Alexon Oliveira (the reporter) tested v2 of this series and gave his Tested-by on the cover letter. Repeating it here, as patchwork does not carry a tag from the cover letter over to the individual patches.
He tested on a ThinkPad P1 Gen 7 (Meteor Lake) with a Lenovo ThinkPad Thunderbolt 4 Dock (Synaptics VMM6212) and a Samsung TV, with Fedora 44 kernel 7.2.15 and drm_display_helper.ko and i915.ko rebuilt with both patches. The "Failed to get ACT after 3000 ms" errors, 3 without the patches, are gone, and the new TV connector is connected 2.48 s after the HPD drop instead of 8.44 s. https://gitlab.freedesktop.org/drm/i915/kernel/-/work_items/17741 For reference, the commit message of v2: > drm/i915/dp_mst: Skip link retrain and the ACT wait for disconnected MST > streams > > When a dock drops the upstream HPD, for instance when a TV behind it is > powered on, the MST topology is torn down and the MST connectors get > unregistered, while the CRTCs stay bound to them in the atomic state until > userspace disables them. > > After the HPD is back, the link status of the link reads as all zeros, so > intel_dp_retrain_link() selects the pipe still bound to the old MST > connector and forces a modeset on it, while the topology is being > rediscovered. drm_dp_add_payload_part1() fails, as the port is not in the > topology anymore, but the enable sequence carries on and waits for an ACT > at the branch that is never triggered, since nothing was written to DPCD > 0x1c0-0x1c2: > > [drm] *ERROR* Failed to get ACT after 3000 ms, last status: 04 > > The same wait happens when userspace disables the stale pipe afterwards, > as drm_dp_remove_payload_part1() doesn't remove the payload of a port that > is not in the topology. The new connector only gets registered after these > waits are over. > > Fix this by: > > - Not retraining the link if the MST sinks of all the active pipes are > gone. If any stream is still connected all the pipes are kept, as the > link only gets retrained after all the streams on it are disabled. > > - Only waiting for the ACT at the branch if the payload was allocated to > the DFP on enable, and if drm_dp_remove_payload_part1() removed it from > the branch on disable. The ACT generation and the wait for it in the > source are left as they were. > > Link: https://gitlab.freedesktop.org/drm/i915/kernel/-/work_items/17741 > Signed-off-by: Xiao Lu <[email protected]> Tested-by: Alexon Oliveira <[email protected]>
