Thank you for your contribution! Sashiko AI review found 2 potential issues to consider: - [High] drm/bridge: dw-dp: dw_dp_hpd_work() retrains link before parsing sink DPCD - [Medium] drm/bridge: dw-dp: unlocked link state check and TOCTOU in dw_dp_link_retrain() --
commit ecb979561e6b36e93516c81e2368f1af68040e73 Author: Yury Smirnov <[email protected]> drm/bridge: synopsys: dw-dp: Retrain on hotplug with pre-enabled link This commit retrains an already-enabled DisplayPort link on a long HPD pulse when a sink is detected before sending the DRM hotplug event. > diff --git a/drivers/gpu/drm/bridge/synopsys/dw-dp.c > b/drivers/gpu/drm/bridge/synopsys/dw-dp.c > index bf3a53463394f..b245d62009139 100644 > --- a/drivers/gpu/drm/bridge/synopsys/dw-dp.c > +++ b/drivers/gpu/drm/bridge/synopsys/dw-dp.c > @@ -2394,6 +2394,12 @@ static void dw_dp_hpd_work(struct work_struct *work) > dev_warn(dp->dev, "Retrain link failed\n"); > } > } else { > + if (dp->link.enabled && dw_dp_hpd_detect(dp)) { > + ret = dw_dp_link_retrain(dp); > + if (ret) > + dev_warn(dp->dev, "Retrain link failed\n"); > + } > + > drm_helper_hpd_irq_event(dp->bridge.dev); > } > } [Severity: High] Should dw_dp_link_retrain() run after drm_helper_hpd_irq_event() rather than before it? Connector detection in drm_helper_hpd_irq_event() invokes dw_dp_bridge_detect() -> dw_dp_hpd_detect_link() -> dw_dp_link_parse(), which calls dw_dp_link_reset() (zeroing link->rate, link->lanes, and link->caps) and then reads the sink's DPCD capabilities. If a previous unplug detection failed after dw_dp_link_reset() (for example, on a DP branch device or USB-C adapter when sink_count dropped to 0 and dw_dp_link_parse() returned -ENODEV), dp->link.rate and dp->link.lanes are still 0 when the sink is replugged. Calling dw_dp_link_retrain() first then runs dw_dp_phy_configure(dp, 0, 0, false), putting the PHY into P3 powerdown with all lanes disabled and failing link training before dw_dp_link_parse() has re-read the DPCD. Even when dp->link was not zeroed, retraining before drm_helper_hpd_irq_event() trains against stale DPCD capabilities without transitioning the replugged sink to DP_SET_POWER_D0 via drm_dp_link_power_up(), and the subsequent dw_dp_link_parse() call in drm_helper_hpd_irq_event() immediately resets dp->link and overwrites any downgraded link->rate. [Severity: Medium] Can dw_dp_link_retrain() race with a concurrent atomic disable and train a link whose PHY has already been powered off? dw_dp_hpd_work() checks dp->link.enabled and dw_dp_link_retrain() checks dw_dp_needs_link_retrain() before acquiring dev->mode_config.connection_mutex, and neither is rechecked after acquiring the lock: CPU0 (kworker) CPU1 (modeset) dw_dp_hpd_work() if (dp->link.enabled && ...) dw_dp_link_retrain() dw_dp_needs_link_retrain() dw_dp_bridge_atomic_disable() dw_dp_link_disable() dw_dp_phy_xmit_enable(dp, 0) phy_power_off(dp->phy) link->enabled = false drm_modeset_lock(&connection_mutex) dw_dp_link_train() dw_dp_phy_configure() <-- configures powered-off PHY Rechecking dp->link.enabled and dw_dp_needs_link_retrain() after locking connection_mutex would close this window. -- Sashiko AI review ยท https://sashiko.dev/#/patchset/20261114-synopsys-dw-dp-improvements-v12-0-d42d3dc5b...@collabora.com?part=23
