iframe-proxy

FireBurn · 2026-06-30T14:14:10Z

EVDI PR notes — drm-next (7.1) build support + USB-unplug teardown fix

PR-ready material for DisplayLink/evdi. Branch fix-usb-unplug-teardown
(in evdi/, based on origin/main) carries three clean, signed-off commits:

Commit 1 is first so the tree builds on a 7.1 kernel before the behavioural
commits (bisectable on every kernel; the shim is inert below 7.1).

1. drm-next (Linux 7.1) build support

drm-next renamed the global atomic-state object and its allocator helpers:

struct drm_atomic_state          -> struct drm_atomic_commit
drm_atomic_state_alloc/clear/put -> drm_atomic_commit_alloc/clear/put

The atomic-helper vtable callbacks (.atomic_flush/.atomic_update/...) and the
state accessors (drm_atomic_get_{new,old}_*_state()) now take a
struct drm_atomic_commit *. evdi still uses the historical drm_atomic_state
names throughout evdi_modeset.c, so against >= 7.1 it fails with incompatible
pointer / function-pointer types.

The commit aliases the old identifiers to the new ones for
KERNEL_VERSION(7, 1, 0)+ in evdi_drm_drv.h. Because evdi refers to the struct
and the three renamed helpers by name everywhere, a handful of whole-token
#defines cascade through the whole module; the old names are gone from the
tree, so the macros cannot collide. Authored by Mike Lothian (predates this
work).

2 + 3. USB-unplug teardown

When the dock was physically removed, evdi's USB-removal notifier
(evdi_platform_drv_usb()) was supposed to unlink and destroy the
dynamically-created evdi platform/DRM device. It never did — drm_dev_unplug()
never ran, no remove uevent was emitted, the DRM node lingered as a ghost, and
/sys/module/evdi/refcnt climbed and never fell (reboot required). Wayland
compositors (KWin) and Xorg, which rely on the hot-unplug uevent, were left
holding a stale fd.

Two independent bugs, either alone enough to block teardown — so the path had
never worked:

Wrong notifier action constant (commit 2). The notifier is installed with
usb_register_notify(), so the USB core delivers USB_DEVICE_ADD /
USB_DEVICE_REMOVE (drivers/usb/core/notify.c). The handler instead tested
action != BUS_NOTIFY_DEL_DEVICE, and BUS_NOTIFY_DEL_DEVICE == 1 == USB_DEVICE_ADD, so the body ran on every add and returned early on every
remove — it never reacted to an unplug. Fixed by testing
USB_DEVICE_REMOVE.
Dead destroy branch (commit 3). Even once the handler ran, destroy was
gated on pdev->dev.parent == &usb_dev->dev, but evdi platform devices are
created with .parent = NULL (evdi_platform_drv_create_new_device()); the
USB link lives in evdi_platform_device_data::parent. The condition was never
true. Fixed by making evdi_platform_device_unlink_if_linked_with() return
bool (did it detach from this parent) and destroying when it did,
restricted to dynamic devices (i >= evdi_initial_device_count) so the
statically pre-allocated pool (for static-device-list clients like Xorg) is
only detached.

Teardown chain restored: evdi_platform_dev_destroy() →
platform_device_unregister() → evdi_platform_device_remove() →
evdi_drm_device_remove() → drm_dev_unplug() → udev remove uevent.

Verification (HW, 2026-06-30)

Dell D6000 dock, kernel 7.1.0-rc5-drm+, initial_device_count=0, KWin/Wayland +
DisplayLinkManager driving two heads.

Before: unplug produced only change uevents; evdi.0/evdi.1 and their
device symlinks persisted; refcnt stuck (23 → 23).
After: dmesg Detached from parent device → Parent USB removed. Removing evdi.0/evdi.1 → Evdi platform_device destroy → Evdi drm_device removed.;
udev remove (not just change) on card2/card3 + connectors;
/sys/devices/platform/evdi.* gone; /dev/dri/ back to the real GPUs;
refcnt 23 → 0 (no leak).

module/evdi.ko builds clean across all three commits (sha verified identical to
the tested module).

checkpatch (`--strict`): 0 errors on all three

Remaining items are non-blocking:

Commit 1: LINUX_VERSION_CODE should be avoided — inherent to any kernel
version-compat shim; evdi is an out-of-tree module built on KERNEL_VERSION()
guards throughout.
Commits 2 & 3: Co-authored-by: is not a kernel-standard trailer (the kernel
recognises Co-developed-by:, which needs its own DCO sign-off). Drop the
trailer if a spotless run is wanted.
Commit 3: 2 advisory CHECKs ("Alignment should match open parenthesis") on the
evdi_platform_device_unlink_if_linked_with signature — pre-existing upstream
indentation; the patch only changes the return keyword (void→bool, same
width), introducing no new misalignment.

drm-next (Linux 7.1) renamed the global atomic-state object and its allocator helpers: struct drm_atomic_state -> struct drm_atomic_commit drm_atomic_state_alloc/clear/put -> drm_atomic_commit_alloc/clear/put The atomic-helper vtable callbacks (.atomic_flush/.atomic_update/...) and the state accessors (drm_atomic_get_{new,old}_*_state()) now take a struct drm_atomic_commit *. evdi still uses the historical drm_atomic_state names throughout evdi_modeset.c, so it fails to build against >= 7.1 with incompatible pointer and function-pointer types. Alias the old identifiers to the new ones for KERNEL_VERSION(7, 1, 0) and later. The old names are fully gone from the tree, so these whole-token macros cannot collide with anything still in the headers. Signed-off-by: Mike Lothian <mike@fireburn.co.uk>

@DaTa

evdi_platform_drv_usb() is installed with usb_register_notify(), so the USB core delivers it the USB_DEVICE_ADD / USB_DEVICE_REMOVE notifications (see usb_notify_add_device()/usb_notify_remove_device() in drivers/usb/core/notify.c), with the struct usb_device passed in @DaTa. The handler instead compared @action against the bus-notifier constant BUS_NOTIFY_DEL_DEVICE. That enum value happens to equal USB_DEVICE_ADD (1), so the notifier body ran on every USB device *addition* and returned early on every *removal*. As a result evdi never reacted to its parent USB device (the dock) being unplugged. Test the correct USB_DEVICE_REMOVE constant so the handler runs when the parent USB device is removed. Signed-off-by: Mike Lothian <mike@fireburn.co.uk> Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

When the parent USB device was unplugged the removal notifier detached the evdi platform device from it but never destroyed it. The teardown was gated on: if (pdev->dev.parent == &usb_dev->dev) but evdi platform devices are created with .parent = NULL (see evdi_platform_drv_create_new_device()); the USB relationship is tracked in evdi_platform_device_data::parent, not in dev.parent. The condition was therefore never true, so the DRM device was never unregistered, drm_dev_unplug() never ran, and userspace (Wayland compositors, Xorg) received no hot-unplug uevent and leaked the stale device. Have evdi_platform_device_unlink_if_linked_with() report whether it actually detached the device from this USB parent, and destroy the device when it did. Restrict the destroy to dynamically created devices (i >= evdi_initial_device_count); the statically pre-allocated devices are only detached, preserving their intended lifetime for static-list clients. Signed-off-by: Mike Lothian <mike@fireburn.co.uk> Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

mkaspryk-synaptics

@FireBurn
LGTM.
Minor comments.
Please rename the PR to something more descriptive (I don't like 'Fixes' name) and update copyrights in all changed files.

mkaspryk-synaptics · 2026-10-07T08:21:23Z

+/*
+ * drm-next (>= 7.1.0) renamed the global atomic-state object and its helpers:
+ *   struct drm_atomic_state              -> struct drm_atomic_commit
+ *   drm_atomic_state_alloc/clear/put     -> drm_atomic_commit_alloc/clear/put
+ * The atomic-helper vtable callbacks (.atomic_flush/.atomic_update/...) and the
+ * state accessors (drm_atomic_get_{new,old}_*_state) now take a
+ * struct drm_atomic_commit *. evdi's source still uses the historical names, so
+ * alias them. The old identifiers are fully gone from the tree, so these
+ * whole-token macros are collision-free.
+ */


Please remove this comment.
The commit description is sufficient in this case.
Please rename the commit to 'Add drm_atomic_commit compatibility aliases for Linux 7.1+' or something similar to this one.

