Commit 1bda647ce61b ("drm/msm/dp: move link-level teardown from
display_disable to display_unprepare") moved msm_dp_ctrl_off_link() and
the PHY exit into msm_dp_display_unprepare(). Unlike
msm_dp_display_disable(), that function does not check power_on.

When link training fails, msm_dp_ctrl_on_link() has already turned the
link clocks and the PHY off, and .atomic_enable() returns with power_on
still false. The next disable now runs msm_dp_ctrl_off_link() anyway,
which writes REG_DP_MAINLINK_CTRL and pulses DP_SW_RESET with the link
clocks off. On an ASUS Zenbook A16 (Glymur, eDP) the SoC then resets.

Skip the link teardown when power_on is false, as before that commit,
but still drop the runtime PM reference.

Tested on the A16 by forcing link training to fail and then blanking
the display: the SoC reset 2/2 times without this patch and 0/3 with
it.

Fixes: 1bda647ce61b ("drm/msm/dp: move link-level teardown from display_disable 
to display_unprepare")
Signed-off-by: Burke Cates <[email protected]>
---

Notes:
    Test base: v7.4-rc1 with out-of-tree Glymur platform, board and eDP
    PHY patches; drivers/gpu/drm/msm/dp as in v7.4-rc1 plus a local
    knob that fails every training attempt. The reset is silent (no
    pstore), so which of the two accesses triggers it is not known.
    
    As before 1bda647ce61b, a failure after successful training still
    leaves the link up. Xilin Wu's hdr10 v2 09/20 handles these paths
    more generally; this is the minimal fix for -rc. Happy to test his
    approach instead.
    
https://lore.kernel.org/all/[email protected]/
    
    The bug was found and this patch drafted with an AI coding
    assistant (Claude Code, Claude Opus 5.5); I reviewed it and ran the
    tests.

 drivers/gpu/drm/msm/dp/dp_display.c | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/drivers/gpu/drm/msm/dp/dp_display.c 
b/drivers/gpu/drm/msm/dp/dp_display.c
index 9299ce616fdb..76784cc45228 100644
--- a/drivers/gpu/drm/msm/dp/dp_display.c
+++ b/drivers/gpu/drm/msm/dp/dp_display.c
@@ -1528,6 +1528,14 @@ static void msm_dp_display_unprepare(struct 
msm_dp_display_private *dp)
 {
        struct msm_dp *msm_dp_display = &dp->msm_dp_display;
 
+       /*
+        * Failed link training leaves the link clocks off. Avoid accessing
+        * link registers after a failed enable, but still release the
+        * runtime-PM reference below.
+        */
+       if (!msm_dp_display->power_on)
+               goto out;
+
        /* dongle is still connected but sinks are disconnected */
        if (dp->link->sink_count == 0)
                msm_dp_link_psm_config(dp->link, &dp->panel->link_info, true);
@@ -1542,6 +1550,7 @@ static void msm_dp_display_unprepare(struct 
msm_dp_display_private *dp)
 
        msm_dp_display->power_on = false;
 
+out:
        pm_runtime_put_sync(&msm_dp_display->pdev->dev);
 
        drm_dbg_dp(dp->drm_dev, "type=%d Done\n", 
msm_dp_display->connector_type);

base-commit: 6628e7d4c6b25c8d18f1345e3a20d21829aae7e4
-- 
2.54.0

Reply via email to