Hi Andrés Thanks for doing the bisecting
On Sun, 22 Feb 2026 at 22:56, Andrés Pérez <[email protected]> wrote: > > # OVERVIEW > > Since kernel v6.16.1, switching from an X11 session to a text VT and back > freezes the display on a ThinkPad P15 Gen 2. The system remains responsive > over SSH; only the display is frozen. Bisecting identified commit > d1b618e7954802fe ("media: uvcvideo: Do not turn on the camera for some > ioctls") as the trigger. Reverting the logic change in that commit > fixes VT switching > on v6.16.1, v6.17.9, and v6.18.9, but that is not an actual solution. Wayland > compositors (e.g., river and sway) are not affected. > > Last good: v6.15.9 > First bad: v6.16.1 > Bisect result: d1b618e7954802fe media: uvcvideo: Do not turn on the > camera for some ioctls > > ## Hardware: Lenovo ThinkPad P15 Gen 2i (20YQ0031US) > CPU: Intel Core i7-11800H (Tiger Lake-H) > iGPU: Intel UHD Graphics (TGL GT1) > dGPU: NVIDIA T1200 (not involved in eDP output; driver: nvidia-open) > Display: 15.6" 1920x1080 eDP, 10 bpc capable (EDID 1.4) > Webcam: Integrated Camera on PCH xHCI (Bus 003 Port 004) > Firmware: LENOVO N37ET61W (1.97) > OS: Arch Linux, Nix home-manager, X11 + xmonad, no display manager > > ## Symptoms and reproduction steps: > 1. Boot, start X11 on tty1 (startx). > 2. Switch to tty2 (Ctrl+Alt+F2): works. > 3. Switch back to tty1 (Ctrl+Alt+F1): display freezes. > - Frozen on the last frame shown before switching away. > - System is fully responsive over SSH. > - Other VTs switch normally between each other as long as X11 is > not active on them. > - Killing X does not recover the display. A reboot is required. > > # DEBUG ANALYSIS > > On v6.16.1, the VT switch back to X triggers a full modeset due to pipe > configuration mismatches detected by intel_pipe_config_compare: > > [drm:intel_pipe_config_compare] fastset requirement not met in pipe_bpp > (expected 30, found 24) > [drm:intel_pipe_config_compare] fastset requirement not met in dp_m_n > (expected link 269484/524288, found link 336855/524288) > [drm:intel_pipe_config_compare] fastset requirement not met in dpll_hw_state > (expected cfgcr0: 0xe001a5, found cfgcr0: 0x1c2) > [drm:intel_pipe_config_compare] fastset requirement not met in port_clock > (expected 270000, found 216000) > [drm:intel_atomic_check] forcing full modeset > > On v6.15.9, the same VT switch shows no such messages. > no pipe_config_compare runs, no modeset, no freeze. > > # BISECT AND VERIFICATION > > The bisect converged on d1b618e7954802fe in the uvcvideo driver. This > commit adds a switch statement to uvc_v4l2_unlocked_ioctl that allows > certain V4L2 IOCTLS to call video_ioctl2 directly without first calling > uvc_pm_get/uvc_pm_put. Prior to this commit, all ioctls called uvc_pm_get > before video_ioctl2. > > ## VT switching verification across kernel versions: > > v6.12.74 arch pkg: WORKS > v6.15.9 arch pkg: WORKS > v6.15.9 from source: WORKS > v6.16.1 with d1b618e reverted: WORKS > v6.17.9 with PM wrapping restored: WORKS > v6.18.9 with PM wrapping restored: WORKS > > v6.16.1 from source: FREEZES > v6.16.1 arch pkg: FREEZES > v6.17.9 arch pkg: FREEZES > v6.18.9 from source: FREEZES > v6.18.9 arch pkg: FREEZES > > ## Things that do not eliminate the freeze > > - module_blacklist=uvcvideo on boot > - CONFIG_USB_VIDEO_CLASS=n (compiled out) This is puzzling me a bit... You are saying that if you do not build the uvc driver, the freeze is still happening? Am I understanding this correctly? > - i915.enable_psr=0 > - Bypassing intel_vrr_transcoder_enable/disable (no-op) > - xrandr --output eDP-1 --set "max bpc" 10 > - Xorg config FBDepth 30 (No effect on pipe_bpp) > > ## Workaround patch > > Reverting the optimization from d1b618e to restore the unconditional > uvc_pm_get/put wrapping for all ioctls. This is not a proper fix. > > diff --git a/drivers/media/usb/uvc/uvc_v4l2.c > b/drivers/media/usb/uvc/uvc_v4l2.c > index 9e4a251eca88..15057b47ec4f 100644 > --- a/drivers/media/usb/uvc/uvc_v4l2.c > +++ b/drivers/media/usb/uvc/uvc_v4l2.c > @@ -1199,33 +1199,12 @@ static long uvc_v4l2_unlocked_ioctl(struct file *file, > unsigned int converted_cmd = v4l2_translate_cmd(cmd); > int ret; > > - /* The following IOCTLs need to turn on the camera. */ > - switch (converted_cmd) { > - case UVCIOC_CTRL_MAP: > - case UVCIOC_CTRL_QUERY: > - case VIDIOC_G_CTRL: > - case VIDIOC_G_EXT_CTRLS: > - case VIDIOC_G_INPUT: > - case VIDIOC_QUERYCTRL: > - case VIDIOC_QUERYMENU: > - case VIDIOC_QUERY_EXT_CTRL: > - case VIDIOC_S_CTRL: > - case VIDIOC_S_EXT_CTRLS: > - case VIDIOC_S_FMT: > - case VIDIOC_S_INPUT: > - case VIDIOC_S_PARM: > - case VIDIOC_TRY_EXT_CTRLS: > - case VIDIOC_TRY_FMT: > - ret = uvc_pm_get(handle->stream->dev); > - if (ret) > - return ret; > - ret = video_ioctl2(file, cmd, arg); > - uvc_pm_put(handle->stream->dev); > + ret = uvc_pm_get(handle->stream->dev); > + if (ret) > return ret; > - } > - > - /* The other IOCTLs can run with the camera off. */ > - return video_ioctl2(file, cmd, arg); > + ret = video_ioctl2(file, cmd, arg); > + uvc_pm_put(handle->stream->dev); > + return ret; > } > > const struct v4l2_ioctl_ops uvc_ioctl_ops = { > > Andrés > -- Ricardo Ribalda
