Skip to content

drm/asahi: G17 kill flag order and keep_graph() doc comment - #550

Merged
DjDeveloperr merged 2 commits into
integration/aurorafrom
fix/g17-kill-flag-order
Oct 28, 2026
Merged

DjDeveloperr merged 2 commits into
integration/aurorafrom
fix/g17-kill-flag-order

Conversation

@DjDeveloperr

Copy link
Copy Markdown
Contributor

Mirror of two review follow-ups on omacom/linux-aurora#244 (commits 5389bde8 and 8a7ef5d6 there), so the G17 driver stays the same in both trees.

  • give keep_graph() its doc comment back. LEAKED_GRAPHS sat between keep_graph()'s doc comment and the function, so rustdoc attached the comment to the static.
  • mark a G17 context killed right after the kill is published. Both kill paths set the kill-published flag only after a fallible step that follows the irrevocable publication: record_render_kill(...)? on the render-only path and .ok_or(EIO)? on the compute path. Neither can fail today, since the candidate checks rule those cases out under the same device mutex. The flag is now set as soon as publish_control() succeeds, so the at-most-once rule from drm/asahi: kill a G17 context at most once #513 doesn't depend on that.

No behaviour change in practice. Both picks applied without conflicts.

Testing: asahi KUnit on this branch (minimal arm64 config with DRM_ASAHI): 181 tests, 180 pass, 1 skip. On omacom, the full KUnit suite with these commits passes (1718 tests, 1715 pass, 3 skip).

LEAKED_GRAPHS was added between keep_graph()'s doc comment and the
function, so rustdoc attached "Keeps `graph` until the firmware
stops..." to the static. Move the static above that comment.

Signed-off-by: DjDeveloperr <hey@djdev.me>
(cherry picked from omacom/linux-aurora commit 5389bde85ad5dfde4936bc6517afb83cdcd1fdab)
Both kill paths set the context's kill-published flag only after a
fallible step that follows the irrevocable publication:
record_render_kill(...)? on the render-only path, and the compute queue
lookup's .ok_or(EIO)? on the compute path. Neither can fail today,
because the candidate checks rule those cases out under the same
device mutex, but the at-most-once rule should not depend on that.

Set the flag as soon as publish_control() succeeds.

Signed-off-by: DjDeveloperr <hey@djdev.me>
(cherry picked from omacom/linux-aurora commit 8a7ef5d6ae9fdb44826c866195070b490f05843d)
@DjDeveloperr

Copy link
Copy Markdown
Contributor Author

Qualified on the Neo (J700). Test module: these two commits cherry-picked onto integration/aurora 2dd77ab (which includes #555), built against the int1028 kit core, release 7.3.0-ARCH-neoint, Mesa aurora-wip f3639d31. Regression gate PERFORMANCE-REGRESSION-PASS (cross-queue order/data, memory pressure, mixed vkstress); targeted Vulkan CTS at 6 workers: 4687 pass / 681 not supported / 0 fail; desktop alive; no new kernel errors.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant