cix_dptx_drm.c > b/drivers/gpu/drm/cix/dptx/cix_dptx_drm.c > new file mode 100644 > index 0000000000000..0ddcd76691e03 > --- /dev/null > +++ b/drivers/gpu/drm/cix/dptx/cix_dptx_drm.c [ ... ] > +static int cix_dp_cal_bpc(struct cix_dp *dp, > + struct drm_connector_state *connector_state, int bpc) > +{ > + struct drm_display_info *info = > + &connector_state->connector->display_info; > + > + if (connector_state->colorspace >= DRM_MODE_COLORIMETRY_BT2020_CYCC && > + connector_state->colorspace <= DRM_MODE_COLORIMETRY_BT2020_YCC) > + bpc = 10; > + > + if (info->bpc && bpc > info->bpc) > + bpc = info->bpc; > + > + if (connector_state->max_requested_bpc && > + bpc > connector_state->max_requested_bpc) > + bpc = connector_state->max_requested_bpc; > + > + return bpc; > +} [Severity: Medium] When a BT.2020 colorspace is set, doesn't overwriting bpc = 10 in cix_dp_cal_bpc() prevent cix_dp_pick_format_bpc() from stepping down to 8 or 6 bpc when link bandwidth cannot fit 10 bpc? Each fallback call to compute_available_clock_rate() with curr_bpc = 8 or 6 gets forced back to 10. Also, in cix_dp_pick_config(), passing max(info->bpc, 6) as suggest_bpc selects 6 bpc instead of 8 bpc whenever an EDID leaves info->bpc as 0. [ ... ] > +static void cix_dp_adjust_mode(struct drm_display_mode *mode, > + struct drm_display_mode *adjusted_mode) > +{ > + int diff = mode->htotal - mode->hsync_end; > + > + /* > + * Cix DP requires horizontal backporch to be greater than 12. > + * This limitation may not be compatible with the sink device. > + * > + * Everything below is derived from @mode instead of accumulating on > + * @adjusted_mode, so that repeated atomic checks of the same state > + * produce the same result. > + */ > + if (diff < CIX_DPTX_MIN_H_BACKPORCH) { > + adjusted_mode->htotal = mode->htotal + > + CIX_DPTX_MIN_H_BACKPORCH - diff; > + adjusted_mode->clock = (int64_t)mode->clock * > + adjusted_mode->htotal / mode->htotal; > + } [Severity: Medium] Will the plain 64-bit division in (int64_t)mode->clock * adjusted_mode->htotal / mode->htotal fail to link with an undefined reference to __divdi3 / __aeabi_ldivmod on 32-bit architectures? DRM_CIX_DP does not depend on 64BIT, so this needs div_s64() or mul_u32_u32() with div_u64(). > + > + /* The sync polarities are inverted with respect to the source mode. */ > + if (mode->flags & DRM_MODE_FLAG_PVSYNC) > + adjusted_mode->flags &= ~DRM_MODE_FLAG_PVSYNC; > + else > + adjusted_mode->flags |= DRM_MODE_FLAG_PVSYNC; > + > + if (mode->flags & DRM_MODE_FLAG_PHSYNC) > + adjusted_mode->flags &= ~DRM_MODE_FLAG_PHSYNC; > + else > + adjusted_mode->flags |= DRM_MODE_FLAG_PHSYNC; > +} [ ... ] > +static int cix_dp_pick_format_bpc(struct cix_dp *dp, > + struct drm_connector_state *conn_state, > + struct drm_display_mode *adjusted_mode, > + int info_formats, int suggest_bpc, > + int *color_format, int *bpc, int *bpp) > +{ > + struct drm_display_info *info = > + &conn_state->connector->display_info; > + int curr_bpc; > + int format; > + int i; > + > + for (i = 0; i < ARRAY_SIZE(fmt_prio); i++) { > + format = fmt_prio[i].format; > + curr_bpc = fmt_prio[i].bpc ? fmt_prio[i].bpc : suggest_bpc; > + > + /* Only use 420_only flag for YUV420 format */ > + if (!(info_formats & BIT(format)) && > + !(format == DRM_OUTPUT_COLOR_FORMAT_YCBCR420 &&\n+ > drm_mode_is_420_only(info, adjusted_mode))) > + continue; [Severity: High] Does inverting the sync polarity flags on adjusted_mode in cix_dp_adjust_mode() break drm_mode_is_420_only(info, adjusted_mode) and YCbCr 4:2:0 format selection? cix_dp_bridge_atomic_check() calls cix_dp_adjust_mode() before cix_dp_pick_config(). Because cix_dp_adjust_mode() flips DRM_MODE_FLAG_PVSYNC and DRM_MODE_FLAG_PHSYNC on adjusted_mode (and may alter htotal and clock), drm_match_cea_mode(adjusted_mode) inside drm_mode_is_420_only() fails to match the CEA mode table and returns 0. Even when a mode is 4:2:0-only, cix_dp_pick_format_bpc() iterates fmt_prio[] from RGB444 and YCBCR422 without checking drm_mode_is_420_only() first, so any 4:2:0-only mode whose clock fits on the link at RGB or YCbCr 4:2:2 picks a non-4:2:0 format. Conversely, when info->color_formats has DRM_OUTPUT_COLOR_FORMAT_YCBCR420 set, fmt_prio[4] can pick YCBCR420 on modes where the sink does not support 4:2:0, and without checking that adjusted_mode->clock <= CIX_DP_MAX_FREQ (since YCbCr 4:2:0 forces 1 pixel per cycle). [ ... ] > +static void cix_dp_bridge_mode_set(struct cix_dp *dp, > + struct drm_crtc_state *crtc_state, > + struct drm_connector_state *conn_state) > +{ [ ... ] > + cix_conn->config.format = cix_dp_drm_format_to_dp(color_format); > + cix_conn->config.colorspace = conn_state->colorspace; > + cix_conn->config.bpc = bpc; > + cix_conn->config.bpp = bpp; > + dp->pixel_per_cycle = cix_dp_pick_ppc(adjusted_mode, > + cix_conn->config.format); > + > + link_rate = cix_dp_pick_link_rate(dp, adjusted_mode, bpp); > + bw_code = drm_dp_link_rate_to_bw_code(link_rate); > + > + dp->ops->mode_configure(dp, adjusted_mode->clock, bw_code, bpp, false); [ ... ] > +} [ ... ] > +static void > +cix_dp_bridge_atomic_enable(struct drm_bridge *bridge, > + struct drm_atomic_commit *state) > +{ [ ... ] > + cix_dp_bridge_mode_set(dp, crtc_state, conn_state); > + > + ret = dp->ops->encoder_enable(dp, dp->dp_panel.source_id, true); > + if (ret) > + dev_err(dp->dev, "failed to enable DP output: %d\n", ret); > +} [Severity: Medium] Should cix_dp_bridge_mode_set() run under dp->session_lock and after cix_dp_prepare() has read the sink's DPCD? cix_dp_bridge_atomic_enable() calls cix_dp_bridge_mode_set() unlocked before dp->ops->encoder_enable() -> cix_dp_prepare(). This mutates dp->connector.config, dp->pixel_per_cycle, and dp->mode without holding dp->session_lock (racing with cix_dp_hpd_irq_work_func() -> cix_dp_train_loop()), and computes link_rate and lane_cnt from dp->link_config before cix_dp_prepare() -> cix_dp_handle_connect() / cix_dp_core_on() updates dp->link_config from the sink's DPCD. -- Sashiko AI review ยท https://sashiko.dev/#/patchset/[email protected]?part=2
