Skip to content

drm/asahi: advertise SOFT_FAULTS on T6050 G17C - #2551

Closed
erikwb wants to merge 1 commit into
integration/aurorafrom
j716c/g17-m5-soft-faults
Closed

erikwb wants to merge 1 commit into
integration/aurorafrom
j716c/g17-m5-soft-faults

Conversation

@erikwb

@erikwb erikwb commented Nov 20, 2026 •

Copy link
Copy Markdown
Contributor

Advertise DRM_ASAHI_FEATURE_SOFT_FAULTS on T6050 GpuVariant::C (0x6050, M5 Max) unless overridden via aurora.fault_control.

Summary

  • Both render and compute queues on T6050 G17C already configure hardware soft-fault zeroing (m5::queue::ConfigProfile::soft_fault_enabled and QueueConfig::UNK_66), so unmapped GPU reads return zero and unmapped writes are discarded without raising a GPU fault.
  • Gate supports_soft_faults() strictly on self.cfg.chip_id == 0x6050 && self.cfg.identity.gpu_variant == crate::hw::GpuVariant::C && fault_control == 0xb, leaving G17S, G17D, and other variants excluded until independently verified.
  • Document in asahi_core.rs and g17/mod.rs that on G17 aurora.fault_control overrides only the UAPI advertisement without altering queue configuration.

Verification

  • Built cleanly with both CONFIG_KUNIT=y and CONFIG_KUNIT=n.
  • Verified on J716c (T6050 G17C) across 20,246 Vulkan robustness and sparse buffer CTS tests (vertex_attribute_fetch, storage_buffer, uniform_buffer, vertex_access, buffer_access, null_descriptor, sparse_resources.buffer): 20,246/20,246 PASS, zero GPU faults or recoveries.
  • Verified with Cyberpunk 2077 T2 benchmark on J716c (paired with Borealis soft-fault robustness lowering, eliminating software buffer bounds overhead).

Created with Antigravity (reviewed with gpt-6.2-sol).

On T6050 G17C (M5 Max), both render and compute queues configure hardware
soft-fault zeroing (`m5::queue::ConfigProfile::soft_fault_enabled` and
`QueueConfig::UNK_66`), so unmapped GPU reads return zero and unmapped
writes are discarded without raising a GPU fault.

Advertise `DRM_ASAHI_FEATURE_SOFT_FAULTS` on T6050 `GpuVariant::C`
(`0x6050`, M5 Max) unless overridden via `aurora.fault_control`, and
document that on G17 `aurora.fault_control` overrides only the UAPI
advertisement without changing queue configuration. Other T6050 variants
remain excluded until independently verified.

Verified on J716c (T6050 G17C) across 20,246 Vulkan robustness and sparse
buffer CTS tests (`vertex_attribute_fetch`, `storage_buffer`,
`uniform_buffer`, `vertex_access`, `buffer_access`, `null_descriptor`,
`sparse_resources.buffer`) with 20,246/20,246 passing and zero GPU faults
or recoveries. Built cleanly with both CONFIG_KUNIT=y and CONFIG_KUNIT=n.

Developed with Claude Code.
Signed-off-by: Erik Bourget <erik@ebourget.net>
@RyanTheTide

Copy link
Copy Markdown
Contributor

Thanks for the J716c robustness and sparse runs, 20,246 clean cases is a strong result.

This overlaps with linux#2081, which turns DRM_ASAHI_FEATURE_SOFT_FAULTS on for every GPU the G17 backend drives. The reason there is the same one this PR gives: the driver programs the same soft-fault setting (QueueConfig::unk_66, the word the M5 firmware calls soft_fault_enabled) into every G17 queue. #2081 already has robustness rows on an M4 Pro and an A18 Pro. This PR changes the same function to

self.cfg.chip_id == 0x6050
    && self.cfg.identity.gpu_variant == crate::hw::GpuVariant::C

(drivers/gpu/drm/asahi/g17/mod.rs:2581-2583). We select behaviour by a hardware property, not by the list of machines it was tested on. Here that property is the queue's soft-fault setting, and G17G and G17S share it.

Could you run your CTS set on #2081's head on the J716c and post it there? If it passes, this PR can close. If G17C needs something #2081 doesn't do, a note on what differs would help. We can add rows on our M1 Pro, M2, M4 Pro and A18 Pro.

@erikwb

erikwb commented Nov 21, 2026

Copy link
Copy Markdown
Contributor Author

Note

🤖 Claude Opus 5.5 responding on behalf of erikwb

@RyanTheTide, the row you asked for: #2081's head (50c9683) on the J716c (M5 Max, G17C, 27.2 firmware), in a kernel built from the J716c's current base, with current aurora-wip Mesa.

  • dEQP-VK.robustness.*, 12,000 evenly spaced cases: 5,952 pass, 6,048 not supported, 0 failures.
  • dEQP-VK.sparse_resources.*, 8,000 evenly spaced cases: 3,714 pass, 4,286 not supported, 0 failures.
  • No GPU faults or recoveries in dmesg.

#2081 covers G17C, so I'll close this PR.

@erikwb erikwb closed this Nov 21, 2026
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.

2 participants