@FireBurn I've noticed that this is already handled in the compatibility work from #572, which addresses the drm_atomic_state -> drm_atomic_commit rename without the version-specific alias. I think we should stick with the existing approach. Please drop this commit.

mkaspryk-synaptics · 2026-10-07T08:37:45Z

 	if (!usb_dev)
 		return 0;
-	if (action != BUS_NOTIFY_DEL_DEVICE)
+	if (action != USB_DEVICE_REMOVE)


LGTM.
One thing - please rename the commit to 'React to USB_DEVICE_REMOVE in the USB unplug notifier'

mkaspryk-synaptics · 2026-10-07T08:48:17Z

+		if (evdi_platform_device_unlink_if_linked_with(pdev, &usb_dev->dev) &&
+		    i >= evdi_initial_device_count) {


LGTM.
Please rename the commit to 'Destroy dynamic devices when their USB parent is removed'

FireBurn and others added 3 commits June 30, 2026 14:56

jonathan-conder mentioned this pull request Jul 7, 2026

Workshop breaks when a symlink can't be resolved canonical/workshop#924

Open

mkaspryk-synaptics force-pushed the main branch from fb2ff77 to 558af9c Compare October 2, 2026 10:39

mkaspryk-synaptics reviewed Oct 7, 2026

View reviewed changes

#	Commit	What
1	`evdi: support the drm_atomic_state -> drm_atomic_commit rename`	builds evdi against Linux 7.1 / drm-next
2	`evdi: react to USB_DEVICE_REMOVE in the USB unplug notifier`	unplug bug #1
3	`evdi: destroy dynamic devices when their USB parent is removed`	unplug bug #2

Sunbelt Computer Software

PL/B Language Development and Support

Navigation Menu

Repository navigation

Fixes - #581

Fixes#581
FireBurn wants to merge 3 commits into
DisplayLink:mainfrom
FireBurn:fixes

FireBurn commented Jun 30, 2026 •

edited

Loading

Uh oh!

mkaspryk-synaptics left a comment •

edited

Loading

Uh oh!

mkaspryk-synaptics Oct 7, 2026

Uh oh!

mkaspryk-synaptics Oct 12, 2026

Uh oh!

mkaspryk-synaptics Oct 7, 2026

Uh oh!

mkaspryk-synaptics Oct 7, 2026

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

		if (evdi_platform_device_unlink_if_linked_with(pdev, &usb_dev->dev) &&
		i >= evdi_initial_device_count) {

Sunbelt Computer Software

PL/B Language Development and Support

Conversation

FireBurn commented Jun 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

EVDI PR notes — drm-next (7.1) build support + USB-unplug teardown fix

1. drm-next (Linux 7.1) build support

2 + 3. USB-unplug teardown

Verification (HW, 2026-06-30)

checkpatch (--strict): 0 errors on all three

Uh oh!

mkaspryk-synaptics left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mkaspryk-synaptics Oct 7, 2026

Choose a reason for hiding this comment

Uh oh!

mkaspryk-synaptics Oct 12, 2026

Choose a reason for hiding this comment

Uh oh!

mkaspryk-synaptics Oct 7, 2026

Choose a reason for hiding this comment

Uh oh!

mkaspryk-synaptics Oct 7, 2026

Choose a reason for hiding this comment

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

FireBurn commented Jun 30, 2026 •

edited

Loading

checkpatch (`--strict`): 0 errors on all three

mkaspryk-synaptics left a comment •

edited

Loading